sched : reintroduce less synchronizations during split compute (#20793 )

* CUDA: Improve performance via less synchronizations between token (#17795) * Adds CPU-to-CUDA copy capability to ggml_backend_cuda_cpy_tensor_async() * Adds function to relax sync requirements between input copies on supported backends (CUDA for now) * Exchanges synchronous copy with async copy function. * Adds macro guards to allow compilation in non-CUDA builds * Reworked backend detection in ggml-backend.cpp to avoid linking conflicts * Relax requirement of checks in async CUDA copies from backend and buffer type to just buffer type, to avoid linking issues * Minor cleanup * Makes opt-in to relax use of explicit syncs more general. Backends like vulkan which require a synchronization between HtoD copies and graph execution could also adopt this change now. * Reintroduces stricter check for CPU->CUDA backend async copy via GGML_DEVICE_TYPE_CPU. * Corrects initialization of ggml_backend_sync_mode in ggml_backend_sched_split initialization * Simplifies synchronizations to adhere to `saaasg` pattern. * Apply suggestion from @ggerganov (src->buffer to buf_src) Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Apply suggestion from @ggerganov (src->buffer to buf_src) v2 Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> * Apply suggestions from @johannesgaessler code review Co-authored-by: Johannes Gäßler <johannesg@5d6.de> * Adds single-GPU synchronizations to multi-GPU settings to fix hip backend pipeline parallel bugs. * Scheduler Hardening: Exclude hip/MUSA from copy_from_host CPU split -> GPU split optimization * Scheduler Hardening: Re-adding original additional synchronizations for non-async backends * Adds disclaimer to hip/musa exclusion of copy_from_host. Highlights that it is out of precaution, but that no perf-impact is visible, and that it can be revisited separately anytime. --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
devops : add llama in all docker images (#25035 )
2026-06-26 23:57:40 +02:00 · 2026-06-26 17:18:30 +03:00 · 2026-06-26 15:15:48 +02:00 · 2026-06-26 14:36:03 +02:00 · 2026-06-26 15:07:19 +03:00 · 2026-06-26 15:04:42 +03:00
300 changed files with 18414 additions and 13109 deletions
@@ -145,7 +145,7 @@ ENTRYPOINT ["/app/tools.sh"]
 # ==============================================================================
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 ENTRYPOINT [ "/app/llama-cli" ]

@@ -156,7 +156,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 HEALTHCHECK --interval=5m CMD [ "curl", "-f", "http://localhost:8080/health" ]

@@ -104,7 +104,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -115,7 +115,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -113,7 +113,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -124,7 +124,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -141,7 +141,7 @@ ENTRYPOINT ["/app/tools.sh"]
 FROM base AS light

 COPY --from=build /app/lib/ /app
-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -153,7 +153,7 @@ FROM base AS server
 ENV LLAMA_ARG_HOST=0.0.0.0

 COPY --from=build /app/lib/ /app
-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -115,7 +115,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -126,7 +126,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -1,12 +1,12 @@
-ARG OPENVINO_VERSION_MAJOR=2026.2
-ARG OPENVINO_VERSION_FULL=2026.2.0.21903.52ddc073857
+ARG OPENVINO_VERSION_MAJOR=2026.2.1
+ARG OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3
 ARG UBUNTU_VERSION=24.04

 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
-ARG IGC_VERSION=v2.34.4
-ARG IGC_VERSION_FULL=2_2.34.4+21428
-ARG COMPUTE_RUNTIME_VERSION=26.18.38308.1
-ARG COMPUTE_RUNTIME_VERSION_FULL=26.18.38308.1-0
+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 IGDGMM_VERSION=22.10.0

 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
@@ -214,7 +214,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app/
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app/

 WORKDIR /app

@@ -225,7 +225,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app/
+COPY --from=build /app/full/llama /app/full/llama-server /app/

 WORKDIR /app

@@ -127,7 +127,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -138,7 +138,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -124,7 +124,7 @@ WORKDIR /llama.cpp/bin

 # Copy llama.cpp binaries and libraries
 COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
-COPY --from=collector /llama.cpp/bin/llama-cli /llama.cpp/bin/llama-completion /llama.cpp/bin
+COPY --from=collector /llama.cpp/bin/llama /llama.cpp/bin/llama-cli /llama.cpp/bin/llama-completion /llama.cpp/bin

 ENTRYPOINT [ "/llama.cpp/bin/llama-cli" ]

@@ -138,7 +138,7 @@ WORKDIR /llama.cpp/bin

 # Copy llama.cpp binaries and libraries
 COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
-COPY --from=collector /llama.cpp/bin/llama-server /llama.cpp/bin
+COPY --from=collector /llama.cpp/bin/llama /llama.cpp/bin/llama-server /llama.cpp/bin

 EXPOSE 8080

@@ -107,7 +107,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -118,7 +118,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -97,7 +97,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light

-COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
+COPY --from=build /app/full/llama /app/full/llama-cli /app/full/llama-completion /app

 WORKDIR /app

@@ -108,7 +108,7 @@ FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

-COPY --from=build /app/full/llama-server /app
+COPY --from=build /app/full/llama /app/full/llama-server /app

 WORKDIR /app

@@ -35,8 +35,20 @@ AMD ZenDNN:
 documentation:
    - changed-files:
        - any-glob-to-any-file:
+            - "**/*.md"
            - docs/**
            - media/**
+examples:
+    - all:
+        - changed-files:
+            - any-glob-to-any-file:
+                - app/**
+                - examples/**
+                - tools/**
+            - all-globs-to-all-files:
+                - '!tools/server/**'
+                - '!tools/mtmd/**'
+                - '!tools/ui/**'
 testing:
    - changed-files:
        - any-glob-to-any-file:
@@ -47,28 +59,12 @@ build:
            - cmake/**
            - CMakeLists.txt
            - CMakePresets.json
-examples:
-    - changed-files:
-        - any-glob-to-any-file:
-            - examples/**
-            - tools/**
 devops:
    - changed-files:
        - any-glob-to-any-file:
            - .devops/**
            - .github/**
            - ci/**
-python:
-    - changed-files:
-        - any-glob-to-any-file:
-            - "**/*.py"
-            - requirements/**
-            - gguf-py/**
-            - .flake8
-script:
-    - changed-files:
-        - any-glob-to-any-file:
-            - scripts/**
 android:
    - changed-files:
        - any-glob-to-any-file:
@@ -81,9 +77,20 @@ server:
    - changed-files:
        - any-glob-to-any-file:
            - tools/server/**
-
-
-
+mtmd:
+    - changed-files:
+        - any-glob-to-any-file:
+            - tools/mtmd/**
+conversion:
+    - changed-files:
+        - any-glob-to-any-file:
+            - conversion/**
+            - convert_*.py
+            - gguf-py/**
+vendor:
+    - changed-files:
+        - any-glob-to-any-file:
+            - vendor/**
 ggml:
    - changed-files:
        - any-glob-to-any-file:
@@ -68,8 +68,8 @@ jobs:

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"

    steps:
      - name: Clone
@@ -96,8 +96,8 @@ jobs:

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"

    steps:
      - name: Clone
@@ -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"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"

    steps:
      - name: Clone
@@ -96,8 +96,8 @@ 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"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"

    steps:
      - name: Clone
@@ -266,8 +266,8 @@ jobs:

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"

    steps:
      - name: Clone
@@ -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"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"

    steps:
      - name: Set OpenVINO version output
@@ -506,8 +506,11 @@ jobs:
          cmake -B build/ReleaseOV -G Ninja \
            -DCMAKE_BUILD_TYPE=Release \
            -DGGML_OPENVINO=ON \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }}
-          cmake --build build/ReleaseOV --config Release -j $(nproc)
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build/ReleaseOV --config Release --parallel

      - name: ccache-clear
        uses: ./.github/actions/ccache-clear
@@ -521,8 +524,26 @@ jobs:
      - name: Pack artifacts
        id: pack_artifacts
        run: |
-          cp LICENSE ./build/ReleaseOV/bin/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./build/ReleaseOV/bin .
+          dest=./build/ReleaseOV/bin
+          OPENVINO_ROOT=./openvino_toolkit
+          ov_lib="$OPENVINO_ROOT/runtime/lib/intel64"
+
+          # Bundle OpenVINO runtime libs + TBB. Binaries built with RPATH=$ORIGIN
+          # load these siblings without setupvars.sh / LD_LIBRARY_PATH.
+          cp -P "$ov_lib"/libopenvino.so* \
+                "$ov_lib"/libopenvino_c.so* \
+                "$ov_lib"/libopenvino_*_plugin.so \
+                "$ov_lib"/libopenvino_intel_npu_compiler*.so \
+                "$OPENVINO_ROOT"/runtime/3rdparty/tbb/lib/*.so* \
+                "$dest"
+          cp -P /usr/lib/x86_64-linux-gnu/libOpenCL.so.1* "$dest" 2>/dev/null || true
+          cp "$ov_lib"/cache.json "$dest" 2>/dev/null || true
+
+          # OpenVINO licensing
+          cp -r "$OPENVINO_ROOT"/docs/licensing "$dest"/openvino-licensing
+
+          cp LICENSE "$dest"
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C "$dest" .

      - name: Upload artifacts
        uses: actions/upload-artifact@v6
@@ -538,8 +559,8 @@ 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"
-      OPENVINO_VERSION_FULL: "2026.2.0.21903.52ddc073857"
+      OPENVINO_VERSION_MAJOR: "2026.2.1"
+      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"

    steps:
      - name: Set OpenVINO version output
@@ -607,7 +628,9 @@ jobs:
            -A x64 ^
            -DCMAKE_BUILD_TYPE=Release ^
            -DGGML_OPENVINO=ON ^
-            -DCMAKE_TOOLCHAIN_FILE=C:\vcpkg\scripts\buildsystems\vcpkg.cmake
+            -DLLAMA_BUILD_BORINGSSL=ON ^
+            -DCMAKE_TOOLCHAIN_FILE=C:\vcpkg\scripts\buildsystems\vcpkg.cmake ^
+            ${{ env.CMAKE_ARGS }}

          cmake --build build\ReleaseOV --config Release -- /m

@@ -624,8 +647,29 @@ jobs:
        id: pack_artifacts
        shell: powershell
        run: |
-          Copy-Item LICENSE .\build\ReleaseOV\bin\
-          7z a -snl llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip .\build\ReleaseOV\bin\*
+          # Locate the extracted OpenVINO toolkit root (same pattern as the Build step).
+          $OPENVINO_ROOT = (Get-ChildItem -Directory openvino_toolkit | Select-Object -First 1).FullName
+          if (-not $OPENVINO_ROOT) {
+            Write-Error "OpenVINO toolkit folder not found under .\openvino_toolkit"
+            exit 1
+          }
+
+          $dest = ".\build\ReleaseOV\bin\Release"
+
+          $ovBin = Join-Path $OPENVINO_ROOT 'runtime\bin\intel64\Release'
+          Copy-Item -Path (Join-Path $ovBin '*.dll')       -Destination $dest -Force
+          Copy-Item -Path (Join-Path $ovBin 'cache.json')  -Destination $dest -Force
+
+          $tbbBin = Join-Path $OPENVINO_ROOT 'runtime\3rdparty\tbb\bin'
+          Copy-Item -Path (Join-Path $tbbBin 'tbb*.dll') -Destination $dest -Force
+
+          # OpenVINO licensing
+          $licensingDest = Join-Path $dest 'openvino-licensing'
+          New-Item -ItemType Directory -Force -Path $licensingDest | Out-Null
+          Copy-Item -Path (Join-Path $OPENVINO_ROOT 'docs\licensing\*') -Destination $licensingDest -Recurse -Force
+
+          Copy-Item LICENSE $dest
+          7z a -snl llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip $dest\*

      - name: Upload artifacts
        uses: actions/upload-artifact@v6
@@ -222,6 +222,16 @@ if (LLAMA_BUILD_APP)
    add_subdirectory(app)
 endif()

+# Standalone libmtmd build without pulling in the rest of the tools/ tree.
+# Useful when packaging just the mtmd library for language bindings (e.g. an
+# Apple XCFramework, or a WASM build). When the full tools build is enabled,
+# mtmd is already built by the tools/ subdirectory above; this hook only fires
+# when LLAMA_BUILD_TOOLS is OFF to avoid double-adding the target.
+option(LLAMA_BUILD_MTMD "llama: build tools/mtmd library standalone" OFF)
+if (LLAMA_BUILD_MTMD AND NOT (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS))
+    add_subdirectory(tools/mtmd)
+endif()
+
 #
 # install
 #
@@ -10,7 +10,7 @@
 # ggml-org/ggml-rpc         : rgerganov
 # ggml-org/ggml-sycl        : arthw
 # ggml-org/ggml-vulkan      : 0cc4m, jeffbolznv
-# ggml-org/ggml-webgpu      : reeselevine
+# ggml-org/ggml-webgpu      : reeselevine, yomaytk
 # ggml-org/ggml-zdnn        : taronaeo
 # ggml-org/llama-common     : ggerganov, aldehir, angt, danbev, ngxson, pwilkin
 # ggml-org/llama-mtmd       : ngxson
@@ -142,7 +142,9 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
 - [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview)
 - [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
- [x] [LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38)
+- [x] [Liquid LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2)
+- [x] [Liquid LFM2.5 models](https://huggingface.co/collections/LiquidAI/lfm25)
+- [x] [Liquid Nanos](https://huggingface.co/collections/LiquidAI/liquid-nanos)
 - [x] [Hunyuan models](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7)
 - [x] [BailingMoeV2 (Ring/Ling 2.0) models](https://huggingface.co/collections/inclusionAI/ling-v2-68bf1dd2fc34c306c1fa6f86)
 - [x] [Mellum models](https://huggingface.co/JetBrains/models?search=mellum)
@@ -1,6 +1,6 @@
 set(TARGET llama-app)

-add_executable(${TARGET} llama.cpp)
+add_executable(${TARGET} llama.cpp download.cpp)
 set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama)

 target_link_libraries(${TARGET} PRIVATE
@@ -0,0 +1,71 @@
+#include "arg.h"
+#include "common.h"
+#include "download.h"
+#include "log.h"
+
+#include <cstdio>
+#include <filesystem>
+
+static void print_usage(int /*argc*/, char ** argv) {
+    printf(
+        "\nexamples:\n"
+        "  %s -hf ggml-org/gemma-3-4b-it-qat-GGUF\n"
+        "  %s -hf ggml-org/gemma-3-4b-it-qat-GGUF:Q4_K_M\n"
+        "  %s -hf ggml-org/models -hff model.gguf\n"
+        "  %s -mu https://example.com/model.gguf -m model.gguf\n"
+        "\n",
+        argv[0], argv[0], argv[0], argv[0]
+    );
+}
+
+int llama_download(int argc, char ** argv);
+
+int llama_download(int argc, char ** argv) {
+    common_init();
+
+    common_params params;
+    params.verbosity = LOG_LEVEL_ERROR;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_DOWNLOAD, print_usage)) {
+        return 1;
+    }
+
+    const bool has_source = !params.model.hf_repo.empty() || !params.model.url.empty() ||
+                            !params.model.path.empty()    || !params.model.docker_repo.empty();
+    if (!has_source) {
+        fprintf(stderr, "error: no model source specified (use --hf-repo, --model-url, --model or --docker-repo)\n");
+        return 1;
+    }
+
+    try {
+        common_models_handler handler = common_models_handler_init(params, LLAMA_EXAMPLE_DOWNLOAD);
+        common_models_handler_apply(handler, params);
+    } catch (const std::exception & e) {
+        fprintf(stderr, "error: %s\n", e.what());
+        return 1;
+    }
+
+    if (!params.models_preset.empty()) {
+        // -hf pointed at a preset repo: print the preset path and stop
+        printf("%s\n", params.models_preset.c_str());
+        return 0;
+    }
+    if (params.model.path.empty()) {
+        fprintf(stderr, "error: model download failed\n");
+        return 1;
+    }
+    if (!std::filesystem::exists(params.model.path)) {
+        fprintf(stderr, "error: model file does not exist: %s\n", params.model.path.c_str());
+        return 1;
+    }
+
+    printf("%s\n", params.model.path.c_str());
+    if (!params.mmproj.path.empty()) {
+        printf("%s\n", params.mmproj.path.c_str());
+    }
+    if (!params.speculative.draft.mparams.path.empty()) {
+        printf("%s\n", params.speculative.draft.mparams.path.c_str());
+    }
+
+    return 0;
+}
@@ -19,6 +19,7 @@ int llama_batched_bench(int argc, char ** argv);
 int llama_fit_params(int argc, char ** argv);
 int llama_quantize(int argc, char ** argv);
 int llama_perplexity(int argc, char ** argv);
+int llama_download(int argc, char ** argv);

 // Self-update is only supported for binaries built with llama-install.sh
 static int llama_update(int argc, char ** argv) {
@@ -61,6 +62,7 @@ static const command cmds[] = {
    {"serve",         "HTTP API server",                                    {"server"},   false,         llama_server       },
    {"cli",           "Command-line interactive interface",                 {"client"},   false,         llama_cli          },
    {"update",        "Update llama to the latest release",                 {},           UPDATE_HIDDEN, llama_update       },
+    {"download",      "Download a model",                                   {"get"},      false,         llama_download     },
    {"completion",    "Text completion",                                    {"complete"}, true,          llama_completion   },
    {"bench",         "Benchmark prompt processing and text generation",    {},           true,          llama_bench        },
    {"batched-bench", "Benchmark batched decoding performance",             {},           true,          llama_batched_bench},
@@ -13,6 +13,7 @@ LLAMA_BUILD_EXAMPLES=OFF
 LLAMA_BUILD_TOOLS=OFF
 LLAMA_BUILD_TESTS=OFF
 LLAMA_BUILD_SERVER=OFF
+LLAMA_BUILD_MTMD=ON
 GGML_METAL=ON
 GGML_METAL_EMBED_LIBRARY=ON
 GGML_BLAS_DEFAULT=ON
@@ -39,6 +40,7 @@ COMMON_CMAKE_ARGS=(
    -DLLAMA_BUILD_TOOLS=${LLAMA_BUILD_TOOLS}
    -DLLAMA_BUILD_TESTS=${LLAMA_BUILD_TESTS}
    -DLLAMA_BUILD_SERVER=${LLAMA_BUILD_SERVER}
+    -DLLAMA_BUILD_MTMD=${LLAMA_BUILD_MTMD}
    -DGGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY}
    -DGGML_BLAS_DEFAULT=${GGML_BLAS_DEFAULT}
    -DGGML_METAL=${GGML_METAL}
@@ -126,6 +128,8 @@ setup_framework_structure() {
    cp ggml/include/ggml-cpu.h     ${header_path}
    cp ggml/include/ggml-blas.h    ${header_path}
    cp ggml/include/gguf.h         ${header_path}
+    cp tools/mtmd/mtmd.h           ${header_path}
+    cp tools/mtmd/mtmd-helper.h    ${header_path}

    # Create module map (common for all platforms)
    cat > ${module_path}module.modulemap << EOF
@@ -247,6 +251,7 @@ combine_static_libraries() {
        "${base_dir}/${build_dir}/ggml/src/${release_dir}/libggml-cpu.a"
        "${base_dir}/${build_dir}/ggml/src/ggml-metal/${release_dir}/libggml-metal.a"
        "${base_dir}/${build_dir}/ggml/src/ggml-blas/${release_dir}/libggml-blas.a"
+        "${base_dir}/${build_dir}/tools/mtmd/${release_dir}/libmtmd.a"
    )

    # Create temporary directory for processing
@@ -410,6 +415,7 @@ cmake -B build-ios-sim -G Xcode \
    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
    -DLLAMA_OPENSSL=OFF \
+    -DMTMD_VIDEO=OFF \
    -S .
 cmake --build build-ios-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet

@@ -424,6 +430,7 @@ cmake -B build-ios-device -G Xcode \
    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
    -DLLAMA_OPENSSL=OFF \
+    -DMTMD_VIDEO=OFF \
    -S .
 cmake --build build-ios-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet

@@ -450,6 +457,7 @@ cmake -B build-visionos -G Xcode \
    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
    -DLLAMA_OPENSSL=OFF \
    -DLLAMA_BUILD_SERVER=OFF \
+    -DMTMD_VIDEO=OFF \
    -S .
 cmake --build build-visionos --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet

@@ -465,6 +473,7 @@ cmake -B build-visionos-sim -G Xcode \
    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
    -DLLAMA_OPENSSL=OFF \
    -DLLAMA_BUILD_SERVER=OFF \
+    -DMTMD_VIDEO=OFF \
    -S .
 cmake --build build-visionos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet

@@ -481,6 +490,7 @@ cmake -B build-tvos-sim -G Xcode \
    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
    -DLLAMA_OPENSSL=OFF \
+    -DMTMD_VIDEO=OFF \
    -S .
 cmake --build build-tvos-sim --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet

@@ -496,6 +506,7 @@ cmake -B build-tvos-device -G Xcode \
    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
    -DLLAMA_OPENSSL=OFF \
+    -DMTMD_VIDEO=OFF \
    -S .
 cmake --build build-tvos-device --config Release -j $(sysctl -n hw.logicalcpu) -- -quiet

@@ -80,8 +80,6 @@ add_library(${TARGET}
    http.h
    imatrix-loader.cpp
    imatrix-loader.h
-    json-partial.cpp
-    json-partial.h
    json-schema-to-grammar.cpp
    llguidance.cpp
    log.cpp
@@ -297,58 +297,6 @@ struct handle_model_result {
    std::string preset_path;
 };

-static handle_model_result common_params_handle_model(struct common_params_model & model,
-                                                      const common_download_opts & opts) {
-    handle_model_result result;
-
-    if (!model.docker_repo.empty()) {
-        model.path = common_docker_resolve_model(model.docker_repo);
-    } else if (!model.hf_repo.empty()) {
-        // If -m was used with -hf, treat the model "path" as the hf_file to download
-        if (model.hf_file.empty() && !model.path.empty()) {
-            model.hf_file = model.path;
-            model.path = "";
-        }
-        common_download_opts hf_opts = opts;
-        auto download_result = common_download_model(model, hf_opts);
-
-        if (!download_result.preset_path.empty()) {
-            result.found_preset = true;
-            result.preset_path = download_result.preset_path;
-            return result; // skip everything else if preset.ini is used
-        }
-
-        if (download_result.model_path.empty()) {
-            throw std::runtime_error("failed to download model from Hugging Face");
-        }
-
-        model.path = download_result.model_path;
-
-        if (!download_result.mmproj_path.empty()) {
-            result.found_mmproj = true;
-            result.mmproj.path  = download_result.mmproj_path;
-        }
-
-        if (!download_result.mtp_path.empty()) {
-            result.found_mtp = true;
-            result.mtp.path  = download_result.mtp_path;
-        }
-    } else if (!model.url.empty()) {
-        if (model.path.empty()) {
-            auto f = string_split<std::string>(model.url, '#').front();
-            f = string_split<std::string>(f, '?').front();
-            model.path = fs_get_cache_file(string_split<std::string>(f, '/').back());
-        }
-
-        auto download_result = common_download_model(model, opts);
-        if (download_result.model_path.empty()) {
-            throw std::runtime_error("failed to download model from " + model.url);
-        }
-    }
-
-    return result;
-}
-
 const std::vector<ggml_type> kv_cache_types = {
    GGML_TYPE_F32,
    GGML_TYPE_F16,
@@ -393,72 +341,241 @@ static bool parse_bool_value(const std::string & value) {
 }

 //
-// CLI argument parsing functions
+// common_models_handler
 //

-bool common_params_handle_models(common_params & params, llama_example curr_ex) {
-    const bool spec_type_draft_mtp = std::find(params.speculative.types.begin(),
-                                         params.speculative.types.end(),
-                                         COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
+static std::string get_default_local_path(const std::string & url) {
+    auto f = string_split<std::string>(url, '#').front();
+    f = string_split<std::string>(f, '?').front();
+    return fs_get_cache_file(string_split<std::string>(f, '/').back());
+}

+common_models_handler common_models_handler_init(const common_params & params, llama_example curr_ex) {
+    common_download_hf_plan plan;
+    common_download_hf_plan plan_spec;
+    common_download_hf_plan plan_voc;
    common_download_opts opts;
+
+    const bool spec_type_draft_mtp = std::find(params.speculative.types.begin(),
+                                        params.speculative.types.end(),
+                                        COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
+
+    // only download mmproj if the current example is using it
+    bool use_mmproj = false;
+    for (const auto & ex : mmproj_examples) {
+        if (curr_ex == ex) {
+            use_mmproj = true;
+            break;
+        }
+    }
+
    opts.bearer_token    = params.hf_token;
    opts.offline         = params.offline;
-    opts.skip_download   = params.skip_download;
    opts.download_mtp    = spec_type_draft_mtp;
-    opts.download_mmproj = !params.no_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty();
+    opts.download_mmproj = use_mmproj && !params.no_mmproj
+                        && params.mmproj.path.empty() && params.mmproj.url.empty();

-    // sub-models (draft, mmproj, vocoder) are explicitly specified by the user,
-    // so we should not auto-discover mtp/mmproj siblings for them
-    common_download_opts sub_opts = opts;
-    sub_opts.download_mtp    = false;
-    sub_opts.download_mmproj = false;
+    if (!params.model.hf_repo.empty()) {
+        plan = common_download_get_hf_plan(params.model, opts);
+    }

-    try {
-        auto res = common_params_handle_model(params.model, opts);
-        if (res.found_preset) {
-            if (!params.models_preset.empty()) {
-                throw std::invalid_argument("cannot use both --models-preset and -hf with a preset.ini file");
+    if (!params.speculative.draft.mparams.hf_repo.empty()) {
+        plan_spec = common_download_get_hf_plan(params.speculative.draft.mparams, opts);
+    }
+
+    if (!params.vocoder.model.hf_repo.empty()) {
+        plan_voc = common_download_get_hf_plan(params.vocoder.model, opts);
+    }
+
+    return common_models_handler{plan, plan_spec, plan_voc, opts};
+}
+
+bool common_models_handler_is_preset_repo(const common_models_handler & handler) {
+    return !handler.plan.preset.url.empty();
+}
+
+static std::vector<common_download_task> build_url_tasks(const common_params_model & model, common_download_opts opts) {
+    auto parts = common_download_get_all_parts(model.url);
+    std::vector<common_download_task> tasks;
+
+    // single-part: download straight to model.path if the user gave one (-m), else the cache default
+    if (parts.size() == 1) {
+        common_download_task task;
+        task.url        = parts[0];
+        task.local_path = model.path.empty() ? get_default_local_path(parts[0]) : model.path;
+        task.opts       = opts;
+        tasks.push_back(std::move(task));
+        return tasks;
+    }
+
+    // multi-part: place each part under the user's -m directory (if given), else the cache default
+    std::string base_dir;
+    if (!model.path.empty()) {
+        auto pos = model.path.rfind('/');
+        base_dir = pos == std::string::npos ? std::string(".") : model.path.substr(0, pos);
+    }
+
+    for (const auto & part : parts) {
+        common_download_task task;
+        task.url  = part;
+        task.opts = opts;
+
+        std::string local = get_default_local_path(part);
+        if (!base_dir.empty()) {
+            auto pos = local.rfind('/');
+            std::string name = pos == std::string::npos ? local : local.substr(pos + 1);
+            local = base_dir + "/" + name;
+        }
+        task.local_path = local;
+        tasks.push_back(std::move(task));
+    }
+    return tasks;
+}
+
+void common_models_handler_apply(common_models_handler & handler, common_params & params, common_download_callback * callback) {
+    std::vector<common_download_task> tasks;
+
+    auto & plan      = handler.plan;
+    auto & plan_spec = handler.plan_spec;
+    auto & plan_voc  = handler.plan_voc;
+
+    auto opts = handler.opts; // copy
+    opts.callback = callback;
+
+    // handle plain "url" if needed
+    auto handle_url = [&](common_params_model & model) {
+        if (!model.url.empty()) {
+            if (model.path.empty()) {
+                model.path = get_default_local_path(model.url);
            }
+        }
+    };
+    handle_url(params.model);
+    handle_url(params.mmproj);
+    handle_url(params.vocoder.model);
+    handle_url(params.speculative.draft.mparams);
+
+    // optionally, if docker repo is set, resolve it
+    if (!params.model.docker_repo.empty()) {
+        params.model.url  = common_docker_resolve_model(params.model.docker_repo);
+        params.model.path = get_default_local_path(params.model.url);
+    }
+
+    // handle plain "url" tasks (non-hf)
+    if (!params.model.url.empty()) {
+        auto url_tasks = build_url_tasks(params.model, opts);
+        // the first part is what gets loaded, so point params.model.path at it
+        if (!url_tasks.empty()) {
+            std::string first_path = url_tasks.front().local_path;
+            url_tasks.front().on_done = [&]() { params.model.path = first_path; };
+        }
+        for (auto & task : url_tasks) {
+            tasks.push_back(std::move(task));
+        }
+    }
+    if (!params.mmproj.url.empty()) {
+        common_download_task task;
+        task.url        = params.mmproj.url;
+        task.local_path = params.mmproj.path;
+        task.opts       = opts;
+        tasks.push_back(task);
+    }
+    if (!params.vocoder.model.url.empty()) {
+        common_download_task task;
+        task.url        = params.vocoder.model.url;
+        task.local_path = params.vocoder.model.path;
+        task.opts       = opts;
+        tasks.push_back(task);
+    }
+    if (!params.speculative.draft.mparams.url.empty()) {
+        common_download_task task;
+        task.url        = params.speculative.draft.mparams.url;
+        task.local_path = params.speculative.draft.mparams.path;
+        task.opts       = opts;
+        tasks.push_back(task);
+    }
+
+    // handle hf_plan tasks
+    auto add_tasks = [&opts, &tasks](const hf_cache::hf_files & model_files, common_params_model & model) {
+        for (size_t i = 0; i < model_files.size(); ++i) {
+            auto & model_file = model_files[i];
+            bool is_first = (i == 0);
+            tasks.emplace_back(model_file, opts, [&, is_first]() {
+                if (is_first) {
+                    // only use first part as model path
+                    model.path = hf_cache::finalize_file(model_file);
+                } else {
+                    hf_cache::finalize_file(model_file);
+                }
+            });
+        }
+    };
+    if (!plan.model_files.empty()) {
+        add_tasks(plan.model_files, params.model);
+    }
+    if (!plan.mmproj.local_path.empty()) {
+        tasks.emplace_back(plan.mmproj, opts, [&]() {
+            params.mmproj.path = hf_cache::finalize_file(plan.mmproj);
+        });
+    }
+    if (!plan.mtp.local_path.empty()) {
+        tasks.emplace_back(plan.mtp, opts, [&]() {
+            // only fall back to the discovered MTP head when no draft was explicitly provided
+            if (params.speculative.draft.mparams.empty()) {
+                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan.mtp);
+            } else {
+                hf_cache::finalize_file(plan.mtp);
+            }
+        });
+    }
+    if (!plan.preset.local_path.empty()) {
+        tasks.emplace_back(plan.preset, opts, [&]() {
            // if HF repo is a preset repo, we simply run server in router mode with the preset.ini file
            params.models_preset_hf = params.model.hf_repo; // only for showing a warning
-            params.models_preset    = res.preset_path;
+            params.models_preset    = hf_cache::finalize_file(plan.preset);
            params.model = common_params_model{}; // make sure to clear model, so server starts in router mode
-            return true;
-        }
+        });
+    }

-        if (params.no_mmproj) {
-            params.mmproj = {};
-        } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
-            // optionally, handle mmproj model when -hf is specified
-            params.mmproj = res.mmproj;
-        }
-        // only download mmproj if the current example is using it
-        for (const auto & ex : mmproj_examples) {
-            if (curr_ex == ex) {
-                common_params_handle_model(params.mmproj, sub_opts);
-                break;
+    // handle plan_spec (e.g. --spec-draft-hf)
+    if (!plan_spec.model_files.empty()) {
+        add_tasks(plan_spec.model_files, params.speculative.draft.mparams);
+    }
+
+    // handle vocoder plan (e.g. --hf-repo-v)
+    if (!plan_voc.model_files.empty()) {
+        add_tasks(plan_voc.model_files, params.vocoder.model);
+    }
+
+    // run all tasks in parallel
+    if (!params.offline) {
+        // if duplicated files are found, only download once (but still call on_done for each task)
+        std::unordered_map<std::string, common_download_task *> unique_tasks;
+        for (auto & task : tasks) {
+            auto it = unique_tasks.find(task.local_path);
+            if (it == unique_tasks.end()) {
+                unique_tasks[task.local_path] = &task;
            }
        }
-
-        // when --spec-type mtp is set and no draft model was provided explicitly,
-        // fall back to the MTP head discovered alongside the -hf model
-        if (spec_type_draft_mtp && res.found_mtp &&
-            params.speculative.draft.mparams.path.empty() &&
-            params.speculative.draft.mparams.hf_repo.empty() &&
-            params.speculative.draft.mparams.url.empty()) {
-            params.speculative.draft.mparams.path = res.mtp.path;
+        std::vector<common_download_task> unique_tasks_vec;
+        for (auto & pair : unique_tasks) {
+            unique_tasks_vec.push_back(*pair.second);
+        }
+        common_download_run_tasks(unique_tasks_vec);
+    }
+
+    // download successful, update params with the downloaded paths
+    for (const auto & task : tasks) {
+        if (task.on_done) {
+            task.on_done();
        }
-        common_params_handle_model(params.speculative.draft.mparams, sub_opts);
-        common_params_handle_model(params.vocoder.model,             sub_opts);
-        return true;
-    } catch (const common_skip_download_exception &) {
-        return false;
-    } catch (const std::exception &) {
-        throw;
    }
 }

+//
+// CLI argument parsing functions
+//
+
 static bool common_params_parse_ex(int argc, char ** argv, common_params_context & ctx_arg) {
    common_params & params = ctx_arg.params;

@@ -584,17 +701,22 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
        throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
    }

-    // export_graph_ops loads only metadata
-    const bool skip_model_download = ctx_arg.ex == LLAMA_EXAMPLE_EXPORT_GRAPH_OPS;
+    const bool skip_model_download =
+        // server will call common_params_handle_models() later, so we skip it here
+        ctx_arg.ex == LLAMA_EXAMPLE_SERVER ||
+        // download calls common_params_handle_models() itself and prints the paths
+        ctx_arg.ex == LLAMA_EXAMPLE_DOWNLOAD ||
+        // export_graph_ops loads only metadata
+        ctx_arg.ex == LLAMA_EXAMPLE_EXPORT_GRAPH_OPS;

    if (!skip_model_download) {
        // handle model and download
-        common_params_handle_models(params, ctx_arg.ex);
+        common_models_handler handler = common_models_handler_init(params, ctx_arg.ex);
+        common_models_handler_apply(handler, params);

        // model is required (except for server)
        // TODO @ngxson : maybe show a list of available models in CLI in this case
        if (params.model.path.empty()
-                && ctx_arg.ex != LLAMA_EXAMPLE_SERVER
                && !params.usage
                && !params.completion) {
            throw std::invalid_argument("error: --model is required\n");
@@ -662,15 +784,19 @@ static void common_params_print_usage(common_params_context & ctx_arg) {
            common_options.push_back(&opt);
        }
    }
-    printf("----- common params -----\n\n");
-    print_options(common_options);
-    printf("\n\n----- sampling params -----\n\n");
-    print_options(sampling_options);
-    printf("\n\n----- speculative params -----\n\n");
-    print_options(spec_options);
-    // TODO: maybe convert enum llama_example to string
-    printf("\n\n----- example-specific params -----\n\n");
-    print_options(specific_options);
+    bool first = true;
+    auto print_section = [&](const char * header, std::vector<common_arg *> & options) {
+        if (options.empty()) {
+            return;
+        }
+        printf("%s----- %s -----\n\n", first ? "" : "\n\n", header);
+        first = false;
+        print_options(options);
+    };
+    print_section("common params",           common_options);
+    print_section("sampling params",         sampling_options);
+    print_section("speculative params",      spec_options);
+    print_section("example-specific params", specific_options);
 }

 static void common_params_print_completion(common_params_context & ctx_arg) {
@@ -1070,7 +1196,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     * - if both {LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_*,} are set, we will prioritize the LLAMA_EXAMPLE_* matching current example
     */
    auto add_opt = [&](common_arg arg) {
-        if ((arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) && !arg.is_exclude(ex)) {
+        // download only exposes the handful of args explicitly tagged for it
+        const bool inherit_common = ex != LLAMA_EXAMPLE_DOWNLOAD;
+        if ((arg.in_example(ex) || (inherit_common && arg.in_example(LLAMA_EXAMPLE_COMMON))) && !arg.is_exclude(ex)) {
            ctx_arg.options.push_back(std::move(arg));
        }
    };
@@ -1081,7 +1209,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params) {
            params.usage = true;
        }
-    ));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}));
    add_opt(common_arg(
        {"--version"},
        "show version and build info",
@@ -2203,7 +2331,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params, bool value) {
            params.no_mmproj = !value;
        }
-    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_AUTO"));
+    ).set_examples({LLAMA_EXAMPLE_MTMD, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_MMPROJ_AUTO"));
    add_opt(common_arg(
        {"--mmproj-offload"},
        {"--no-mmproj-offload"},
@@ -2602,14 +2730,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params, const std::string & value) {
            params.model.path = value;
        }
-    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}).set_env("LLAMA_ARG_MODEL"));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_MODEL"));
    add_opt(common_arg(
        {"-mu", "--model-url"}, "MODEL_URL",
        "model download url (default: unused)",
        [](common_params & params, const std::string & value) {
            params.model.url = value;
        }
-    ).set_env("LLAMA_ARG_MODEL_URL"));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_MODEL_URL"));
    add_opt(common_arg(
        { "-dr", "--docker-repo" }, "[<repo>/]<model>[:quant]",
        "Docker Hub model repository. repo is optional, default to ai/. quant is optional, default to :latest.\n"
@@ -2618,7 +2746,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params, const std::string & value) {
            params.model.docker_repo = value;
        }
-    ).set_env("LLAMA_ARG_DOCKER_REPO"));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_DOCKER_REPO"));
    add_opt(common_arg(
        {"-hf", "-hfr", "--hf-repo"}, "<user>/<model>[:quant]",
        "Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.\n"
@@ -2628,14 +2756,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params, const std::string & value) {
            params.model.hf_repo = value;
        }
-    ).set_env("LLAMA_ARG_HF_REPO"));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_HF_REPO"));
    add_opt(common_arg(
        {"-hff", "--hf-file"}, "FILE",
        "Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)",
        [](common_params & params, const std::string & value) {
            params.model.hf_file = value;
        }
-    ).set_env("LLAMA_ARG_HF_FILE"));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("LLAMA_ARG_HF_FILE"));
    add_opt(common_arg(
        {"-hfv", "-hfrv", "--hf-repo-v"}, "<user>/<model>[:quant]",
        "Hugging Face model repository for the vocoder model (default: unused)",
@@ -2656,7 +2784,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params, const std::string & value) {
            params.hf_token = value;
        }
-    ).set_env("HF_TOKEN"));
+    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD}).set_env("HF_TOKEN"));
+    add_opt(common_arg(
+        {"--mtp"},
+        "also download the multi-token prediction (MTP) head, if available (default: unused)",
+        [](common_params & params) {
+            params.speculative.types.push_back(COMMON_SPECULATIVE_TYPE_DRAFT_MTP);
+        }
+    ).set_examples({LLAMA_EXAMPLE_DOWNLOAD}));
    add_opt(common_arg(
        {"--context-file"}, "FNAME",
        "file to load context from (use comma-separated values to specify multiple files)",
@@ -3613,6 +3748,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        "draft model for speculative decoding (default: unused)",
        [](common_params & params, const std::string & value) {
            params.speculative.draft.mparams.path = value;
+            params.speculative.draft.mparams.hf_file = value; // will be used if --spec-draft-hf is set
        }
    ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_MODEL"));
    add_opt(common_arg(
@@ -1,12 +1,14 @@
 #pragma once

 #include "common.h"
+#include "download.h"

 #include <set>
 #include <map>
 #include <string>
 #include <vector>
 #include <cstring>
+#include <memory>

 // pseudo-env variable to identify preset-only arguments
 #define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
@@ -129,11 +131,21 @@ bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<com
 // see: https://github.com/ggml-org/llama.cpp/issues/18163
 void common_params_add_preset_options(std::vector<common_arg> & args);

-// populate model paths (main model, mmproj, etc) from -hf if necessary
-// return true if the model is ready to use
-// throw an exception if there is an error that prevents the model from being used (e.g. network error, model not found, etc)
-// if params.skip_download is true, no downloads will be attempted. return false if the model is invalid or missing (e.g. ETag check failed)
-bool common_params_handle_models(common_params & params, llama_example curr_ex);
+struct common_models_handler {
+    common_download_hf_plan plan;
+    common_download_hf_plan plan_spec;
+    common_download_hf_plan plan_voc;
+    common_download_opts opts;
+};
+
+// initialize downloading opts and hf_plan if needed, but does not download anything yet
+common_models_handler common_models_handler_init(const common_params & params, llama_example curr_ex);
+
+// check if the model is a preset repo (i.e. has a preset file)
+bool common_models_handler_is_preset_repo(const common_models_handler & handler);
+
+// download and update params with the downloaded model path
+void common_models_handler_apply(common_models_handler & handler, common_params & params, common_download_callback * callback = nullptr);

 // initialize argument parser context - used by test-arg-parser and preset
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
@@ -395,10 +395,11 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
                                           arguments.name_suffix) +
                           arguments.value_prefix +
                           (schema_info.resolves_to_string(param_schema) ?
-                                p.tool_arg_string_value(until_suffix) :
-                                p.tool_arg_json_value(p.schema(
-                                    p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, false))) +
-                           p.tool_arg_close(p.literal(arguments.value_suffix)));
+                                p.ac(p.tool_arg_string_value(until_suffix) +
+                                    p.tool_arg_close(p.literal(arguments.value_suffix)), arguments.value_suffix) :
+                                (p.tool_arg_json_value(p.schema(
+                                    p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, false)) +
+                                    p.tool_arg_close(p.literal(arguments.value_suffix)))));

            auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
            if (is_required) {
@@ -90,41 +90,93 @@ std::string common_chat_msg::render_content(const std::string & delimiter) const
    return text;
 }

-std::vector<common_chat_msg_span> common_chat_split_by_role(const std::string & prompt, const std::vector<common_chat_msg_delimiter> & delims) {
-    if (delims.empty() || prompt.empty()) {
-        return {};
+common_chat_role common_chat_role_from_string(const std::string & role) {
+    if (role == "system")    { return COMMON_CHAT_ROLE_SYSTEM;    }
+    if (role == "assistant") { return COMMON_CHAT_ROLE_ASSISTANT; }
+    if (role == "user")      { return COMMON_CHAT_ROLE_USER;      }
+    if (role == "tool")      { return COMMON_CHAT_ROLE_TOOL;      }
+    return COMMON_CHAT_ROLE_UNKNOWN;
+}
+
+const char * common_chat_role_to_string(common_chat_role role) {
+    switch (role) {
+        case COMMON_CHAT_ROLE_SYSTEM:    return "system";
+        case COMMON_CHAT_ROLE_ASSISTANT: return "assistant";
+        case COMMON_CHAT_ROLE_USER:      return "user";
+        case COMMON_CHAT_ROLE_TOOL:      return "tool";
+        case COMMON_CHAT_ROLE_UNKNOWN:   return "";
+    }
+    return "";
+}
+
+json common_chat_msg_delimiters::to_json() const {
+    json result = json::array();
+    for (const auto & d : delimiters) {
+        result.push_back({
+            { "role",      common_chat_role_to_string(d.role) },
+            { "delimiter", d.delimiter                        },
+        });
+    }
+    return result;
+}
+
+common_chat_msg_delimiters common_chat_msg_delimiters_parse(const json & delimiters) {
+    common_chat_msg_delimiters result;
+
+    if (!delimiters.is_array()) {
+        return result;
    }

-    auto parser = build_peg_parser([&](common_peg_parser_builder & p) {
-        std::vector<std::string>       all_delims;
-        std::vector<common_peg_parser> tagged_messages;
-
-        all_delims.reserve(delims.size());
-        tagged_messages.reserve(delims.size());
-        for (const auto & d : delims) {
-            all_delims.push_back(d.delimiter);
+    result.delimiters.reserve(delimiters.size());
+    for (const auto & d : delimiters) {
+        if (!d.is_object()) {
+            continue;
        }
-
-        auto any_delim = p.until_one_of(all_delims);
-        for (const auto & d : delims) {
-            tagged_messages.push_back(p.tag(d.role, p.literal(d.delimiter) + any_delim));
-        }
-
-        return any_delim + p.zero_or_more(p.choice(tagged_messages)) + p.end();
-    });
-
-    common_peg_parse_context ctx(prompt);
-    const auto result = parser.parse(ctx);
-    if (!result.success()) {
-        return {};
+        result.delimiters.push_back({
+            common_chat_role_from_string(d.value("role", std::string())),
+            d.value("delimiter", std::string()),
+        });
    }

-    std::vector<common_chat_msg_span> spans;
-    ctx.ast.visit(result, [&](const common_peg_ast_node & node) {
-        if (!node.tag.empty()) {
-            spans.push_back({ node.tag, node.start, node.end - node.start });
+    return result;
+}
+
+void common_chat_msg_delimiters::tokenize(const llama_vocab * vocab) {
+    for (auto & d : delimiters) {
+        d.tokens = common_tokenize(vocab, d.delimiter, false, true);
+    }
+}
+
+common_chat_msg_spans common_chat_msg_delimiters::split(const llama_tokens & tokens, const std::map<size_t, size_t> & skips) const {
+    std::vector<std::pair<common_chat_role, size_t>> matches;
+
+    auto skip = skips.begin();
+    for (size_t i = 0; i < tokens.size();) {
+        if (skip != skips.end() && i == skip->first) {
+            i += skip->second;
+            ++skip;
+            continue;
        }
-    });
+        for (const auto & d : delimiters) {
+            if (i + d.tokens.size() > tokens.size()) {
+                continue;
+            }
+            if (std::equal(d.tokens.begin(), d.tokens.end(), tokens.begin() + i)) {
+                matches.emplace_back(d.role, i);
+                break;
+            }
+        }
+        i++;
+    }
+
+    matches.emplace_back(COMMON_CHAT_ROLE_UNKNOWN, tokens.size());
+
+    common_chat_msg_spans spans;
+    for (size_t i = 0; i + 1 < matches.size(); i++) {
+        const auto & curr = matches[i];
+        const auto & next = matches[i + 1];
+        spans.add(curr.first, curr.second, next.second - curr.second);
+    }

    return spans;
 }
@@ -1081,13 +1133,13 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp

    data.prompt            = prompt;
    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, /* messages_override= */ adjusted_messages);
-    data.message_spans = common_chat_split_by_role(prompt, {
-        { "assistant", "<|start|>assistant" },
-        { "user",      "<|start|>user"      },
-        { "system",    "<|start|>developer" },
-        { "system",    "<|start|>system"    },
-        { "tool",      "<|start|>functions" },
-    });
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|start|>assistant" },
+        { COMMON_CHAT_ROLE_USER,      "<|start|>user"      },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>developer" },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>system"    },
+        { COMMON_CHAT_ROLE_TOOL,      "<|start|>functions" },
+    };

    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
    data.supports_thinking = true;
@@ -1228,10 +1280,10 @@ static common_chat_params common_chat_params_init_gemma4(const common_chat_templ
        data.prompt += data.generation_prompt;
    }

-    data.message_spans = common_chat_split_by_role(data.prompt, {
-        { "user",      "<|turn>user\n"  },
-        { "assistant", "<|turn>model\n" },
-    });
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_USER,      "<|turn>user"  },
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|turn>model" },
+    };

    data.format            = COMMON_CHAT_FORMAT_PEG_GEMMA4;
    data.supports_thinking  = true;
@@ -2030,15 +2082,15 @@ static common_chat_params common_chat_params_init_cohere2moe(const common_chat_t
        RESULT_START, RESULT_END,
    };

-    // Split the rendered prompt into per-role message spans. Tool results are rendered with the
+    // Declare per-role message delimiters. Tool results are rendered with the
    // system token followed by <|START_TOOL_RESULT|>, so the "tool" delimiter must be listed before
    // the plain "system" one (it is a strict superset, and the role split tries delimiters in order).
-    data.message_spans = common_chat_split_by_role(data.prompt, {
-        { "assistant", GEN_PREFIX },
-        { "user",      TURN_START + USER },
-        { "tool",      TURN_START + SYSTEM + RESULT_START },
-        { "system",    TURN_START + SYSTEM },
-    });
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, GEN_PREFIX },
+        { COMMON_CHAT_ROLE_USER,      TURN_START + USER },
+        { COMMON_CHAT_ROLE_TOOL,      TURN_START + SYSTEM + RESULT_START },
+        { COMMON_CHAT_ROLE_SYSTEM,    TURN_START + SYSTEM },
+    };

    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
@@ -2526,17 +2578,15 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
        autoparser.analyze_template(tmpl);
        auto auto_params = autoparser::peg_generator::generate_parser(tmpl, params, autoparser);

-        std::vector<common_chat_msg_delimiter> delimiters;
+        common_chat_msg_delimiters delimiters;
        if (!autoparser.assistant_start.empty()) {
-            delimiters.push_back({ "assistant", autoparser.assistant_start });
+            delimiters.add(COMMON_CHAT_ROLE_ASSISTANT, autoparser.assistant_start);
        }
        if (!autoparser.user_start.empty()) {
-            delimiters.push_back({ "user", autoparser.user_start });
+            delimiters.add(COMMON_CHAT_ROLE_USER, autoparser.user_start);
        }

-        if (!delimiters.empty()) {
-            auto_params.message_spans = common_chat_split_by_role(auto_params.prompt, delimiters);
-        }
+        auto_params.message_delimiters = std::move(delimiters);

        auto_params.supports_thinking = autoparser.reasoning.mode != autoparser::reasoning_mode::NONE;
        if (auto_params.supports_thinking) {
@@ -2708,5 +2758,9 @@ common_chat_msg common_chat_peg_parse(const common_peg_arena &          src_pars
 std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates) {
    GGML_ASSERT(chat_templates != nullptr);
    GGML_ASSERT(chat_templates->template_default != nullptr);
+    if (chat_templates->template_tool_use != nullptr) {
+        // take the more expressive template when available
+        return chat_templates->template_tool_use->caps.to_map();
+    }
    return chat_templates->template_default->caps.to_map();
 }
@@ -143,15 +143,75 @@ struct common_chat_msg_diff {
    }
 };

+enum common_chat_role {
+    COMMON_CHAT_ROLE_UNKNOWN,
+    COMMON_CHAT_ROLE_SYSTEM,
+    COMMON_CHAT_ROLE_ASSISTANT,
+    COMMON_CHAT_ROLE_USER,
+    COMMON_CHAT_ROLE_TOOL
+};
+
+common_chat_role common_chat_role_from_string(const std::string & role);
+const char *     common_chat_role_to_string(common_chat_role role);
+
 struct common_chat_msg_span {
-    std::string role;
+    common_chat_role role = COMMON_CHAT_ROLE_UNKNOWN;
    std::size_t pos = 0;
    std::size_t len = 0;
+
+    bool valid() const {
+        return role != COMMON_CHAT_ROLE_UNKNOWN;
+    }
+};
+
+struct common_chat_msg_spans {
+    std::vector<common_chat_msg_span> spans;
+
+    void add(common_chat_role role, size_t pos, size_t len) {
+        spans.push_back({ role, pos, len });
+    }
+
+    bool is_user_start(int32_t pos) const {
+        for (auto it = spans.begin(); it != spans.end(); ++it) {
+            if (it->role == COMMON_CHAT_ROLE_USER && pos == (int32_t) it->pos) {
+                return true;
+            }
+        }
+        return false;
+    }
+
+    int32_t last_user_message_pos() const {
+        for (auto it = spans.rbegin(); it != spans.rend(); ++it) {
+            if (it->role == COMMON_CHAT_ROLE_USER) {
+                return (int32_t) it->pos;
+            }
+        }
+        return -1;
+    }
 };

 struct common_chat_msg_delimiter {
-    std::string role;
-    std::string delimiter;
+    common_chat_role role = COMMON_CHAT_ROLE_UNKNOWN;
+    std::string      delimiter;
+    llama_tokens     tokens = {};
+};
+
+struct common_chat_msg_delimiters {
+    std::vector<common_chat_msg_delimiter> delimiters;
+
+    common_chat_msg_delimiters() = default;
+    common_chat_msg_delimiters(std::initializer_list<common_chat_msg_delimiter> delims) : delimiters(delims) {}
+
+    void add(common_chat_role role, const std::string & delimiter) {
+        delimiters.push_back({ role, delimiter });
+    }
+
+    void tokenize(const llama_vocab * vocab);
+
+    // split tokens into message spans. skips maps a start index to a length of a region to jump over without matching
+    common_chat_msg_spans split(const llama_tokens & tokens, const std::map<size_t, size_t> & skips = {}) const;
+
+    nlohmann::ordered_json to_json() const;
 };

 struct common_chat_tool {
@@ -219,7 +279,7 @@ struct common_chat_params {
    std::vector<std::string>            preserved_tokens;
    std::vector<std::string>            additional_stops;
    std::string                         parser;
-    std::vector<common_chat_msg_span>   message_spans;
+    common_chat_msg_delimiters          message_delimiters;
 };

 // per-message parsing syntax
@@ -325,5 +385,4 @@ struct common_chat_prompt_preset {

 common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates);

-std::vector<common_chat_msg_span> common_chat_split_by_role(const std::string & prompt, const std::vector<common_chat_msg_delimiter> & delims);
-
+common_chat_msg_delimiters common_chat_msg_delimiters_parse(const nlohmann::ordered_json & delimiters);
@@ -96,6 +96,7 @@ enum llama_example {
    LLAMA_EXAMPLE_FIT_PARAMS,
    LLAMA_EXAMPLE_RESULTS,
    LLAMA_EXAMPLE_EXPORT_GRAPH_OPS,
+    LLAMA_EXAMPLE_DOWNLOAD,

    LLAMA_EXAMPLE_COUNT,
 };
@@ -290,13 +291,13 @@ struct common_params_sampling {
 };

 struct common_params_model {
-    std::string path        = ""; // model local path                                       // NOLINT
-    std::string url         = ""; // model url to download                                  // NOLINT
-    std::string hf_repo     = ""; // HF repo                                                // NOLINT
-    std::string hf_file     = ""; // HF file                                                // NOLINT
-    std::string docker_repo = ""; // Docker repo                                            // NOLINT
+    std::string path        = ""; // model local path
+    std::string url         = ""; // model url to download
+    std::string hf_repo     = ""; // HF repo
+    std::string hf_file     = ""; // HF file
+    std::string docker_repo = ""; // Docker repo

-    std::string get_name() {
+    std::string get_name() const {
        if (!hf_repo.empty()) {
            return hf_repo;
        }
@@ -305,6 +306,10 @@ struct common_params_model {
        }
        return path;
    }
+
+    bool empty() const {
+        return get_name().empty();
+    }
 };

 // draft-model-based speculative decoding parameters
@@ -367,7 +372,7 @@ struct common_params_speculative {
    common_params_speculative_ngram_cache ngram_cache;

    bool has_dft() const {
-        return !draft.mparams.path.empty() || !draft.mparams.hf_repo.empty();
+        return !draft.mparams.empty();
    }

    uint32_t need_n_rs_seq() const {
@@ -519,7 +524,6 @@ struct common_params {
    int32_t control_vector_layer_start = -1; // layer range for control vector
    int32_t control_vector_layer_end   = -1; // layer range for control vector
    bool    offline                    = false;
-    bool    skip_download              = false; // skip model file downloading

    int32_t ppl_stride      = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
    int32_t ppl_output_type = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
@@ -609,7 +613,7 @@ struct common_params {
    bool    cache_prompt        = true;  // whether to enable prompt caching
    bool    cache_idle_slots    = true;  // save and clear idle slots upon starting a new task
    int32_t n_ctx_checkpoints   = 32;    // max number of context checkpoints per slot
-    int32_t checkpoint_min_step = 256;   // minimum spacing between context checkpoints
+    int32_t checkpoint_min_step = 8192;  // minimum spacing between context checkpoints
    int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.

    std::string hostname      = "127.0.0.1";
@@ -292,10 +292,6 @@ static int common_download_file_single_online(const std::string & url,

    const bool file_exists = std::filesystem::exists(path);

-    if (!file_exists && opts.skip_download) {
-        return -2; // file is missing and download is disabled
-    }
-
    if (file_exists && skip_etag) {
        LOG_DBG("%s: using cached file: %s\n", __func__, path.c_str());
        return 304; // 304 Not Modified - fake cached response
@@ -362,9 +358,6 @@ static int common_download_file_single_online(const std::string & url,
            return 304; // 304 Not Modified - fake cached response
        }
        // pass this point, the file exists but is different from the server version, so we need to redownload it
-        if (opts.skip_download) {
-            return -2; // special code to indicate that the download was skipped due to etag mismatch
-        }
        if (remove(path.c_str()) != 0) {
            LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
            return -1;
@@ -691,19 +684,8 @@ static void list_available_gguf_files(const hf_cache::hf_files & files) {
    }
 }

-struct hf_plan {
-    hf_cache::hf_file primary;
-    hf_cache::hf_files model_files;
-    hf_cache::hf_file mmproj;
-    hf_cache::hf_file mtp;
-    hf_cache::hf_file preset; // if set, only this file is downloaded
-};
-
-static hf_plan get_hf_plan(const common_params_model  & model,
-                           const common_download_opts & opts,
-                           bool download_mmproj,
-                           bool download_mtp) {
-    hf_plan plan;
+common_download_hf_plan common_download_get_hf_plan(const common_params_model & model, const common_download_opts & opts) {
+    common_download_hf_plan plan;
    hf_cache::hf_files all;

    auto [repo, tag] = common_download_split_repo_tag(model.hf_repo);
@@ -752,125 +734,49 @@ static hf_plan get_hf_plan(const common_params_model  & model,
    plan.primary = primary;
    plan.model_files = get_split_files(all, primary);

-    if (download_mmproj) {
+    if (opts.download_mmproj) {
        plan.mmproj = find_best_mmproj(all, primary.path);
    }
-
-    if (download_mtp) {
+    if (opts.download_mtp) {
        plan.mtp = find_best_mtp(all, primary.path);
    }

    return plan;
 }

-struct download_task {
-    std::string url;
-    std::string path;
-};
-
-static std::vector<download_task> get_url_tasks(const common_params_model & model) {
-    auto split = get_gguf_split_info(model.url);
-
-    if (split.count <= 1) {
-        return {{model.url, model.path}};
-    }
-
-    auto filename = split.prefix;
-    if (auto pos = split.prefix.rfind('/'); pos != std::string::npos) {
-        filename = split.prefix.substr(pos + 1);
-    }
-
-    auto parent_path = std::filesystem::path(model.path).parent_path();
-    auto prefix_path = (parent_path / filename).string();
-
-    std::vector<download_task> tasks;
-    for (int i = 1; i <= split.count; i++) {
-        auto suffix = string_format("-%05d-of-%05d.gguf", i, split.count);
-        tasks.push_back({split.prefix + suffix, prefix_path + suffix});
-    }
-    return tasks;
-}
-
-common_download_model_result common_download_model(const common_params_model  & model,
-                                                   const common_download_opts & opts) {
-    common_download_model_result result;
-    std::vector<download_task> tasks;
-    hf_plan hf;
-
-    bool download_mmproj = opts.download_mmproj;
-    bool download_mtp = opts.download_mtp;
-    bool is_hf = !model.hf_repo.empty();
-
-    if (is_hf) {
-        hf = get_hf_plan(model, opts, download_mmproj, download_mtp);
-        if (!hf.preset.path.empty()) {
-            // if preset.ini exists, only download that file alone
-            tasks.push_back({hf.preset.url, hf.preset.local_path});
-        } else {
-            for (const auto & f : hf.model_files) {
-                tasks.push_back({f.url, f.local_path});
-            }
-            if (!hf.mmproj.path.empty()) {
-                tasks.push_back({hf.mmproj.url, hf.mmproj.local_path});
-            }
-            if (!hf.mtp.path.empty()) {
-                tasks.push_back({hf.mtp.url, hf.mtp.local_path});
-            }
-        }
-    } else if (!model.url.empty()) {
-        tasks = get_url_tasks(model);
-    } else {
-        result.model_path = model.path;
-        return result;
-    }
-
-    if (tasks.empty()) {
-        return result;
-    }
-
+void common_download_run_tasks(const std::vector<common_download_task> & tasks) {
    std::vector<std::future<int>> futures;
    for (const auto & task : tasks) {
        futures.push_back(std::async(std::launch::async,
-            [&task, &opts, is_hf]() {
-                return common_download_file_single(task.url, task.path, opts, is_hf);
+            [&task]() {
+                return common_download_file_single(task.url, task.local_path, task.opts, task.is_hf);
            }
        ));
    }

-    for (auto & f : futures) {
-        int status = f.get();
-        if (status == -2 && opts.skip_download) {
-            throw common_skip_download_exception();
-        }
+    for (size_t i = 0; i < futures.size(); ++i) {
+        std::string url = tasks[i].url;
+        int status = futures[i].get();
        bool is_ok = is_http_status_ok(status);
        if (!is_ok) {
-            return {};
+            throw std::runtime_error(string_format("Download '%s' failed with status code: %d", url.c_str(), status));
        }
    }
+}

-    if (is_hf) {
-        if (!hf.preset.path.empty()) {
-            // if preset.ini is used, do not set other paths
-            result.preset_path = hf_cache::finalize_file(hf.preset);
-        } else {
-            for (const auto & f : hf.model_files) {
-                hf_cache::finalize_file(f);
-            }
-            result.model_path = hf.primary.final_path;
+std::vector<std::string> common_download_get_all_parts(const std::string & url) {
+    auto split = get_gguf_split_info(url);

-            if (!hf.mmproj.path.empty()) {
-                result.mmproj_path = hf_cache::finalize_file(hf.mmproj);
-            }
-
-            if (!hf.mtp.path.empty()) {
-                result.mtp_path = hf_cache::finalize_file(hf.mtp);
-            }
-        }
-    } else {
-        result.model_path = model.path;
+    if (split.count <= 1) {
+        return {url};
    }

-    return result;
+    std::vector<std::string> parts;
+    for (int i = 1; i <= split.count; i++) {
+        auto suffix = string_format("-%05d-of-%05d.gguf", i, split.count);
+        parts.push_back(split.prefix + suffix);
+    }
+    return parts;
 }

 //
@@ -1,7 +1,10 @@
 #pragma once

+#include "hf-cache.h"
+
 #include <string>
 #include <vector>
+#include <functional>

 struct common_params_model;

@@ -47,66 +50,40 @@ struct common_cached_model_info {
    }
 };

-// Options for common_download_model and common_download_file_single
+// Options for common_download_file_single
 struct common_download_opts {
    std::string bearer_token;
    common_header_list headers;
    bool offline = false;
-    bool skip_download = false; // if true, only validation is performed, common_skip_download_exception may be thrown if the file is missing or invalid
    bool download_mmproj = false;
    bool download_mtp = false;
    common_download_callback * callback = nullptr;
 };

-// Result of common_download_model
-struct common_download_model_result {
-    std::string model_path;
-    std::string mmproj_path;
-    std::string mtp_path;
-    std::string preset_path;
+struct common_download_task {
+    common_download_opts opts;
+    std::string url;
+    std::string local_path;
+    std::function<void()> on_done;
+    bool is_hf = false;
+
+    common_download_task() = default;
+    common_download_task(hf_cache::hf_file f,
+            const common_download_opts & opts,
+            std::function<void()> on_done = nullptr)
+        : opts(opts), url(f.url), local_path(f.local_path), on_done(on_done), is_hf(true) {}
 };

-// throw if the file is missing or invalid (e.g. ETag check failed)
-struct common_skip_download_exception : public std::runtime_error {
-    common_skip_download_exception() : std::runtime_error("skip download") {}
-};
+void common_download_run_tasks(const std::vector<common_download_task> & tasks);

-// Download model from HuggingFace repo or URL
-//
-// input (via model struct):
-// - model.hf_repo: HF repo with optional tag, see common_download_split_repo_tag
-// - model.hf_file: specific file in the repo (requires hf_repo)
-// - model.url: simple download (used if hf_repo is empty)
-// - model.path: local file path
-//
-// tag matching (for HF repos without model.hf_file):
-// - if tag is specified, searches for GGUF matching that quantization
-// - if no tag, searches for Q4_K_M, then Q4_0, then first available GGUF
-//
-// split GGUF: multi-part files like "model-00001-of-00003.gguf" are automatically
-// detected and all parts are downloaded
-//
-// caching:
-// - HF repos: uses HuggingFace cache
-// - URLs: uses ETag-based caching
-//
-// when opts.offline=true, no network requests are made
-// when download_mmproj=true, searches for mmproj in same directory as model or any parent directory
-// then with the closest quantization bits
-// when download_mtp=true, applies the same sibling search for an MTP-head GGUF
-//
-// returns result with model_path, mmproj_path and mtp_path (empty when not found / on failure)
-common_download_model_result common_download_model(
-    const common_params_model & model,
-    const common_download_opts & opts = {}
-);
+// if url is a multi-part GGUF file, returns all parts, otherwise returns the single file
+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();

 // download single file from url to local path
 // returns status code or -1 on error
-// returns -2 if the download was skipped due to ETag mismatch (file outdated, skip_download=true)
 // skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)
 int common_download_file_single(const std::string & url,
                                const std::string & path,
@@ -123,3 +100,12 @@ std::string common_docker_resolve_model(const std::string & docker);
 // - if tag is present, removes only files matching that tag (and orphaned blobs)
 // returns true if anything was removed
 bool common_download_remove(const std::string & hf_repo_with_tag);
+
+struct common_download_hf_plan {
+    hf_cache::hf_file primary;
+    hf_cache::hf_files model_files;
+    hf_cache::hf_file mmproj;
+    hf_cache::hf_file mtp;
+    hf_cache::hf_file preset; // if set, only this file is downloaded
+};
+common_download_hf_plan common_download_get_hf_plan(const common_params_model & model, const common_download_opts & opts);
@@ -1,324 +0,0 @@
-#include "json-partial.h"
-
-#include "log.h"
-
-#include <nlohmann/json.hpp>
-
-#include <string>
-#include <regex>
-
-using json = nlohmann::ordered_json;
-
-enum common_json_stack_element_type {
-    COMMON_JSON_STACK_ELEMENT_OBJECT,
-    COMMON_JSON_STACK_ELEMENT_KEY,
-    COMMON_JSON_STACK_ELEMENT_ARRAY,
-};
-
-struct common_json_stack_element {
-    common_json_stack_element_type type;
-    std::string key;
-};
-
-bool common_json_parse(
-    const std::string & input,
-    const std::string & healing_marker,
-    common_json & out)
-{
-    std::string::const_iterator it = input.begin();
-    const auto end = input.end();
-    return common_json_parse(it, end, healing_marker, out);
-}
-
-bool common_json_parse(
-    std::string::const_iterator & it,
-    const std::string::const_iterator & end,
-    const std::string & healing_marker,
-    common_json & out)
-{
-    // // https://json.nlohmann.me/features/parsing/sax_interface/
-    struct json_error_locator : public nlohmann::json_sax<json> {
-        std::size_t position;
-        bool found_error;
-        std::string last_token;
-        std::string exception_message;
-        std::vector<common_json_stack_element> stack;
-
-        json_error_locator() : position(0), found_error(false) {}
-
-        bool parse_error(std::size_t position, const std::string & last_token, const json::exception & ex) override { // NOLINT
-            this->position = position - 1;
-            this->found_error = true;
-            this->last_token = last_token;
-            this->exception_message = ex.what();
-            return false;
-        }
-        void close_value() {
-            if (!stack.empty() && (stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY)) {
-                stack.pop_back();
-            }
-        }
-        bool null() override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool boolean(bool) override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool number_integer(number_integer_t) override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool number_unsigned(number_unsigned_t) override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool number_float(number_float_t, const string_t &) override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool string(string_t &) override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool binary(binary_t &) override { // NOLINT
-            close_value();
-            return true;
-        }
-        bool start_object(std::size_t) override { // NOLINT
-            stack.push_back({COMMON_JSON_STACK_ELEMENT_OBJECT, ""});
-            return true;
-        }
-        bool end_object() override {
-            GGML_ASSERT(!stack.empty() && stack.back().type == COMMON_JSON_STACK_ELEMENT_OBJECT);
-            stack.pop_back();
-            close_value();
-            return true;
-        }
-        bool key(string_t & key) override { // NOLINT
-            stack.push_back({COMMON_JSON_STACK_ELEMENT_KEY, key});
-            return true;
-        }
-        bool start_array(std::size_t) override { // NOLINT
-            stack.push_back({COMMON_JSON_STACK_ELEMENT_ARRAY, ""});
-            return true;
-        }
-        bool end_array() override {
-            GGML_ASSERT(!stack.empty() && stack.back().type == COMMON_JSON_STACK_ELEMENT_ARRAY);
-            stack.pop_back();
-            close_value();
-            return true;
-        }
-    };
-    json_error_locator err_loc;
-    auto start = it;
-    json::sax_parse(it, end, &err_loc);
-
-    if (err_loc.found_error) {
-        it = start;
-        auto temptative_end = it + err_loc.position;
-        // LOG_DBG("Error at position %zu (is_end = %s): %s\n", err_loc.position, temptative_end == end ? "true" : "false", err_loc.exception_message.c_str());
-
-        auto input = std::string(it, temptative_end);
-        try {
-            out.json = json::parse(input);
-            // out.json = json::parse(it, temptative_end);
-            it = temptative_end;
-            return true;
-        } catch (const std::exception & ex) {
-            // No, needs healing.
-            LOG_DBG("Failed to parse up to error: %s: <<<%s>>>\n", ex.what(), std::string(it, temptative_end).c_str());
-        }
-        auto can_parse = [](const std::string & str) {
-            try {
-                auto _ = json::parse(str); // NOLINT
-                return true;
-            } catch (const std::exception &) {
-                return false;
-            }
-        };
-        if (!healing_marker.empty() && !err_loc.stack.empty()) {
-            std::string str(it, temptative_end);
-            auto last_non_sp_pos = str.find_last_not_of(" \n\r\t");
-            if (last_non_sp_pos == std::string::npos) {
-                throw std::runtime_error("Cannot heal a truncated JSON that stopped in an unknown location");
-            }
-            auto last_non_sp_char = str[last_non_sp_pos];
-            // Used to detect stops on a number, which may not be complete.
-            auto was_maybe_number = [&]() {
-                if (!str.empty() && std::isspace(str.back())) {
-                    return false;
-                }
-                return std::isdigit(last_non_sp_char) ||
-                    last_non_sp_char == '.' ||
-                    last_non_sp_char == 'e' ||
-                    last_non_sp_char == 'E' ||
-                    last_non_sp_char == '-';
-            };
-
-            std::string closing;
-            for (size_t i = err_loc.stack.size(); i > 0; i--) {
-                auto & el = err_loc.stack[i - 1];
-                if (el.type == COMMON_JSON_STACK_ELEMENT_OBJECT) {
-                    closing += "}";
-                } else if (el.type == COMMON_JSON_STACK_ELEMENT_ARRAY) {
-                    closing += "]";
-                } else if (el.type != COMMON_JSON_STACK_ELEMENT_KEY) {
-                    throw std::runtime_error("Unexpected stack element type");
-                }
-            }
-
-            // Matches a potentially partial unicode escape sequence, e.g. \u, \uX, \uXX, \uXXX, \uXXXX
-            static const std::regex partial_unicode_regex(R"(\\u(?:[0-9a-fA-F](?:[0-9a-fA-F](?:[0-9a-fA-F](?:[0-9a-fA-F])?)?)?)?$)");
-
-            auto is_high_surrogate = [&](const std::string & s) {
-                // Check if a partial of a high surrogate (U+D800-U+DBFF)
-                return s.length() >= 4 &&
-                    s[0] == '\\' && s[1] == 'u' &&
-                    std::tolower(s[2]) == 'd' &&
-                    (s[3] == '8' || s[3] == '9' || std::tolower(s[3]) == 'a' || std::tolower(s[3]) == 'b');
-            };
-
-            // Initialize the unicode marker to a low surrogate to handle the edge case
-            // where a high surrogate (U+D800-U+DBFF) is immediately followed by a
-            // backslash (\)
-            std::string unicode_marker_padding = "udc00";
-            std::smatch last_unicode_seq;
-
-            if (std::regex_search(str, last_unicode_seq, partial_unicode_regex)) {
-                std::smatch second_last_seq;
-                std::string prelude = str.substr(0, last_unicode_seq.position());
-
-                // Pad the escape sequence with 0s until it forms a complete sequence of 6 characters
-                unicode_marker_padding = std::string(6 - last_unicode_seq.length(), '0');
-
-                if (is_high_surrogate(last_unicode_seq.str())) {
-                    // If the sequence is a partial match for a high surrogate, add a low surrogate (U+DC00-U+UDFF)
-                    unicode_marker_padding += "\\udc00";
-                } else if (std::regex_search(prelude, second_last_seq, partial_unicode_regex)) {
-                    if (is_high_surrogate(second_last_seq.str())) {
-                        // If this follows a high surrogate, pad it to be a low surrogate
-                        if (last_unicode_seq.length() == 2) {
-                            unicode_marker_padding = "dc00";
-                        } else if (last_unicode_seq.length() == 3) {
-                            unicode_marker_padding = "c00";
-                        } else {
-                            // The original unicode_marker_padding is already padded with 0s
-                        }
-                    }
-                }
-            }
-
-            const auto & magic_seed = out.healing_marker.marker = healing_marker;//"$llama.cpp.json$";
-
-            if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY) {
-                // We're inside an object value
-                if (last_non_sp_char == ':' && can_parse(str + "1" + closing)) {
-                    // Was about to create an object value
-                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
-                } else if (can_parse(str + ": 1" + closing)) {
-                    str += (out.healing_marker.json_dump_marker = ":\"" + magic_seed) + "\"" + closing;
-                } else if (last_non_sp_char == '{' && can_parse(str + closing)) {
-                    // Was about to create an object
-                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\": 1" + closing;
-                } else if (can_parse(str + "\"" + closing)) {
-                    // Was inside an object value string
-                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\"" + closing;
-                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
-                    // Was inside an object value string after an escape
-                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
-                } else if (can_parse(str + unicode_marker_padding + "\"" + closing)) {
-                    // Was inside an object value string after a partial unicode escape
-                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\"" + closing;
-                } else {
-                    // find last :
-                    auto last_pos = str.find_last_of(':');
-                    if (last_pos == std::string::npos) {
-                        throw std::runtime_error("Cannot heal a truncated JSON that stopped in an unknown location");
-                    }
-                    // Cutting back to opening : for object value
-                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
-                }
-            } else if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_ARRAY) {
-                if ((last_non_sp_char == ',' || last_non_sp_char == '[') && can_parse(str + "1" + closing)) {
-                    // Was about to create an array value
-                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
-                } else if (can_parse(str + "\"" + closing)) {
-                    // Was inside an array value string
-                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\"" + closing;
-                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
-                    // Was inside an array value string after an escape
-                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
-                } else if (can_parse(str + unicode_marker_padding + "\"" + closing)) {
-                    // Was inside an array value string after a partial unicode escape
-                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\"" + closing;
-                } else if (!was_maybe_number() && can_parse(str + ", 1" + closing)) {
-                    // Had just finished a value
-                    str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\"" + closing;
-                } else {
-                    auto last_pos = str.find_last_of("[,");
-                    if (last_pos == std::string::npos) {
-                        throw std::runtime_error("Cannot heal a truncated JSON array stopped in an unknown location");
-                    }
-                    // Cutting back to last [ or , for array value
-                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
-                }
-            } else if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_OBJECT) {
-                if ((last_non_sp_char == '{' && can_parse(str + closing)) ||
-                        (last_non_sp_char == ',' && can_parse(str + "\"\": 1" + closing))) {
-                    // Was about to create an object key+value
-                    str += (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\": 1" + closing;
-                } else if (!was_maybe_number() && can_parse(str + ",\"\": 1" + closing)) {
-                    // Was about to create an object key+value
-                    str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\": 1" + closing;
-                } else if (can_parse(str + "\": 1" + closing)) {
-                    // Was inside an object key string
-                    str += (out.healing_marker.json_dump_marker = magic_seed) + "\": 1" + closing;
-                } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\": 1" + closing)) {
-                    // Was inside an object key string after an escape
-                    str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\": 1" + closing;
-                } else if (can_parse(str + unicode_marker_padding + "\": 1" + closing)) {
-                    // Was inside an object key string after a partial unicode escape
-                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\": 1" + closing;
-                } else {
-                    auto last_pos = str.find_last_of(':');
-                    if (last_pos == std::string::npos) {
-                        throw std::runtime_error("Cannot heal a truncated JSON object stopped in an unknown location");
-                    }
-                    // fprintf(stderr, "Cutting back to last : for object key+value\n");
-                    str = str.substr(0, last_pos + 1) + (out.healing_marker.json_dump_marker = "\"" + magic_seed) + "\"" + closing;
-                }
-            } else {
-                throw std::runtime_error("Cannot heal a truncated JSON object stopped in an unknown location");
-            }
-            // fprintf(stderr, "HEALED:\nSTRING <<<\n%s\n>>>\n\nmagic_cut: <<<\n%s\n>>>\n\n", str.c_str(), out.healing_marker.json_dump_marker.c_str());
-            out.json = json::parse(str);
-            it = temptative_end;
-            return true;
-        }
-        // handle unclosed top-level primitive
-        if (err_loc.position != 0 && !healing_marker.empty() && err_loc.stack.empty()) {
-            std::string str(it, temptative_end);
-            const auto & magic_seed = out.healing_marker.marker = healing_marker;
-            if (can_parse(str + "\"")) {
-                // Was inside an string
-                str += (out.healing_marker.json_dump_marker = magic_seed) + "\"";
-            } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"")) {
-                // Was inside an string after an escape
-                str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"";
-            } else {
-                // TODO: handle more unclosed top-level primitive if the stack was empty but we got an error (e.g. "tru", "\"", etc...)
-                // fprintf(stderr, "Closing: TODO\n");
-                return false;
-            }
-            out.json = json::parse(str);
-            it = temptative_end;
-            return true;
-        }
-        return false;
-    }
-    out.json = json::parse(it, end);
-    it = end;
-    return true;
-}
@@ -1,39 +0,0 @@
-#pragma once
-
-// TODO: use json_fwd.hpp when possible
-#include <nlohmann/json.hpp>
-
-// Healing marker (empty if the JSON was fully parsed / wasn't healed).
-struct common_healing_marker {
-    // Raw marker.
-    std::string marker;
-
-    // Cutting the `common_json.json.dump()` string at the (only) occurrence of this marker should yield the original partial JSON string (modulo spaces / if it had the same dump format).
-    std::string json_dump_marker;
-};
-
-// Represents a parsed JSON object, with its optional healing marker (a JSON dump fragment that can be used to find the position of healing in the JSON dump string)
-struct common_json {
-    nlohmann::ordered_json json;
-
-    common_healing_marker healing_marker;
-};
-
-// Parse the JSON string, healing (closing) any partial JSON if `healing_marker` is not empty.
-//
-// Healing completes partial JSON strings by adding a (possibly modified) healing marker, then whatever is needed to close the JSON.
-// This allows to parse the resulting healed JSON string, yet be able to cut it again if needed at the healing marker.
-// (this is used when parsing JSON outputs from the models, then crafting partial JSONs for the partial tool calls in OAI format).
-//
-// For instance, parsing `{` with a healing marker `foo` will produce a healed JSON `{"foo":1}`, w/ json_dump_marker = `"foo"` (which can be used to break the JSON again).
-bool common_json_parse(
-    const std::string & input,
-    const std::string & healing_marker,
-    common_json & out);
-
-// Parse the JSON string (see overload above), but advancing an iterator to the end of the input when the (potentially partial) parsing succeeds.
-bool common_json_parse(
-    std::string::const_iterator & it,
-    const std::string::const_iterator & end,
-    const std::string & healing_marker,
-    common_json & out);
@@ -921,6 +921,10 @@ struct parser_executor {
    common_peg_parse_result operator()(const common_peg_gbnf_parser & p) {
        return arena.parse(p.child, ctx, start_pos);
    }
+
+    common_peg_parse_result operator()(const common_peg_ac_parser & p) {
+        return arena.parse(p.child, ctx, start_pos);
+    }
 };

 common_peg_parse_result common_peg_arena::parse(common_peg_parse_context & ctx, size_t start) const {
@@ -989,7 +993,8 @@ void common_peg_arena::resolve_refs() {
                                 std::is_same_v<T, common_peg_not_parser> ||
                                 std::is_same_v<T, common_peg_tag_parser> ||
                                 std::is_same_v<T, common_peg_atomic_parser> ||
-                                 std::is_same_v<T, common_peg_gbnf_parser>) {
+                                 std::is_same_v<T, common_peg_gbnf_parser> ||
+                                 std::is_same_v<T, common_peg_ac_parser>) {
                p.child = resolve_ref(p.child);
            } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
                p.child = resolve_ref(p.child);
@@ -1070,6 +1075,8 @@ std::string common_peg_arena::dump_impl(common_peg_parser_id
            return "Atomic(" + dump_impl(p.child, visited) + ")";
        } else if constexpr (std::is_same_v<T, common_peg_gbnf_parser>) {
            return "Gbnf(" + p.grammar + ", " + dump_impl(p.child, visited) + ")";
+        } else if constexpr (std::is_same_v<T, common_peg_ac_parser>) {
+            return "Ac(" + string_join(p.delimiters, " | ") + ", " + dump_impl(p.child, visited) + ")";
        } else if constexpr (std::is_same_v<T, common_peg_any_parser>) {
            return "Any";
        } else if constexpr (std::is_same_v<T, common_peg_space_parser>) {
@@ -1479,6 +1486,13 @@ common_peg_parser common_peg_parser_builder::json_member(const std::string & key
    });
 }

+common_peg_parser common_peg_parser_builder::ac(const common_peg_parser & p, const std::vector<std::string> & delimiters) {
+    if (delimiters.empty()) {
+        throw std::runtime_error("ac parser requires at least one delimiter");
+    }
+    return add(common_peg_ac_parser{p, delimiters});
+}
+
 static std::string gbnf_escape_char_class(uint32_t c) {
    if (c == '-' || c == ']' || c == '[' || c == '\\') {
        return "\\" + std::string(1, (char) c);
@@ -1529,14 +1543,22 @@ static std::string gbnf_escape_char_class(uint32_t c) {
    return std::string(buf);
 }

-// GBNF grammar matching strings that contain no string in `strings` as a
-// substring. Emits the complement of an Aho-Corasick automaton DFA and returns
-// the start state rule name.
-//
-// ref: https://github.com/ggml-org/llama.cpp/pull/24839
-static std::string gbnf_excluding_grammar(const common_grammar_builder & builder,
-                                          const std::string &            prefix,
-                                          const std::vector<std::string> & strings) {
+static std::string gbnf_char_class(const std::vector<uint32_t> & chars, bool negate) {
+    std::string s = negate ? "[^" : "[";
+    for (uint32_t ch : chars) {
+        s += gbnf_escape_char_class(ch);
+    }
+    return s + "]";
+}
+
+static std::string gbnf_ac_grammar(
+    const common_grammar_builder &   builder,
+    const std::string &              prefix,
+    const std::vector<std::string> & strings,
+    const std::function<std::string(const std::vector<uint32_t> &,
+                                    const std::map<size_t, std::vector<uint32_t>> &,
+                                    const std::vector<uint32_t> &,
+                                    const std::function<std::string(size_t)> &)> & build_rule) {
    aho_corasick ac(strings);

    auto state_name = [&](size_t s) -> std::string {
@@ -1548,42 +1570,30 @@ static std::string gbnf_excluding_grammar(const common_grammar_builder & builder
        return prefix + "-" + num;
    };

-    auto char_class = [](const std::vector<uint32_t> & chars, bool negate) {
-        std::string s = negate ? "[^" : "[";
-        for (uint32_t ch : chars) {
-            s += gbnf_escape_char_class(ch);
-        }
-        return s + "]";
-    };
-
    for (size_t q = 0; q < ac.num_states(); q++) {
        if (ac.is_terminal(q)) {
-            continue; // match states are dropped
+            continue; // match states
        }

        std::map<size_t, std::vector<uint32_t>> buckets;
-        std::vector<uint32_t> excluded;
+        std::vector<uint32_t> completing;  // chars that complete a delimiter
+        std::vector<uint32_t> specific;    // chars with an explicit transition
        for (uint32_t c : ac.alphabet) {
            size_t d = ac.next(q, c);
            if (ac.is_terminal(d)) {
-                excluded.push_back(c); // completes a forbidden string -> omit
+                completing.push_back(c);
+                specific.push_back(c);
            } else if (d != 0) {
                buckets[d].push_back(c); // specific non-root destination
-                excluded.push_back(c);
+                specific.push_back(c);
            }
        }

-        std::string rhs = "|"; // every state is accepting
-        for (const auto & [d, chars] : buckets) {
-            rhs += " " + char_class(chars, false) + " " + state_name(d) + " |";
-        }
-        rhs += " " + char_class(excluded, true) + " " + state_name(0);
-
-        builder.add_rule(state_name(q), rhs);
+        builder.add_rule(state_name(q), build_rule(completing, buckets, specific, state_name));
    }

    // An empty delimiter makes the start state terminal. Emit an entry rule
-    // that matches nothing so the returned reference stays valid.
+    // that matches the empty string so the returned reference stays valid.
    if (ac.is_terminal(0)) {
        builder.add_rule(prefix, "|");
    }
@@ -1591,6 +1601,54 @@ static std::string gbnf_excluding_grammar(const common_grammar_builder & builder
    return state_name(0);
 }

+// GBNF grammar matching strings that contain no string in `strings` as a
+// substring. Emits the complement of an Aho-Corasick automaton DFA and returns
+// the start state rule name.
+//
+// ref: https://github.com/ggml-org/llama.cpp/pull/24839
+static std::string gbnf_excluding_grammar(const common_grammar_builder & builder,
+                                          const std::string &            prefix,
+                                          const std::vector<std::string> & strings) {
+    return gbnf_ac_grammar(builder, prefix, strings,
+        [](const std::vector<uint32_t> & /*completing*/,
+           const std::map<size_t, std::vector<uint32_t>> & buckets,
+           const std::vector<uint32_t> & specific,
+           const std::function<std::string(size_t)> & state_name) {
+            // every state is accepting and completing chars get no
+            // alternative, so a forbidden string can never be matched
+            std::string rhs = "|";
+            for (const auto & [d, chars] : buckets) {
+                rhs += " " + gbnf_char_class(chars, false) + " " + state_name(d) + " |";
+            }
+            rhs += " " + gbnf_char_class(specific, true) + " " + state_name(0);
+            return rhs;
+        });
+}
+
+// GBNF grammar matching everything up to and including the first occurrence of
+// any string in `strings`. Emits the Aho-Corasick automaton DFA and returns
+// the start state rule name.
+static std::string gbnf_including_grammar(const common_grammar_builder & builder,
+                                          const std::string &            prefix,
+                                          const std::vector<std::string> & strings) {
+    return gbnf_ac_grammar(builder, prefix, strings,
+        [](const std::vector<uint32_t> & completing,
+           const std::map<size_t, std::vector<uint32_t>> & buckets,
+           const std::vector<uint32_t> & specific,
+           const std::function<std::string(size_t)> & state_name) {
+            std::vector<std::string> alts;
+            if (!completing.empty()) {
+                alts.push_back(gbnf_char_class(completing, false)); // terminate on match
+            }
+            for (const auto & [d, chars] : buckets) {
+                alts.push_back(gbnf_char_class(chars, false) + " " + state_name(d));
+            }
+            // every other character keeps scanning from the start state
+            alts.push_back(gbnf_char_class(specific, true) + " " + state_name(0));
+            return string_join(alts, " | ");
+        });
+}
+
 static std::set<std::string> collect_reachable_rules(
    const common_peg_arena & arena,
    const common_peg_parser_id & rule
@@ -1628,6 +1686,7 @@ static std::set<std::string> collect_reachable_rules(
                                 std::is_same_v<T, common_peg_tag_parser> ||
                                 std::is_same_v<T, common_peg_atomic_parser> ||
                                 std::is_same_v<T, common_peg_gbnf_parser> ||
+                                 std::is_same_v<T, common_peg_ac_parser> ||
                                 std::is_same_v<T, common_peg_schema_parser>) {
                visit(p.child);
            } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
@@ -1822,6 +1881,8 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
                return to_gbnf(p.child);
            } else if constexpr (std::is_same_v<T, common_peg_gbnf_parser>) {
                return p.grammar;
+            } else if constexpr (std::is_same_v<T, common_peg_ac_parser>) {
+                return gbnf_including_grammar(builder, "ac-" + std::to_string(id), p.delimiters);
            } else {
                static_assert(is_always_false_v<T>);
            }
@@ -1958,6 +2019,8 @@ static nlohmann::json serialize_parser_variant(const common_peg_parser_variant &
            };
        } else if constexpr (std::is_same_v<T, common_peg_gbnf_parser>) {
            return json{{"type", "gbnf"}, {"child", p.child}, {"grammar", p.grammar}};
+        } else if constexpr (std::is_same_v<T, common_peg_ac_parser>) {
+            return json{{"type", "ac"}, {"child", p.child}, {"delimiters", p.delimiters}};
        }
    }, variant);
 }
@@ -2130,6 +2193,16 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
        };
    }

+    if (type == "ac") {
+        if (!j.contains("child") || !j.contains("delimiters") || !j["delimiters"].is_array() || j["delimiters"].empty()) {
+            throw std::runtime_error("ac parser requires 'child' and a non-empty 'delimiters' array");
+        }
+        return common_peg_ac_parser{
+            j["child"].get<common_peg_parser_id>(),
+            j["delimiters"].get<std::vector<std::string>>(),
+        };
+    }
+
    throw std::runtime_error("Unknown parser type: " + type);
 }

@@ -275,6 +275,11 @@ struct common_peg_gbnf_parser {
    std::string grammar;
 };

+struct common_peg_ac_parser {
+    common_peg_parser_id child;
+    std::vector<std::string> delimiters;
+};
+
 // Variant holding all parser types
 using common_peg_parser_variant = std::variant<
    common_peg_epsilon_parser,
@@ -296,7 +301,8 @@ using common_peg_parser_variant = std::variant<
    common_peg_ref_parser,
    common_peg_atomic_parser,
    common_peg_tag_parser,
-    common_peg_gbnf_parser
+    common_peg_gbnf_parser,
+    common_peg_ac_parser
 >;

 class common_peg_arena {
@@ -514,6 +520,13 @@ class common_peg_parser_builder {
    // the child's grammar. Parsing delegates entirely to the child.
    common_peg_parser gbnf(const common_peg_parser & p, const std::string & grammar) { return add(common_peg_gbnf_parser{p, grammar}); }

+    // Wraps a child parser but emits a GBNF grammar built from the Aho-Corasick
+    // automaton of `delimiters`, matching everything up to and including the
+    // first delimiter. Parsing delegates entirely to the child, which is
+    // responsible for consuming the delimiter (e.g. until(D) + literal(D)).
+    common_peg_parser ac(const common_peg_parser & p, const std::vector<std::string> & delimiters);
+    common_peg_parser ac(const common_peg_parser & p, const std::string & delimiter) { return ac(p, std::vector<std::string>{delimiter}); }
+
    void set_root(const common_peg_parser & p);

    common_peg_arena build();
@@ -46,6 +46,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "DbrxForCausalLM": "dbrx",
    "DeciLMForCausalLM": "deci",
    "DeepseekForCausalLM": "deepseek",
+    "DeepseekOCRForCausalLM": "deepseek",
    "DeepseekV2ForCausalLM": "deepseek",
    "DeepseekV3ForCausalLM": "deepseek",
    "DeepseekV32ForCausalLM": "deepseek",
@@ -96,6 +97,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "GraniteMoeHybridForCausalLM": "granite",
    "GraniteMoeSharedForCausalLM": "granite",
    "GraniteSpeechForConditionalGeneration": "granite",
+    "GraniteSpeechPlusForConditionalGeneration": "granite",
    "Grok1ForCausalLM": "grok",
    "GrokForCausalLM": "grok",
    "GroveMoeForCausalLM": "grovemoe",
@@ -123,6 +125,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "LLaDAModelLM": "llada",
    "LLaMAForCausalLM": "llama",
    "Lfm25AudioTokenizer": "lfm2",
+    "Lfm2BidirectionalModel": "lfm2",
    "Lfm2ForCausalLM": "lfm2",
    "Lfm2Model": "lfm2",
    "Lfm2MoeForCausalLM": "lfm2",
@@ -133,6 +136,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "LlamaModel": "llama",
    "Eagle3DraftModel": "llama",
    "Eagle3Speculator": "llama",
+    "Eagle3LlamaForCausalLM": "llama",
    "LlamaForCausalLMEagle3": "llama",
    "LlavaForConditionalGeneration": "llama",
    "LlavaStableLMEpochForCausalLM": "stablelm",
@@ -231,6 +235,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "UMT5ForConditionalGeneration": "t5",
    "UMT5Model": "t5",
    "UltravoxModel": "ultravox",
+    "UnlimitedOCRForCausalLM": "deepseek",
    "VLlama3ForCausalLM": "llama",
    "VoxtralForConditionalGeneration": "llama",
    "WavTokenizerDec": "wavtokenizer",
@@ -261,6 +266,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "GlmasrModel": "ultravox",
    "Granite4VisionForConditionalGeneration": "granite",
    "GraniteSpeechForConditionalGeneration": "granite",
+    "GraniteSpeechPlusForConditionalGeneration": "granite",
    "HunYuanVLForConditionalGeneration": "hunyuan",
    "Idefics3ForConditionalGeneration": "smolvlm",
    "InternVisionModel": "internvl",
@@ -296,6 +302,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "StepVLForConditionalGeneration": "step3",
    "Step3p7ForConditionalGeneration": "step3",
    "UltravoxModel": "ultravox",
+    "UnlimitedOCRForCausalLM": "deepseek",
    "VoxtralForConditionalGeneration": "ultravox",
    "YoutuVLForConditionalGeneration": "youtuvl",
 }
@@ -14,7 +14,7 @@ from .base import MmprojModel, ModelBase, TextModel, gguf, logger
 from .qwen import QwenModel


-@ModelBase.register("DeepseekOCRForCausalLM")
+@ModelBase.register("DeepseekOCRForCausalLM", "UnlimitedOCRForCausalLM")
 class DeepseekOCRVisionModel(MmprojModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
@@ -205,6 +205,8 @@ class DeepseekModel(TextModel):
@ModelBase.register(
    "DeepseekV2ForCausalLM",
    "DeepseekV3ForCausalLM",
+    "DeepseekOCRForCausalLM",
+    "UnlimitedOCRForCausalLM",
    "KimiVLForConditionalGeneration",
    "KimiK25ForConditionalGeneration",
    "YoutuForCausalLM",
@@ -224,7 +226,7 @@ class DeepseekV2Model(TextModel):
        self.origin_hf_arch = hparams.get('architectures', [None])[0]

        # special handling for Deepseek OCR
-        if self.origin_hf_arch in ("DeepseekOCRForCausalLM", "DeepseekOCR2ForCausalLM"):
+        if self.origin_hf_arch in ("DeepseekOCRForCausalLM", "DeepseekOCR2ForCausalLM", "UnlimitedOCRForCausalLM"):
            self.model_arch = gguf.MODEL_ARCH.DEEPSEEK2OCR
            self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[self.model_arch]
            self.gguf_writer.add_architecture()
@@ -350,6 +352,12 @@ class DeepseekV2Model(TextModel):

        self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])

+        # Unlimited-OCR sliding window; written for metadata, the decoder ignores it (full MHA)
+        if is_ocr:
+            sliding_window = hparams.get("sliding_window_size") or hparams.get("sliding_window")
+            if sliding_window:
+                self.gguf_writer.add_sliding_window(sliding_window)
+
        if (rope_mscale_all := self.rope_parameters.get("mscale_all_dim")) is not None:
            # [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
            # note: for legacy reasons, this is not consistent with the other usages of self.gguf_writer.add_rope_scaling_yarn_log_mul
@@ -348,6 +348,34 @@ class GraniteSpeechMmprojModel(MmprojModel):
        yield from super().modify_tensors(data_torch, name, bid)


+@ModelBase.register("GraniteSpeechPlusForConditionalGeneration")
+class GraniteSpeechPlusMmprojModel(GraniteSpeechMmprojModel):
+    """Conversion for GraniteSpeechPlus - extends GraniteSpeech with feature layer concatenation"""
+    has_vision_encoder = False
+    has_audio_encoder = True
+
+    def set_gguf_parameters(self):
+        assert self.hparams_audio is not None
+        super().set_gguf_parameters()
+
+        # Add feature_layer if present in encoder config
+        if feature_layers := self.hparams_audio.get("cat_hidden_layers"):
+            self.gguf_writer.add_audio_feature_layers(feature_layers)
+            logger.info(f"gguf: audio feature_layers = {feature_layers}")
+
+            # Validate projector dimension matches concatenated encoder output
+            hidden_dim = self.hparams_audio["hidden_dim"]
+            expected_dim = hidden_dim * (len(feature_layers) + 1)
+            projector_dim = self.global_config["projector_config"]["encoder_hidden_size"]
+
+            if projector_dim != expected_dim:
+                raise ValueError(
+                    f"Projector encoder_hidden_size ({projector_dim}) does not match "
+                    f"expected concatenated dimension ({expected_dim}). "
+                    f"Expected: hidden_dim ({hidden_dim}) * (len(feature_layers) + 1) = {expected_dim}"
+                )
+
+
@ModelBase.register("Granite4VisionForConditionalGeneration")
 class Granite4VisionMmprojModel(MmprojModel):
    has_vision_encoder = True
@@ -64,11 +64,17 @@ class LFM2Model(TextModel):
        yield from super().modify_tensors(data_torch, name, bid)


-@ModelBase.register("Lfm2Model")
+@ModelBase.register("Lfm2Model", "Lfm2BidirectionalModel")
 class LFM2ColBertModel(LFM2Model):
    model_arch = gguf.MODEL_ARCH.LFM2
    dense_tensor_name = "dense_2"

+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if self.hf_arch == "Lfm2BidirectionalModel":
+            self.gguf_writer.add_causal_attention(False)
+        self._try_set_pooling_type()
+
    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        if not name.startswith(self.dense_tensor_name):
            name = "model." + name
@@ -76,10 +82,11 @@ class LFM2ColBertModel(LFM2Model):
        yield from super().modify_tensors(data_torch, name, bid)

    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
-        # dense tensor is stored in a separate safetensors file
+        # optional dense tensor is stored in a separate safetensors file
        from safetensors.torch import load_file
        tensors_file = self.dir_model / "1_Dense" / "model.safetensors"
-        assert tensors_file.is_file()
+        if not tensors_file.is_file():
+            return
        tensor = load_file(tensors_file)["linear.weight"]
        self.gguf_writer.add_embedding_length_out(tensor.shape[0])
        yield f"{self.dense_tensor_name}.weight", tensor.clone()
@@ -23,6 +23,7 @@ from .base import ModelBase, TextModel, gguf, logger
    "LlavaForConditionalGeneration",
    "VoxtralForConditionalGeneration",
    "LlamaForCausalLMEagle3",
+    "Eagle3LlamaForCausalLM",
    "Eagle3Speculator",
    "Eagle3DraftModel",
    "IQuestCoderForCausalLM",
@@ -114,7 +114,8 @@ class Mamba2Model(TextModel):
            hparams["text_config"] = hparams["llm_config"]
        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
        self.d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
-        self.d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or 2 * self.d_model
+        self.expand = self.find_hparam(["mamba_expand", "expand"], optional=True) or 2
+        self.d_inner = self.find_hparam(["mamba_d_ssm", "intermediate_size", "d_inner"], optional=True) or self.expand * self.d_model
        self.n_group = self.find_hparam(["n_groups"], optional=True) or 1

    def set_vocab(self):
@@ -144,11 +145,9 @@ class Mamba2Model(TextModel):

        rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5

-        # Fail early for models which don't have a block expansion factor of 2
-        # TODO: does this really matter?
        # skip the assertion for FalconH1 Model
        if self.model_arch != gguf.MODEL_ARCH.FALCON_H1:
-            assert self.d_inner == 2 * self.d_model
+            assert self.d_inner == self.expand * self.d_model
            assert self.d_inner % head_dim == 0

        self.gguf_writer.add_context_length(2**20)  # arbitrary value; for those who use the default
@@ -29,7 +29,7 @@ With Termux, you can install and run `llama.cpp` as if the environment were Linu

 ```
 $ apt update && apt upgrade -y
-$ apt install git cmake
+$ apt install git cmake libandroid-spawn
 ```

 Then, follow the [build instructions](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md), specifically for CMake.
@@ -237,8 +237,8 @@ chmod +x ubuntu-llamacpp-ov-install.sh
 # ============================================
 set -euo pipefail

-OPENVINO_VERSION_MAJOR="2026.2"
-OPENVINO_VERSION_FULL="2026.2.0.21903.52ddc073857"
+OPENVINO_VERSION_MAJOR="2026.2.1"
+OPENVINO_VERSION_FULL="2026.2.1.21919.ede283a88e3"

 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` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
+> 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.

 </details>

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

-set "OPENVINO_VERSION_MAJOR=2026.2"
-set "OPENVINO_VERSION_FULL=2026.2.0.21903.52ddc073857"
+set "OPENVINO_VERSION_MAJOR=2026.2.1"
+set "OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3"

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

 > [!NOTE]
-> The script pins OpenVINO `2026.2` 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 `2026.2.1` 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**.

 </details>

@@ -413,6 +413,15 @@ In two device selection modes, the default SYCL backend is level_zero, you can c
 |------------------|----------------------------------------|
 | Single device    | --split-mode none --main-gpu DEVICE_ID |
 | Multiple devices | --split-mode layer (default)           |
+| Multiple devices | --split-mode tensor (tensor parallelism) |
+
+`--split-mode tensor` (tensor parallelism) shards each layer across the selected
+GPUs. It requires flash attention, which is auto-enabled when `--flash-attn` is
+left at its default `auto`, so `--split-mode tensor` works out of the box.
+Passing `--flash-attn off` together with `--split-mode tensor` is rejected at
+context creation. The default `f16` KV cache is recommended. Tensor parallelism
+is currently optimized for 2 GPUs; other device counts fall back to a generic
+all-reduce.

 Examples:

@@ -715,6 +724,15 @@ In two device selection modes, the default SYCL backend is level_zero, you can c
 |------------------|----------------------------------------|
 | Single device    | --split-mode none --main-gpu DEVICE_ID |
 | Multiple devices | --split-mode layer (default)           |
+| Multiple devices | --split-mode tensor (tensor parallelism) |
+
+`--split-mode tensor` (tensor parallelism) shards each layer across the selected
+GPUs. It requires flash attention, which is auto-enabled when `--flash-attn` is
+left at its default `auto`, so `--split-mode tensor` works out of the box.
+Passing `--flash-attn off` together with `--split-mode tensor` is rejected at
+context creation. The default `f16` KV cache is recommended. Tensor parallelism
+is currently optimized for 2 GPUs; other device counts fall back to a generic
+all-reduce.

 Examples:

@@ -24,7 +24,6 @@
            "GGML_LLAMAFILE":   "OFF",
            "GGML_OPENCL":      "ON",
            "GGML_HEXAGON":     "ON",
-            "GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE": "128",
            "LLAMA_OPENSSL":    "OFF"
        }
    },
@@ -47,7 +46,6 @@
            "GGML_LLAMAFILE":   "OFF",
            "GGML_OPENCL":      "ON",
            "GGML_HEXAGON":     "ON",
-            "GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE": "128",
            "LLAMA_OPENSSL":    "OFF"
        }
    },
@@ -73,7 +71,6 @@
            "GGML_LLAMAFILE":   "OFF",
            "GGML_OPENCL":      "OFF",
            "GGML_HEXAGON":     "ON",
-            "GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE": "128",
            "LLAMA_OPENSSL":    "OFF"
        }
    },
@@ -13,6 +13,45 @@ The `llama-server` application supports several implementations of speculative d
 A much smaller model (called the _draft model_) generates drafts.
 A draft model is the most used approach in speculative decoding.

+### EAGLE-3 (`draft-eagle3`)
+
+EAGLE-3 uses a small draft model that reads the target model's hidden states to predict the next tokens, so it
+reaches higher acceptance than a standalone draft model of the same size. The draft is a one-layer transformer
+trained for a specific target model; it shares the target model's tokenizer and, optionally, uses a reduced draft
+vocabulary with its own `lm_head`, which is mapped back using a `d2t` table.
+
+Convert the EAGLE-3 checkpoint with `--target-model-dir` so it inherits the target's tokenizer and the layer
+indices to read. Both the SpecForge `LlamaForCausalLMEagle3` and the vLLM/AngelSlim `Eagle3LlamaForCausalLM`
+checkpoint formats are supported (for example [`AngelSlim/Qwen3-4B_eagle3`](https://huggingface.co/AngelSlim/Qwen3-4B_eagle3)
+for `Qwen/Qwen3-4B`):
+
+```bash
+python convert_hf_to_gguf.py AngelSlim/Qwen3-4B_eagle3 \
+    --target-model-dir Qwen/Qwen3-4B --outtype bf16 --outfile Qwen3-4B-eagle3.gguf
+
+llama-server -m Qwen3-4B.gguf -md Qwen3-4B-eagle3.gguf --spec-type draft-eagle3
+```
+
+Supported EAGLE-3 draft models include:
+
+- [yuhuili/EAGLE3-LLaMA3.1-Instruct-8B](https://huggingface.co/yuhuili/EAGLE3-LLaMA3.1-Instruct-8B)
+- [yuhuili/EAGLE3-LLaMA3.3-Instruct-70B](https://huggingface.co/yuhuili/EAGLE3-LLaMA3.3-Instruct-70B)
+- [RedHatAI/gemma-4-31B-it-speculator.eagle3](https://huggingface.co/RedHatAI/gemma-4-31B-it-speculator.eagle3)
+- [RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3](https://huggingface.co/RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3)
+- [Tengyunw/qwen3_8b_eagle3](https://huggingface.co/Tengyunw/qwen3_8b_eagle3)
+- [Tengyunw/qwen3_30b_moe_eagle3](https://huggingface.co/Tengyunw/qwen3_30b_moe_eagle3)
+- [AngelSlim/Qwen3-1.7B_eagle3](https://huggingface.co/AngelSlim/Qwen3-1.7B_eagle3)
+- [AngelSlim/Qwen3-4B_eagle3](https://huggingface.co/AngelSlim/Qwen3-4B_eagle3)
+- [AngelSlim/Qwen3-8B_eagle3](https://huggingface.co/AngelSlim/Qwen3-8B_eagle3)
+- [AngelSlim/Qwen3-14B_eagle3](https://huggingface.co/AngelSlim/Qwen3-14B_eagle3)
+- [AngelSlim/Qwen3-32B_eagle3](https://huggingface.co/AngelSlim/Qwen3-32B_eagle3)
+- [AngelSlim/Qwen3-a3B_eagle3](https://huggingface.co/AngelSlim/Qwen3-a3B_eagle3)
+- [RedHatAI/gpt-oss-20b-speculator.eagle3](https://huggingface.co/RedHatAI/gpt-oss-20b-speculator.eagle3)
+- [lmsys/EAGLE3-gpt-oss-120b-bf16](https://huggingface.co/lmsys/EAGLE3-gpt-oss-120b-bf16)
+- [nvidia/gpt-oss-120b-Eagle3-long-context](https://huggingface.co/nvidia/gpt-oss-120b-Eagle3-long-context)
+
+For the full and up-to-date list of supported models, see #18039.
+
 ### n-gram Cache (`ngram-cache`)

 An n-gram is a sequence of n tokens. The n-gram cache implementation maintains statistics about short n-gram sequences.
@@ -108,7 +147,7 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
 ### General Speculative Parameters

 ```
--spec-type [none|draft-simple|draft-mtp|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]
+--spec-type [none|draft-simple|draft-eagle3|draft-mtp|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]
                                        comma-separated list of types of speculative decoding to use
                                        (default: none)
                                        (env: LLAMA_ARG_SPEC_TYPE)
@@ -247,6 +286,7 @@ Specifies a comma-separated list of speculative decoding types to use.
 |------|-------------|
 | `none` | No speculative decoding (default) |
 | `draft-simple` | Use a simple draft model for speculation |
+| `draft-eagle3` | Use an EAGLE-3 draft model that reads the target's hidden states |
 | `draft-mtp` | Use Multi Token Prediction (MTP) heads from the main model |
 | `ngram-cache` | Use n-gram cache lookup |
 | `ngram-simple` | Use simple n-gram pattern matching |
@@ -5,7 +5,7 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 15)
-set(GGML_VERSION_PATCH 2)
+set(GGML_VERSION_PATCH 3)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")

 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -266,7 +266,6 @@ set   (GGML_OPENCL_TARGET_VERSION "300" CACHE STRING
                                            "ggml: OpenCL API version to target")

 option(GGML_HEXAGON                         "ggml: enable Hexagon backend"                    OFF)
-set(GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE 128 CACHE STRING "ggml: quantize group size (32, 64, or 128)")

 # toolchain for vulkan-shaders-gen
 set   (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN "" CACHE FILEPATH "ggml: toolchain file for vulkan-shaders-gen")
@@ -27,6 +27,14 @@ GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int de
 // split tensor buffer that splits matrices by rows across multiple devices
 GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);

+// Tensor parallelism (--split-mode tensor): comm_init/free/allreduce_tensor
+// trio queried by the meta-backend via ggml_backend_reg_get_proc_address.
+// See typedefs in ggml/include/ggml-backend.h. Mirrors the CUDA backend's
+// pattern (ggml_backend_cuda_comm_*).
+GGML_BACKEND_API void * ggml_backend_sycl_comm_init(ggml_backend_t * backends, size_t n_backends);
+GGML_BACKEND_API void   ggml_backend_sycl_comm_free(void * comm_ctx);
+GGML_BACKEND_API bool   ggml_backend_sycl_comm_allreduce_tensor(void * comm_ctx, struct ggml_tensor ** tensors);
+
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);

@@ -1551,6 +1551,8 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
        int split_backend_id = split->backend_id;
        ggml_backend_t split_backend = sched->backends[split_backend_id];

+        ggml_backend_synchronize(split_backend);
+
        // 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]);
@@ -1561,15 +1563,15 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                    ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
-                } else {
+                } else if (!split_backend->iface.cpy_tensor_async) {
                    ggml_backend_synchronize(split_backend);
                }
-                ggml_backend_tensor_copy(input, input_cpy);
+                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
            } else {
                // wait for the split backend to finish using the input before overwriting it
                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
                    ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
-                } else {
+                } else if (!split_backend->iface.cpy_tensor_async) {
                    ggml_backend_synchronize(split_backend);
                }

@@ -1674,6 +1676,8 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
            }
        }

+        ggml_backend_synchronize(split_backend);
+
        if (!sched->callback_eval) {
            enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
            if (ec != GGML_STATUS_SUCCESS) {
@@ -3688,8 +3688,6 @@ static void ggml_compute_forward_norm_f32(

    GGML_ASSERT(ggml_are_same_shape(src0, dst));

-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-
    const int ith = params->ith;
    const int nth = params->nth;

@@ -3703,25 +3701,49 @@ static void ggml_compute_forward_norm_f32(
    for (int64_t i03 = 0; i03 < ne03; i03++) {
        for (int64_t i02 = 0; i02 < ne02; i02++) {
            for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
-                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                const char * x = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                char * y = (char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3;

-                float sum = 0.0;
-                ggml_vec_sum_f32(ne00, &sum, x);
-                float mean = sum/ne00;
+                if (nb00 == sizeof(float) && nb0 == sizeof(float)) {
+                    const float * xf = (const float *) x;

-                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
-                float variance = 0;
+                    float sum = 0.0;
+                    ggml_vec_sum_f32(ne00, &sum, xf);
+                    float mean = sum/ne00;
+
+                    float * yf = (float *) y;
+                    float variance = 0;

 #ifdef GGML_USE_ACCELERATE
-                mean = -mean;
-                vDSP_vsadd(x, 1, &mean, y, 1, ne00);
-                vDSP_measqv(y, 1, &variance, ne00);
+                    mean = -mean;
+                    vDSP_vsadd(xf, 1, &mean, yf, 1, ne00);
+                    vDSP_measqv(yf, 1, &variance, ne00);
 #else
-                variance = ggml_vec_cvar_f32(ne00, y, x, mean);
+                    variance = ggml_vec_cvar_f32(ne00, yf, xf, mean);
 #endif //GGML_USE_ACCELERATE

-                const float scale = 1.0f/sqrtf(variance + eps);
-                ggml_vec_scale_f32(ne00, y, scale);
+                    const float scale = 1.0f/sqrtf(variance + eps);
+                    ggml_vec_scale_f32(ne00, yf, scale);
+                } else {
+                    float sum = 0.0;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        sum += *(const float *) (x + i00*nb00);
+                    }
+                    const float mean = sum/ne00;
+
+                    float variance = 0.0f;
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const float v = *(const float *) (x + i00*nb00) - mean;
+                        *(float *) (y + i00*nb0) = v;
+                        variance += v * v;
+                    }
+                    variance /= ne00;
+
+                    const float scale = 1.0f/sqrtf(variance + eps);
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        *(float *) (y + i00*nb0) *= scale;
+                    }
+                }
            }
        }
    }
@@ -4142,8 +4164,6 @@ static void ggml_compute_forward_l2_norm_f32(

    GGML_ASSERT(ggml_are_same_shape(src0, dst));

-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-
    const int ith = params->ith;
    const int nth = params->nth;

@@ -4158,20 +4178,27 @@ static void ggml_compute_forward_l2_norm_f32(
    for (int64_t i03 = 0; i03 < ne03; i03++) {
        for (int64_t i02 = 0; i02 < ne02; i02++) {
            for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
-                const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                const char * x = (const char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;

                ggml_float sum = 0.0;
                for (int64_t i00 = 0; i00 < ne00; i00++) {
-                    sum += (ggml_float)(x[i00] * x[i00]);
+                    const float xi = *(const float *) (x + i00*nb00);
+                    sum += (ggml_float)(xi * xi);
                }

-                float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
-
-                memcpy(y, x, ne00 * sizeof(float));
-
                const float scale = 1.0f/fmaxf(sqrtf(sum), eps);

-                ggml_vec_scale_f32(ne00, y, scale);
+                char * y = (char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3;
+
+                if (nb00 == sizeof(float) && nb0 == sizeof(float)) {
+                    memcpy(y, x, ne00 * sizeof(float));
+                    ggml_vec_scale_f32(ne00, (float *) y, scale);
+                } else {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const float xi = *(const float *) (x + i00*nb00);
+                        *(float *) (y + i00*nb0) = xi * scale;
+                    }
+                }
            }
        }
    }
@@ -75,12 +75,12 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
            ay1 = GGML_F32_VEC_LOAD(y + i);
            sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
        }
-        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
+        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmla on available elements only
        if (np2 < n) {
            svbool_t pg = svwhilelt_b32(np2, n);
            ax1 = svld1_f32(pg, x + np2);
            ay1 = svld1_f32(pg, y + np2);
-            sum1 = svmad_f32_m(pg, ax1, ay1, sum1);
+            sum1 = svmla_f32_m(pg, sum1, ax1, ay1);
        }
        // reduce sum1,sum2 to sum1
        GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);
@@ -34,26 +34,26 @@ template <float (*bin_op)(const float, const float),
 static __global__ void k_bin_bcast(const src0_t *         src0,
                                   const src1_t *         src1,
                                   dst_t *                dst,
-                                   const int              ne0,
-                                   const int              ne1,
-                                   const int              ne2,
+                                   const uint32_t         ne0,
+                                   const uint32_t         ne1,
+                                   const uint32_t         ne2,
                                   const uint3            ne3,
                                   const uint3            ne10,
                                   const uint3            ne11,
                                   const uint3            ne12,
                                   const uint3            ne13,
-                                 /*const int              s0,*/
-                                   const int              s1,
-                                   const int              s2,
-                                   const int              s3,
-                                   const int              s00,
-                                   const int              s01,
-                                   const int              s02,
-                                   const int              s03,
-                                   const int              s10,
-                                   const int              s11,
-                                   const int              s12,
-                                   const int              s13,
+                                 /*const uint32_t         s0,*/
+                                   const uint32_t         s1,
+                                   const uint32_t         s2,
+                                   const uint32_t         s3,
+                                   const uint32_t         s00,
+                                   const uint32_t         s01,
+                                   const uint32_t         s02,
+                                   const uint32_t         s03,
+                                   const uint32_t         s10,
+                                   const uint32_t         s11,
+                                   const uint32_t         s12,
+                                   const uint32_t         s13,
                                   src1_ptrs... src1s) {
    ggml_cuda_pdl_lc();
    const uint32_t i0s = blockDim.x * blockIdx.x + threadIdx.x;
@@ -61,7 +61,7 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
    const uint32_t i2  = fastdiv((blockDim.z * blockIdx.z + threadIdx.z), ne3);
    const uint32_t i3  = (blockDim.z * blockIdx.z + threadIdx.z) - (i2 * ne3.z);

-    if (i0s >= (uint32_t)ne0 || i1 >= (uint32_t)ne1 || i2 >= (uint32_t)ne2 || i3 >= ne3.z) {
+    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3.z) {
        return;
    }

@@ -69,25 +69,32 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
    const uint32_t i12 = fastmodulo(i2, ne12);
    const uint32_t i13 = fastmodulo(i3, ne13);

-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
+    const size_t i_src0 = size_t( i3)*s03 + size_t( i2)*s02 + size_t( i1)*s01;
+    const size_t i_src1 = size_t(i13)*s13 + size_t(i12)*s12 + size_t(i11)*s11;
+    const size_t i_dst  = size_t( i3)*s3  + size_t( i2)*s2  + size_t( i1)*s1;

    const src0_t * src0_row = src0 ? (src0 + i_src0) : nullptr;
    dst_t * dst_row = dst + i_dst;

+    const uint32_t s0 = blockDim.x * gridDim.x;
+
    ggml_cuda_pdl_sync();
-    for (int i0 = i0s; i0 < ne0; i0 += blockDim.x * gridDim.x) {
+    for (uint32_t i0 = i0s; i0 < ne0; i0 += s0) {
        const uint32_t i10 = fastmodulo(i0, ne10);

-        float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
+        float result = src0_row ? (float) src0_row[size_t(i0)*s00] : 0.0f;
        if constexpr (sizeof...(src1_ptrs) > 0) {
-            result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
+            result = (..., (result = bin_op(result, (float)src1s[i_src1 + size_t(i10)*s10])));
        } else {
-            result = bin_op(result, (float)src1[i_src1 + i10*s10]);
+            result = bin_op(result, (float)src1[i_src1 + size_t(i10)*s10]);
        }

        dst_row[i0] = (dst_t) result;
+
+        // protect i0 from overflow
+        if (ne0 - i0 <= s0) {
+           break;
+        }
    }
 }

@@ -110,19 +117,19 @@ static __global__ void k_bin_bcast_unravel(const src0_t *         src0,
                                           const uint3            ne12,
                                           const uint3            ne13,
                                         /*const int              s0,*/
-                                           const int              s1,
-                                           const int              s2,
-                                           const int              s3,
-                                           const int              s00,
-                                           const int              s01,
-                                           const int              s02,
-                                           const int              s03,
-                                           const int              s10,
-                                           const int              s11,
-                                           const int              s12,
-                                           const int              s13,
+                                           const uint32_t         s1,
+                                           const uint32_t         s2,
+                                           const uint32_t         s3,
+                                           const uint32_t         s00,
+                                           const uint32_t         s01,
+                                           const uint32_t         s02,
+                                           const uint32_t         s03,
+                                           const uint32_t         s10,
+                                           const uint32_t         s11,
+                                           const uint32_t         s12,
+                                           const uint32_t         s13,
                                           src1_ptrs... src1s) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+    const uint32_t i  = blockDim.x*blockIdx.x + threadIdx.x;

    const uint32_t i3 = fastdiv(i, prod_012);
    const uint32_t i2 = fastdiv(i - i3 * prod_012.z, prod_01);
@@ -133,25 +140,25 @@ static __global__ void k_bin_bcast_unravel(const src0_t *         src0,
        return;
    }

-    const int i11 = fastmodulo(i1, ne11);
-    const int i12 = fastmodulo(i2, ne12);
-    const int i13 = fastmodulo(i3, ne13);
+    const uint32_t i11 = fastmodulo(i1, ne11);
+    const uint32_t i12 = fastmodulo(i2, ne12);
+    const uint32_t i13 = fastmodulo(i3, ne13);

-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
+    const size_t i_src0 = size_t( i3)*s03 + size_t( i2)*s02 + size_t( i1)*s01;
+    const size_t i_src1 = size_t(i13)*s13 + size_t(i12)*s12 + size_t(i11)*s11;
+    const size_t i_dst  = size_t( i3)*s3  + size_t( i2)*s2  + size_t( i1)*s1;

    const src0_t * src0_row = src0 ? (src0 + i_src0) : nullptr;
    dst_t * dst_row = dst + i_dst;

-    const int i10 = fastmodulo(i0, ne10);
+    const uint32_t i10 = fastmodulo(i0, ne10);

    ggml_cuda_pdl_sync();
-    float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
+    float result = src0_row ? (float) src0_row[size_t(i0)*s00] : 0.0f;
    if constexpr (sizeof...(src1_ptrs) > 0) {
-        result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
+        result = (..., (result = bin_op(result, (float)src1s[i_src1 + size_t(i10)*s10])));
    } else {
-        result = bin_op(result, (float)src1[i_src1 + i10*s10]);
+        result = bin_op(result, (float)src1[i_src1 + size_t(i10)*s10]);
    }

    dst_row[i0] = (dst_t) result;
@@ -248,6 +255,31 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
        size_t s02 = nb02 / sizeof(src0_t);
        size_t s03 = nb03 / sizeof(src0_t);

+        GGML_ASSERT(ne0 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(ne1 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(ne2 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(ne3 <= std::numeric_limits<uint32_t>::max());
+
+      //GGML_ASSERT(s0  <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s1  <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s2  <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s3  <= std::numeric_limits<uint32_t>::max());
+
+        GGML_ASSERT(s00 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s01 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s02 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s03 <= std::numeric_limits<uint32_t>::max());
+
+        GGML_ASSERT(s10 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s11 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s12 <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(s13 <= std::numeric_limits<uint32_t>::max());
+
+        GGML_ASSERT(cne1[0] <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(cne1[1] <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(cne1[2] <= std::numeric_limits<uint32_t>::max());
+        GGML_ASSERT(cne1[3] <= std::numeric_limits<uint32_t>::max());
+
        GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
        GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
        GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
@@ -263,6 +295,8 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
        GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
        GGML_ASSERT(nb13 % sizeof(src1_t) == 0);

+        GGML_ASSERT(ne2 * ne3 <= std::numeric_limits<unsigned int>::max());
+
        const int block_size = 128;

        int64_t hne0 = std::max(ne0 / 2LL, 1LL);
@@ -281,7 +315,13 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
        const uint3 ne13 = init_fastdiv_values((uint32_t) cne1[3]);

        if (block_nums.z > 65535 || block_nums.y > 65535) {
-            int         block_num  = (ne0 * ne1 * ne2 * ne3 + block_size - 1) / block_size;
+            int64_t     block_num   = (ne0 * ne1 * ne2 * ne3 + block_size - 1) / block_size;
+
+            GGML_ASSERT(block_num              <= std::numeric_limits<uint32_t>::max());
+            GGML_ASSERT(block_num * block_size <= std::numeric_limits<uint32_t>::max());
+            GGML_ASSERT(ne0 * ne1              <= std::numeric_limits<uint32_t>::max());
+            GGML_ASSERT(ne0 * ne1 * ne2        <= std::numeric_limits<uint32_t>::max());
+
            const uint3 prod_012    = init_fastdiv_values((uint32_t) (ne0 * ne1 * ne2));
            const uint3 prod_01     = init_fastdiv_values((uint32_t) (ne0 * ne1));
            const uint3 ne0_fastdiv = init_fastdiv_values((uint32_t) ne0);
@@ -298,6 +338,10 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
                    s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
            }
        } else {
+            GGML_ASSERT(int64_t(block_nums.x) * block_dims.x <= std::numeric_limits<uint32_t>::max());
+            GGML_ASSERT(int64_t(block_nums.y) * block_dims.y <= std::numeric_limits<uint32_t>::max());
+            GGML_ASSERT(int64_t(block_nums.z) * block_dims.z <= std::numeric_limits<uint32_t>::max());
+
            const uint3 ne3_fastdiv = init_fastdiv_values((uint32_t) ne3);
            {
                const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
@@ -53,10 +53,10 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const
    const int64_t nmat = ne / (ne00 * ne01);
    const int64_t n = ne00 * ne01;

-    const int x = blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.x;
-    const int y = blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
-    const int tx = blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.x;  // transpose block offset
-    const int ty = blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
+    const int64_t x  = (int64_t) blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.x;
+    const int64_t y  = (int64_t) blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
+    const int64_t tx = (int64_t) blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.x;  // transpose block offset
+    const int64_t ty = (int64_t) blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.y;

    __shared__ float tile[2][CUDA_CPY_TILE_DIM_2D][CUDA_CPY_TILE_DIM_2D+1];
    int cur_tile_buf = 0;
@@ -197,7 +197,7 @@ static void ggml_cpy_scalar_contiguous_cuda(
 cudaStream_t stream) {

    const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream);
    ggml_cuda_kernel_launch(cpy_scalar_contiguous<src_t, dst_t>, launch_params, cx, cdst, ne);
 }
@@ -208,6 +208,14 @@ static void ggml_cpy_scalar_cuda(
    const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t nb00, const int64_t nb01, const int64_t nb02,
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {

+    const auto launch_scalar_generic = [&]() {
+        const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+        GGML_ASSERT(num_blocks <= INT_MAX);
+        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream);
+        ggml_cuda_kernel_launch(cpy_scalar<cpy_1_scalar<src_t, dst_t>>, launch_params,
+            cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+    };
+
    if (transposed) {
        GGML_ASSERT(ne == ne00*ne01*ne02);  // ne[3] is 1 assumed
        int64_t ne00n, ne01n, ne02n;
@@ -224,20 +232,18 @@ static void ggml_cpy_scalar_cuda(
        int64_t grid_x = (ne01n + CUDA_CPY_TILE_DIM_2D - 1) / CUDA_CPY_TILE_DIM_2D;
        int64_t grid_y = (ne00n + CUDA_CPY_TILE_DIM_2D - 1) / CUDA_CPY_TILE_DIM_2D;
        int64_t grid_z = (ne/(ne01n*ne00n) + CUDA_CPY_BLOCK_NM - 1) / CUDA_CPY_BLOCK_NM;
-        GGML_ASSERT(grid_x < UINT_MAX);
-        GGML_ASSERT(grid_y < USHRT_MAX);
-        GGML_ASSERT(grid_z < USHRT_MAX);
-        dim3 dimGrid(grid_x, grid_y, grid_z);
-        dim3 dimBlock(CUDA_CPY_TILE_DIM_2D, CUDA_CPY_BLOCK_ROWS, 1);
-        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(dimGrid, dimBlock, 0, stream);
-        ggml_cuda_kernel_launch(cpy_scalar_transpose<dst_t>, launch_params,
-            cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        GGML_ASSERT(grid_x <= INT_MAX);
+        if (grid_y > USHRT_MAX || grid_z > USHRT_MAX) {
+            launch_scalar_generic();
+        } else {
+            dim3 dimGrid(grid_x, grid_y, grid_z);
+            dim3 dimBlock(CUDA_CPY_TILE_DIM_2D, CUDA_CPY_BLOCK_ROWS, 1);
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(dimGrid, dimBlock, 0, stream);
+            ggml_cuda_kernel_launch(cpy_scalar_transpose<dst_t>, launch_params,
+                cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        }
    } else {
-        const int64_t num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-        GGML_ASSERT(num_blocks < UINT_MAX);
-        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params((dim3)num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream);
-        ggml_cuda_kernel_launch(cpy_scalar<cpy_1_scalar<src_t, dst_t>>, launch_params,
-            cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+        launch_scalar_generic();
    }
 }

@@ -248,7 +254,7 @@ static void ggml_cpy_f32_q8_0_cuda(

    GGML_ASSERT(ne % QK8_0 == 0);
    const int64_t num_blocks = ne / QK8_0;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
    cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -259,7 +265,7 @@ static void ggml_cpy_q8_0_f32_cuda(
    const int64_t nb03, const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13, cudaStream_t stream) {

    const int64_t num_blocks = ne;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
    cpy_q_f32<cpy_blck_q8_0_f32, QK8_0><<<num_blocks, 1, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -271,7 +277,7 @@ static void ggml_cpy_f32_q4_0_cuda(

    GGML_ASSERT(ne % QK4_0 == 0);
    const int64_t num_blocks = ne / QK4_0;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
    cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -284,7 +290,7 @@ static void ggml_cpy_q4_0_f32_cuda(
    const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
    cudaStream_t stream) {
    const int64_t num_blocks = ne;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
    cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0><<<num_blocks, 1, 0, stream>>>(
        cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
@@ -297,7 +303,7 @@ static void ggml_cpy_f32_q4_1_cuda(

    GGML_ASSERT(ne % QK4_1 == 0);
    const int64_t num_blocks = ne / QK4_1;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
    cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -310,7 +316,7 @@ static void ggml_cpy_q4_1_f32_cuda(
    const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
    cudaStream_t stream) {
    const int64_t num_blocks = ne;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
    cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1><<<num_blocks, 1, 0, stream>>>(
        cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
         ne10, ne11, ne12, nb10, nb11, nb12, nb13);
@@ -323,7 +329,7 @@ static void ggml_cpy_f32_q5_0_cuda(

    GGML_ASSERT(ne % QK5_0 == 0);
    const int64_t num_blocks = ne / QK5_0;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
    cpy_f32_q<cpy_blck_f32_q5_0, QK5_0><<<num_blocks, 1, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -336,7 +342,7 @@ static void ggml_cpy_q5_0_f32_cuda(
    const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
    cudaStream_t stream) {
    const int64_t num_blocks = ne;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
    cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0><<<num_blocks, 1, 0, stream>>>(
        cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
@@ -349,7 +355,7 @@ static void ggml_cpy_f32_q5_1_cuda(

    GGML_ASSERT(ne % QK5_1 == 0);
    const int64_t num_blocks = ne / QK5_1;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
    cpy_f32_q<cpy_blck_f32_q5_1, QK5_1><<<num_blocks, 1, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -362,7 +368,7 @@ static void ggml_cpy_q5_1_f32_cuda(
    const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
    cudaStream_t stream) {
    const int64_t num_blocks = ne;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
    cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1><<<num_blocks, 1, 0, stream>>>(
        cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
        ne10, ne11, ne12, nb10, nb11, nb12, nb13);
@@ -375,7 +381,7 @@ static void ggml_cpy_f32_iq4_nl_cuda(

    GGML_ASSERT(ne % QK4_NL == 0);
    const int64_t num_blocks = ne / QK4_NL;
-    GGML_ASSERT(num_blocks < UINT_MAX);
+    GGML_ASSERT(num_blocks <= INT_MAX);
    cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL><<<num_blocks, 1, 0, stream>>>
        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
 }
@@ -3192,11 +3192,24 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
    ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
    ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;

-    if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
+    // Enables async copies from CPU to CUDA, instead of only CUDA-to-CUDA
+    // Excluding this path for HIP and MUSA as a precaution.
+    // According to the summary in https://github.com/ggml-org/llama.cpp/pull/20793#issuecomment-4275794315, this change is not beneficial for hip anyways.
+    // Additionally, there is a lot of anectodal evidence that hip/musa stream behavior might not always 1:1 match CUDA behavior.
+    // e.g. https://github.com/ROCm/rocm-systems/issues/5109
+    // It thus makes sense to exclude this path for HIP and MUSA. This PR was not aimed these backends, the majority of testing happened on CUDA.
+    // This can be revisited in the future if enabling copy_from_host benefits hip/MUSA, and if the PR author can extensively test on these backends.
+#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA)
+    const bool copy_from_host = false;
+#else
+    const bool copy_from_host = ggml_backend_buffer_is_host(buf_src) && ggml_backend_dev_type(backend_src->device) == GGML_BACKEND_DEVICE_TYPE_CPU;
+#endif
+
+    if (!(copy_from_host || ggml_backend_is_cuda(backend_src)) || !ggml_backend_is_cuda(backend_dst)) {
        return false;
    }

-    if (!ggml_backend_buffer_is_cuda(buf_src) || !ggml_backend_buffer_is_cuda(buf_dst)) {
+    if (!(copy_from_host || ggml_backend_buffer_is_cuda(buf_src)) || !ggml_backend_buffer_is_cuda(buf_dst)) {
        return false;
    }

@@ -3207,14 +3220,17 @@ static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_
    ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *) buf_src->context;
    ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *) buf_dst->context;

-    if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
+    if ((copy_from_host && cuda_ctx_dst->device != buf_ctx_dst->device) ||
+        !copy_from_host && (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device)) {
 #ifndef NDEBUG
        GGML_LOG_DEBUG("%s: backend and buffer devices do not match\n", __func__);
 #endif // NDEBUG
        return false;
    }

-    if (backend_src != backend_dst) {
+    if (copy_from_host) {
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyHostToDevice, cuda_ctx_dst->stream()));
+    } else if (backend_src != backend_dst) {
        // copy on src stream
        if (cuda_ctx_src->device == cuda_ctx_dst->device) {
            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
@@ -5334,7 +5350,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
        case GGML_OP_NORM:
        case GGML_OP_RMS_NORM:
        case GGML_OP_L2_NORM:
-            return true;
+            return ggml_is_contiguous_rows(op->src[0]);
        case GGML_OP_RMS_NORM_BACK:
            return ggml_is_contiguous(op->src[0]);
            break;
@@ -2,6 +2,28 @@

 #include <cstdint>

+static __global__ void k_compute_out_prod_ptrs(
+        const float * src0_d, const float * src1_d, float * dst_d,
+        const float ** ptrs_a, const float ** ptrs_b, float ** ptrs_c,
+        const int64_t ne2, const int64_t ne3,
+        const int64_t dps2, const int64_t dps3,
+        const size_t s02, const size_t s03,
+        const size_t s12, const size_t s13,
+        const size_t s2,  const size_t s3) {
+    const int64_t i2 = blockIdx.x*blockDim.x + threadIdx.x;
+    const int64_t i3 = blockIdx.y*blockDim.y + threadIdx.y;
+
+    if (i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    const int64_t idx = i3*ne2 + i2;
+
+    ptrs_a[idx] = src0_d + (i3/dps3)*s03 + (i2/dps2)*s02;
+    ptrs_b[idx] = src1_d +  i3      *s13 +  i2      *s12;
+    ptrs_c[idx] = dst_d  +  i3      *s3  +  i2      *s2;
+}
+
 void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];
@@ -67,18 +89,39 @@ void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
                        &beta,  dst_d  +  i3     *s3,  ldc, s2,
                        batch_count));
        }
+    } else if (ne2 > 1 || ne3 > 1) {
+        // dps2 > 1 (src0 broadcast along dim 2 with non-uniform stride) or multiple GEMMs
+        // along dim 3: compute per-GEMM pointers on the device and use a single batched GEMM.
+        GGML_ASSERT(ne3 > 0);
+        GGML_ASSERT(ne2 <= (int64_t) std::numeric_limits<int>::max() / ne3);
+        const int batch_count = (int) (ne2 * ne3);
+
+        ggml_cuda_pool_alloc<const float *> ptrs_a(ctx.pool(), batch_count);
+        ggml_cuda_pool_alloc<const float *> ptrs_b(ctx.pool(), batch_count);
+        ggml_cuda_pool_alloc<      float *> ptrs_c(ctx.pool(), batch_count);
+
+        const dim3 block_dims(16, 16);
+        const dim3 grid_dims((ne2 + block_dims.x - 1)/block_dims.x, (ne3 + block_dims.y - 1)/block_dims.y);
+        k_compute_out_prod_ptrs<<<grid_dims, block_dims, 0, stream>>>(
+            src0_d, src1_d, dst_d,
+            ptrs_a.get(), ptrs_b.get(), ptrs_c.get(),
+            ne2, ne3, dps2, dps3, s02, s03, s12, s13, s2, s3);
+        CUDA_CHECK(cudaGetLastError());
+
+        CUBLAS_CHECK(
+            cublasSgemmBatched(handle, CUBLAS_OP_N, src1_cublas_op,
+                    ne0, ne1, ne01,
+                    &alpha, ptrs_a.get(), lda,
+                            ptrs_b.get(), ldb,
+                    &beta,  ptrs_c.get(), ldc,
+                    batch_count));
    } else {
-        // Fallback: ne2 == 1 (no batching benefit) or dps2 > 1 (src0 broadcast along dim 2
-        // with non-uniform stride; would need cublasSgemmBatched with pointer arrays).
-        for (int64_t i3 = 0; i3 < ne3; ++i3) {
-            for (int64_t i2 = 0; i2 < ne2; ++i2) {
-                CUBLAS_CHECK(
-                    cublasSgemm(handle, CUBLAS_OP_N, src1_cublas_op,
-                            ne0, ne1, ne01,
-                            &alpha, src0_d + (i3/dps3)*s03 + (i2/dps2)*s02, lda,
-                                    src1_d +  i3      *s13 +  i2      *s12, ldb,
-                            &beta,  dst_d  +  i3      *s3  +  i2      *s2,  ldc));
-            }
-        }
+        // ne2 == 1 && ne3 == 1: single GEMM
+        CUBLAS_CHECK(
+            cublasSgemm(handle, CUBLAS_OP_N, src1_cublas_op,
+                    ne0, ne1, ne01,
+                    &alpha, src0_d, lda,
+                            src1_d, ldb,
+                    &beta,  dst_d,  ldc));
    }
 }
@@ -48,6 +48,7 @@
 #define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
 #define cublasSetStream hipblasSetStream
 #define cublasSgemm hipblasSgemm
+#define cublasSgemmBatched hipblasSgemmBatched
 #define cublasSgemmStridedBatched hipblasSgemmStridedBatched
 #define cublasStatus_t hipblasStatus_t
 #define cublasOperation_t hipblasOperation_t
@@ -32,6 +32,7 @@
 #define cublasSetMathMode mublasSetMathMode
 #define cublasSetStream mublasSetStream
 #define cublasSgemm mublasSgemm
+#define cublasSgemmBatched mublasSgemmBatched
 #define cublasSgemmStridedBatched mublasSgemmStridedBatched
 #define cublasStatus_t mublasStatus_t
 #define cublasOperation_t mublasOperation_t
@@ -25,7 +25,6 @@ include(ExternalProject)
 option(GGML_HEXAGON_HTP_DEBUG  "ggml-hexagon: enable HTP debug output" OFF)
 option(GGML_HEXAGON_FA_EXP2_HF "ggml-hexagon: use FP16 exp2 polynomial in FA softmax instead of F32 exp round-trip" OFF)
 set(GGML_HEXAGON_HTP_CERT  "$ENV{HEXAGON_HTP_CERT}" CACHE PATH "ggml-hexagon: enable HTP library signing using certificate")
-set(GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE 128 CACHE STRING "ggml-hexagon: quantize group size (32, 64, or 128)")

 add_library(htp_iface OBJECT
    ${CMAKE_CURRENT_BINARY_DIR}/htp_iface_stub.c)
@@ -72,15 +71,12 @@ function(build_htp_skel V)
            -DHEXAGON_SDK_ROOT=${HEXAGON_SDK_ROOT}
            -DHEXAGON_TOOLS_ROOT=${HEXAGON_TOOLS_ROOT}
            -DHEXAGON_HTP_DEBUG=${GGML_HEXAGON_HTP_DEBUG}
-            -DGGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE=${GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE}
            -DDSP_VERSION=${V}
            -DPREBUILT_LIB_DIR="toolv19_${V}")
    list(APPEND HTP_SKELS ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-${V}.so)
    set(HTP_SKELS ${HTP_SKELS} PARENT_SCOPE)
 endfunction()

-build_htp_skel(v68)
-build_htp_skel(v69)
 build_htp_skel(v73)
 build_htp_skel(v75)
 build_htp_skel(v79)
@@ -5,10 +5,12 @@
 #include "ggml-backend-impl.h"
 #include "ggml-common.h"

+#include <algorithm>
 #include <string>
 #include <vector>
 #include <stdio.h>
 #include "htp-ops.h"
+#include "htp/matmul-ops.h"

 struct htp_opnode {
    ggml_tensor * node = nullptr;
@@ -17,6 +19,13 @@ struct htp_opnode {

    htp_op_code opcode = HTP_OP_INVALID;

+    std::vector<ggml_tensor *> extra_dsts;
+
+    int32_t kernel_params[HTP_OP_MAX_KERN_PARAMS] = {0};
+
+    htp_opnode(ggml_tensor * node = nullptr, std::vector<ggml_tensor *> fused = {}, htp_op_code opcode = HTP_OP_INVALID, std::vector<ggml_tensor *> extra_dsts = {})
+        : node(node), fused(std::move(fused)), opcode(opcode), extra_dsts(std::move(extra_dsts)) {}
+
    ggml_op op() const {
        return node->op;
    }
@@ -25,6 +34,26 @@ struct htp_opnode {
        return fused.empty() ? node : fused.back();
    }

+    void add_fused(ggml_tensor * t, bool extra_dst = false) {
+        fused.push_back(t);
+        if (extra_dst) {
+            extra_dsts.push_back(t);
+        }
+    }
+
+    std::vector<const ggml_tensor *> get_outputs() const {
+        std::vector<const ggml_tensor *> res;
+        if (extra_dsts.empty()) {
+            res.push_back(dst());
+        } else {
+            res.push_back(node);
+            for (const auto * x : extra_dsts) {
+                res.push_back(x);
+            }
+        }
+        return res;
+    }
+
    const ggml_tensor * src0() const {
        return node->src[0];
    }
@@ -37,10 +66,6 @@ struct htp_opnode {
        return ggml_op_is_empty(node->op);
    }

-    void add_fused(ggml_tensor * t) {
-        fused.push_back(t);
-    }
-
    bool stackable() const {
        switch (this->op()) {
            case GGML_OP_MUL_MAT:
@@ -131,87 +156,117 @@ struct htp_opformat {
    char types[16 * GGML_MAX_SRC];
    char buffs[64 * GGML_MAX_SRC];
    char names[64 * GGML_MAX_SRC];
+    char kparams[128];

-    int format_tensor_dims(char * str, const struct ggml_tensor * t) {
+    int format_tensor_dims(char * str, size_t max_size, const struct ggml_tensor * t) {
        if (!t) {
-            return sprintf(str, "NONE");
+            return snprintf(str, max_size, "NONE");
        }
        if (t->ne[2] == 1 && t->ne[3] == 1) {
-            return sprintf(str, "%d:%d", (int) t->ne[0], (int) t->ne[1]);
+            return snprintf(str, max_size, "%d:%d", (int) t->ne[0], (int) t->ne[1]);
        } else {
-            return sprintf(str, "%d:%d:%d:%d", (int) t->ne[0], (int) t->ne[1], (int) t->ne[2], (int) t->ne[3]);
+            return snprintf(str, max_size, "%d:%d:%d:%d", (int) t->ne[0], (int) t->ne[1], (int) t->ne[2], (int) t->ne[3]);
        }
    }

-    void format_op_dims(char * str, const htp_opnode & node) {
+    void format_op_dims(char * str, size_t max_size, const htp_opnode & node) {
        char * p = str;
+        char * p_end = str + max_size;
        auto inputs = node.get_inputs();

        if (!inputs.empty()) {
-            p += format_tensor_dims(p, inputs[0]);
+            p += std::min((size_t)format_tensor_dims(p, p_end - p, inputs[0]), (size_t)(p_end - p));

            for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += format_tensor_dims(p, inputs[i]);
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)format_tensor_dims(p, p_end - p, inputs[i]), (size_t)(p_end - p));
+                }
            }

-            p += sprintf(p, " -> ");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
        }

        char self[64];
-        format_tensor_dims(self, node.dst());
-        p += sprintf(p, "%s", self);
+        format_tensor_dims(self, sizeof(self), node.dst());
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", self), (size_t)(p_end - p));
+        }
    }

-    int format_tensor_strides(char * str, const struct ggml_tensor * t) {
+    int format_tensor_strides(char * str, size_t max_size, const struct ggml_tensor * t) {
        if (!t) {
-            return sprintf(str, "NONE");
+            return snprintf(str, max_size, "NONE");
        }
        const char * c = ggml_is_contiguous(t) ? "" : "!";

        if (t->ne[2] == 1 && t->ne[3] == 1) {
-            return sprintf(str, "%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], c);
+            return snprintf(str, max_size, "%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], c);
        } else {
-            return sprintf(str, "%zu:%zu:%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], (size_t) t->nb[2], (size_t) t->nb[3], c);
+            return snprintf(str, max_size, "%zu:%zu:%zu:%zu%s", (size_t) t->nb[0], (size_t) t->nb[1], (size_t) t->nb[2], (size_t) t->nb[3], c);
        }
    }

-    void format_op_strides(char * str, const htp_opnode & node) {
+    void format_op_strides(char * str, size_t max_size, const htp_opnode & node) {
        char * p = str;
+        char * p_end = str + max_size;
        auto inputs = node.get_inputs();

        if (!inputs.empty()) {
-            p += format_tensor_strides(p, inputs[0]);
+            p += std::min((size_t)format_tensor_strides(p, p_end - p, inputs[0]), (size_t)(p_end - p));

            for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += format_tensor_strides(p, inputs[i]);
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)format_tensor_strides(p, p_end - p, inputs[i]), (size_t)(p_end - p));
+                }
            }

-            p += sprintf(p, " -> ");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
        }

        char self[64];
-        format_tensor_strides(self, node.dst());
-        p += sprintf(p, "%s", self);
+        format_tensor_strides(self, sizeof(self), node.dst());
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", self), (size_t)(p_end - p));
+        }
    }

-    void format_op_types(char * str, const htp_opnode & node) {
+    void format_op_types(char * str, size_t max_size, const htp_opnode & node) {
        char * p = str;
+        char * p_end = str + max_size;
        auto inputs = node.get_inputs();

        if (!inputs.empty()) {
-            p += sprintf(p, "%s", inputs[0] ? ggml_type_name(inputs[0]->type) : "NONE");
-
-            for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += sprintf(p, "%s", inputs[i] ? ggml_type_name(inputs[i]->type) : "NONE");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, "%s", inputs[0] ? ggml_type_name(inputs[0]->type) : "NONE"), (size_t)(p_end - p));
            }

-            p += sprintf(p, " -> ");
+            for (size_t i = 1; i < inputs.size(); i++) {
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, "%s", inputs[i] ? ggml_type_name(inputs[i]->type) : "NONE"), (size_t)(p_end - p));
+                }
+            }
+
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
        }

-        p += sprintf(p, "%s", ggml_type_name(node.dst()->type));
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", ggml_type_name(node.dst()->type)), (size_t)(p_end - p));
+        }
    }

    const char * tensor_buff_name(const struct ggml_tensor * t) {
@@ -221,51 +276,102 @@ struct htp_opformat {
        return "NONE";
    }

-    void format_op_buffs(char * str, const htp_opnode & node) {
+    void format_op_buffs(char * str, size_t max_size, const htp_opnode & node) {
        char * p = str;
+        char * p_end = str + max_size;
        auto inputs = node.get_inputs();

        if (!inputs.empty()) {
-            p += sprintf(p, "%s", tensor_buff_name(inputs[0]));
-
-            for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += sprintf(p, "%s", tensor_buff_name(inputs[i]));
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, "%s", tensor_buff_name(inputs[0])), (size_t)(p_end - p));
            }

-            p += sprintf(p, " -> ");
+            for (size_t i = 1; i < inputs.size(); i++) {
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, "%s", tensor_buff_name(inputs[i])), (size_t)(p_end - p));
+                }
+            }
+
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
        }

-        p += sprintf(p, "%s", tensor_buff_name(node.dst()));
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", tensor_buff_name(node.dst())), (size_t)(p_end - p));
+        }
    }

-    void format_op_names(char * str, const htp_opnode & node) {
+    void format_op_names(char * str, size_t max_size, const htp_opnode & node) {
        char * p = str;
+        char * p_end = str + max_size;
        auto inputs = node.get_inputs();

        if (!inputs.empty()) {
-            p += sprintf(p, "%s", inputs[0] ? inputs[0]->name : "NONE");
-
-            for (size_t i = 1; i < inputs.size(); i++) {
-                p += sprintf(p, " x ");
-                p += sprintf(p, "%s", inputs[i] ? inputs[i]->name : "NONE");
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, "%s", inputs[0] ? inputs[0]->name : "NONE"), (size_t)(p_end - p));
            }

-            p += sprintf(p, " -> ");
+            for (size_t i = 1; i < inputs.size(); i++) {
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, " x "), (size_t)(p_end - p));
+                }
+                if (p < p_end) {
+                    p += std::min((size_t)snprintf(p, p_end - p, "%s", inputs[i] ? inputs[i]->name : "NONE"), (size_t)(p_end - p));
+                }
+            }
+
+            if (p < p_end) {
+                p += std::min((size_t)snprintf(p, p_end - p, " -> "), (size_t)(p_end - p));
+            }
        }

-        p += sprintf(p, "%s", node.dst()->name);
+        if (p < p_end) {
+            p += std::min((size_t)snprintf(p, p_end - p, "%s", node.dst()->name), (size_t)(p_end - p));
+        }
+    }
+    void format_kernel_params(char * str, size_t max_size, const htp_opnode & node) {
+        if (node.opcode == HTP_OP_MUL_MAT || node.opcode == HTP_OP_MUL_MAT_ID ||
+            node.opcode == HTP_OP_MUL_MAT_QKV || node.opcode == HTP_OP_MUL_MAT_FFN) {
+            const auto * kparams = (const struct htp_mm_kernel_params *) node.kernel_params;
+            const char * path = "unknown";
+            int32_t type = kparams->kernel_type;
+            if (type == HTP_MM_KERNEL_HMX_2D || type == HTP_MM_KERNEL_HMX_F16_BATCHED) {
+                path = "hmx-tiled";
+            } else if (type == HTP_MM_KERNEL_HVX_F16_F16_VTCM || type == HTP_MM_KERNEL_HVX_F32_F32_VTCM ||
+                       type == HTP_MM_KERNEL_HVX_QUANT_ROW    || type == HTP_MM_KERNEL_HVX_QUANT_BLOCK) {
+                path = "hvx-tiled";
+            } else if (type == HTP_MM_KERNEL_HVX_F16_F16_DDR  || type == HTP_MM_KERNEL_HVX_F16_F32_DDR ||
+                       type == HTP_MM_KERNEL_HVX_F32_F32_DDR  || type == HTP_MM_KERNEL_HVX_F32_F16_DDR ||
+                       type == HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT) {
+                path = "hvx-flat";
+            }
+            snprintf(str, max_size, "%s vtcm %d", path, (int) kparams->vtcm_size);
+        } else {
+            snprintf(str, max_size, "----");
+        }
    }

    void format(const htp_opnode & node) {
-        format_op_dims(dims, node);
-        format_op_strides(strides, node);
-        format_op_types(types, node);
-        format_op_buffs(buffs, node);
-        format_op_names(names, node);
+        format_op_dims(dims, sizeof(dims), node);
+        format_op_strides(strides, sizeof(strides), node);
+        format_op_types(types, sizeof(types), node);
+        format_op_buffs(buffs, sizeof(buffs), node);
+        format_op_names(names, sizeof(names), node);
+        format_kernel_params(kparams, sizeof(kparams), node);
    }

-    htp_opformat() {}
+    htp_opformat() {
+        strides[0] = '\0';
+        dims[0]    = '\0';
+        types[0]   = '\0';
+        buffs[0]   = '\0';
+        names[0]   = '\0';
+        kparams[0] = '\0';
+    }
    htp_opformat(const htp_opnode & node) { format(node); }
 };

@@ -19,43 +19,9 @@ add_library(${HTP_LIB} SHARED
    htp_iface_skel.c
    worker-pool.c
    hex-dma.c
-)
-
-target_compile_definitions(${HTP_LIB} PRIVATE
-    $<IF:$<BOOL:${HEXAGON_HTP_DEBUG}>,HTP_DEBUG=1,NDEBUG=1>
-    $<IF:$<BOOL:${HEXAGON_HTP_DEBUG}>,FARF_HIGH=1,>
-    FP32_QUANTIZE_GROUP_SIZE=${GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE})
-
-if (GGML_HEXAGON_FA_EXP2_HF)
-    message(STATUS "ggml-htp: HMX_FA_USE_EXP2_HF=1 (use FP16 exp2 polynomial in FA softmax)")
-    target_compile_definitions(${HTP_LIB} PRIVATE HMX_FA_USE_EXP2_HF=1)
-endif()
-
-# HMX acceleration: available on v73+ architectures
-set(HTP_HMX_VERSIONS v73 v75 v79 v81)
-list(FIND HTP_HMX_VERSIONS ${DSP_VERSION} _hmx_idx)
-
-if (_hmx_idx GREATER_EQUAL 0)
-    target_sources(${HTP_LIB} PRIVATE
-        hmx-flash-attn-ops.c
-        hmx-matmul-ops.c
-        hmx-queue.c
-    )
-
-    # -mhmx enables HMX instruction set (needed by files that include hmx-utils.h)
-    set_source_files_properties(
-        hmx-flash-attn-ops.c
-        hmx-matmul-ops.c
-        hmx-queue.c
-        PROPERTIES COMPILE_OPTIONS "-mhmx"
-    )
-
-    target_compile_definitions(${HTP_LIB} PRIVATE HTP_HAS_HMX=1)
-endif()
-
-build_idl(htp_iface.idl ${HTP_LIB})
-
-target_sources(${HTP_LIB} PRIVATE
+    hmx-queue.c
+    flash-attn-ops.c
+    hmx-flash-attn-ops.c
    matmul-ops.c
    binary-ops.c
    unary-ops.c
@@ -63,7 +29,6 @@ target_sources(${HTP_LIB} PRIVATE
    softmax-ops.c
    act-ops.c
    rope-ops.c
-    flash-attn-ops.c
    set-rows-ops.c
    get-rows-ops.c
    cpy-ops.c
@@ -79,6 +44,17 @@ target_sources(${HTP_LIB} PRIVATE
    pad-ops.c
 )

+target_compile_definitions(${HTP_LIB} PRIVATE
+    $<IF:$<BOOL:${HEXAGON_HTP_DEBUG}>,HTP_DEBUG=1,NDEBUG=1>
+    $<IF:$<BOOL:${HEXAGON_HTP_DEBUG}>,FARF_HIGH=1,>)
+
+if (GGML_HEXAGON_FA_EXP2_HF)
+    message(STATUS "ggml-htp: HMX_FA_USE_EXP2_HF=1 (use FP16 exp2 polynomial in FA softmax)")
+    target_compile_definitions(${HTP_LIB} PRIVATE HMX_FA_USE_EXP2_HF=1)
+endif()
+
+build_idl(htp_iface.idl ${HTP_LIB})
+
 set_target_properties(${HTP_LIB} PROPERTIES EXPORT_COMPILE_COMMANDS ON)

 install(TARGETS ${HTP_LIB})
@@ -3,7 +3,7 @@ if (HEXAGON_TOOLCHAIN_INCLUDED)
 endif()
 set(HEXAGON_TOOLCHAIN_INCLUDED true)

-#Cross Compiling for Hexagon
+# Cross Compiling for Hexagon
 set(HEXAGON TRUE)
 set(CMAKE_SYSTEM_NAME QURT)
 set(CMAKE_SYSTEM_PROCESSOR Hexagon)
@@ -14,7 +14,6 @@ set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
 set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
 set(CUSTOM_RUNELF_PATH "")

-#To fix backward compatibility with EAI addon.
 if (NOT HEXAGON_SDK_ROOT)
    set(HEXAGON_SDK_ROOT $ENV{HEXAGON_SDK_ROOT})
 endif()
@@ -31,7 +30,6 @@ endif()
 file(TO_CMAKE_PATH "${HEXAGON_TOOLS_ROOT}" HEXAGON_TOOLS_ROOT)
 file(TO_CMAKE_PATH "${HEXAGON_SDK_ROOT}"   HEXAGON_SDK_ROOT)

-#Get the Binary extension of the Hexagon Toolchain
 if(CMAKE_HOST_SYSTEM_NAME STREQUAL Windows)
    set(HEXAGON_TOOLCHAIN_SUFFIX .exe)
 endif()
@@ -48,12 +46,12 @@ set(CMAKE_TRY_COMPILE_PLATFORM_VARIABLES
    HEXAGON_TOOLS_ROOT
 )

-#QURT Related includes and linker flags
+# QURT Related includes and linker flags
 set(V_ARCH ${HEXAGON_ARCH})
 set(_QURT_INSTALL_DIR "${HEXAGON_SDK_ROOT}/rtos/qurt/ADSP${V_ARCH}MP${V_ARCH_EXTN}")
 set(_QURT_INSTALL_DIR "${HEXAGON_SDK_ROOT}/rtos/qurt/compute${V_ARCH}${V_ARCH_EXTN}")

-if( ${TREE} MATCHES PAKMAN )
+if (${TREE} MATCHES PAKMAN)
    set(_QURT_INSTALL_DIR "${QURT_IMAGE_DIR}/compute${V_ARCH}${V_ARCH_EXTN}")
 endif()
 message(DEBUG "_QURT_INSTALL_DIR:${_QURT_INSTALL_DIR}")
@@ -83,11 +81,9 @@ set(QURT_START_LINK_LIBS
    )
 STRING(REPLACE ";" " " QURT_START_LINK_LIBS "${QURT_START_LINK_LIBS}")

-set(QURT_END_LINK_LIBS
-    ${TARGET_DIR}/fini.o
-    )
+set(QURT_END_LINK_LIBS ${TARGET_DIR}/fini.o)

-#Non QURT related includes and linker flags
+# Non QURT related includes and linker flags

 set(TARGET_DIR_NOOS "${HEXAGON_TOOLCHAIN}/Tools/target/hexagon/lib/${HEXAGON_ARCH}")

@@ -99,8 +95,10 @@ if (NOT NO_WRAP_MEM_API)
    set(WRAP_MEMALIGN -Wl,--wrap=memalign)
 endif()

+set(ARCH_FLAGS "-mcpu=${V_ARCH} -m${V_ARCH} -mhvx=${V_ARCH} -mhmx")
+
 set(PIC_SHARED_LD_FLAGS
-    -mcpu=${V_ARCH} -m${V_ARCH} -mhvx=${V_ARCH}
+    ${ARCH_FLAGS}
    -G0
    -fpic
    -Wl,-Bsymbolic
@@ -120,13 +118,13 @@ STRING(REPLACE ";" " " PIC_SHARED_LD_FLAGS "${PIC_SHARED_LD_FLAGS}")

 set(HEXAGON_PIC_SHARED_LINK_OPTIONS "${PIC_SHARED_LD_FLAGS}")

-#System include paths
+# System include paths
 include_directories(SYSTEM ${HEXAGON_SDK_ROOT}/incs)
 include_directories(SYSTEM ${HEXAGON_SDK_ROOT}/incs/stddef)
 include_directories(SYSTEM ${HEXAGON_SDK_ROOT}/ipc/fastrpc/incs)

-#LLVM toolchain setup
-#Compiler paths, options and architecture
+# LLVM toolchain setup
+# Compiler paths, options and architecture
 set(CMAKE_C_COMPILER ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-clang${HEXAGON_TOOLCHAIN_SUFFIX})
 set(CMAKE_CXX_COMPILER ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-clang++${HEXAGON_TOOLCHAIN_SUFFIX})
 set(CMAKE_AR ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-ar${HEXAGON_TOOLCHAIN_SUFFIX})
@@ -137,8 +135,8 @@ set(CMAKE_PREFIX_PATH ${HEXAGON_TOOLCHAIN}/Tools/target/hexagon)
 set(CMAKE_SHARED_LIBRARY_SONAME_C_FLAG   "-Wl,-soname,")
 set(CMAKE_SHARED_LIBRARY_SONAME_CXX_FLAG "-Wl,-soname,")

-#Compiler Options
-set(COMMON_FLAGS "-mcpu=hexagon${V_ARCH} -m${V_ARCH} -mhvx=${V_ARCH} -fvectorize -flto -Wall -Werror -fno-zero-initialized-in-bss -G0 -fdata-sections -fpic ${XQF_ARGS}")
+# Compiler Options
+set(COMMON_FLAGS "${ARCH_FLAGS} -fvectorize -flto -Wall -Werror -fno-zero-initialized-in-bss -G0 -fdata-sections -fpic ${XQF_ARGS}")

 set(CMAKE_CXX_FLAGS_DEBUG          "${COMMON_FLAGS} -O0 -D_DEBUG -g")
 set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${COMMON_FLAGS} -O2 -g")
@@ -18,7 +18,8 @@
 #include "htp-ctx.h"
 #include "htp-ops.h"
 #include "htp-ops.h"
-#include "hmx-ops.h"
+
+int hmx_flash_attn_ext(struct htp_ops_context * octx);

 // Must be multiple of 32
 #define FLASH_ATTN_BLOCK_SIZE (32 * 2)
@@ -633,7 +634,6 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
        return HTP_STATUS_NO_SUPPORT;
    }

-#ifdef HTP_HAS_HMX
    // HMX path: head_dim multiple of 64, F16 KV, and no sinks
    if (k->type == HTP_TYPE_F16 && v->type == HTP_TYPE_F16 && k->ne[0] % 64 == 0 && v->ne[0] % 64 == 0 && octx->src[4] == NULL) {
        int ret = hmx_flash_attn_ext(octx);
@@ -642,7 +642,6 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
        }
        // VTCM too small or other failure -> fall through to HVX path
    }
-#endif

    struct htp_fa_context factx;
    factx.octx = octx;
@@ -0,0 +1,80 @@
+#ifndef HEX_COMMON_H
+#define HEX_COMMON_H
+
+#include <stdint.h>
+#include <stddef.h>
+#include <stdbool.h>
+
+#ifndef SIZE_MAX
+#define SIZE_MAX ((size_t)-1)
+#endif
+
+#ifndef MAX
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+#endif
+
+#ifndef MIN
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#endif
+
+static inline uint32_t hex_ceil_pow2(uint32_t x) {
+    if (x <= 1) { return 1; }
+    int p = 2;
+    x--;
+    while (x >>= 1) { p <<= 1; }
+    return p;
+}
+
+static inline size_t hmx_ceil_div(size_t num, size_t den) {
+    return (num + den - 1) / den;
+}
+
+static inline int32_t hex_is_aligned(const void * addr, uint32_t align) {
+    return ((size_t) addr & (align - 1)) == 0;
+}
+
+static inline size_t hex_align_up(size_t v, size_t align) {
+    return hmx_ceil_div(v, align) * align;
+}
+
+static inline size_t hex_align_down(size_t v, size_t align) {
+    return (v / align) * align;
+}
+
+static inline int32_t hex_is_one_chunk(void * addr, uint32_t n, uint32_t chunk_size) {
+    uint32_t left_off  = (size_t) addr & (chunk_size - 1);
+    uint32_t right_off = left_off + n;
+    return right_off <= chunk_size;
+}
+
+static inline uint32_t hex_round_up(uint32_t n, uint32_t m) {
+    return m * ((n + m - 1) / m);
+}
+
+static inline size_t hex_smin(size_t a, size_t b) {
+    return a < b ? a : b;
+}
+
+static inline size_t hex_smax(size_t a, size_t b) {
+    return a > b ? a : b;
+}
+
+static inline void hex_swap_ptr(void ** p1, void ** p2) {
+    void * t = *p1;
+    *p1      = *p2;
+    *p2      = t;
+}
+
+static inline bool hex_mul_overflow(size_t a, size_t b, size_t *out) {
+    if (a != 0 && b > SIZE_MAX / a) return true;
+    *out = a * b;
+    return false;
+}
+
+static inline bool hex_add_overflow(size_t a, size_t b, size_t *out) {
+    if (a > SIZE_MAX - b) return true;
+    *out = a + b;
+    return false;
+}
+
+#endif // HEX_COMMON_H
@@ -5,6 +5,7 @@
 #include <hexagon_types.h>
 #include <stdbool.h>
 #include <stdint.h>
+#include "hex-utils.h"

 #include "hex-profile.h"

@@ -127,13 +128,8 @@ static inline dma_ptr dma_make_ptr(void *dst, const void *src)
    return p;
 }

-#if __HVX_ARCH__ < 73
-static const uint32_t dma_src_l2_bypass_on = 1;
-static const uint32_t dma_dst_l2_bypass_on = 0;
-#else
 static const uint32_t dma_src_l2_bypass_on = 1;
 static const uint32_t dma_dst_l2_bypass_on = 1;
-#endif

 static inline bool dma_queue_push_single_1d(dma_queue * q, dma_ptr dptr, size_t size) {
    if (((q->push_idx + 1) & q->idx_mask) == q->pop_idx) {
@@ -11,14 +11,7 @@

 #include "hex-fastdiv.h"
 #include "hex-dump.h"
-
-#ifndef MAX
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-#endif
-
-#ifndef MIN
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-#endif
+#include "hex-common.h"

 static inline uint64_t hex_get_cycles() {
    uint64_t cycles = 0;
@@ -32,54 +25,6 @@ static inline uint64_t hex_get_pktcnt() {
    return pktcnt;
 }

-static inline uint32_t hex_ceil_pow2(uint32_t x) {
-    if (x <= 1) { return 1; }
-    int p = 2;
-    x--;
-    while (x >>= 1) { p <<= 1; }
-    return p;
-}
-
-static inline size_t hmx_ceil_div(size_t num, size_t den) {
-    return (num + den - 1) / den;
-}
-
-static inline int32_t hex_is_aligned(const void * addr, uint32_t align) {
-    return ((size_t) addr & (align - 1)) == 0;
-}
-
-static inline size_t hex_align_up(size_t v, size_t align) {
-    return hmx_ceil_div(v, align) * align;
-}
-
-static inline size_t hex_align_down(size_t v, size_t align) {
-    return (v / align) * align;
-}
-
-static inline int32_t hex_is_one_chunk(void * addr, uint32_t n, uint32_t chunk_size) {
-    uint32_t left_off  = (size_t) addr & (chunk_size - 1);
-    uint32_t right_off = left_off + n;
-    return right_off <= chunk_size;
-}
-
-static inline uint32_t hex_round_up(uint32_t n, uint32_t m) {
-    return m * ((n + m - 1) / m);
-}
-
-static inline size_t hex_smin(size_t a, size_t b) {
-    return a < b ? a : b;
-}
-
-static inline size_t hex_smax(size_t a, size_t b) {
-    return a > b ? a : b;
-}
-
-static inline void hex_swap_ptr(void ** p1, void ** p2) {
-    void * t = *p1;
-    *p1      = *p2;
-    *p2      = t;
-}
-
 static inline void hex_l2fetch(const void * p, uint32_t width, uint32_t stride, uint32_t height) {
    const uint64_t control = Q6_P_combine_RR(stride, Q6_R_combine_RlRl(width, height));
    Q6_l2fetch_AP((void *) p, control);
@@ -49,7 +49,7 @@
 // g_br = hex_align_up(gqa_factor * Br, 32) replaces Br for all Q/O/S/P/D dimensions.
 // Layout: Q + O_ping + O_pong + K_dma*2 + V_dma*2 + K_tile + V_tile + S + P + D + vectors + scales
 // Mask is DMA'd into a VTCM buffer (Br rows per KV block) to avoid DDR reads in softmax.
-static size_t hmx_fa_compute_vtcm_usage(size_t gqa_factor, size_t DK, size_t DV, size_t Br, size_t Bc, size_t n_threads, bool use_pipeline) {
+static size_t hmx_fa_compute_vtcm_usage(size_t gqa_factor, size_t DK, size_t DV, size_t Br, size_t Bc, size_t n_threads, bool pipeline) {
    const size_t g_br         = hex_align_up(gqa_factor * Br, HMX_FP16_TILE_N_ROWS);
    const size_t q_tile_size  = hex_align_up(g_br * DK * sizeof(__fp16), 4096);    // Q:  [g_br, DK]
    const size_t o_tile_size  = hex_align_up(g_br * DV * sizeof(__fp16), 4096);    // O:  [g_br, DV] x2 ping-pong
@@ -70,7 +70,7 @@ static size_t hmx_fa_compute_vtcm_usage(size_t gqa_factor, size_t DK, size_t DV,
           + k_dma_size  * 2               // K DMA x2
           + v_dma_size  * 2               // V DMA x2
           + k_tile_size * 1               // K tiles
-           + v_tile_size * (use_pipeline ? 2 : 1) // V tiles (double-buffered if pipelining)
+           + v_tile_size * (pipeline ? 2 : 1) // V tiles (double-buffered if pipelining)
           + s_tile_size * 2               // S + P
           + d_tile_size * 1               // D (diagonal matrix)
           + col_vec_size * 4              // m_vec, l_vec, s_rowmax, p_rowsum
@@ -290,7 +290,7 @@ static const int16_t d_tile_scatter_offsets[64] __attribute__((aligned(128))) =

 struct hmx_fa_context {
    const struct htp_ops_context * octx;
-    bool         use_pipeline;  // true when n_kv_blocks >= FA_MIN_KV_BLOCKS && n_threads >= 2
+    bool         pipeline;  // true when n_kv_blocks >= FA_MIN_KV_BLOCKS && n_threads >= 2
    uint32_t     n_threads;

    // Op parameters
@@ -409,7 +409,7 @@ static void fa_v_interleave_thread(unsigned int n, unsigned int i, void * data)
        return;
    }

-    __fp16 * v_tiles_dest = factx->use_pipeline ? factx->vtcm_v_tiles[args->buf_idx] : factx->vtcm_v_tiles[0];
+    __fp16 * v_tiles_dest = factx->pipeline ? factx->vtcm_v_tiles[args->buf_idx] : factx->vtcm_v_tiles[0];

    struct htp_thread_trace * tr = factx->octx->ctx ? &factx->octx->ctx->trace[i] : NULL;
    htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, start);
@@ -1312,13 +1312,13 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
    const size_t g_br = hex_align_up(G * Br, HMX_FP16_TILE_N_ROWS);

    const uint32_t n_kv_blocks  = (nek1 + Bc - 1) / Bc;
-    const bool     use_pipeline = (n_kv_blocks >= FA_MIN_KV_BLOCKS && n_threads_init >= 2);
+    const bool     pipeline = (n_kv_blocks >= FA_MIN_KV_BLOCKS && n_threads_init >= 2);

    // Bypass thread pool dispatch for small prompts/non-pipelined prefill by setting n_threads = 1
-    const uint32_t n_threads = use_pipeline ? n_threads_init : 1;
+    const uint32_t n_threads = pipeline ? n_threads_init : 1;

    FARF(HIGH, "hmx-fa: neq1=%u nek1=%u DK=%u DV=%u G=%u Br=%zu Bc=%zu g_br=%zu n_kv_blocks=%u pipeline=%d vtcm=%zu",
-         neq1, nek1, DK, DV, G, Br, Bc, g_br, n_kv_blocks, use_pipeline, vtcm_budget);
+         neq1, nek1, DK, DV, G, Br, Bc, g_br, n_kv_blocks, pipeline, vtcm_budget);

    // ======== Build context ========
    struct hmx_fa_context factx;
@@ -1339,7 +1339,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
    factx.n_kv_blocks    = n_kv_blocks;
    factx.is_q_fp32      = (q->type == HTP_TYPE_F32);
    factx.is_dst_fp32    = (dst->type == HTP_TYPE_F32);
-    factx.use_pipeline   = use_pipeline;
+    factx.pipeline   = pipeline;
    factx.mask_broadcast = (mask != NULL && mask->ne[2] == 1);

    // Extract op parameters (mutable during softcap adjustment, then stored as const in factx)
@@ -1405,7 +1405,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
    factx.vtcm_v_fp16[1]      = (__fp16 *) vtcm_seq_alloc(&vtcm_cur, v_dma_bytes);
    factx.vtcm_k_tiles        = (__fp16 *) vtcm_seq_alloc(&vtcm_cur, k_tile_bytes);
    factx.vtcm_v_tiles[0]     = (__fp16 *) vtcm_seq_alloc(&vtcm_cur, v_tile_bytes);
-    if (use_pipeline) {
+    if (pipeline) {
        factx.vtcm_v_tiles[1] = (__fp16 *) vtcm_seq_alloc(&vtcm_cur, v_tile_bytes);
    } else {
        factx.vtcm_v_tiles[1] = NULL;
@@ -1456,7 +1456,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
    // ======== HMX lock strategy ========
    // Pipeline: queue thread auto-acquires HMX lock on first push; released by suspend.
    // Fallback: main thread holds the lock (original behavior).
-    if (!factx.use_pipeline) {
+    if (!factx.pipeline) {
        HAP_compute_res_hmx_lock(ctx->vtcm_rctx);
    }

@@ -1550,7 +1550,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                const size_t k_src_stride = size_k_row_padded / sizeof(__fp16);
                const size_t v_src_stride = size_v_row_padded / sizeof(__fp16);

-                if (factx.use_pipeline) {
+                if (factx.pipeline) {
                    // ==================================================================
                    // Pipeline path: HVX phases ‖ HMX queue worker
                    // ==================================================================
@@ -1780,7 +1780,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                    fa_build_d_diag_inv_l(&factx, n_row_tiles, n_row_tiles_g_br);

                    // HMX: O_final = diag(1/l) @ O_prev
-                    if (factx.use_pipeline) {
+                    if (factx.pipeline) {
                        on_job.o_curr           = o_tile_curr;
                        on_job.o_prev           = o_tile_prev;
                        on_job.d_tiles          = factx.vtcm_d_tiles;
@@ -1826,7 +1826,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
        }  // end KV head loop
    }  // end batch loop

-    if (factx.use_pipeline) {
+    if (factx.pipeline) {
        hmx_queue_suspend(ctx->hmx_queue);
    } else {
        HAP_compute_res_hmx_unlock(ctx->vtcm_rctx);
@@ -1,6 +0,0 @@
-// HMX operations compiled as a single translation unit.
-// This allows interprocedural optimizations within HMX ops without requiring global HTP LTO.
-
-#include "hmx-queue.c"
-#include "hmx-matmul-ops.c"
-#include "hmx-flash-attn-ops.c"
@@ -1,88 +0,0 @@
-// HMX operation entry-point declarations.
-// Ported from htp-ops-lib/include/dsp/ops.h (renamed, benchmark kernels removed). (https://github.com/haozixu/htp-ops-lib)
-
-#ifndef HMX_OPS_H
-#define HMX_OPS_H
-
-#include <stddef.h>
-#include <stdint.h>
-
-#include "htp-ops.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-typedef struct {
-    float        *dst;
-    const float  *activation;
-    const __fp16 *permuted_weight;
-    int           m;
-    int           k;
-    int           n;
-    int           act_stride;
-    int           weight_stride;
-    int           dst_stride;
-    int           ne02;
-    int           ne03;
-    int           ne12;
-    int           ne13;
-    size_t        src0_nb2;
-    size_t        src0_nb3;
-    size_t        src1_nb2;
-    size_t        src1_nb3;
-    size_t        dst_nb2;
-    size_t        dst_nb3;
-} hmx_matmul_f16_f32_batched_params_t;
-
-// HMX matrix multiplication — tile-permuted FP16 weights, FP32 activation/output
-// act_stride: activation row stride in elements (= k for contiguous, or
-//             nb[1]/sizeof(float) for permuted tensors like attention Q).
-// weight_stride: weight row stride in elements (= k for compact weights, or
-//                nb[1]/sizeof(__fp16) for permuted KV-cache views used by QK).
-int hmx_matmul_f16_f32(struct htp_context *ctx,
-                                float *restrict dst,
-                                const float *activation,
-                                const __fp16 *permuted_weight,
-                                int m, int k, int n,
-                                int act_stride,
-                                int weight_stride);
-
-// Batched F16 wrapper over hmx_mat_mul_f16_f32.
-// Batch semantics match ggml_mul_mat(): src0 broadcasts to src1 in dims 2/3.
-int hmx_matmul_f16_f32_batched(struct htp_context *ctx, const hmx_matmul_f16_f32_batched_params_t *params);
-
-// HMX matrix multiplication — all supported weight types (F16/F32/Q4_0/Q4_1/Q8_0/IQ4_NL/MXFP4)
-int hmx_matmul_2d_f32(struct htp_context *ctx,
-                                      float *restrict dst,
-                                      const float *activation,
-                                      const uint8_t *permuted_weight,
-                                      int m, int k, int n,
-                                      int act_stride,
-                                      int weight_stride,
-                                      int weight_type);
-
-struct mmid_row_mapping;
-
-int hmx_matmul_id_2d_f32(struct htp_context *ctx,
-                                         float *restrict dst,
-                                         const float *activation,
-                                         const uint8_t *permuted_weight,
-                                         int m, int k, int n,
-                                         int ne11,
-                                         size_t act_nb1, size_t act_nb2,
-                                         size_t dst_nb1, size_t dst_nb2,
-                                         int weight_stride,
-                                         int weight_type,
-                                         const struct mmid_row_mapping *matrix_rows,
-                                         int cur_a,
-                                         int mapping_stride);
-
-// HMX flash attention
-int hmx_flash_attn_ext(struct htp_ops_context * octx);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif // HMX_OPS_H
@@ -13,7 +13,9 @@
 #include <stdint.h>
 #include <stdbool.h>

+#ifndef HTP_MAX_NTHREADS
 #define HTP_MAX_NTHREADS 10
+#endif
 #define HTP_MAX_MMAPS    16

 // Memory mapping
@@ -42,9 +44,13 @@ struct htp_ops_context {

    enum htp_op_code    op; // FIXME: rename to opcode
    int32_t             op_params[HTP_OP_MAX_PARAMS];
+    int32_t             kernel_params[HTP_OP_MAX_KERN_PARAMS];

    const struct htp_tensor * src[HTP_OP_MAX_INPUTS];
-    const struct htp_tensor * dst;
+    union {
+        const struct htp_tensor * dst;
+        const struct htp_tensor * dsts[HTP_OP_MAX_OUTPUTS];
+    };

    // TODO convert these to an array
    struct htp_spad src0_spad;
@@ -87,13 +93,13 @@ struct htp_context {

    struct htp_ops_context octx;

-#ifdef HTP_HAS_HMX
    struct hmx_queue *     hmx_queue; // Async HMX queue for pipeline overlap
-#endif
 };

 int op_matmul(struct htp_ops_context * octx);
 int op_matmul_id(struct htp_ops_context * octx);
+int op_matmul_qkv(struct htp_ops_context * octx);
+int op_matmul_ffn(struct htp_ops_context * octx);
 int op_binary(struct htp_ops_context * octx);
 int op_unary(struct htp_ops_context * octx);
 int op_sum_rows(struct htp_ops_context * octx);
@@ -28,18 +28,19 @@ enum htp_data_type {
    HTP_TYPE_MXFP4  = 39,

    // types used internally for repack, dyn.quant, etc
-    HTP_TYPE_Q4_0x4x2 = 200,
-    HTP_TYPE_Q4_1x4x2,
-    HTP_TYPE_Q8_0x4x2,
-    HTP_TYPE_MXFP4x4x2,
+    HTP_TYPE_Q4_0_TILED = 200,
+    HTP_TYPE_Q4_1_TILED,
+    HTP_TYPE_Q8_0_TILED,
+    HTP_TYPE_MXFP4_TILED,

    HTP_TYPE_INVALID
 };

 // Constats for internal types
-#define QK_Q4_0x4x2  256  // 4x Q4_0  blocks packed with next 4x Q4_0 blocks (size in bytes 128)
-#define QK_Q8_0x4x2  256  // 4x Q8_0  blocks concat with next 4x Q8_0 blocks
-#define QK_MXFP4x4x2 256  // 4x MXFP4 blocks concat with next 4x MXFP4 blocks
+#define QK_Q4_0_TILED  256  // 32x32 Q4_0 tiled layout
+#define QK_Q8_0_TILED  128  // 32x32 Q8_0 tiled layout
+#define QK_MXFP4_TILED 256  // 32x32 MXFP4 tiled layout
+


 // Mask to enable various stages of the Ops.
@@ -57,6 +58,8 @@ enum htp_op_code {
    HTP_OP_DIV = 3,
    HTP_OP_MUL_MAT,
    HTP_OP_MUL_MAT_ID,
+    HTP_OP_MUL_MAT_QKV,
+    HTP_OP_MUL_MAT_FFN,
    HTP_OP_RMS_NORM,
    HTP_OP_RMS_NORM_MUL,
    HTP_OP_UNARY_SILU,
@@ -99,7 +102,9 @@ enum htp_op_code {

 #define HTP_OP_MAX_DIMS    4    // aka GGML_MAX_DIMS
 #define HTP_OP_MAX_INPUTS  6    // aka GGML_MAX_SRCS
+#define HTP_OP_MAX_OUTPUTS 4
 #define HTP_OP_MAX_PARAMS  16   // aka GGML_MAX_OP_PARAMS
+#define HTP_OP_MAX_KERN_PARAMS 32

 #define HTP_OP_MAX_BUFS    16
 #define HTP_OP_MAX_REQS    256
@@ -142,8 +147,10 @@ struct htp_op_desc {
    uint32_t opcode;                    // GGML/HTP Op
    uint32_t flags;                     // Op flags
    int32_t  params[HTP_OP_MAX_PARAMS]; // Params for the op, e.g. epsilon of RMS norm
+    int32_t  kernel_params[HTP_OP_MAX_KERN_PARAMS]; // generic blob for host-precomputed parameters
    uint16_t src[HTP_OP_MAX_INPUTS];    // Input tensors indices
-    uint16_t dst;                       // Output tensor index
+    uint16_t dst[HTP_OP_MAX_OUTPUTS];   // Output tensor indices
+    uint16_t pad[2];                    // padding to align to 64 bits
 };

 #ifndef HTP_MAX_NTHREADS
@@ -11,12 +11,13 @@ struct htp_iface_pmu_conf {
 };

 interface htp_iface : remote_handle64 {
-    AEEResult start(in uint32 sess_id, in uint64 dsp_queue_id, in uint32 n_hvx, in uint32 use_hmx, in uint64 max_vmem);
+    AEEResult start(in uint32 sess_id, in uint64 dsp_queue_id, in uint32 n_hvx, in uint32 n_hmx, in uint64 max_vmem);
    AEEResult stop();
    AEEResult mmap(in uint32 fd, in uint32 size);
    AEEResult munmap(in uint32 fd);
    AEEResult profiler(in uint32 mode, in htp_iface_pmu_conf pmu);
    AEEResult etm(in uint32 enable);
+    AEEResult hwinfo(rout uint32 n_threads, rout uint32 n_hvx, rout uint32 n_hmx, rout uint64 vtcm_size);
 };

 #endif /* HTP_IDL */
@@ -170,25 +170,7 @@ static inline HVX_VectorPair hvx_vec_f16_to_f32(HVX_Vector v) {
 }
 #endif

-/* Q6_Vsf_equals_Vw is only available on v73+.*/
-#if __HVX_ARCH__ < 73
-static inline HVX_Vector hvx_vec_i32_to_qf32(HVX_Vector const in)
-{
-    HVX_Vector const vzero = Q6_V_vzero();
-    HVX_VectorPred is_zero = Q6_Q_vcmp_eq_VwVw(in, vzero);
-    HVX_Vector lshift = Q6_Vw_vnormamt_Vw(in);
-    HVX_Vector normalized = Q6_Vw_vasl_VwVw(in, lshift);
-    HVX_Vector vexp = Q6_Vw_vsub_VwVw(Q6_V_vsplat_R(0x7f + 30), lshift);
-    HVX_Vector mant = Q6_V_vand_VV(Q6_V_vsplat_R(0xFFFFFF00), normalized);
-    HVX_Vector ret = Q6_V_vmux_QVV(is_zero, vzero, Q6_Vw_vadd_VwVw(mant, vexp));
-    return ret;
-}

-static inline HVX_Vector Q6_Vsf_equals_Vw(HVX_Vector const in)
-{
-    return Q6_Vsf_equals_Vqf32(hvx_vec_i32_to_qf32(in));
-}
-#endif

 static inline HVX_Vector hvx_vec_i16_from_hf_rnd_sat(HVX_Vector vin) {
    // This looks complicated.
@@ -305,4 +287,17 @@ static inline HVX_Vector hvx_vec_mul_f32_f32(HVX_Vector a, HVX_Vector b) {

 #endif // __HVX_ARCH__ < 79

+static inline HVX_Vector hvx_vec_load_act_tile(const uint8_t * y_q, uint32_t kt, HVX_Vector * v_act_all) {
+    if (kt % 4 == 0) {
+        *v_act_all = hvx_vmem(y_q + kt * 32);
+        return *v_act_all;
+    } else if (kt % 4 == 1) {
+        return Q6_V_vror_VR(*v_act_all, 32);
+    } else if (kt % 4 == 2) {
+        return Q6_V_vror_VR(*v_act_all, 64);
+    } else {
+        return Q6_V_vror_VR(*v_act_all, 96);
+    }
+}
+
 #endif /* HVX_BASE_H */
@@ -361,7 +361,7 @@ static void vtcm_free(struct htp_context * ctx) {
 static void htp_packet_callback(dspqueue_t queue, int error, void * context);
 static void htp_error_callback(dspqueue_t queue, int error, void * context);

-AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_queue_id, uint32 n_hvx, uint32 use_hmx, uint64_t max_vmem) {
+AEEResult htp_iface_start(remote_handle64 handle, uint32_t sess_id, uint64_t dsp_queue_id, uint32_t n_hvx, uint32_t n_hmx, uint64_t max_vmem) {
    struct htp_context * ctx = (struct htp_context *) handle;

    if (!ctx) {
@@ -395,10 +395,9 @@ AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_que
        return AEE_ENOMEMORY;
    }

-#ifdef HTP_HAS_HMX
-    ctx->hmx_enabled = use_hmx;
+    ctx->hmx_enabled = n_hmx;
    ctx->hmx_queue   = NULL;
-    if (use_hmx) {
+    if (n_hmx) {
        ctx->hmx_queue = hmx_queue_create(16, ctx->vtcm_rctx);
        if (ctx->hmx_queue) {
            ctx->hmx_queue->trace = &ctx->trace[HTP_MAX_NTHREADS];
@@ -407,8 +406,7 @@ AEEResult htp_iface_start(remote_handle64 handle, uint32 sess_id, uint64 dsp_que
            ctx->hmx_enabled = false;
        }
    }
-    FARF(HIGH, "HMX %s (use_hmx=%d)", ctx->hmx_enabled ? "enabled" : "disabled", use_hmx);
-#endif
+    FARF(HIGH, "HMX %s (n_hmx=%d)", ctx->hmx_enabled ? "enabled" : "disabled", n_hmx);

    qurt_sysenv_max_hthreads_t hw_threads;
    qurt_sysenv_get_max_hw_threads(&hw_threads);
@@ -481,13 +479,11 @@ AEEResult htp_iface_stop(remote_handle64 handle) {
        dma_queue_delete(ctx->dma[i]);
    }

-#ifdef HTP_HAS_HMX
    if (ctx->hmx_queue) {
        hmx_queue_delete(ctx->hmx_queue);
        ctx->hmx_queue = NULL;
    }
    ctx->hmx_enabled = false;
-#endif

    vtcm_free(ctx);

@@ -500,6 +496,36 @@ AEEResult htp_iface_stop(remote_handle64 handle) {
    return AEE_SUCCESS;
 }

+AEEResult htp_iface_hwinfo(remote_handle64 handle, uint32_t * n_threads, uint32_t * n_hvx, uint32_t * n_hmx, uint64_t * vtcm_size) {
+    (void)handle;
+    if (!n_threads || !n_hvx || !n_hmx || !vtcm_size) {
+        return AEE_EBADPARM;
+    }
+
+    qurt_sysenv_max_hthreads_t hw_threads;
+    qurt_sysenv_get_max_hw_threads(&hw_threads);
+    uint32_t hw_nhvx = (qurt_hvx_get_units() >> 8) & 0xFF;
+
+    uint32_t n_hvx_val = hw_nhvx;
+    if (n_hvx_val > hw_threads.max_hthreads) {
+        n_hvx_val = hw_threads.max_hthreads;
+    }
+    if (n_hvx_val > HTP_MAX_NTHREADS) {
+        n_hvx_val = HTP_MAX_NTHREADS;
+    }
+
+    // for now we force n_threads == n_hvx
+    *n_threads = n_hvx_val;
+    *n_hvx     = n_hvx_val;
+    *n_hmx     = 1;
+
+    uint32_t vtcm_sz = 8 * 1024 * 1024; // 8MB default fallback
+    HAP_compute_res_query_VTCM(0, (unsigned int *)&vtcm_sz, NULL, NULL, NULL);
+    *vtcm_size = vtcm_sz;
+
+    return AEE_SUCCESS;
+}
+
 static void htp_error_callback(dspqueue_t queue, int error, void * context) {
    // No errors expected on the DSP.
    FARF(ERROR, "Error callback: 0x%08x", (unsigned) error);
@@ -554,6 +580,12 @@ static int execute_op(struct htp_ops_context * octx) {
        case HTP_OP_MUL_MAT_ID:
            return op_matmul_id(octx);

+        case HTP_OP_MUL_MAT_QKV:
+            return op_matmul_qkv(octx);
+
+        case HTP_OP_MUL_MAT_FFN:
+            return op_matmul_ffn(octx);
+
        case HTP_OP_MUL:
        case HTP_OP_ADD:
        case HTP_OP_SUB:
@@ -762,8 +794,9 @@ static void prep_tensors(struct htp_context *ctx, struct htp_buf_desc *bufs, str
    }
 }

-static void proc_op_req(struct htp_ops_context * octx, struct htp_tensor *tens, uint32_t idx, struct htp_op_desc * op) {
+static int proc_op_req(struct htp_ops_context * octx, struct htp_tensor *tens, uint32_t idx, struct htp_op_desc * op) {
    memcpy(octx->op_params, op->params, sizeof(octx->op_params));
+    memcpy(octx->kernel_params, op->kernel_params, sizeof(octx->kernel_params));
    octx->flags = op->flags;
    octx->op    = op->opcode;

@@ -785,22 +818,41 @@ static void proc_op_req(struct htp_ops_context * octx, struct htp_tensor *tens,
            src->ne[0], src->ne[1], src->ne[3], src->ne[3]);
    }

-    // Prep output tensor
-    struct htp_tensor *dst = tens + op->dst;
+    // Prep output tensors
+    for (uint32_t i = 0; i < HTP_OP_MAX_OUTPUTS; i++) {
+        uint16_t dst_idx = op->dst[i];
+        if (dst_idx == 0xffff) {
+            octx->dsts[i] = NULL;
+            continue;
+        }
+        struct htp_tensor *dst = tens + dst_idx;
+        octx->dsts[i] = dst;

-    octx->dst = dst;
+        FARF(HIGH, "prep-dst[%u] #%u: data %p size %u : %u:%u:%u:%u", i, dst_idx, (void*) dst->data, dst->size,
+            dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]);
+    }

-    FARF(HIGH, "prep-dst #%u: data %p size %u : %u:%u:%u:%u", op->dst, (void*) dst->data, dst->size,
-        dst->ne[0], dst->ne[1], dst->ne[3], dst->ne[3]);
+    int status = execute_op(octx);

-    (void) execute_op(octx);
+    octx->src0_spad.src = NULL;
+    octx->src1_spad.src = NULL;
+    octx->src2_spad.src = NULL;
+    octx->src3_spad.src = NULL;
+    octx->dst_spad.src  = NULL;

    // flush buffers on output
-    hex_l2flush((void *) dst->data, dst->size);
-    dst->flags |= HTP_TENSOR_FLUSHED;
+    for (uint32_t i = 0; i < HTP_OP_MAX_OUTPUTS; i++) {
+        if (octx->dsts[i]) {
+            struct htp_tensor *dst = (struct htp_tensor *)octx->dsts[i];
+            hex_l2flush((void *) dst->data, dst->size);
+            dst->flags |= HTP_TENSOR_FLUSHED;

-    FARF(HIGH, "post-dst #%u: data %p size %u : %u:%u:%u:%u", op->dst, (void*) dst->data, dst->size,
-        dst->ne[0], dst->ne[1], dst->ne[3], dst->ne[3]);
+            FARF(HIGH, "post-dst[%u] #%u: data %p size %u : %u:%u:%u:%u", i, op->dst[i], (void*) dst->data, dst->size,
+                dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3]);
+        }
+    }
+
+    return status;
 }

 #define DSPQUEUE_POLL_TIMEOUT_USEC 100
@@ -892,20 +944,26 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
            }
        }

+        int      op_status = HTP_STATUS_OK;
+        uint32_t op_wakeup = n_ops / 2; // half-way throgh the batch
+
        for (uint32_t i=0; i < n_ops; i++) {
            struct profile_data prof;

-            if (i == (n_ops-1)) {
-                // wake up the host before starting the last op
+            if (i == op_wakeup) {
                dspqueue_write_early_wakeup_noblock(queue, 0, 0);
            }

            profile_start(ctx->profiler, &prof);

-            proc_op_req(octx, tens, i, &ops[i]);
+            op_status = proc_op_req(octx, tens, i, &ops[i]);

            profile_stop(ctx->profiler, &prof);

+            if (op_status != HTP_STATUS_OK) {
+                break;
+            }
+
            if (ctx->profiler) {
                pds[i].opcode = ops[i].opcode;
                pds[i].usecs  = prof.usecs;
@@ -919,7 +977,7 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {

        struct htp_opbatch_rsp rsp;
        rsp.id        = req.id;
-        rsp.status    = HTP_STATUS_OK;
+        rsp.status    = op_status;
        rsp.n_bufs    = n_bufs;
        rsp.n_tensors = n_tens;
        rsp.n_ops     = n_ops;
@@ -0,0 +1,508 @@
+#ifndef HTP_MATMUL_OPS_H
+#define HTP_MATMUL_OPS_H
+
+#include <stdint.h>
+#include <stddef.h>
+#include "htp-ops.h"
+#include "hex-fastdiv.h"
+#include "hex-common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// --- HMX Tile Constraints ---
+#define HTP_MM_HMX_TILE_N_COLS 32
+#define HTP_MM_HMX_TILE_N_ROWS 32
+#define HTP_MM_HMX_TILE_SIZE   (32 * 32 * sizeof(__fp16)) // 2048 bytes
+#define HTP_MM_HMX_TILE_N_ELMS 1024
+#define HTP_MM_HMX_MIN_NROWS   4
+
+// --- Weight Repacked Tile Sizes ---
+#define HTP_MM_WEIGHT_TILE_SIZE_Q4_0   576
+#define HTP_MM_WEIGHT_TILE_SIZE_Q4_1   640
+#define HTP_MM_WEIGHT_TILE_SIZE_Q8_0   1088
+#define HTP_MM_WEIGHT_TILE_SIZE_IQ4_NL 576
+#define HTP_MM_WEIGHT_TILE_SIZE_MXFP4  544
+
+// --- Weight Repacked Aligned Tile Sizes ---
+#define HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_Q4_0   640
+#define HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_Q4_1   640
+#define HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_Q8_0   1152
+#define HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_IQ4_NL 640
+#define HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_MXFP4  640
+
+// --- Activation Tiled Block Sizes (including padding) ---
+#define HTP_MM_ACT_TILE_SIZE_Q8_0      1152
+#define HTP_MM_ACT_TILE_SIZE_Q8_1      1280
+
+#define HTP_MM_MAX_PREFETCH 16
+
+// --- Solver Cost Model Penalty Weights (HMX-specific) ---
+#define HTP_MM_HMX_COST_W_DEQUANT 3 // cost penalty for quantized weight loading/dequantization
+#define HTP_MM_HMX_COST_A_CONVERT 2 // cost penalty for activation loading/conversion
+
+// --- DMA Activation Transfer Configuration ---
+#define HTP_MM_DMA_ACT_ROWS_PER_STEP 2
+#define HTP_MM_DMA_ACT_MULTIPLIER    4
+
+enum htp_mm_kernel_type {
+    HTP_MM_KERNEL_UNSUPPORTED = 0,
+
+    // HMX paths
+    HTP_MM_KERNEL_HMX_2D,
+    HTP_MM_KERNEL_HMX_F16_BATCHED,
+
+    // HVX floating-point paths
+    HTP_MM_KERNEL_HVX_F16_F16_VTCM,
+    HTP_MM_KERNEL_HVX_F16_F16_DDR,
+    HTP_MM_KERNEL_HVX_F16_F32_DDR,
+
+    HTP_MM_KERNEL_HVX_F32_F32_VTCM,
+    HTP_MM_KERNEL_HVX_F32_F32_DDR,
+    HTP_MM_KERNEL_HVX_F32_F16_DDR,
+
+    // HVX quantized paths
+    HTP_MM_KERNEL_HVX_QUANT_ROW,      // standard row-wise parallel quantization
+    HTP_MM_KERNEL_HVX_QUANT_BLOCK,    // parallel block-wise quantization
+    HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT, // row-wise fallback flat quantization
+};
+
+// Op-specific struct for precomputed matmul params
+struct htp_mm_kernel_params {
+    int32_t  kernel_type;        // enum htp_mm_kernel_type
+    int32_t  pipeline;           // 1 = pipelined execution, 0 = standard
+    int32_t  m_chunk;            // Row chunk size (M chunk)
+    int32_t  n_chunk;            // Col chunk size (N chunk)
+    int32_t  n_threads;          // Number of threads to spawn
+    int32_t  n_act_threads;      // Number of threads for activation preparation
+    int32_t  n_hmx;              // 1 = use HMX, 0 = use HVX
+    int32_t  n_prefetch;         // Prefetch lookahead buffers/rows in VTCM
+    int32_t  tile_size;          // Weight tile size
+    int32_t  aligned_tile_size;  // Aligned weight tile size (padded to 128)
+    int32_t  src1_row_size;      // Row size for quantized activation
+    int32_t  vtcm_size;          // Total required scratchpad size in VTCM
+    int32_t  vtcm_src0_size;     // src0 scratchpad size in VTCM
+    int32_t  vtcm_src1_size;     // src1 scratchpad size in VTCM
+    int32_t  vtcm_src2_size;     // src2 scratchpad size in VTCM (fused only)
+    int32_t  vtcm_src3_size;     // src3 scratchpad size in VTCM (fused only)
+    int32_t  vtcm_dst_size;      // dst scratchpad size in VTCM
+
+    // Precomputed division values
+    struct fastdiv_values div_ne12_ne1;
+    struct fastdiv_values div_ne1;
+    struct fastdiv_values div_r2;
+    struct fastdiv_values div_r3;
+    struct fastdiv_values div_ne11;
+};
+
+#if defined(__cplusplus)
+static_assert(sizeof(struct htp_mm_kernel_params) <= 128, "htp_matmul_kernel_params is too large for kernel_params blob");
+#else
+_Static_assert(sizeof(struct htp_mm_kernel_params) <= 128, "htp_matmul_kernel_params is too large for kernel_params blob");
+#endif
+
+struct mmid_row_mapping {
+    uint32_t i1;
+    uint32_t i2;
+};
+
+// Search for optimal (mc, nc) chunk sizes within VTCM budget.
+static inline int htp_mm_hmx_compute_chunks(size_t   vtcm_total,
+                              size_t   overhead,
+                              size_t   per_n_cost,
+                              size_t   per_m_cost,
+                              size_t   per_mn_cost,
+                              size_t   m,
+                              size_t   n,
+                              size_t   m_block_cost,
+                              size_t   n_block_cost,
+                              size_t * m_chunk_out,
+                              size_t * n_chunk_out,
+                              size_t * total_out) {
+    if (m == 0 || n == 0) return -1;
+    if (vtcm_total <= overhead) return -1;
+    if (per_n_cost == 0 || per_m_cost == 0 || per_mn_cost == 0) return -1;
+
+    const size_t usable = vtcm_total - overhead;
+
+    size_t best_cost = SIZE_MAX;
+    size_t best_mn   = 0;
+    size_t best_m = 0, best_n = 0;
+
+    const size_t n_max = hex_align_down((size_t)n, HTP_MM_HMX_TILE_N_COLS);
+    for (size_t nc = n_max; nc >= HTP_MM_HMX_TILE_N_COLS; nc -= HTP_MM_HMX_TILE_N_COLS) {
+        size_t n_fixed = 0, ncmn = 0, mc_denom = 0;
+        if (hex_mul_overflow(nc, per_n_cost, &n_fixed)) continue;
+        if (n_fixed >= usable) goto next_nc;
+
+        if (hex_mul_overflow(nc, per_mn_cost, &ncmn)) goto next_nc;
+        if (hex_add_overflow(per_m_cost, ncmn, &mc_denom) || mc_denom == 0) goto next_nc;
+
+        {
+            size_t remain = usable - n_fixed;
+            size_t mc = remain / mc_denom;
+            mc = hex_align_down(mc, HTP_MM_HMX_TILE_N_ROWS);
+            mc = hex_smin(mc, m);
+
+            if (mc == 0) {
+                goto next_nc;
+            }
+
+            size_t mblocks = ((size_t) m + mc - 1) / mc;
+            size_t nblocks = ((size_t) n + nc - 1) / nc;
+            size_t cost    = mblocks * m_block_cost + nblocks * n_block_cost;
+            size_t mn      = mc * nc;
+            if (cost < best_cost || (cost == best_cost && mn > best_mn)) {
+                best_cost = cost;
+                best_mn   = mn;
+                best_m    = mc;
+                best_n    = nc;
+            }
+        }
+
+next_nc:
+        if (nc == HTP_MM_HMX_TILE_N_COLS) break;  // avoid size_t underflow
+    }
+
+    if (best_m == 0 || best_n == 0) return -1;
+
+    // Compute exact total (with overflow checks)
+    size_t t0 = 0, t1 = 0, t2 = 0, mn = 0, total = 0;
+    if (hex_mul_overflow(best_n, per_n_cost, &t0)) return -1;
+    if (hex_mul_overflow(best_m, per_m_cost, &t1)) return -1;
+    if (hex_mul_overflow(best_m, best_n, &mn))     return -1;
+    if (hex_mul_overflow(mn, per_mn_cost, &t2))    return -1;
+    if (hex_add_overflow(t0, t1, &total))          return -1;
+    if (hex_add_overflow(total, t2, &total))       return -1;
+    if (hex_add_overflow(total, overhead, &total)) return -1;
+
+    *m_chunk_out = best_m;
+    *n_chunk_out = best_n;
+    *total_out   = total;
+    return 0;
+}
+
+// --- Tile Size Helpers ---
+static inline uint32_t htp_mm_get_weight_tile_size(int weight_type) {
+    switch (weight_type) {
+        case HTP_TYPE_Q4_0:
+        case HTP_TYPE_IQ4_NL:
+            return HTP_MM_WEIGHT_TILE_SIZE_Q4_0;
+        case HTP_TYPE_Q4_1:
+            return HTP_MM_WEIGHT_TILE_SIZE_Q4_1;
+        case HTP_TYPE_Q8_0:
+            return HTP_MM_WEIGHT_TILE_SIZE_Q8_0;
+        case HTP_TYPE_MXFP4:
+            return HTP_MM_WEIGHT_TILE_SIZE_MXFP4;
+        default:
+            return 0;
+    }
+}
+
+static inline uint32_t htp_mm_get_weight_aligned_tile_size(int weight_type) {
+    switch (weight_type) {
+        case HTP_TYPE_Q4_0:
+        case HTP_TYPE_IQ4_NL:
+            return HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_Q4_0;
+        case HTP_TYPE_Q4_1:
+            return HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_Q4_1;
+        case HTP_TYPE_Q8_0:
+            return HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_Q8_0;
+        case HTP_TYPE_MXFP4:
+            return HTP_MM_WEIGHT_ALIGNED_TILE_SIZE_MXFP4;
+        default:
+            return 0;
+    }
+}
+
+// --- Activation/Row Size Helpers ---
+static inline size_t htp_mm_q8_0_tiled_row_size(uint32_t ne) {
+    const uint32_t ne_padded = ((ne + 127) / 128) * 128;
+    const uint32_t nb_32 = ne_padded / 32;
+    return nb_32 * HTP_MM_ACT_TILE_SIZE_Q8_0;
+}
+
+static inline size_t htp_mm_q8_1_tiled_row_size(uint32_t ne) {
+    const uint32_t ne_padded = ((ne + 127) / 128) * 128;
+    const uint32_t nb_32 = ne_padded / 32;
+    return nb_32 * HTP_MM_ACT_TILE_SIZE_Q8_1;
+}
+
+static inline size_t htp_mm_q8_0_flat_row_size(uint32_t ne) {
+    const uint32_t quants_size = hex_align_up(ne, 128);
+    const uint32_t num_scales = (ne + 31) / 32;
+    const uint32_t scales_size = hex_align_up(num_scales * 2, 128);
+    return quants_size + scales_size;
+}
+
+static inline size_t htp_mm_q8_1_flat_row_size(uint32_t ne) {
+    const uint32_t quants_size = hex_align_up(ne, 128);
+    const uint32_t num_scales = (ne + 31) / 32;
+    const uint32_t scales_size = hex_align_up(num_scales * 4, 128);
+    return quants_size + scales_size;
+}
+
+static inline size_t htp_mm_get_tiled_row_stride(int weight_type, uint32_t k) {
+    uint32_t nb = (k + QK_Q4_0_TILED - 1) / QK_Q4_0_TILED;
+    switch (weight_type) {
+        case HTP_TYPE_Q4_0:
+        case HTP_TYPE_IQ4_NL:
+        case HTP_TYPE_Q4_1:
+        case HTP_TYPE_Q8_0:
+        case HTP_TYPE_MXFP4:
+            return (size_t) nb * htp_mm_get_weight_tile_size(weight_type);
+        case HTP_TYPE_F16:
+            return (size_t) k * sizeof(__fp16);
+        case HTP_TYPE_F32:
+            return (size_t) k * sizeof(float);
+        default:
+            return 0;
+    }
+}
+
+static inline size_t htp_mm_round_up(size_t n, size_t m) {
+    return ((n + m - 1) / m) * m;
+}
+
+static inline bool htp_mm_hmx_pipeline(uint32_t m) {
+    return m > 32;
+}
+
+static inline void htp_mm_hmx_get_2d_chunk_costs(
+    int wtype, uint32_t k, bool pipeline, uint32_t aligned_tile_size,
+    size_t * size_per_n_out, size_t * size_per_m_out, size_t * size_per_mn_out
+) {
+    const bool is_quant = (wtype != HTP_TYPE_F16 && wtype != HTP_TYPE_F32);
+    const size_t row_stride = htp_mm_get_tiled_row_stride(wtype, k);
+    const size_t vec_dot_size = k * sizeof(uint16_t);
+    const uint32_t n_k_tiles = k / HTP_MM_HMX_TILE_N_COLS;
+    const size_t qweight_row_stride = is_quant ? (size_t)(n_k_tiles * aligned_tile_size) / 32 : 0;
+
+    *size_per_n_out = (pipeline ? 2 : 1) * (is_quant ? qweight_row_stride : row_stride) +
+                      (pipeline ? 2 * vec_dot_size : vec_dot_size);
+    *size_per_m_out = vec_dot_size;
+    *size_per_mn_out = (pipeline ? 2 : 1) * sizeof(uint16_t);
+}
+
+static inline void htp_mm_hmx_get_batched_chunk_costs(
+    uint32_t k, uint32_t group_size,
+    size_t * size_per_n_out, size_t * size_per_m_out, size_t * size_per_mn_out
+) {
+    const size_t vec_dot_size = k * sizeof(uint16_t);
+    *size_per_n_out = 3 * vec_dot_size;
+    *size_per_m_out = group_size * vec_dot_size;
+    *size_per_mn_out = sizeof(uint16_t);
+}
+
+static inline size_t htp_mm_hmx_get_2d_vtcm_size(
+    int wtype, uint32_t k, size_t mc, size_t nc, bool pipeline, uint32_t act_threads, uint32_t aligned_tile_size
+) {
+    const uint32_t n_k_tiles = k / HTP_MM_HMX_TILE_N_COLS;
+    const bool is_quant = (wtype != HTP_TYPE_F16 && wtype != HTP_TYPE_F32);
+    const size_t row_stride = htp_mm_get_tiled_row_stride(wtype, k);
+    const size_t vec_dot_size = k * sizeof(uint16_t);
+
+    const size_t act_f32_size = htp_mm_round_up(act_threads * 4 * k * sizeof(float), HTP_MM_HMX_TILE_SIZE);
+    size_t weight_area_size = is_quant
+        ? htp_mm_round_up((nc / 32) * n_k_tiles * aligned_tile_size, HTP_MM_HMX_TILE_SIZE)
+        : htp_mm_round_up(nc * row_stride, HTP_MM_HMX_TILE_SIZE);
+    if (pipeline) {
+        weight_area_size *= 2;
+    }
+    const size_t act_area_size    = htp_mm_round_up(mc * vec_dot_size, HTP_MM_HMX_TILE_SIZE);
+    const size_t output_area_size = htp_mm_round_up(mc * nc * sizeof(uint16_t), HTP_MM_HMX_TILE_SIZE);
+
+    size_t scratch0_size = htp_mm_round_up(nc * vec_dot_size, HTP_MM_HMX_TILE_SIZE);
+    size_t scratch1_size = pipeline ? scratch0_size : 0;
+    size_t scratch2_size = pipeline ? output_area_size : 0;
+
+    return weight_area_size + act_area_size + act_f32_size + output_area_size +
+           scratch0_size + scratch1_size + scratch2_size + 256;
+}
+
+static inline size_t htp_mm_hmx_get_batched_vtcm_size(
+    int wtype, uint32_t k, size_t mc, size_t nc, uint32_t group_size, bool use_dma_activation, bool pipeline, uint32_t act_threads) {
+    (void)wtype;
+    (void)pipeline;
+    const size_t vec_dot_size     = k * sizeof(uint16_t);
+    const size_t f32_scratch_size = use_dma_activation
+        ? htp_mm_round_up(act_threads * 4 * k * sizeof(float), HTP_MM_HMX_TILE_SIZE) : 0;
+
+    const size_t act_head_stride   = mc * k;
+    const size_t weight_area_size  = htp_mm_round_up(nc * vec_dot_size, HTP_MM_HMX_TILE_SIZE);
+    const size_t act_area_size     = htp_mm_round_up(group_size * act_head_stride * sizeof(uint16_t), HTP_MM_HMX_TILE_SIZE);
+    const size_t output_area_size  = htp_mm_round_up(group_size * mc * nc * sizeof(uint16_t), HTP_MM_HMX_TILE_SIZE);
+    const size_t scratch_area_size = htp_mm_round_up(nc * vec_dot_size, HTP_MM_HMX_TILE_SIZE);
+
+    return weight_area_size + act_area_size + output_area_size +
+           2 * scratch_area_size + 256 + f32_scratch_size;
+}
+
+static inline size_t htp_mm_hvx_get_vtcm_sizes(
+    int kernel_type,
+    int wtype,
+    uint32_t ne10,       // k
+    uint32_t src1_nrows, // m_total (or act_nrows)
+    uint32_t n_threads,
+    size_t dst_row_size,
+    size_t src0_row_size,
+    size_t src1_row_size,
+    uint32_t n_prefetch,
+    size_t * vtcm_src0_size_out,
+    size_t * vtcm_src1_size_out,
+    size_t * vtcm_dst_size_out
+) {
+    size_t vtcm_src0_size = 0;
+    size_t vtcm_src1_size = 0;
+    size_t vtcm_dst_size  = 0;
+
+    const bool is_repack = (wtype == HTP_TYPE_Q4_0 || wtype == HTP_TYPE_Q4_1 ||
+                            wtype == HTP_TYPE_Q8_0 || wtype == HTP_TYPE_IQ4_NL ||
+                            wtype == HTP_TYPE_MXFP4);
+
+    const size_t src0_row_size_padded = htp_mm_round_up(src0_row_size, 128);
+    const size_t dst_nrows = (src1_nrows > 1) ? 0 : 1;
+
+    switch (kernel_type) {
+        case HTP_MM_KERNEL_HVX_F16_F16_VTCM: {
+            size_t f16_src1_row_size = htp_mm_round_up(ne10 * 2, 128);
+            vtcm_src1_size = htp_mm_round_up(f16_src1_row_size * src1_nrows, 256);
+            vtcm_src0_size = htp_mm_round_up(n_prefetch * src0_row_size_padded, 256) * n_threads;
+            vtcm_dst_size  = dst_nrows > 0 ? htp_mm_round_up(dst_row_size, 128) * n_threads : 0;
+            break;
+        }
+        case HTP_MM_KERNEL_HVX_F16_F32_DDR:
+        case HTP_MM_KERNEL_HVX_F16_F16_DDR:
+        case HTP_MM_KERNEL_HVX_F32_F32_DDR:
+        case HTP_MM_KERNEL_HVX_F32_F16_DDR: {
+            vtcm_src0_size = htp_mm_round_up(n_prefetch * src0_row_size, 256) * n_threads;
+            vtcm_src1_size = htp_mm_round_up(n_prefetch * src1_row_size, 256) * n_threads;
+            vtcm_dst_size  = dst_nrows > 0 ? htp_mm_round_up(dst_row_size, 128) * n_threads : 0;
+            break;
+        }
+        case HTP_MM_KERNEL_HVX_F32_F32_VTCM: {
+            size_t f32_src1_row_size = htp_mm_round_up(ne10 * 4, 128);
+            vtcm_src1_size = htp_mm_round_up(f32_src1_row_size * src1_nrows, 256);
+            vtcm_src0_size = htp_mm_round_up(n_prefetch * src0_row_size_padded, 256) * n_threads;
+            vtcm_dst_size  = dst_nrows > 0 ? htp_mm_round_up(dst_row_size, 128) * n_threads : 0;
+            break;
+        }
+        case HTP_MM_KERNEL_HVX_QUANT_BLOCK:
+        case HTP_MM_KERNEL_HVX_QUANT_ROW: {
+            size_t q_src1_row_size = (wtype == HTP_TYPE_Q4_1) ? htp_mm_q8_1_tiled_row_size(ne10) : htp_mm_q8_0_tiled_row_size(ne10);
+
+            vtcm_dst_size  = dst_nrows > 0 ? htp_mm_round_up(dst_row_size, 128) : 0;
+            vtcm_src0_size = htp_mm_round_up(n_prefetch * src0_row_size_padded, 256);
+            vtcm_src1_size = htp_mm_round_up(q_src1_row_size * src1_nrows, 256);
+
+            // src0 spad is also used in dynamic quantizer to store padded src1 rows
+            size_t src1_row_size_padded = htp_mm_round_up(q_src1_row_size, QK_Q8_0_TILED * sizeof(float));
+            if (vtcm_src0_size < src1_row_size_padded) {
+                vtcm_src0_size = src1_row_size_padded;
+            }
+
+            vtcm_src0_size = vtcm_src0_size * n_threads;
+            vtcm_dst_size  = vtcm_dst_size * n_threads;
+
+            if (is_repack) {
+                uint32_t aligned_tile_size = htp_mm_get_weight_aligned_tile_size(wtype);
+                uint32_t n_k_tiles = ne10 / 32;
+                uint32_t tile_row_size = n_k_tiles * aligned_tile_size;
+                size_t repacked_vtcm_size = htp_mm_round_up(n_prefetch * tile_row_size, 256);
+                if (repacked_vtcm_size < src1_row_size_padded) {
+                    repacked_vtcm_size = src1_row_size_padded;
+                }
+                vtcm_src0_size = repacked_vtcm_size * n_threads;
+            }
+            break;
+        }
+        case HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT: {
+            size_t q_src1_row_size = (wtype == HTP_TYPE_Q4_1) ? htp_mm_q8_1_flat_row_size(ne10) : htp_mm_q8_0_flat_row_size(ne10);
+
+            vtcm_dst_size  = dst_nrows > 0 ? htp_mm_round_up(dst_row_size, 128) : 0;
+            vtcm_src0_size = htp_mm_round_up(n_prefetch * src0_row_size_padded, 256);
+            vtcm_src1_size = htp_mm_round_up(q_src1_row_size * src1_nrows, 256);
+
+            size_t src1_row_size_padded = htp_mm_round_up(q_src1_row_size, 256);
+            if (vtcm_src0_size < src1_row_size_padded) {
+                vtcm_src0_size = src1_row_size_padded;
+            }
+
+            vtcm_src0_size = vtcm_src0_size * n_threads;
+            vtcm_dst_size  = vtcm_dst_size * n_threads;
+
+            if (is_repack) {
+                uint32_t aligned_tile_size = htp_mm_get_weight_aligned_tile_size(wtype);
+                uint32_t n_k_tiles = ne10 / 32;
+                uint32_t tile_row_size = n_k_tiles * aligned_tile_size;
+                size_t repacked_vtcm_size = htp_mm_round_up(n_prefetch * tile_row_size, 256);
+                if (repacked_vtcm_size < src1_row_size_padded) {
+                    repacked_vtcm_size = src1_row_size_padded;
+                }
+                vtcm_src0_size = repacked_vtcm_size * n_threads;
+            }
+            break;
+        }
+        default:
+            break;
+    }
+
+    *vtcm_src0_size_out = vtcm_src0_size;
+    *vtcm_src1_size_out = vtcm_src1_size;
+    *vtcm_dst_size_out  = vtcm_dst_size;
+
+    return vtcm_src0_size + vtcm_src1_size + vtcm_dst_size;
+}
+
+static inline size_t htp_mm_hvx_id_get_vtcm_sizes(
+    int wtype,
+    uint32_t ne10,                // k
+    uint32_t src1_nrows,
+    uint32_t n_threads,
+    size_t src0_row_size,    // nb01
+    uint32_t n_prefetch,
+    size_t * vtcm_src0_size_out,
+    size_t * vtcm_src1_size_out
+) {
+    const bool is_repack = (wtype == HTP_TYPE_Q4_0 || wtype == HTP_TYPE_Q4_1 ||
+                            wtype == HTP_TYPE_Q8_0 || wtype == HTP_TYPE_IQ4_NL ||
+                            wtype == HTP_TYPE_MXFP4);
+
+    const size_t src0_row_size_padded = htp_mm_round_up(src0_row_size, 128);
+    const size_t src1_row_size = (wtype == HTP_TYPE_Q4_1) ? htp_mm_q8_1_tiled_row_size(ne10)
+                                                          : htp_mm_q8_0_tiled_row_size(ne10);
+
+    size_t src0_sz_per_thread = htp_mm_round_up(n_prefetch * src0_row_size_padded, 256);
+    size_t src1_sz            = htp_mm_round_up(src1_row_size * src1_nrows, 256);
+
+    // src0 spad also holds temporary transposed src1 columns during dynamic quantization.
+    const size_t src1_row_size_padded = htp_mm_round_up(src1_row_size, QK_Q8_0_TILED * sizeof(float));
+    if (src0_sz_per_thread < src1_row_size_padded) {
+        src0_sz_per_thread = src1_row_size_padded;
+    }
+
+    if (is_repack) {
+        const uint32_t aligned_tile_size = htp_mm_get_weight_aligned_tile_size(wtype);
+        const uint32_t n_k_tiles    = ne10 / 32;
+        const uint32_t tile_row_size = n_k_tiles * aligned_tile_size;
+        size_t repacked_vtcm_size = htp_mm_round_up(n_prefetch * tile_row_size, 256);
+        if (repacked_vtcm_size < src1_row_size_padded) {
+            repacked_vtcm_size = src1_row_size_padded;
+        }
+        src0_sz_per_thread = repacked_vtcm_size;
+    }
+
+    const size_t vtcm_src0_size = src0_sz_per_thread * n_threads;
+
+    *vtcm_src0_size_out = vtcm_src0_size;
+    *vtcm_src1_size_out = src1_sz;
+
+    return vtcm_src0_size + src1_sz;
+}
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // HTP_MATMUL_OPS_H
@@ -14,8 +14,6 @@ Drivers_Dir = 13
 1 = %DiskId%

 [SourceDisksFiles]
-libggml-htp-v68.so = 1
-libggml-htp-v69.so = 1
 libggml-htp-v73.so = 1
 libggml-htp-v75.so = 1
 libggml-htp-v79.so = 1
@@ -28,8 +26,6 @@ ExcludeFromSelect = *
 CopyFiles=Drivers_Dir

 [Drivers_Dir]
-libggml-htp-v68.so,,,0x10 ;COPYFLG_NO_OVERWRITE
-libggml-htp-v69.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v73.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v75.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v79.so,,,0x10 ;COPYFLG_NO_OVERWRITE
@@ -850,6 +850,7 @@ struct ggml_backend_opencl_context {
        ref_count--;
        if (ref_count == 0) {
 #ifdef GGML_OPENCL_PROFILING
+            flush_profiling_batch();
            write_profiling_info();
            profiling_results.clear();
 #endif
@@ -10152,14 +10153,8 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const
    float eps;
    memcpy(&eps, dst->op_params, sizeof(float));

-    const int ne00 = src0 ? src0->ne[0] : 0;
-    const int ne01 = src0 ? src0->ne[1] : 0;
-    const int ne02 = src0 ? src0->ne[2] : 0;
-    const int ne03 = src0 ? src0->ne[3] : 0;
-
-    const cl_ulong nb01 = src0 ? src0->nb[1] : 0;
-    const cl_ulong nb02 = src0 ? src0->nb[2] : 0;
-    const cl_ulong nb03 = src0 ? src0->nb[3] : 0;
+    GGML_TENSOR_LOCALS(int,      ne0, src0, ne);
+    GGML_TENSOR_LOCALS(cl_ulong, nb0, src0, nb);

    const int nth = MIN(64, ne00);

@@ -10173,11 +10168,12 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const
    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),       &ne01));
    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),       &ne02));
    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),       &ne03));
-    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong),  &nb01));
-    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong),  &nb02));
-    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),  &nb03));
-    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(float),     &eps));
-    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float)*nth, NULL));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong),  &nb00));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong),  &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),  &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),  &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float),     &eps));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(float)*nth, NULL));

    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
    size_t local_work_size[] = {(size_t)nth, 1, 1};
@@ -174,7 +174,7 @@ __kernel void kernel_gemv_noshuffle_q8_0_f32(
        regA.s6 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
        regA.s7 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;

-        dequantizeBlockAccum_ns_sgbroadcast_1(totalSum, regA, regS, regB);
+        dequantizeBlockAccum_ns_sgbroadcast_1(totalSum, regA, convert_float(regS), regB);
    }

    // reduction in local memory, assumes #wave=4
@@ -24,6 +24,7 @@ kernel void kernel_norm(
        int ne01,
        int ne02,
        int ne03,
+        ulong nb00,
        ulong nb01,
        ulong nb02,
        ulong nb03,
@@ -43,7 +44,8 @@ kernel void kernel_norm(
    // parallel sum
    sum[get_local_id(0)] = 0.0f;
    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
-        sum[get_local_id(0)] += x[i00];
+        // this kernel handles float, nb00/4 translates byte offset to element offset
+        sum[get_local_id(0)] += x[i00*nb00/4];
    }
    // reduce
    barrier(CLK_LOCAL_MEM_FENCE);
@@ -60,7 +62,8 @@ kernel void kernel_norm(
    global float * y = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
    sum[get_local_id(0)] = 0.0f;
    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
-        y[i00] = x[i00] - mean;
+        // this kernel handles float, nb00/4 translates byte offset to element offset
+        y[i00] = x[i00*nb00/4] - mean;
        sum[get_local_id(0)] += y[i00] * y[i00];
    }

@@ -1270,77 +1270,14 @@ void GgmlOvDecoder::visit_subgraph(std::function<void(std::shared_ptr<GgmlDecode
 }

 std::string GgmlOvDecoder::compute_op_type(const ggml_tensor * node) {
-    static const std::map<ggml_op, std::string> ops = {
-        {GGML_OP_NONE,            "GGML_OP_NONE"           },
-        {GGML_OP_ACC,             "GGML_OP_ACC"            },
-        {GGML_OP_ADD,             "GGML_OP_ADD"            },
-        {GGML_OP_ADD1,            "GGML_OP_ADD1"           },
-        {GGML_OP_ADD_ID,          "GGML_OP_ADD_ID"         },
-        {GGML_OP_CONCAT,          "GGML_OP_CONCAT"         },
-        {GGML_OP_CONT,            "GGML_OP_CONT"           },
-        {GGML_OP_DIV,             "GGML_OP_DIV"            },
-        {GGML_OP_DUP,             "GGML_OP_DUP"            },
-        {GGML_OP_GET_ROWS,        "GGML_OP_GET_ROWS"       },
-        {GGML_OP_MUL,             "GGML_OP_MUL"            },
-        {GGML_OP_MUL_MAT,         "GGML_OP_MUL_MAT"        },
-        {GGML_OP_MUL_MAT_ID,      "GGML_OP_MUL_MAT_ID"     },
-        {GGML_OP_PERMUTE,         "GGML_OP_PERMUTE"        },
-        {GGML_OP_RESHAPE,         "GGML_OP_RESHAPE"        },
-        {GGML_OP_RMS_NORM,        "GGML_OP_RMS_NORM"       },
-        {GGML_OP_NORM,            "GGML_OP_NORM"           },
-        {GGML_OP_ROPE,            "GGML_OP_ROPE"           },
-        {GGML_OP_SCALE,           "GGML_OP_SCALE"          },
-        {GGML_OP_SOFT_MAX,        "GGML_OP_SOFT_MAX"       },
-        {GGML_OP_SUM_ROWS,        "GGML_OP_SUM_ROWS"       },
-        {GGML_OP_SUB,             "GGML_OP_SUB"            },
-        {GGML_OP_TRANSPOSE,       "GGML_OP_TRANSPOSE"      },
-        {GGML_OP_VIEW,            "GGML_OP_VIEW"           },
-        {GGML_OP_SET_ROWS,        "GGML_OP_SET_ROWS"       },
-        {GGML_OP_CPY,             "GGML_OP_CPY"            },
-        {GGML_OP_FLASH_ATTN_EXT,  "GGML_OP_FLASH_ATTN_EXT" },
-        {GGML_OP_L2_NORM,         "GGML_OP_L2_NORM"        },
-        {GGML_OP_CLAMP,           "GGML_OP_CLAMP"          },
-        {GGML_OP_PAD,             "GGML_OP_PAD"            },
-        {GGML_OP_SSM_CONV,        "GGML_OP_SSM_CONV"       },
-        {GGML_OP_GATED_DELTA_NET, "GGML_OP_GATED_DELTA_NET"},
-        {GGML_OP_ARGSORT,         "GGML_OP_ARGSORT"        },
-        {GGML_OP_REPEAT,          "GGML_OP_REPEAT"         },
-        {GGML_OP_IM2COL,          "GGML_OP_IM2COL"         }
-    };
-    static const std::map<ggml_unary_op, std::string> unary_ops = {
-        {GGML_UNARY_OP_ABS,         "GGML_UNARY_OP_ABS"        },
-        {GGML_UNARY_OP_SGN,         "GGML_UNARY_OP_SGN"        },
-        {GGML_UNARY_OP_NEG,         "GGML_UNARY_OP_NEG"        },
-        {GGML_UNARY_OP_STEP,        "GGML_UNARY_OP_STEP"       },
-        {GGML_UNARY_OP_TANH,        "GGML_UNARY_OP_TANH"       },
-        {GGML_UNARY_OP_ELU,         "GGML_UNARY_OP_ELU"        },
-        {GGML_UNARY_OP_RELU,        "GGML_UNARY_OP_RELU"       },
-        {GGML_UNARY_OP_SIGMOID,     "GGML_UNARY_OP_SIGMOID"    },
-        {GGML_UNARY_OP_GELU,        "GGML_UNARY_OP_GELU"       },
-        {GGML_UNARY_OP_GELU_QUICK,  "GGML_UNARY_OP_GELU_QUICK" },
-        {GGML_UNARY_OP_SILU,        "GGML_UNARY_OP_SILU"       },
-        {GGML_UNARY_OP_SOFTPLUS,    "GGML_UNARY_OP_SOFTPLUS"   },
-        {GGML_UNARY_OP_HARDSWISH,   "GGML_UNARY_OP_HARDSWISH"  },
-        {GGML_UNARY_OP_HARDSIGMOID, "GGML_UNARY_OP_HARDSIGMOID"},
-        {GGML_UNARY_OP_EXP,         "GGML_UNARY_OP_EXP"        },
-        {GGML_UNARY_OP_COUNT,       "GGML_UNARY_OP_COUNT"      }
-    };
-    static const std::map<ggml_glu_op, std::string> glu_ops = {
-        {GGML_GLU_OP_SWIGLU, "GGML_GLU_OP_SWIGLU"},
-        {GGML_GLU_OP_GEGLU,  "GGML_GLU_OP_GEGLU" },
-        {GGML_GLU_OP_REGLU,  "GGML_GLU_OP_REGLU" }
-    };
-
    switch (node->op) {
    case GGML_OP_UNARY:
-        return unary_ops.at(ggml_get_unary_op(node));
+        return std::string("GGML_UNARY_OP_") + ggml_unary_op_name(ggml_get_unary_op(node));
    case GGML_OP_GLU:
-        return glu_ops.at(ggml_get_glu_op(node));
+        return std::string("GGML_GLU_OP_") + ggml_glu_op_name(ggml_get_glu_op(node));
    default:
-        return ops.at(node->op);
+        return std::string("GGML_OP_") + ggml_op_name(node->op);
    }
-    static const std::string unknown_op = "UNKNOWN_GGML_OP";
-    return unknown_op;
 }

 const std::string & GgmlOvDecoder::get_op_type(int node_idx) const {
@@ -17,6 +17,22 @@ namespace frontend {
 namespace ggml {
 namespace op {

+static ov::Output<ov::Node> reshape_add_id_input_to_2d(const ov::Output<ov::Node> & input,
+                                                       const ov::PartialShape & input_shape,
+                                                       const std::vector<int> & dims) {
+    const auto actual_shape = input.get_partial_shape();
+    if (actual_shape.rank().is_static() && actual_shape.rank().get_length() == 2) {
+        return input;
+    }
+
+    if (input_shape.rank().is_static() && input_shape.rank().get_length() == 2) {
+        return input;
+    }
+
+    auto shape = std::make_shared<ov::op::v3::ShapeOf>(input, ov::element::i64);
+    return std::make_shared<ov::op::v1::Reshape>(input, get_dimensions(shape, dims), false);
+}
+
 OutputVector translate_add_id(const NodeContext & context) {
    num_inputs_check(context, 3, 3);

@@ -28,11 +44,9 @@ OutputVector translate_add_id(const NodeContext & context) {
    //   input: [1, n_token, n_used, n_embd]
    //   bias:  [1, 1, n_expert, n_embd]
    //   ids:   [1, 1, n_token, n_used]
-    auto bias_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(bias, ov::element::i64);
-    auto ids_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(ids, ov::element::i64);
-
-    bias = std::make_shared<ov::op::v1::Reshape>(bias, get_dimensions(bias_shape_4d, {2, 3}), false);
-    ids = std::make_shared<ov::op::v1::Reshape>(ids, get_dimensions(ids_shape_4d, {2, 3}), false);
+    // Model bias constants may already be stored as [n_expert, n_embd].
+    bias = reshape_add_id_input_to_2d(bias, context.get_input_shape(1), {2, 3});
+    ids = reshape_add_id_input_to_2d(ids, context.get_input_shape(2), {2, 3});

    if (ids.get_element_type() != ov::element::i32 && ids.get_element_type() != ov::element::i64) {
        ids = std::make_shared<ov::op::v0::Convert>(ids, ov::element::i32);
@@ -3,8 +3,11 @@
 #include "../utils.h"

 #include <cstdint>
+#include <limits>
 #include <memory>
 #include <openvino/core/node_output.hpp>
+#include <openvino/op/add.hpp>
+#include <openvino/op/clamp.hpp>
 #include <openvino/op/constant.hpp>
 #include <openvino/op/multiply.hpp>
 #include <openvino/op/sigmoid.hpp>
@@ -15,7 +18,7 @@ namespace frontend {
 namespace ggml {
 namespace op {

-OutputVector translate_glu_swiglu(const NodeContext & context) {
+static std::pair<ov::Output<ov::Node>, ov::Output<ov::Node>> get_glu_inputs(const NodeContext & context) {
    num_inputs_check(context, 1, 2);

    ov::Output<ov::Node> src0;
@@ -52,6 +55,12 @@ OutputVector translate_glu_swiglu(const NodeContext & context) {
        std::swap(src0, src1);
    }

+    return {src0, src1};
+}
+
+OutputVector translate_glu_swiglu(const NodeContext & context) {
+    auto [src0, src1] = get_glu_inputs(context);
+
    auto sigmoid = std::make_shared<ov::op::v0::Sigmoid>(src0);
    auto silu = std::make_shared<ov::op::v1::Multiply>(src0, sigmoid);
    auto res = std::make_shared<ov::op::v1::Multiply>(silu, src1);
@@ -59,6 +68,27 @@ OutputVector translate_glu_swiglu(const NodeContext & context) {
    return rename_outputs_with_suffix({res}, context.get_name());
 }

+OutputVector translate_glu_swiglu_oai(const NodeContext & context) {
+    auto [src0, src1] = get_glu_inputs(context);
+
+    const int32_t * params = context.get_output_op_params();
+    const float alpha = reinterpret_cast<const float *>(params)[2];
+    const float limit = reinterpret_cast<const float *>(params)[3];
+
+    auto gate = std::make_shared<ov::op::v0::Clamp>(src0, -std::numeric_limits<float>::infinity(), limit);
+    auto alpha_const = ov::op::v0::Constant::create(ov::element::f32, {}, {alpha});
+    auto scaled_gate = std::make_shared<ov::op::v1::Multiply>(gate, alpha_const);
+    auto sigmoid = std::make_shared<ov::op::v0::Sigmoid>(scaled_gate);
+    auto out_glu = std::make_shared<ov::op::v1::Multiply>(gate, sigmoid);
+
+    auto up = std::make_shared<ov::op::v0::Clamp>(src1, -limit, limit);
+    auto one = ov::op::v0::Constant::create(ov::element::f32, {}, {1.0f});
+    auto up_plus_one = std::make_shared<ov::op::v1::Add>(up, one);
+    auto res = std::make_shared<ov::op::v1::Multiply>(out_glu, up_plus_one);
+
+    return rename_outputs_with_suffix({res}, context.get_name());
+}
+
 }  // namespace op
 }  // namespace ggml
 }  // namespace frontend
@@ -2,23 +2,135 @@
 #include "../op_table.h"
 #include "../utils.h"

+#include <cstdint>
+#include <cstring>
+#include <limits>
 #include <memory>
+#include <openvino/op/bitwise_and.hpp>
+#include <openvino/op/bitwise_right_shift.hpp>
 #include <openvino/op/broadcast.hpp>
 #include <openvino/op/concat.hpp>
 #include <openvino/op/constant.hpp>
 #include <openvino/op/convert.hpp>
 #include <openvino/op/gather.hpp>
 #include <openvino/op/matmul.hpp>
+#include <openvino/op/multiply.hpp>
 #include <openvino/op/reshape.hpp>
 #include <openvino/op/shape_of.hpp>
-#include <openvino/op/squeeze.hpp>
+#include <openvino/op/slice.hpp>
 #include <openvino/op/unsqueeze.hpp>
+#include <vector>

 namespace ov {
 namespace frontend {
 namespace ggml {
 namespace op {

+namespace {
+
+std::shared_ptr<ov::op::v0::Constant> const_i64(const std::vector<int64_t> & values) {
+    return ov::op::v0::Constant::create(ov::element::i64, ov::Shape{values.size()}, values);
+}
+
+ov::Output<ov::Node> slice_axis(const ov::Output<ov::Node> & input, int64_t axis, int64_t begin, int64_t end) {
+    return std::make_shared<ov::op::v8::Slice>(input, const_i64({begin}), const_i64({end}), const_i64({1}),
+                                              const_i64({axis}));
+}
+
+ov::Output<ov::Node> translate_mul_mat_id_mxfp4_packed(const NodeContext & context,
+                                                       ov::Output<ov::Node> expert_weights,
+                                                       ov::Output<ov::Node> activations,
+                                                       ov::Output<ov::Node> ids) {
+    auto packed_shape = expert_weights.get_partial_shape().to_shape();
+    FRONT_END_OP_CONVERSION_CHECK(packed_shape.size() == 5 && packed_shape[4] == 17,
+                                  "Expected packed MXFP4 expert weights with shape [1, n_expert, m, k_blocks, 17]");
+
+    const int64_t n_expert = static_cast<int64_t>(packed_shape[1]);
+    const int64_t rows = static_cast<int64_t>(packed_shape[2]);
+    const int64_t k_blocks = static_cast<int64_t>(packed_shape[3]);
+    const int64_t qk = 32;
+    const int64_t cols = k_blocks * qk;
+
+    auto packed_shape_4d = const_i64({n_expert, rows, k_blocks, 17});
+    expert_weights = std::make_shared<ov::op::v1::Reshape>(expert_weights, packed_shape_4d, false);
+
+    auto activations_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(activations, ov::element::i64);
+    auto ids_shape_4d = std::make_shared<ov::op::v3::ShapeOf>(ids, ov::element::i64);
+    auto activations_shape_3d = get_dimensions(activations_shape_4d, {1, 2, 3});
+    auto ids_shape_2d = get_dimensions(ids_shape_4d, {2, 3});
+
+    activations = std::make_shared<ov::op::v1::Reshape>(activations, activations_shape_3d, false);
+    ids = std::make_shared<ov::op::v1::Reshape>(ids, ids_shape_2d, false);
+    if (ids.get_element_type() != ov::element::i32 && ids.get_element_type() != ov::element::i64) {
+        ids = std::make_shared<ov::op::v0::Convert>(ids, ov::element::i32);
+    }
+
+    auto gather_axis = ov::op::v0::Constant::create(ov::element::i32, ov::Shape{}, {0});
+
+    static const std::vector<float> f4e2m1_lut = {0.0f,  0.5f,  1.0f,  1.5f,  2.0f,  3.0f,  4.0f,  6.0f,
+                                                  -0.0f, -0.5f, -1.0f, -1.5f, -2.0f, -3.0f, -4.0f, -6.0f};
+    std::vector<float> e8m0_lut(256);
+    for (size_t i = 0; i < e8m0_lut.size(); ++i) {
+        uint32_t bits = static_cast<uint32_t>(i) << 23;
+        memcpy(&e8m0_lut[i], &bits, sizeof(float));
+    }
+    e8m0_lut[0] = std::numeric_limits<float>::min() / 2.0f;
+    e8m0_lut[255] = std::numeric_limits<float>::quiet_NaN();
+
+    auto f4_lut = ov::op::v0::Constant::create(ov::element::f32, ov::Shape{f4e2m1_lut.size()}, f4e2m1_lut);
+    auto scale_lut = ov::op::v0::Constant::create(ov::element::f32, ov::Shape{e8m0_lut.size()}, e8m0_lut);
+
+    auto selected_packed_weights = std::make_shared<ov::op::v8::Gather>(expert_weights, ids, gather_axis);
+    auto scale_byte = slice_axis(selected_packed_weights, 4, 0, 1);
+    auto qs = slice_axis(selected_packed_weights, 4, 1, 17);
+    auto low = std::make_shared<ov::op::v13::BitwiseAnd>(
+        qs, ov::op::v0::Constant::create(ov::element::u8, ov::Shape{}, {0x0F}), ov::op::AutoBroadcastType::NUMPY);
+    auto high_shift = std::make_shared<ov::op::v15::BitwiseRightShift>(
+        qs, ov::op::v0::Constant::create(ov::element::u8, ov::Shape{}, {4}), ov::op::AutoBroadcastType::NUMPY);
+    auto nibbles = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{low, high_shift}, 4);
+    auto nibble_indices = std::make_shared<ov::op::v0::Convert>(nibbles, ov::element::i32);
+    auto weights_f32 = std::make_shared<ov::op::v8::Gather>(f4_lut, nibble_indices, gather_axis);
+
+    auto scale_indices = std::make_shared<ov::op::v0::Convert>(scale_byte, ov::element::i32);
+    auto scales_f32 = std::make_shared<ov::op::v8::Gather>(scale_lut, scale_indices, gather_axis);
+    ov::Output<ov::Node> selected_weights = std::make_shared<ov::op::v1::Multiply>(weights_f32, scales_f32,
+                                                                                  ov::op::AutoBroadcastType::NUMPY);
+
+    auto ids_shape = std::make_shared<ov::op::v3::ShapeOf>(ids, ov::element::i64);
+    auto selected_weights_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{get_dimensions(ids_shape, {0, 1}), const_i64({rows, cols})}, 0);
+    selected_weights = std::make_shared<ov::op::v1::Reshape>(selected_weights, selected_weights_target_dims, false);
+
+    auto activations_shape = std::make_shared<ov::op::v3::ShapeOf>(activations, ov::element::i64);
+    ov::Output<ov::Node> acts_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{
+            get_dimensions(activations_shape, {0}),
+            get_dimensions(ids_shape, {1}),
+            get_dimensions(activations_shape, {2}),
+        },
+        0);
+    ov::Output<ov::Node> acts_broadcasted =
+        std::make_shared<ov::op::v3::Broadcast>(activations, acts_target_dims, ov::op::BroadcastType::BIDIRECTIONAL);
+
+    auto activations_expanded = std::make_shared<ov::op::v0::Unsqueeze>(acts_broadcasted, const_i64({2}));
+    ov::Output<ov::Node> result =
+        std::make_shared<ov::op::v0::MatMul>(activations_expanded, selected_weights, false, true);
+
+    auto batch_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+    auto row_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {rows});
+    auto result_target_dims = std::make_shared<ov::op::v0::Concat>(
+        ov::OutputVector{batch_dim, get_dimensions(ids_shape, {0, 1}), row_dim}, 0);
+    result = std::make_shared<ov::op::v1::Reshape>(result, result_target_dims, false);
+
+    const auto output_type = context.get_output_type();
+    if (result.get_element_type() != output_type) {
+        result = std::make_shared<ov::op::v0::Convert>(result, output_type);
+    }
+    return result;
+}
+
+}  // namespace
+
 OutputVector translate_mul_mat_id(const NodeContext & context) {
    num_inputs_check(context, 3, 3);

@@ -26,6 +138,12 @@ OutputVector translate_mul_mat_id(const NodeContext & context) {
    auto activations = process_view_input_new(context, 1);
    auto ids = process_view_input_new(context, 2);

+    if (expert_weights.get_element_type() == ov::element::u8 && expert_weights.get_partial_shape().rank().is_static() &&
+        expert_weights.get_partial_shape().rank().get_length() == 5) {
+        return rename_outputs_with_suffix({translate_mul_mat_id_mxfp4_packed(context, expert_weights, activations, ids)},
+                                          context.get_name());
+    }
+
    // OpenVINO sees GGML tensors in reversed dimension order:
    //   weights: [1, n_expert, m, k]
    //   activations: [1, n_tokens, n_used_or_1, k]
@@ -6,12 +6,16 @@
 #include <cstdint>
 #include <cstring>
 #include <memory>
+#include <openvino/op/broadcast.hpp>
 #include <openvino/frontend/exception.hpp>
 #include <openvino/op/add.hpp>
+#include <openvino/op/concat.hpp>
 #include <openvino/op/constant.hpp>
 #include <openvino/op/convert.hpp>
 #include <openvino/op/multiply.hpp>
 #include <openvino/op/reshape.hpp>
+#include <openvino/op/shape_of.hpp>
+#include <openvino/op/slice.hpp>
 #include <openvino/op/softmax.hpp>
 #include <vector>

@@ -20,12 +24,31 @@ namespace frontend {
 namespace ggml {
 namespace op {

+static bool is_static_one(const ov::Dimension & dim) {
+    return dim.is_static() && dim.get_length() == 1;
+}
+
+static bool same_static_dim(const ov::Dimension & lhs, const ov::Dimension & rhs) {
+    return lhs.is_static() && rhs.is_static() && lhs.get_length() == rhs.get_length();
+}
+
+static bool is_attention_sinks_input_shape(const ov::PartialShape & candidate, const ov::PartialShape & logits_shape) {
+    if (candidate.rank().is_dynamic() || logits_shape.rank().is_dynamic() || candidate.rank().get_length() != 4 ||
+        logits_shape.rank().get_length() != 4) {
+        return false;
+    }
+
+    return is_static_one(candidate[0]) && is_static_one(candidate[1]) && is_static_one(candidate[2]) &&
+           same_static_dim(candidate[3], logits_shape[1]);
+}
+
 // Reimplementation of GGML_OP_SOFT_MAX semantics for OpenVINO backend:
 // 1) logits = src0 * scale
 // 2) logits += mask (if provided)
-// 3) softmax over the last dimension
+// 3) append attention sinks as hidden logits (if provided)
+// 4) softmax over the last dimension and remove the hidden sink column
 OutputVector translate_soft_max(const NodeContext & context) {
-    num_inputs_check(context, 1, 2);
+    num_inputs_check(context, 1, 3);

    float scale = 1.0f;
    float max_bias = 0.0f;
@@ -33,6 +56,11 @@ OutputVector translate_soft_max(const NodeContext & context) {
    memcpy(&max_bias, (float *) context.get_output_op_params() + 1, sizeof(float));

    ov::Output<ov::Node> logits = context.get_input(0);
+    const bool second_input_is_sinks =
+        context.get_input_size() == 2 && is_attention_sinks_input_shape(context.get_input_shape(1), context.get_output_shape());
+    const bool has_mask = context.get_input_size() > 1 && !second_input_is_sinks;
+    const bool has_sinks = second_input_is_sinks || context.get_input_size() > 2;
+    const size_t sinks_input_idx = second_input_is_sinks ? 1 : 2;

    // Apply scale first: logits = src0 * scale
    if (scale != 1.0f) {
@@ -41,12 +69,12 @@ OutputVector translate_soft_max(const NodeContext & context) {
        logits = std::make_shared<ov::op::v1::Multiply>(logits, scale_const);
    }

-    FRONT_END_CHECK_IMPLEMENTED(!(max_bias > 0.0f && context.get_input_size() < 2),
+    FRONT_END_CHECK_IMPLEMENTED(!(max_bias > 0.0f && !has_mask),
                                "OpenVINO softmax ALiBi path requires mask input");

    // Optional mask add: logits += mask
    // For max_bias > 0 (ALiBi), apply per-head slope to mask before adding.
-    if (context.get_input_size() > 1) {
+    if (has_mask) {
        ov::Output<ov::Node> mask = context.get_input(1);

        // For stateful
@@ -94,8 +122,40 @@ OutputVector translate_soft_max(const NodeContext & context) {
        logits = std::make_shared<ov::op::v1::Add>(logits, mask);
    }

+    ov::Output<ov::Node> softmax_input = logits;
+    if (has_sinks) {
+        ov::Output<ov::Node> sinks = context.get_input(sinks_input_idx);
+        if (sinks.get_element_type() != logits.get_element_type()) {
+            sinks = std::make_shared<ov::op::v0::Convert>(sinks, logits.get_element_type());
+        }
+
+        auto sink_shape = ov::op::v0::Constant::create(ov::element::i64, {4}, {1, -1, 1, 1});
+        auto sinks_4d = std::make_shared<ov::op::v1::Reshape>(sinks, sink_shape, false);
+
+        auto logits_shape = std::make_shared<ov::op::v3::ShapeOf>(logits, ov::element::i64);
+        auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+        auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto three = ov::op::v0::Constant::create(ov::element::i64, {1}, {3});
+        auto four = ov::op::v0::Constant::create(ov::element::i64, {1}, {4});
+        auto shape_axis = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
+
+        auto sink_prefix_shape = std::make_shared<ov::op::v8::Slice>(logits_shape, zero, three, one, shape_axis);
+        auto sink_last_dim = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
+        auto sink_broadcast_shape = std::make_shared<ov::op::v0::Concat>(
+            ov::OutputVector{sink_prefix_shape, sink_last_dim}, 0);
+        auto sink_column = std::make_shared<ov::op::v3::Broadcast>(sinks_4d, sink_broadcast_shape,
+                                                                   ov::op::BroadcastType::BIDIRECTIONAL);
+        softmax_input = std::make_shared<ov::op::v0::Concat>(ov::OutputVector{logits, sink_column}, 3);
+
+        auto softmax_with_sink = std::make_shared<ov::op::v8::Softmax>(softmax_input, -1);
+        auto original_last_dim = std::make_shared<ov::op::v8::Slice>(logits_shape, three, four, one, shape_axis);
+        auto res = std::make_shared<ov::op::v8::Slice>(softmax_with_sink, zero, original_last_dim, one, three);
+
+        return rename_outputs_with_suffix({res}, context.get_name());
+    }
+
    // Softmax along last dimension (equivalent to ggml softmax over ne[0]).
-    auto res = std::make_shared<ov::op::v8::Softmax>(logits, -1);
+    auto res = std::make_shared<ov::op::v8::Softmax>(softmax_input, -1);

    return rename_outputs_with_suffix({res}, context.get_name());
 }
@@ -47,6 +47,7 @@ std::unordered_map<std::string, CreatorFunction> get_supported_ops() {
        {"GGML_UNARY_OP_TANH",      op::translate_1to1_match_1_input<v0::Tanh>     },
        {"GGML_OP_VIEW",            op::translate_view                             },
        {"GGML_GLU_OP_SWIGLU",      op::translate_glu_swiglu                       },
+        {"GGML_GLU_OP_SWIGLU_OAI",  op::translate_glu_swiglu_oai                   },
        {"GGML_GLU_OP_GEGLU",       op::translate_glu_geglu                        },
        {"GGML_OP_SET_ROWS",        op::translate_set_rows                         },
        {"GGML_OP_CPY",             op::translate_cpy                              },
@@ -32,6 +32,7 @@ GGML_OP_CONVERTER(translate_soft_max);
 GGML_OP_CONVERTER(translate_transpose);
 GGML_OP_CONVERTER(translate_view);
 GGML_OP_CONVERTER(translate_glu_swiglu);
+GGML_OP_CONVERTER(translate_glu_swiglu_oai);
 GGML_OP_CONVERTER(translate_glu_geglu);
 GGML_OP_CONVERTER(translate_set_rows);
 GGML_OP_CONVERTER(translate_cpy);
@@ -293,6 +293,11 @@ inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_t
             (sycl::ext::oneapi::bfloat16 *) dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, ne0, ne1, ne2,
             ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2, nb3, ggml_is_contiguous(src0),
             ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1), main_stream);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_BF16) {
+        op()((const sycl::ext::oneapi::bfloat16 *) src0->data, (const float *) src1->data,
+             (sycl::ext::oneapi::bfloat16 *) dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, ne0, ne1, ne2,
+             ne3, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb0, nb1, nb2, nb3, ggml_is_contiguous(src0),
+             ggml_is_contiguous(src1), ggml_is_permuted(src0), ggml_is_permuted(src1), main_stream);
 #endif
    } else {
        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__, ggml_type_name(dst->type),
@@ -103,8 +103,8 @@ void ggml_sycl_op_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
    // allocate packed arrays: A_packed (k x m), B_packed (k x n)
    ggml_sycl_pool_alloc<float> A_packed_alloc(ctx.pool());
    ggml_sycl_pool_alloc<float> B_packed_alloc(ctx.pool());
-    A_packed_alloc.alloc((size_t) knl_n_total * patch_total * sizeof(float));
-    B_packed_alloc.alloc((size_t) knl_n_total * oc * sizeof(float));
+    A_packed_alloc.alloc((size_t) knl_n_total * patch_total);
+    B_packed_alloc.alloc((size_t) knl_n_total * oc);

    float * A_packed = A_packed_alloc.get();
    float * B_packed = B_packed_alloc.get();
@@ -115,10 +115,16 @@ void ggml_sycl_op_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {

    // Combined kernel: im2col -> pack A, and pack B simultaneously
    const char * src1_base = (const char *) src1->data;
+    const char * src0_base = (const char *) src0->data;
    const int64_t src1_nb0 = src1->nb[0];
    const int64_t src1_nb1 = src1->nb[1];
    const int64_t src1_nb2 = src1->nb[2];
    const int64_t src1_nb3 = src1->nb[3];
+    const int64_t src1_w = src1->ne[0];
+    const int64_t src1_h = src1->ne[1];
+    const int64_t src1_d = src1->ne[2];
+
+    const bool src0_is_f32 = (src0->type == GGML_TYPE_F32);

    // Compute correct strides for src0 as (knl_n_total, oc) matrix
    const int64_t src0_packed_nb0 = kernel_type_size;
@@ -165,7 +171,7 @@ void ggml_sycl_op_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
        const int64_t sz = dst_z * s2 + kz * d2 - p2;

        float val = 0.0f;
-        if (sx >= 0 && sx < src1->ne[0] && sy >= 0 && sy < src1->ne[1] && sz >= 0 && sz < src1->ne[2]) {
+        if (sx >= 0 && sx < src1_w && sy >= 0 && sy < src1_h && sz >= 0 && sz < src1_d) {
            const int64_t channel_idx = batch_idx * c + ic;
            const char * ptr = src1_base + sx * src1_nb0 + sy * src1_nb1 + sz * src1_nb2 + channel_idx * src1_nb3;
            val = *(const float *) ptr;
@@ -184,9 +190,9 @@ void ggml_sycl_op_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {

        const int64_t row = t % k;
        const int64_t col = t / k;
-        const char * src_ptr = (const char *) src0->data + row * src0_packed_nb0 + col * src0_packed_nb1;
+        const char * src_ptr = src0_base + row * src0_packed_nb0 + col * src0_packed_nb1;
        float v;
-        if (src0->type == GGML_TYPE_F32) {
+        if (src0_is_f32) {
            v = *(const float *) src_ptr;
        } else {
            v = sycl::vec<sycl::half, 1>(*(const sycl::half *) src_ptr).convert<float, sycl::rounding_mode::automatic>()[0];
@@ -43,14 +43,44 @@ static __dpct_inline__ T op_sgn(T x) {
    return x > static_cast<T>(0.f) ? static_cast<T>(1.f) : ((x < static_cast<T>(0.f) ? static_cast<T>(-1.f) : static_cast<T>(0.f)));
 }

+
 template<typename T>
 static __dpct_inline__ T op_abs(T x) {
-    return sycl::fabs(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::fabs(x);  // or experimental namespace if needed
+    } else {
+        return sycl::fabs(x);
+    }
+}
+
+template<typename T>
+static __dpct_inline__ T op_expm1(T x) {
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return static_cast<sycl::ext::oneapi::bfloat16>(
+            sycl::expm1(static_cast<float>(x))
+        );
+    } else {
+        return sycl::expm1(x);
+    }
 }

 template<typename T>
 static __dpct_inline__ T op_elu(T x) {
-    return (x > static_cast<T>(0.f)) ? x : sycl::expm1(x);
+    return (x > static_cast<T>(0.f)) ? x : op_expm1(x);
+}
+
+template<typename T>
+static __dpct_inline__ T op_tanh(T x) {
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        constexpr int ver = __INTEL_LLVM_COMPILER;
+#if defined(__INTEL_LLVM_COMPILER) && (__INTEL_LLVM_COMPILER >= 20260000)
+            return sycl::ext::oneapi::experimental::tanh(x);
+#else
+            return static_cast<T>(sycl::tanh(static_cast<float>(x)));
+#endif
+    } else {
+        return sycl::tanh(x);
+    }
 }

 template<typename T>
@@ -59,74 +89,106 @@ static __dpct_inline__ T op_gelu(T x) {
    const T SQRT_2_OVER_PI = static_cast<T>(0.79788456080286535587989211986876f);
    return static_cast<T>(0.5f) * x *
           (static_cast<T>(1.0f) +
-            sycl::tanh(SQRT_2_OVER_PI * x * (static_cast<T>(1.0f) + GELU_COEF_A * x * x)));
+            op_tanh(SQRT_2_OVER_PI * x * (static_cast<T>(1.0f) + GELU_COEF_A * x * x)));
+}
+
+template<typename T>
+static __dpct_inline__ T op_exp(T x) {
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::exp(x);
+    } else {
+        return sycl::exp(x);
+    }
 }

 template<typename T>
 static __dpct_inline__ T op_silu(T x) {
-    return x / (static_cast<T>(1.0f) + sycl::native::exp(-x));
+    return x / (static_cast<T>(1.0f) + op_exp(-x));
 }

 template<typename T>
-static __dpct_inline__ T op_gelu_quick(T x) {
-    const T GELU_QUICK_COEF_LOCAL = static_cast<T>(-1.702f);
-    return x * (static_cast<T>(1.0f) / (static_cast<T>(1.0f) + sycl::native::exp(GELU_QUICK_COEF_LOCAL * x)));
+static __dpct_inline__ T op_erf(T x) {
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return static_cast<sycl::ext::oneapi::bfloat16>(
+            sycl::erf(static_cast<float>(x))
+        );
+    } else {
+        return sycl::erf(x);
+    }
 }

 template<typename T>
 static __dpct_inline__ T op_gelu_erf(T x) {
    const T SQRT_2_INV = static_cast<T>(0.70710678118654752440084436210484f);
-    return static_cast<T>(0.5f) * x * (static_cast<T>(1.0f) + sycl::erf(x * SQRT_2_INV));
+    return static_cast<T>(0.5f) * x * (static_cast<T>(1.0f) + op_erf(x * SQRT_2_INV));
 }

 template<typename T>
-static __dpct_inline__ T op_tanh(T x) {
-    return sycl::tanh(x);
+static __dpct_inline__ T op_gelu_quick(T x) {
+    const T GELU_QUICK_COEF_LOCAL = static_cast<T>(-1.702f);
+    return x * (static_cast<T>(1.0f) / (static_cast<T>(1.0f) + op_exp(GELU_QUICK_COEF_LOCAL * x)));
 }

 template<typename T>
 static __dpct_inline__ T op_relu(T x) {
-    return sycl::fmax(x, static_cast<T>(0));
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::fmax(x, static_cast<T>(0));
+    } else {
+        return sycl::fmax(x, static_cast<T>(0));
+    }
 }

 template<typename T>
 static __dpct_inline__ T op_sigmoid(T x) {
-    return static_cast<T>(1.0f) / (static_cast<T>(1.0f) + sycl::native::exp(-x));
+    return static_cast<T>(1.0f) / (static_cast<T>(1.0f) + op_exp(-x));
 }

 template<typename T>
 static __dpct_inline__ T op_sqrt(T x) {
-    return sycl::sqrt(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::sqrt(x);
+    } else {
+        return sycl::sqrt(x);
+    }
 }

 template<typename T>
 static __dpct_inline__ T op_sin(T x) {
-    return sycl::sin(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::sin(x);
+    } else {
+        return sycl::sin(x);
+    }
 }

 template<typename T>
 static __dpct_inline__ T op_cos(T x) {
-    return sycl::cos(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::cos(x);
+    } else {
+        return sycl::cos(x);
+    }
 }

 template<typename T>
 static __dpct_inline__ T op_hardsigmoid(T x) {
-    return sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::fmin(
+            static_cast<T>(1.0f), sycl::ext::oneapi::experimental::fmax(
+                                      static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+    } else {
+        return sycl::fmin(static_cast<T>(1.0f),
+                          sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+    }
 }

 template<typename T>
 static __dpct_inline__ T op_hardswish(T x) {
-    return x * sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
-}
-
-template<typename T>
-static __dpct_inline__ T op_exp(T x) {
-    return sycl::exp(x);
-}
-
-template<typename T>
-static __dpct_inline__ T op_expm1(T x) {
-    return sycl::expm1(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return x * sycl::ext::oneapi::experimental::fmin(static_cast<T>(1.0f), sycl::ext::oneapi::experimental::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+    } else {
+        return x * sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+    }
 }

 template<typename T>
@@ -134,13 +196,17 @@ static __dpct_inline__ T op_log(T x) {
    if (x <= static_cast<T>(0)) {
        return neg_infinity<T>();
    }
-    return sycl::log(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::log(x);
+    } else {
+        return sycl::log(x);
+    }
 }

 template<typename T>
 static __dpct_inline__ T op_softplus(T x) {
    const float xf = (float) x;
-    const float ax = sycl::fabs(xf);
+    const float ax = op_abs(xf);
    const float m  = sycl::fmax(xf, 0.0f);
    const float y  = m + sycl::log1p(sycl::exp(-ax));
    return (T) y;
@@ -159,8 +225,14 @@ static __dpct_inline__ T op_step(T x) {
 template<typename T>
 static __dpct_inline__ T op_leaky_relu(T x, float negative_slope) {
    T neg_slope_T = static_cast<T>(negative_slope);
-    return sycl::fmax(x, static_cast<T>(0)) +
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::fmax(x, static_cast<T>(0)) +
+           sycl::ext::oneapi::experimental::fmin(x, static_cast<T>(0.0f)) * neg_slope_T;
+
+    } else {
+        return sycl::fmax(x, static_cast<T>(0)) +
           sycl::fmin(x, static_cast<T>(0.0f)) * neg_slope_T;
+    }
 }

 template<typename T>
@@ -175,22 +247,40 @@ static __dpct_inline__ T op_clamp(T x, float min_val, float max_val) {

 template<typename T>
 static __dpct_inline__ T op_floor(T x) {
-    return sycl::floor(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::floor(x);
+    } else {
+        return sycl::floor(x);
+    }
 }

 template<typename T>
 static __dpct_inline__ T op_ceil(T x) {
-    return sycl::ceil(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::ceil(x);
+    } else {
+        return sycl::ceil(x);
+    }
 }

 template<typename T>
 static __dpct_inline__ T op_round(T x) {
-    return sycl::round(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return static_cast<sycl::ext::oneapi::bfloat16>(
+            sycl::round(static_cast<float>(x))
+        );
+    } else {
+        return sycl::round(x);
+    }
 }

 template<typename T>
 static __dpct_inline__ T op_trunc(T x) {
-    return sycl::trunc(x);
+    if constexpr (std::is_same_v<T, sycl::ext::oneapi::bfloat16>) {
+        return sycl::ext::oneapi::experimental::trunc(x);
+    } else {
+        return sycl::trunc(x);
+    }
 }

 template<typename T, typename F>
@@ -339,7 +429,7 @@ static void acc_f32_sycl(const float *x, const float *y, float *dst,
    const int num_blocks = (n_elements + SYCL_ACC_BLOCK_SIZE - 1) / SYCL_ACC_BLOCK_SIZE;
    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE),
                                           sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
-                         [=](sycl::nd_item<3> /*item_ct1*/) {
+                         [=](sycl::nd_item<3> /*item_ct1*/) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                             acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, ne13, s1, s2, s3, offset);
                         });
 }
@@ -354,8 +444,8 @@ static void arange_kernel(T * dst, const int k, T start, T step,

 template<typename KernelInvoker, typename... Args>
 static inline void dispatch_ggml_sycl_op_unary(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16 || dst->src[0]->type == GGML_TYPE_BF16);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_BF16);
    GGML_ASSERT(dst->src[0]->type == dst->type);

    dpct::queue_ptr main_stream = ctx.stream();
@@ -367,6 +457,14 @@ static inline void dispatch_ggml_sycl_op_unary(ggml_backend_sycl_context & ctx,
                kernel_invoker(data_pts.src, data_pts.dst, (int)ggml_nelements(dst->src[0]), main_stream, std::forward<Args>(args)...);
                break;
            }
+#ifdef GGML_SYCL_HAS_BF16
+        case GGML_TYPE_BF16:
+            {
+                auto data_pts = cast_data<sycl::ext::oneapi::bfloat16>(dst);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)ggml_nelements(dst->src[0]), main_stream, std::forward<Args>(args)...);
+                break;
+            }
+#endif
        case GGML_TYPE_F32:
            {
                auto data_pts = cast_data<float>(dst);
@@ -480,7 +578,7 @@ static inline void ggml_sycl_op_unary(
            stream->parallel_for(
                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
                                  sycl::range<1>(256)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                    unary_op_generic_kernel(
                        src, dst_ptr, k_elements,
                        ne0, ne1, ne2, ne3,
@@ -508,7 +606,7 @@ static inline void ggml_sycl_op_arange(ggml_backend_sycl_context & ctx, ggml_ten
    stream->parallel_for(
        sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_ARANGE_BLOCK_SIZE),
                          sycl::range<1>(SYCL_ARANGE_BLOCK_SIZE)),
-        [=](sycl::nd_item<1> item_ct1) {
+        [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
            arange_kernel(dst_ptr, k, start, step, item_ct1);
        });
 }
@@ -602,7 +700,7 @@ static inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor
            stream->parallel_for(
                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_EXP_BLOCK_SIZE),
                                  sycl::range<1>(SYCL_EXP_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                    unary_op_log_kernel(src, dst_ptr, k_elements, item_ct1);
                });
        });
@@ -640,7 +738,7 @@ static inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tenso
            stream->parallel_for(
                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SQRT_BLOCK_SIZE),
                                  sycl::range<1>(SYCL_SQRT_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                    unary_op_sqrt_kernel(src, dst_ptr, k_elements, item_ct1);
                });
        });
@@ -653,7 +751,7 @@ static inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, ggml_tensor
            stream->parallel_for(
                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIN_BLOCK_SIZE),
                                  sycl::range<1>(SYCL_SIN_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                    unary_op_sin_kernel(src, dst_ptr, k_elements, item_ct1);
                });
        });
@@ -666,7 +764,7 @@ static inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, ggml_tensor
            stream->parallel_for(
                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIN_BLOCK_SIZE),
                                  sycl::range<1>(SYCL_SIN_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                    unary_op_cos_kernel(src, dst_ptr, k_elements, item_ct1);
                });
        });
@@ -681,7 +779,7 @@ static inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, ggml
            stream->parallel_for(
                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_RELU_BLOCK_SIZE),
                                  sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                    unary_op_leaky_relu_kernel(src, dst_ptr, k_elements, slope, item_ct1);
                });
        }, negative_slope);
@@ -694,7 +792,7 @@ static inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor
            stream->parallel_for(
                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SQR_BLOCK_SIZE),
                                  sycl::range<1>(SYCL_SQR_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                    unary_op_sqr_kernel(src, dst_ptr, k_elements, item_ct1);
                });
        });
@@ -711,7 +809,7 @@ static inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tens
            stream->parallel_for(
                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE),
                                  sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE)),
-                [=](sycl::nd_item<1> item_ct1) {
+                [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                    clamp(src, dst_ptr, min_arg, max_arg, k_elements, item_ct1);
                });
        }, min_val, max_val);
@@ -774,7 +872,8 @@ static inline void ggml_sycl_op_geglu(ggml_backend_sycl_context & ctx, ggml_tens
        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
            const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
            main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
+                    sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                gated_op_fused_geglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
            });
        });
@@ -785,7 +884,8 @@ static inline void ggml_sycl_op_reglu(ggml_backend_sycl_context & ctx, ggml_tens
        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
            const uint32_t num_blocks = ceil_div((uint32_t)k, SYCL_RELU_BLOCK_SIZE); // Using RELU block size for reglu
            main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),
+                    sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                gated_op_fused_reglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
            });
        });
@@ -796,7 +896,8 @@ static inline void ggml_sycl_op_swiglu(ggml_backend_sycl_context & ctx, ggml_ten
        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
            const uint32_t num_blocks = ceil_div((uint32_t)k, SYCL_SILU_BLOCK_SIZE); // Using SILU block size for swiglu
            main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_SILU_BLOCK_SIZE)),
+                    sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                gated_op_fused_swiglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
            });
        });
@@ -811,7 +912,6 @@ __dpct_inline__ float ggml_sycl_op_swiglu_oai_single(float x, float g, float alp
    return out_glu;
 }

-
 template <typename T>
 static void swiglu_oai_kernel(const T * x, const T * g, T * dst, const int64_t k,
                              const int64_t n, const int64_t o0, const int64_t o1,
@@ -845,7 +945,7 @@ static void swiglu_oai_sycl(const T *       x,
    const int64_t num_blocks = (k + SYCL_GLU_BLOCK_SIZE - 1) / SYCL_GLU_BLOCK_SIZE;
    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GLU_BLOCK_SIZE),
                                           sycl::range<3>(1, 1, SYCL_GLU_BLOCK_SIZE)),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                             swiglu_oai_kernel(x, g, dst, k, n, o0, o1, alpha, limit, item_ct1);
                         });
 }
@@ -899,7 +999,8 @@ static inline void ggml_sycl_op_geglu_erf(ggml_backend_sycl_context & ctx, ggml_
        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
            const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
            main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
+                    sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                gated_op_fused_geglu_erf(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
            });
        });
@@ -910,7 +1011,8 @@ static inline void ggml_sycl_op_geglu_quick(ggml_backend_sycl_context & ctx, ggm
        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
            const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
            main_stream->parallel_for(
-                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
+                    sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                gated_op_fused_geglu_quick(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
            });
        });
@@ -5859,6 +5859,250 @@ static ggml_backend_dev_t ggml_backend_sycl_reg_get_device(ggml_backend_reg_t re
    return ctx->devices[index];
 }

+// ==========================================================================
+// Tensor parallelism (--split-mode tensor) for the SYCL backend.
+//
+// The meta-backend invokes these three entry points via get_proc_address:
+//   * ggml_backend_sycl_comm_init             - one-time per-graph setup
+//   * ggml_backend_sycl_comm_allreduce_tensor - per-allreduce step
+//   * ggml_backend_sycl_comm_free             - tear-down
+//
+// For N=2 (dual-GPU), this is a degenerate ring allreduce with dual paths
+// chosen by tensor size:
+//
+//   * Small (nelem < 32K): FP32 direct memcpy + per-device ADD
+//     kernel. The kernel depends_on() its corresponding memcpy event
+//     so it doesn't read partial data. Both devices run in parallel.
+//
+//   * Large (nelem >= 32K): BF16-compressed. Each device compresses
+//     its FP32 partial to BF16 locally, cross-device memcpys
+//     to the peer (half the PCI bandwidth), where it is decompressed
+//     and added into the local FP32 partial. 6 SYCL submissions per
+//     allreduce (2 compress + 2 memcpy + 2 decompress-add) vs the
+//     4 for the small path, but the bandwidth saving > 6 GB/s PCIe x 2
+//     dominates for larger tensors.
+//
+// Storage: A persistent uint8_t buffer per device, sized to
+// 4 * nelem bytes. Both paths reinterpret the same bytes (small path
+// as nelem floats; large path as outbox + inbox = 2*nelem uint16_t
+// each, using the full 4*nelem byte budget either way). Single
+// alloc+free per device keeps the SYCL pool's strict-LIFO invariant
+// trivial.
+//
+// For non-(N=2 FP32 contiguous) cases, comm_init or comm_allreduce_tensor
+// returns null/false, causing the meta-backend to use its generic
+// butterfly all-reduce fallback.
+// ==========================================================================
+
+struct ggml_backend_sycl_comm_context {
+    std::vector<ggml_backend_t> backends;
+    // ONE persistent per-device byte buffer, 4*nelem bytes.  Both the
+    // FP32 small-tensor path and the BF16 large-tensor path share it
+    // by reinterpreting.
+    std::unique_ptr<ggml_sycl_pool_alloc<uint8_t>> buf0;
+    std::unique_ptr<ggml_sycl_pool_alloc<uint8_t>> buf1;
+    int64_t buf_nelem = 0;
+};
+
+void * ggml_backend_sycl_comm_init(ggml_backend_t * backends, size_t n_backends) try {
+    for (size_t i = 0; i < n_backends; ++i) {
+        if (!ggml_backend_is_sycl(backends[i])) {
+            return nullptr;
+        }
+    }
+
+    // Initial version: N=2 only. For N!=2, returning null makes the
+    // meta-backend skip this backend-specific allreduce entirely.
+    if (n_backends != 2) {
+        return nullptr;
+    }
+
+    auto * ctx = new ggml_backend_sycl_comm_context;
+    ctx->backends.assign(backends, backends + n_backends);
+    auto * sctx0 = (ggml_backend_sycl_context *) backends[0]->context;
+    auto * sctx1 = (ggml_backend_sycl_context *) backends[1]->context;
+    ctx->buf0 = std::make_unique<ggml_sycl_pool_alloc<uint8_t>>(sctx0->pool());
+    ctx->buf1 = std::make_unique<ggml_sycl_pool_alloc<uint8_t>>(sctx1->pool());
+    return ctx;
+}
+catch (const sycl::exception &) { return nullptr; }
+catch (...)                     { return nullptr; }
+
+void ggml_backend_sycl_comm_free(void * comm_ctx_v) {
+    auto * comm_ctx = static_cast<ggml_backend_sycl_comm_context *>(comm_ctx_v);
+    if (comm_ctx == nullptr) {
+        return;
+    }
+
+    // Sync both per-device queues so the pool_alloc destructors don't
+    // return memory still in use by the last kernel.
+    if (comm_ctx->backends.size() == 2) {
+        auto * sctx0 = (ggml_backend_sycl_context *) comm_ctx->backends[0]->context;
+        auto * sctx1 = (ggml_backend_sycl_context *) comm_ctx->backends[1]->context;
+        try {
+            sctx0->stream()->wait();
+            sctx1->stream()->wait();
+        } catch (...) { /* best effort during shutdown */ }
+    }
+
+    delete comm_ctx;
+}
+
+bool ggml_backend_sycl_comm_allreduce_tensor(void * comm_ctx_v, struct ggml_tensor ** tensors) try {
+    if (comm_ctx_v == nullptr) {
+        return false;
+    }
+
+    auto * comm_ctx = static_cast<ggml_backend_sycl_comm_context *>(comm_ctx_v);
+    const size_t n_backends = comm_ctx->backends.size();
+
+    // Fast path: N=2, F32/F16, contiguous, matching shapes.
+    if (n_backends != 2) {
+        return false;
+    }
+    // Accept F32 or F16 inputs natively (types must match). F16 takes the
+    // direct 2-byte memcpy + add path below; other types return false so the
+    // meta-backend uses its generic all-reduce.
+    if (tensors[0]->type != tensors[1]->type) {
+        return false;
+    }
+    if (tensors[0]->type != GGML_TYPE_F32 && tensors[0]->type != GGML_TYPE_F16) {
+        return false;
+    }
+    if (!ggml_is_contiguous(tensors[0]) || !ggml_is_contiguous(tensors[1])) {
+        return false;
+    }
+    if (ggml_nelements(tensors[0]) != ggml_nelements(tensors[1])) {
+        return false;
+    }
+
+    const int64_t nelem  = ggml_nelements(tensors[0]);
+    const size_t  nbytes = ggml_nbytes(tensors[0]);
+    if (nelem == 0) {
+        return true;
+    }
+
+    auto * ctx0 = (ggml_backend_sycl_context *) comm_ctx->backends[0]->context;
+    auto * ctx1 = (ggml_backend_sycl_context *) comm_ctx->backends[1]->context;
+    queue_ptr q0 = ctx0->stream();
+    queue_ptr q1 = ctx1->stream();
+
+    // Grow per-device byte buffers if needed (4 * nelem bytes each).
+    if (comm_ctx->buf_nelem < nelem) {
+        comm_ctx->buf0->realloc(nelem * 4);
+        comm_ctx->buf1->realloc(nelem * 4);
+        comm_ctx->buf_nelem = nelem;
+    }
+    uint8_t * buf0 = comm_ctx->buf0->get();
+    uint8_t * buf1 = comm_ctx->buf1->get();
+
+    // F16 native path: direct 2-byte cross-device copy + add, skipping the
+    // F32 round-trip the meta-backend fallback would force. Cross-device copies
+    // go through dev2dev_memcpy because the two devices are in separate SYCL
+    // contexts (a raw peer-USM q->memcpy would be a silent no-op).
+    if (tensors[0]->type == GGML_TYPE_F16) {
+        sycl::half * f16_out0 = (sycl::half *) tensors[0]->data;
+        sycl::half * f16_out1 = (sycl::half *) tensors[1]->data;
+        sycl::half * f16_tmp0 = (sycl::half *) buf0;
+        sycl::half * f16_tmp1 = (sycl::half *) buf1;
+
+        q0->wait();
+        q1->wait();
+        dev2dev_memcpy(ctx0->device, *q0, ctx1->device, *q1, f16_tmp0, tensors[1]->data, nbytes);
+        dev2dev_memcpy(ctx1->device, *q1, ctx0->device, *q0, f16_tmp1, tensors[0]->data, nbytes);
+
+        q0->submit([&](sycl::handler & h) {
+            h.parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+                f16_out0[i] = (sycl::half) ((float) f16_out0[i] + (float) f16_tmp0[i]);
+            });
+        });
+        q1->submit([&](sycl::handler & h) {
+            h.parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+                f16_out1[i] = (sycl::half) ((float) f16_out1[i] + (float) f16_tmp1[i]);
+            });
+        });
+        return true;
+    }
+
+    float * out0 = (float *) tensors[0]->data;
+    float * out1 = (float *) tensors[1]->data;
+
+    // BF16 threshold: above this, the PCIe savings from halving the
+    // cross-device bytes outweigh the 2 extra compress kernels.
+    // Below: stay on the FP32 fast path.  Threshold mirrors the CUDA
+    // NCCL allreduce pattern for n_backends=2.
+    static constexpr int64_t BF16_THRESHOLD = 32768;
+
+    if (nelem < BF16_THRESHOLD) {
+        // FP32 small path: 4 SYCL submissions per allreduce.
+        float * tmp0 = (float *) buf0;
+        float * tmp1 = (float *) buf1;
+
+        // COMM-D2D-FIX: the two devices are in SEPARATE SYCL contexts, so a raw
+        // q->memcpy of a peer USM pointer is a silent no-op. Route cross-device
+        // copies through dev2dev_memcpy (L0 direct copy / host staging). It is
+        // synchronous, so wait for the local partials to be produced first.
+        q0->wait();
+        q1->wait();
+        dev2dev_memcpy(ctx0->device, *q0, ctx1->device, *q1, tmp0, tensors[1]->data, nbytes);
+        dev2dev_memcpy(ctx1->device, *q1, ctx0->device, *q0, tmp1, tensors[0]->data, nbytes);
+
+        q0->submit([&](sycl::handler & h) {
+            h.parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+                out0[i] += tmp0[i];
+            });
+        });
+        q1->submit([&](sycl::handler & h) {
+            h.parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+                out1[i] += tmp1[i];
+            });
+        });
+        return true;
+    }
+
+    // BF16 large path: 6 SYCL submissions per allreduce, but the
+    // cross-device memcpy is HALF the bytes. Pure bit-shift
+    // conversion (no rounding) — matches ggml's truncating fp32->bf16.
+    uint16_t * outbox0 = (uint16_t *) buf0;
+    uint16_t * inbox0  = outbox0 + nelem;
+    uint16_t * outbox1 = (uint16_t *) buf1;
+    uint16_t * inbox1  = outbox1 + nelem;
+
+    // Phase A: compress each device's local partial in parallel.
+    sycl::event c0 = q0->parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+        outbox0[i] = (uint16_t) (sycl::bit_cast<uint32_t>(out0[i]) >> 16);
+    });
+
+    sycl::event c1 = q1->parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+        outbox1[i] = (uint16_t) (sycl::bit_cast<uint32_t>(out1[i]) >> 16);
+    });
+
+    // Phase B: COMM-D2D-FIX-BF16 cross-device copy of compressed bytes via
+    // dev2dev_memcpy (separate SYCL contexts; sync copy after compress).
+    const size_t bf16_bytes = nelem * sizeof(uint16_t);
+    c0.wait();
+    c1.wait();
+    dev2dev_memcpy(ctx0->device, *q0, ctx1->device, *q1, inbox0, outbox1, bf16_bytes);
+    dev2dev_memcpy(ctx1->device, *q1, ctx0->device, *q0, inbox1, outbox0, bf16_bytes);
+
+    // Phase C: decompress + add into local FP32 partial.
+    q0->submit([&](sycl::handler & h) {
+        h.parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+            out0[i] += sycl::bit_cast<float>(((uint32_t) inbox0[i]) << 16);
+        });
+    });
+
+    q1->submit([&](sycl::handler & h) {
+        h.parallel_for(sycl::range<1>(nelem), [=](sycl::id<1> i) {
+            out1[i] += sycl::bit_cast<float>(((uint32_t) inbox1[i]) << 16);
+        });
+    });
+
+    return true;
+}
+catch (const sycl::exception &) { return false; }
+catch (...)                     { return false; }
+
 static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, const char *name) {
    GGML_UNUSED(reg);

@@ -5866,6 +6110,17 @@ static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, cons
        return (void *)ggml_backend_sycl_split_buffer_type;
    }

+    // Tensor parallelism (--split-mode tensor) entry points.
+    if (strcmp(name, "ggml_backend_comm_init") == 0) {
+        return (void *)ggml_backend_sycl_comm_init;
+    }
+    if (strcmp(name, "ggml_backend_comm_free") == 0) {
+        return (void *)ggml_backend_sycl_comm_free;
+    }
+    if (strcmp(name, "ggml_backend_comm_allreduce_tensor") == 0) {
+        return (void *)ggml_backend_sycl_comm_allreduce_tensor;
+    }
+
    // SYCL doesn't support registering host memory, left here for reference
    // "ggml_backend_register_host_buffer"
    // "ggml_backend_unregister_host_buffer"
@@ -126,7 +126,7 @@ static void soft_max_f32(const float *         x,
            break;
        }

-        const float val = sycl::native::exp(vals[col] - max_val);
+        const float val = sycl::native::exp(sycl::max(vals[col] - max_val, -80.0f));
        tmp += val;
        vals[col] = val;
    }
@@ -154,7 +154,7 @@ static void soft_max_f32(const float *         x,
        tmp = warp_reduce_sum<WARP_SIZE>(tmp);
    }
    if (sinks) {
-        tmp += sycl::native::exp(sinks[i02] - max_val);
+        tmp += sycl::native::exp(sycl::max(sinks[i02] - max_val, -80.0f));
    }
    const float inv_sum = 1.0f / tmp;

@@ -108,6 +108,9 @@ if (Vulkan_FOUND)

    if (GGML_VULKAN_CHECK_RESULTS)
        add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
+        # the result-checking path computes a CPU reference graph via
+        # ggml_graph_compute_with_ctx(), which is defined in ggml-cpu
+        target_link_libraries(ggml-vulkan PRIVATE ggml-cpu)
    endif()

    if (GGML_VULKAN_DEBUG)
@@ -129,6 +132,8 @@ if (Vulkan_FOUND)

    if (GGML_VULKAN_RUN_TESTS)
        add_compile_definitions(GGML_VULKAN_RUN_TESTS)
+        # the test path also calls ggml_graph_compute_with_ctx() (ggml-cpu)
+        target_link_libraries(ggml-vulkan PRIVATE ggml-cpu)
    endif()

    # Set up toolchain for host compilation whether cross-compiling or not
@@ -308,6 +308,7 @@ enum vk_device_architecture {
    AMD_RDNA1,
    AMD_RDNA2,
    AMD_RDNA3,
+    INTEL_XE1,
    INTEL_XE2,
    NVIDIA_PRE_TURING,
    NVIDIA_TURING,
@@ -365,21 +366,26 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
        const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();

        bool subgroup_size_control = false;
+        bool integer_dot_product = false;

        for (const auto& properties : ext_props) {
            if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
                subgroup_size_control = true;
+            } else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0) {
+                integer_dot_product = true;
            }
        }

-        if (!subgroup_size_control) {
+        if (!subgroup_size_control || !integer_dot_product) {
            return vk_device_architecture::OTHER;
        }

        vk::PhysicalDeviceProperties2 props2;
        vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
+        vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR integer_dot_props;

        props2.pNext = &subgroup_size_control_props;
+        subgroup_size_control_props.pNext = &integer_dot_props;
        device.getProperties2(&props2);

        if (subgroup_size_control_props.minSubgroupSize == 16) {
@@ -388,6 +394,9 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
            // https://www.intel.com/content/www/us/en/content-details/824434/2024-intel-tech-tour-xe2-and-lunar-lake-s-gpu.html
            // https://www.intel.com/content/www/us/en/docs/oneapi/optimization-guide-gpu/2025-0/intel-xe-gpu-architecture.html
            return vk_device_architecture::INTEL_XE2;
+        } else if (subgroup_size_control_props.minSubgroupSize == 8 &&
+                 integer_dot_product && integer_dot_props.integerDotProduct4x8BitPackedSignedAccelerated) {
+            return vk_device_architecture::INTEL_XE1;
        }
    } else if (props.vendorID == VK_VENDOR_ID_NVIDIA) {
        const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();
@@ -493,6 +502,20 @@ struct vk_conv2d_pipeline_state {
    }
 };

+struct vk_conv3d_pipeline_state {
+    vk_conv3d_pipeline_state(uint32_t s0, uint32_t s1, uint32_t s2, uint32_t p0, uint32_t p1, uint32_t p2,
+                             uint32_t d0, uint32_t d1, uint32_t d2, uint32_t KW, uint32_t KH, uint32_t KD, uint32_t aligned)
+        : s0(s0), s1(s1), s2(s2), p0(p0), p1(p1), p2(p2), d0(d0), d1(d1), d2(d2), KW(KW), KH(KH), KD(KD), aligned(aligned) {}
+
+    uint32_t s0, s1, s2, p0, p1, p2, d0, d1, d2, KW, KH, KD;
+    uint32_t aligned;
+
+    bool operator<(const vk_conv3d_pipeline_state &b) const {
+        return std::tie(s0, s1, s2, p0, p1, p2, d0, d1, d2, KW, KH, KD, aligned) <
+               std::tie(b.s0, b.s1, b.s2, b.p0, b.p1, b.p2, b.d0, b.d1, b.d2, b.KW, b.KH, b.KD, b.aligned);
+    }
+};
+
 struct vk_solve_tri_pipeline_state {
    vk_solve_tri_pipeline_state(uint32_t N, uint32_t K)
        : N(N), K(K) {}
@@ -685,6 +708,7 @@ struct vk_device_struct {

    bool add_rms_fusion;
    uint32_t partials_binding_alignment;
+    uint32_t max_nodes_per_submit;

    bool shader_64b_indexing;

@@ -777,6 +801,7 @@ struct vk_device_struct {
    vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
    vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
    vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_get_rows_back_f32;
    vk_pipeline pipeline_acc_f32;
    vk_pipeline pipeline_set_f32;

@@ -801,14 +826,10 @@ struct vk_device_struct {
    vk_pipeline pipeline_concat_i8, pipeline_concat_i16, pipeline_concat_i32, pipeline_concat_i64;
    vk_pipeline pipeline_upscale_nearest_f32, pipeline_upscale_bilinear_f32, pipeline_upscale_bicubic_f32, pipeline_upscale_bilinear_antialias_f32;
    vk_pipeline pipeline_scale_f32;
-    vk_pipeline pipeline_sqr_f32;
-    vk_pipeline pipeline_sqrt_f32;
-    vk_pipeline pipeline_sin_f32;
-    vk_pipeline pipeline_cos_f32;
    vk_pipeline pipeline_log[2];
    vk_pipeline pipeline_tri[2];
    vk_pipeline pipeline_diag[2];
-    vk_pipeline pipeline_clamp_f32;
+    vk_pipeline pipeline_clamp[2];
    vk_pipeline pipeline_pad_f32;
    vk_pipeline pipeline_roll_f32;
    vk_pipeline pipeline_repeat_i32, pipeline_repeat_back_f32;
@@ -840,6 +861,10 @@ struct vk_device_struct {
    vk_pipeline pipeline_gelu_quick[2];
    vk_pipeline pipeline_silu[2];
    vk_pipeline pipeline_relu[2];
+    vk_pipeline pipeline_sqr[2];
+    vk_pipeline pipeline_sqrt[2];
+    vk_pipeline pipeline_sin[2];
+    vk_pipeline pipeline_cos[2];
    vk_pipeline pipeline_xielu[2];
    vk_pipeline pipeline_neg[2];
    vk_pipeline pipeline_tanh[2];
@@ -871,7 +896,7 @@ struct vk_device_struct {
    vk_pipeline pipeline_geglu_erf[2];
    vk_pipeline pipeline_geglu_quick[2];

-    vk_pipeline pipeline_leaky_relu_f32;
+    vk_pipeline pipeline_leaky_relu[2];
    vk_pipeline pipeline_silu_back_f32;
    vk_pipeline pipeline_diag_mask_inf_f32;
    vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
@@ -924,6 +949,8 @@ struct vk_device_struct {
    std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv2d_f16_f32[CONV_SHAPE_COUNT];
    std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv_transpose_2d_f32[CONV_SHAPE_COUNT];
    std::map<vk_conv2d_pipeline_state, vk_pipeline> pipeline_conv_transpose_2d_f16_f32[CONV_SHAPE_COUNT];
+    std::map<vk_conv3d_pipeline_state, vk_pipeline> pipeline_conv3d_f32[CONV_SHAPE_COUNT];
+    std::map<vk_conv3d_pipeline_state, vk_pipeline> pipeline_conv3d_f16_f32[CONV_SHAPE_COUNT];
    vk_pipeline pipeline_conv2d_dw_whcn_f32, pipeline_conv2d_dw_whcn_f16_f32;
    vk_pipeline pipeline_conv2d_dw_cwhn_f32, pipeline_conv2d_dw_cwhn_f16_f32;

@@ -1669,6 +1696,41 @@ template <> void init_pushconst_fastdiv(vk_op_conv2d_push_constants &p) {
    init_fastdiv_values(p.OW*p.OH,  p.OWOHmp,  p.OWOHL);
 }

+struct vk_op_conv3d_push_constants {
+    uint32_t OC;
+    uint32_t IC;
+    uint32_t N;
+
+    uint32_t IW;
+    uint32_t IH;
+    uint32_t ID;
+    uint32_t OW;
+    uint32_t OH;
+    uint32_t OD;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
+
+    uint32_t nb1;
+    uint32_t nb2;
+    uint32_t nb3;
+
+    uint32_t OWmp;     uint32_t OWL;
+    uint32_t OWOHmp;   uint32_t OWOHL;
+    uint32_t OWOHODmp; uint32_t OWOHODL;
+};
+
+template <> void init_pushconst_fastdiv(vk_op_conv3d_push_constants &p) {
+    init_fastdiv_values(p.OW,             p.OWmp,     p.OWL);
+    init_fastdiv_values(p.OW*p.OH,        p.OWOHmp,   p.OWOHL);
+    init_fastdiv_values(p.OW*p.OH*p.OD,   p.OWOHODmp, p.OWOHODL);
+}
+
 struct vk_op_conv2d_dw_push_constants {
    uint32_t ne;
    uint32_t batches;
@@ -3784,7 +3846,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
            l_warptile = { 256, 128, 128, 16, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
            l_warptile_mmq = l_warptile_mmq_int = { 256, 128, 128, 32, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
            l_warptile_mmq_int_k = { 256, 128, 128, 32, subgroup_size_16, 64, 1, 4, 2, 1, subgroup_size_16 };
-        } else if (device->vendor_id == VK_VENDOR_ID_INTEL && device->coopmat_support && device->architecture == INTEL_XE2) {
+        } else if (device->vendor_id == VK_VENDOR_ID_INTEL && device->coopmat_support) {
            // Xe2/Xe3 with coopmat enabled - warptile performance tuning
            l_warptile = { 512, 128, 128, 16, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
            l_warptile_mmq = { 512, 128, 128, 32, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
@@ -4074,19 +4136,35 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    }
 #endif

+    auto const &ggml_vk_mul_mm_spec = [](std::vector<uint32_t> spec, bool aligned) {
+        spec.push_back(aligned ? 1u : 0u);
+        return spec;
+    };
+
    const int mul_mat_id_param_count = 5;

 #if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
    if (device->coopmat2) {
+        auto const &ggml_vk_mul_mm_cm2_spec = [](std::vector<uint32_t> spec, bool aligned, bool mul_mat_id) {
+            if (mul_mat_id && spec.size() > 5) {
+                spec.insert(spec.begin() + 5, aligned ? 1u : 0u);
+            } else {
+                spec.push_back(aligned ? 1u : 0u);
+            }
+            if (mul_mat_id && spec.size() == 6) {
+                spec.push_back(32);
+            }
+            return spec;
+        };

        // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, true);   \
-        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(l_ ## WARPTILE, false, PARAMCOUNT == mul_mat_id_param_count), 1, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(m_ ## WARPTILE, false, PARAMCOUNT == mul_mat_id_param_count), 1, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(s_ ## WARPTILE, false, PARAMCOUNT == mul_mat_id_param_count), 1, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(l_ ## WARPTILE, true, PARAMCOUNT == mul_mat_id_param_count), l_align, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(m_ ## WARPTILE, true, PARAMCOUNT == mul_mat_id_param_count), m_align, true);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_cm2_spec(s_ ## WARPTILE, true, PARAMCOUNT == mul_mat_id_param_count), s_align, true);   \

        // Create 2 variants, {f16,f32} accumulator
 #define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
@@ -4161,17 +4239,17 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
        // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
        if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, false), 1, false, true);   \
        if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, false), 1, false, true);   \
        if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, false), 1, false, true);   \
        if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, true), l_align, false, true);   \
        if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, true), m_align, false, true);   \
        if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, true), s_align, false, true);   \

        // Create 2 variants, {f16,f32} accumulator
 #define CREATE_MM2(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
@@ -4284,32 +4362,32 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
        // Selects dot2 SPIR-V variant at runtime when device->dot2_f16 is true
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
        if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _len : NAMELC ## _aligned ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _data : NAMELC ## _aligned ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, true), l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _len : NAMELC ## _aligned ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _data : NAMELC ## _aligned ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, true), m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _len : NAMELC ## _aligned ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2_aligned ## F16ACC ## _data : NAMELC ## _aligned ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _len : NAMELC ## F16ACC ## _len), (device->dot2_f16 ? NAMELC ## _dot2 ## F16ACC ## _data : NAMELC ## F16ACC ## _data), "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, true), s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \

        // bf16 scalar path promotes to f32, no dot2 variant
 #define CREATE_MM_NODOT2(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
        if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, true), l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, true), m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, true), s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \

 #define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
        if (device->mul_mat ## ID ## _l_int[TYPE]) { \
@@ -4474,17 +4552,17 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
        // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID, REQSUBGROUPSIZE) \
        if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, false), 1, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, ggml_vk_mul_mm_spec(l_ ## WARPTILE, true), l_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, ggml_vk_mul_mm_spec(m_ ## WARPTILE, true), m_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
        if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, ggml_vk_mul_mm_spec(s_ ## WARPTILE, true), s_align, false, REQSUBGROUPSIZE > 0, REQSUBGROUPSIZE);   \

 #define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
        if (device->mul_mat ## ID ## _l_int[TYPE]) \
@@ -4641,7 +4719,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    }
    uint32_t rm_iq = 2 * rm_kq;

-    const bool use_subgroups = device->subgroup_arithmetic && device->architecture != vk_device_architecture::AMD_GCN;
+    const bool use_subgroups = device->subgroup_arithmetic;
    // Ensure a subgroup size >= 16 is available
    const bool use_subgroups16 = use_subgroups && subgroup_min_size_16;

@@ -4879,6 +4957,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_NL],  "get_rows_iq4_nl_f32",  get_rows_iq4_nl_f32_len,  get_rows_iq4_nl_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_MXFP4],   "get_rows_mxfp4_f32",   get_rows_mxfp4_f32_len,   get_rows_mxfp4_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_NVFP4],   "get_rows_nvfp4_f32",   get_rows_nvfp4_f32_len,   get_rows_nvfp4_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_back_f32, "get_rows_back_f32", get_rows_back_f32_len, get_rows_back_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {256, 1, 1}, {}, 1, true);

    ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, sizeof(vk_op_flash_attn_split_k_reduce_push_constants), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
@@ -4903,7 +4982,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    }
    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", mul_mat_vec_num_bindings, sizeof(vk_mat_vec_nc_push_constants), {1, 1, 1}, {}, 1);

-    ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);

    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 4, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 0}, 1, true);
@@ -5023,11 +5102,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {

    ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);

-    ggml_vk_create_pipeline(device, device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_sqrt_f32, "sqrt_f32", sqrt_f32_len, sqrt_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_sin_f32, "sin_f32", sin_f32_len, sin_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_cos_f32, "cos_f32", cos_f32_len, cos_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-
    ggml_vk_create_pipeline(device, device->pipeline_log[0], "log_f32", log_f32_len, log_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_log[1], "log_f16", log_f16_len, log_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);

@@ -5037,8 +5111,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    ggml_vk_create_pipeline(device, device->pipeline_diag[0], "diag_f32", diag_f32_len, diag_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_diag[1], "diag_f16", diag_f16_len, diag_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);

-    ggml_vk_create_pipeline(device, device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
-
    ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_pad_push_constants), {512, 1, 1}, {}, 1);

    ggml_vk_create_pipeline(device, device->pipeline_roll_f32, "roll_f32", roll_f32_len, roll_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
@@ -5058,6 +5130,12 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    CREATE_UNARY(gelu_quick)
    CREATE_UNARY(silu)
    CREATE_UNARY(relu)
+    CREATE_UNARY(sqr)
+    CREATE_UNARY(sqrt)
+    CREATE_UNARY(sin)
+    CREATE_UNARY(cos)
+    CREATE_UNARY(clamp)
+    CREATE_UNARY(leaky_relu)
    CREATE_UNARY(xielu)
    CREATE_UNARY(neg)
    CREATE_UNARY(tanh)
@@ -5097,7 +5175,6 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
    CREATE_GLU(geglu_quick)
 #undef CREATE_GLU

-    ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_silu_back_f32, "silu_back_f32", silu_back_f32_len, silu_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);

    ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true);
@@ -5314,7 +5391,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {

    ggml_vk_create_pipeline(device, device->pipeline_opt_step_sgd_f32, "opt_step_sgd_f32", opt_step_sgd_f32_len, opt_step_sgd_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);

-    // conv2d, conv_transpose_2d
+    // conv2d, conv_transpose_2d, conv3d
    for (uint32_t s = 0; s < CONV_SHAPE_COUNT; ++s) {
        // smaller WG for the small-tile fallback gives more concurrent WGs per SM
        uint32_t conv2d_WG_SIZE  = (s == CONV_SHAPE_64x32) ? 128 : 256;
@@ -5377,8 +5454,8 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
            return (conv2d_BS.K * (conv2d_BS.CRS + pad) + conv2d_BS.CRS * (conv2d_BS.NPQ + pad) + csh_elems) * elem_size;
        };

-        // coopmat1 needs to store the output through shared memory, so check up front
-        // whether it'll fit and disable it before applying coopmat1 parameters.
+        // 2D, transpose-2D, and 3D conv use the same KxCRS @ CRSxNPQ shmem
+        // layout. cm1 needs Csh for output, so check before applying cm1 params.
        if (conv2d_use_cm1 && device->properties.limits.maxComputeSharedMemorySize < shmem_req(conv2d_cm1_shmem_pad, true, true)) {
            conv2d_use_cm1 = false;
        }
@@ -5470,6 +5547,53 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
        }
 #undef CREATE_CONV
 #undef CREATE_CONVS
+
+        std::vector<uint32_t> conv3d_spec_constants = { conv2d_WG_SIZE, conv2d_BS.K, conv2d_BS.CRS, conv2d_BS.NPQ, conv2d_TS_K, conv2d_SHMEM_PAD };
+#define CREATE_CONV3D(type_suffix, spv_suffix) \
+        for (auto &c : device->pipeline_conv3d##type_suffix[s]) { \
+            const vk_conv3d_pipeline_state &state = c.first; \
+            std::vector<uint32_t> spec_constants_cpy = conv3d_spec_constants; \
+            spec_constants_cpy.push_back(state.s0); \
+            spec_constants_cpy.push_back(state.s1); \
+            spec_constants_cpy.push_back(state.s2); \
+            spec_constants_cpy.push_back(state.p0); \
+            spec_constants_cpy.push_back(state.p1); \
+            spec_constants_cpy.push_back(state.p2); \
+            spec_constants_cpy.push_back(state.d0); \
+            spec_constants_cpy.push_back(state.d1); \
+            spec_constants_cpy.push_back(state.d2); \
+            spec_constants_cpy.push_back(state.KW); \
+            spec_constants_cpy.push_back(state.KH); \
+            spec_constants_cpy.push_back(state.KD); \
+            spec_constants_cpy.push_back(state.aligned); \
+            spec_constants_cpy.push_back(conv2d_csh_store); \
+            spec_constants_cpy.push_back(conv2d_WM); \
+            spec_constants_cpy.push_back(conv2d_WN); \
+            ggml_vk_create_pipeline( \
+                device, c.second, "conv3d" #type_suffix, \
+                conv3d##type_suffix##spv_suffix##_len, conv3d##type_suffix##spv_suffix##_data, "main", 3, \
+                sizeof(vk_op_conv3d_push_constants), wg_denoms, spec_constants_cpy, 1, true, conv2d_required_subgroup_size != 0, conv2d_required_subgroup_size); \
+        }
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+        if (device->coopmat2) {
+            CREATE_CONV3D(_f32, _cm2)
+            CREATE_CONV3D(_f16_f32, _cm2)
+        } else
+#endif
+#if defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+        if (conv2d_use_cm1) {
+            CREATE_CONV3D(_f32, _cm1)
+            CREATE_CONV3D(_f16_f32, _cm1)
+        } else
+#endif
+        if (conv2d_UNROLL) {
+            CREATE_CONV3D(_f32, _unroll)
+            CREATE_CONV3D(_f16_f32, _unroll)
+        } else {
+            CREATE_CONV3D(_f32, )
+            CREATE_CONV3D(_f16_f32, )
+        }
+#undef CREATE_CONV3D
    }

    ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
@@ -5764,6 +5888,14 @@ static vk_device ggml_vk_get_device(size_t idx) {
        device->subgroup_vote = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
                                (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eVote);

+        // Submit at least every 100 nodes, in case there are workloads without as much matmul.
+        device->max_nodes_per_submit = 100;
+        const char* GGML_VK_MAX_NODES_PER_SUBMIT = getenv("GGML_VK_MAX_NODES_PER_SUBMIT");
+        if (GGML_VK_MAX_NODES_PER_SUBMIT != nullptr) {
+            uint32_t max_nodes_per_submit = std::stoul(GGML_VK_MAX_NODES_PER_SUBMIT);
+            device->max_nodes_per_submit = std::max(max_nodes_per_submit, 1u);
+        }
+
        const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr;

        device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
@@ -6238,9 +6370,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
                break;
            case VK_VENDOR_ID_INTEL: {
                // Current Windows driver does not expose BF16 support.
-                // We only want to use l_warptile if coopmat is available and is Xe2+
-                const bool xe2_with_coopmat = device->coopmat_support && device->architecture == INTEL_XE2;
-                const bool use_l_warptile = (i == GGML_TYPE_BF16) ? (device->coopmat_bf16_support && xe2_with_coopmat) : xe2_with_coopmat;
+                // We only want to use l_warptile if coopmat is available
+                const bool use_l_warptile = (i == GGML_TYPE_BF16) ? (device->coopmat_bf16_support && device->coopmat_support) : device->coopmat_support;
                device->mul_mat_l[i] = use_l_warptile;
                device->mul_mat_id_l[i] = use_l_warptile;
                device->mul_mat_m[i] = true;
@@ -10294,6 +10425,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
            return ctx->device->pipeline_get_rows_f32[src0->type];
        }
        return nullptr;
+    case GGML_OP_GET_ROWS_BACK:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_get_rows_back_f32;
+        }
+        return nullptr;
    case GGML_OP_ACC:
        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
            return ctx->device->pipeline_acc_f32;
@@ -10400,23 +10536,27 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
        }
        return nullptr;
    case GGML_OP_SQR:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_sqr_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_sqr[dst->type == GGML_TYPE_F16];
        }
        return nullptr;
    case GGML_OP_SQRT:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_sqrt_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_sqrt[dst->type == GGML_TYPE_F16];
        }
        return nullptr;
    case GGML_OP_SIN:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_sin_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_sin[dst->type == GGML_TYPE_F16];
        }
        return nullptr;
    case GGML_OP_COS:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_cos_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_cos[dst->type == GGML_TYPE_F16];
        }
        return nullptr;
    case GGML_OP_LOG:
@@ -10438,8 +10578,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
        }
        return nullptr;
    case GGML_OP_CLAMP:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_clamp_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_clamp[dst->type == GGML_TYPE_F16];
        }
        return nullptr;
    case GGML_OP_PAD:
@@ -10807,8 +10948,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
        }
        return nullptr;
    case GGML_OP_LEAKY_RELU:
-        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_leaky_relu_f32;
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_leaky_relu[dst->type == GGML_TYPE_F16];
        }
        return nullptr;
    case GGML_OP_CONV_2D:
@@ -10885,6 +11027,61 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
            }
        }
        return nullptr;
+    case GGML_OP_CONV_3D:
+        if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            const uint32_t OC = (uint32_t)ggml_get_op_params_i32(dst, 11);
+            const uint32_t IC = (uint32_t)ggml_get_op_params_i32(dst, 9);
+            const uint32_t N  = (uint32_t)ggml_get_op_params_i32(dst, 10);
+            const uint32_t NPQ = N * dst->ne[2] * dst->ne[1] * dst->ne[0];
+            const vk_conv_shapes shape = ggml_vk_conv_select_shape(ctx, OC, NPQ);
+
+            const uint32_t KW = (uint32_t)src0->ne[0];
+            const uint32_t KH = (uint32_t)src0->ne[1];
+            const uint32_t KD = (uint32_t)src0->ne[2];
+            const uint32_t s0 = (uint32_t)ggml_get_op_params_i32(dst, 0);
+            const uint32_t s1 = (uint32_t)ggml_get_op_params_i32(dst, 1);
+            const uint32_t s2 = (uint32_t)ggml_get_op_params_i32(dst, 2);
+            const uint32_t p0 = (uint32_t)ggml_get_op_params_i32(dst, 3);
+            const uint32_t p1 = (uint32_t)ggml_get_op_params_i32(dst, 4);
+            const uint32_t p2 = (uint32_t)ggml_get_op_params_i32(dst, 5);
+            const uint32_t d0 = (uint32_t)ggml_get_op_params_i32(dst, 6);
+            const uint32_t d1 = (uint32_t)ggml_get_op_params_i32(dst, 7);
+            const uint32_t d2 = (uint32_t)ggml_get_op_params_i32(dst, 8);
+
+            const uint32_t CRS = IC * KW * KH * KD;
+            const uint32_t BS_K   = vk_conv_block_sizes[shape].K;
+            const uint32_t BS_CRS = vk_conv_block_sizes[shape].CRS;
+            const uint32_t BS_NPQ = vk_conv_block_sizes[shape].NPQ;
+            const uint32_t aligned = ((OC  % BS_K   == 0) &&
+                                      (CRS % BS_CRS == 0) &&
+                                      (NPQ % BS_NPQ == 0)) ? 1u : 0u;
+
+            vk_conv3d_pipeline_state conv3d_pipeline_state(s0, s1, s2, p0, p1, p2, d0, d1, d2, KW, KH, KD, aligned);
+
+            std::map<vk_conv3d_pipeline_state, vk_pipeline> *pipelines = nullptr;
+            if (src0->type == GGML_TYPE_F32) {
+                pipelines = &ctx->device->pipeline_conv3d_f32[shape];
+            } else if (src0->type == GGML_TYPE_F16) {
+                pipelines = &ctx->device->pipeline_conv3d_f16_f32[shape];
+            } else {
+                return nullptr;
+            }
+
+            vk_pipeline pipeline = nullptr;
+
+            {
+                std::lock_guard<std::mutex> guard(ctx->device->compile_mutex);
+                auto it = pipelines->find(conv3d_pipeline_state);
+                if (it != pipelines->end()) {
+                    pipeline = it->second;
+                } else {
+                    (*pipelines)[conv3d_pipeline_state] = pipeline = std::make_shared<vk_pipeline_struct>();
+                }
+            }
+
+            return pipeline;
+        }
+        return nullptr;
    case GGML_OP_ADD1:
        if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
            return ctx->device->pipeline_add1_f16_f16;
@@ -11135,6 +11332,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
        elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
        elements[2] = std::min(elements[2], ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
        break;
+    case GGML_OP_GET_ROWS_BACK:
+        elements = { (uint32_t)dst->ne[0], (uint32_t)dst->ne[1], 1 };
+        elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+        break;
    case GGML_OP_ARGSORT:
        GGML_ASSERT(0);
        break;
@@ -11220,6 +11421,21 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
            GGML_ABORT("invalid push constant type for CONV_2D");
        }
        break;
+    case GGML_OP_CONV_3D:
+        if constexpr (std::is_same_v<PC, vk_op_conv3d_push_constants>) {
+            const uint32_t NPQ = pc.N * pc.OD * pc.OH * pc.OW;
+            const vk_conv_shapes shape = ggml_vk_conv_select_shape(ctx, pc.OC, NPQ);
+            const uint32_t NPQ_blocks = CEIL_DIV(NPQ, vk_conv_block_sizes[shape].NPQ);
+
+            elements = { pc.OC, NPQ_blocks, 1 };
+            if (elements[1] > 512) {
+                elements[2] = CEIL_DIV(elements[1], 512);
+                elements[1] = 512;
+            }
+        } else {
+            GGML_ABORT("invalid push constant type for CONV_3D");
+        }
+        break;
    case GGML_OP_ADD:
    case GGML_OP_SUB:
    case GGML_OP_DIV:
@@ -11236,6 +11452,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
    case GGML_OP_TRI:
    case GGML_OP_DIAG:
    case GGML_OP_CLAMP:
+    case GGML_OP_LEAKY_RELU:
    case GGML_OP_PAD:
    case GGML_OP_ROLL:
    case GGML_OP_REPEAT:
@@ -11380,6 +11597,21 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
    });
 }

+static void ggml_vk_get_rows_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_GET_ROWS_BACK, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2], (uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f, 0,
+    });
+}
+
 static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    const uint32_t src0_type_size = ggml_type_size(src0->type);
    const uint32_t src1_type_size = ggml_type_size(src1->type);
@@ -12087,8 +12319,10 @@ static void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx,

 static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
    float * op_params = (float *)dst->op_params;
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    p.param1 = op_params[0];

-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f, 0.0f, 0.0f });
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_NORM, std::move(p));
 }

 static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
@@ -13118,6 +13352,51 @@ static void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx,
    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, dst->op, std::move(p));
 }

+static void ggml_vk_conv_3d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0,
+                            const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+    GGML_ASSERT(nb00 == sizeof(float) || nb00 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb10 == sizeof(float));
+    GGML_ASSERT(nb0 == sizeof(float));
+
+    vk_op_conv3d_push_constants p{};
+    p.IC = static_cast<uint32_t>(ggml_get_op_params_i32(dst, 9));
+    p.N  = static_cast<uint32_t>(ggml_get_op_params_i32(dst, 10));
+    p.OC = static_cast<uint32_t>(ggml_get_op_params_i32(dst, 11));
+    GGML_ASSERT(src0->ne[3] == (int64_t)p.IC * p.OC);
+    GGML_ASSERT(src1->ne[3] == (int64_t)p.IC * p.N);
+    GGML_ASSERT(dst->ne[3] == (int64_t)p.OC * p.N);
+
+    p.IW = static_cast<uint32_t>(ne10);
+    p.IH = static_cast<uint32_t>(ne11);
+    p.ID = static_cast<uint32_t>(ne12);
+    p.OW = static_cast<uint32_t>(ne0);
+    p.OH = static_cast<uint32_t>(ne1);
+    p.OD = static_cast<uint32_t>(ne2);
+
+    // the shader clamps src addresses to p.IC * p.N * p.IW * p.IH * p.ID - 1 in uint32, so the
+    // total input element count must fit in a uint32.
+    GGML_ASSERT((uint64_t)p.IC * p.N * p.IW * p.IH * p.ID <= 0xFFFFFFFFull);
+
+    p.nb01 = static_cast<uint32_t>(nb01 / nb00);
+    p.nb02 = static_cast<uint32_t>(nb02 / nb00);
+    p.nb03 = static_cast<uint32_t>(nb03 / nb00);
+
+    p.nb11 = static_cast<uint32_t>(nb11 / nb10);
+    p.nb12 = static_cast<uint32_t>(nb12 / nb10);
+    p.nb13 = static_cast<uint32_t>(nb13 / nb10);
+
+    p.nb1 = static_cast<uint32_t>(nb1 / nb0);
+    p.nb2 = static_cast<uint32_t>(nb2 / nb0);
+    p.nb3 = static_cast<uint32_t>(nb3 / nb0);
+
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, nullptr, dst, GGML_OP_CONV_3D, std::move(p));
+}
+
 static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    vk_op_conv2d_dw_push_constants p{};
    p.ne = ggml_nelements(dst);
@@ -13144,7 +13423,10 @@ static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx

 static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
    const float * op_params = (const float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f, 0.0f, 0.0f });
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
+    p.param1 = op_params[0];
+
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, std::move(p));
 }

 #ifdef GGML_VULKAN_RUN_TESTS
@@ -14247,6 +14529,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
    case GGML_OP_GET_ROWS:
        ggml_vk_get_rows(ctx, compute_ctx, src0, src1, node);

+        break;
+    case GGML_OP_GET_ROWS_BACK:
+        ggml_vk_get_rows_back(ctx, compute_ctx, src0, src1, node);
+
        break;
    case GGML_OP_ADD:
        if (ctx->num_additional_fused_ops) {
@@ -14515,6 +14801,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
    case GGML_OP_CONV_TRANSPOSE_2D:
        ggml_vk_conv_2d(ctx, compute_ctx, src0, src1, node);

+        break;
+    case GGML_OP_CONV_3D:
+        ggml_vk_conv_3d(ctx, compute_ctx, src0, src1, node);
+
        break;
    case GGML_OP_CONV_2D_DW:
        ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node);
@@ -15900,8 +16190,6 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
    // Submit after enough work has accumulated, to overlap CPU cmdbuffer generation with GPU execution.
    // Estimate the amount of matmul work by looking at the weight matrix size, and submit every 100MB
    // (and scaled down based on model size, so smaller models submit earlier).
-    // Also submit at least every 100 nodes, in case there are workloads without as much matmul.
-    int nodes_per_submit = 100;
    int submitted_nodes = 0;
    int submit_count = 0;
    uint64_t mul_mat_bytes = 0;
@@ -16127,7 +16415,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg

        // Signal the almost_ready fence when the graph is mostly complete (< 20% remaining)
        bool almost_ready = (cgraph->n_nodes - i) < cgraph->n_nodes / 5;
-        bool submit = (submitted_nodes >= nodes_per_submit) ||
+        bool submit = ((uint32_t)submitted_nodes >= ctx->device->max_nodes_per_submit) ||
                      (mul_mat_bytes_per_submit != 0 && mul_mat_bytes >= mul_mat_bytes_per_submit) ||
                      (i + ctx->num_additional_fused_ops >= last_node) ||
                      (almost_ready && !ctx->almost_ready_fence_pending);
@@ -16964,6 +17252,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                        return false;
                }
            }
+        case GGML_OP_GET_ROWS_BACK:
+            return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
        case GGML_OP_SET_ROWS:
            {
                switch (op->type) {
@@ -17060,12 +17350,11 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
        case GGML_OP_TRANSPOSE:
        case GGML_OP_RMS_NORM:
            return true;
-        case GGML_OP_NORM:
        case GGML_OP_GROUP_NORM:
            return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_NORM:
        case GGML_OP_L2_NORM:
-            return ggml_is_contiguous_rows(op->src[0]) &&
-                   op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
        case GGML_OP_ADD:
        case GGML_OP_SUB:
        case GGML_OP_MUL:
@@ -17084,8 +17373,9 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
        case GGML_OP_SIN:
        case GGML_OP_COS:
        case GGML_OP_CLAMP:
-            return op->src[0]->type == GGML_TYPE_F32;
        case GGML_OP_LEAKY_RELU:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
+                   op->type == op->src[0]->type;
        case GGML_OP_OPT_STEP_ADAMW:
        case GGML_OP_OPT_STEP_SGD:
            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
@@ -17285,6 +17575,13 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                    ggml_is_contiguous(op->src[1]) &&
                    ggml_is_contiguous(op));
            }
+        case GGML_OP_CONV_3D:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
+                op->src[1]->type == GGML_TYPE_F32 &&
+                op->type == GGML_TYPE_F32 &&
+                ggml_is_contiguous(op->src[0]) &&
+                ggml_is_contiguous(op->src[1]) &&
+                ggml_is_contiguous(op);
        default:
            return false;
    }
@@ -17601,9 +17898,9 @@ static bool ggml_vk_device_is_supported(const vk::PhysicalDevice & vkdev) {
 static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch) {
    switch (props.vendorID) {
    case VK_VENDOR_ID_INTEL:
-        // Only allowing Xe2 GPU at the moment since Xe2 GPU can gain significant performance boost,
-        // while some older hardware (ex. Arc A770) has performance regressions
-        return arch == vk_device_architecture::INTEL_XE2;
+        // Only allowing Xe2/Xe3 GPU and integrated Xe GPUs at the moment since older hardware (ex. Arc A770) has performance regressions.
+        return (arch == vk_device_architecture::INTEL_XE2) ||
+            (arch == vk_device_architecture::INTEL_XE1 && props.deviceType == vk::PhysicalDeviceType::eIntegratedGpu && driver_props.driverID == vk::DriverId::eIntelProprietaryWindows);
    case VK_VENDOR_ID_AMD:
        if (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource) {
            // Workaround for AMD proprietary driver reporting support on all GPUs
@@ -17651,6 +17948,8 @@ static uint32_t ggml_vk_intel_shader_core_count(const vk::PhysicalDevice& vkdev)
    case 0xE20B:  // B580
    case 0xE211:  // Pro B60
        return 20;
+    case 0xB080:  // PTL Xe3 LPG 2x6 (12 subslices)
+        return 12;
    default:
        return 0;
    }
@@ -18128,6 +18427,20 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
            const int32_t d0 = tensor->op_params[4];
            const int32_t d1 = tensor->op_params[5];
            tensor_clone = ggml_conv_2d(ggml_ctx, src_clone[0], src_clone[1], s0, s1, p0, p1, d0, d1);
+        } else if (tensor->op == GGML_OP_CONV_3D) {
+            const int32_t s0 = tensor->op_params[0];
+            const int32_t s1 = tensor->op_params[1];
+            const int32_t s2 = tensor->op_params[2];
+            const int32_t p0 = tensor->op_params[3];
+            const int32_t p1 = tensor->op_params[4];
+            const int32_t p2 = tensor->op_params[5];
+            const int32_t d0 = tensor->op_params[6];
+            const int32_t d1 = tensor->op_params[7];
+            const int32_t d2 = tensor->op_params[8];
+            const int32_t IC = tensor->op_params[9];
+            const int32_t N  = tensor->op_params[10];
+            const int32_t OC = tensor->op_params[11];
+            tensor_clone = ggml_conv_3d_direct(ggml_ctx, src_clone[0], src_clone[1], s0, s1, s2, p0, p1, p2, d0, d1, d2, IC, N, OC);
        } else if (tensor->op == GGML_OP_CONV_2D_DW) {
            const int32_t s0 = tensor->op_params[0];
            const int32_t s1 = tensor->op_params[1];
@@ -1,17 +0,0 @@
-#version 450
-
-#include "types.glsl"
-#include "generic_unary_head.glsl"
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
-    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
-}
@@ -158,7 +158,7 @@ const uint32_t Csh_stride = BS_NPQ;
 #ifdef COOPMAT
 const uint32_t Csh_len    = BS_K * Csh_stride;
 #else
-const uint32_t Csh_len    = csh_store != 0 ? BS_K * Csh_stride : 1;
+const uint32_t Csh_len    = csh_store != 0 ? BS_K * Csh_stride : 8; // 8 to workaround compiler bug
 #endif
 shared SHMEM_TYPE Csh[Csh_len];  // K x NPQ
 #endif
--- a/Show More
+++ b/Show More