llama: print memory breakdown on exit (#15860 )

* llama: print memory breakdown on exit
ggml : split graph allocations according to backend max buffer size (#15815 )
2026-06-29 17:17:40 +02:00 · 2025-09-24 16:53:48 +02:00 · 2025-09-24 16:17:49 +02:00 · 2025-09-24 13:42:26 +02:00 · 2025-09-24 10:25:26 +02:00 · 2025-09-24 09:53:47 +03:00
157 changed files with 6432 additions and 4677 deletions
@@ -17,6 +17,7 @@ Checks: >
    clang-analyzer-*,
    -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
    performance-*,
+    -performance-enum-size,
    portability-*,
    -portability-simd-intrinsics,
    misc-*,
@@ -0,0 +1,122 @@
+ARG GCC_VERSION=15.2.0
+ARG UBUNTU_VERSION=24.04
+
+### Build Llama.cpp stage
+FROM --platform=linux/s390x gcc:${GCC_VERSION} AS build
+
+RUN --mount=type=cache,target=/var/cache/apt \
+    --mount=type=cache,target=/var/lib/apt/lists \
+    apt update -y && \
+    apt upgrade -y && \
+    apt install -y --no-install-recommends \
+        git cmake ccache ninja-build \
+        # WARNING: Do not use libopenblas-openmp-dev. libopenblas-dev is faster.
+        libopenblas-dev libcurl4-openssl-dev && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+COPY . .
+
+RUN --mount=type=cache,target=/root/.ccache \
+    --mount=type=cache,target=/app/build \
+    cmake -S . -B build -G Ninja \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DCMAKE_C_COMPILER_LAUNCHER=ccache \
+        -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
+        -DLLAMA_BUILD_TESTS=OFF \
+        -DGGML_BACKEND_DL=OFF \
+        -DGGML_NATIVE=OFF \
+        -DGGML_BLAS=ON \
+        -DGGML_BLAS_VENDOR=OpenBLAS && \
+    cmake --build build --config Release -j $(nproc) && \
+    cmake --install build --prefix /opt/llama.cpp
+
+COPY *.py             /opt/llama.cpp/bin
+COPY .devops/tools.sh /opt/llama.cpp/bin
+
+COPY gguf-py          /opt/llama.cpp/gguf-py
+COPY requirements.txt /opt/llama.cpp/gguf-py
+COPY requirements     /opt/llama.cpp/gguf-py/requirements
+
+
+### Collect all llama.cpp binaries, libraries and distro libraries
+FROM --platform=linux/s390x scratch AS collector
+
+# Copy llama.cpp binaries and libraries
+COPY --from=build /opt/llama.cpp/bin     /llama.cpp/bin
+COPY --from=build /opt/llama.cpp/lib     /llama.cpp/lib
+COPY --from=build /opt/llama.cpp/gguf-py /llama.cpp/gguf-py
+
+
+### Base image
+FROM --platform=linux/s390x ubuntu:${UBUNTU_VERSION} AS base
+
+RUN --mount=type=cache,target=/var/cache/apt \
+    --mount=type=cache,target=/var/lib/apt/lists \
+    apt update -y && \
+    apt install -y --no-install-recommends \
+        # WARNING: Do not use libopenblas-openmp-dev. libopenblas-dev is faster.
+        curl libgomp1 libopenblas-dev && \
+    apt autoremove -y && \
+    apt clean -y && \
+    rm -rf /tmp/* /var/tmp/* && \
+    find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete && \
+    find /var/cache -type f -delete
+
+# Copy llama.cpp libraries
+COPY --from=collector /llama.cpp/lib /usr/lib/s390x-linux-gnu
+
+
+### Full
+FROM --platform=linux/s390x base AS full
+
+ENV PATH="/root/.cargo/bin:${PATH}"
+WORKDIR /app
+
+RUN --mount=type=cache,target=/var/cache/apt \
+    --mount=type=cache,target=/var/lib/apt/lists \
+    apt update -y && \
+    apt install -y \
+        git cmake libjpeg-dev \
+        python3 python3-pip python3-dev && \
+    apt autoremove -y && \
+    apt clean -y && \
+    rm -rf /tmp/* /var/tmp/* && \
+    find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete && \
+    find /var/cache -type f -delete
+
+RUN curl https://sh.rustup.rs -sSf | bash -s -- -y
+
+COPY --from=collector /llama.cpp/bin /app
+COPY --from=collector /llama.cpp/gguf-py /app/gguf-py
+
+RUN pip install --no-cache-dir --break-system-packages \
+        -r /app/gguf-py/requirements.txt
+
+ENTRYPOINT [ "/app/tools.sh" ]
+
+
+### CLI Only
+FROM --platform=linux/s390x base AS light
+
+WORKDIR /llama.cpp/bin
+
+# Copy llama.cpp binaries and libraries
+COPY --from=collector /llama.cpp/bin/llama-cli /llama.cpp/bin
+
+ENTRYPOINT [ "/llama.cpp/bin/llama-cli" ]
+
+
+### Server
+FROM --platform=linux/s390x base AS server
+
+ENV LLAMA_ARG_HOST=0.0.0.0
+
+WORKDIR /llama.cpp/bin
+
+# Copy llama.cpp binaries and libraries
+COPY --from=collector /llama.cpp/bin/llama-server /llama.cpp/bin
+
+EXPOSE 8080
+
+ENTRYPOINT [ "/llama.cpp/bin/llama-server" ]
@@ -6,7 +6,7 @@ on:

 jobs:
  debian-13-riscv64-native: # Bianbu 2.2
-    runs-on: self-hosted
+    runs-on: [self-hosted, RISCV64]

    steps:
      - name: Install prerequisites
@@ -1247,3 +1247,167 @@ jobs:
              -DGGML_CANN=on \
              -DSOC_TYPE=${{ matrix.device }}
          cmake --build build -j $(nproc)
+
+# TODO: simplify the following workflows using a matrix
+# TODO: run lighter CI on PRs and the full CI only on master (if needed)
+  ggml-ci-x64-cpu-low-perf:
+    runs-on: [self-hosted, Linux, X64, CPU, low-perf]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  ggml-ci-arm64-cpu-low-perf:
+    runs-on: [self-hosted, Linux, ARM64, CPU, low-perf]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  ggml-ci-x64-cpu-high-perf:
+    runs-on: [self-hosted, Linux, X64, CPU, high-perf]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  ggml-ci-arm64-cpu-high-perf:
+    runs-on: [self-hosted, Linux, ARM64, CPU, high-perf]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  ggml-ci-x64-nvidia-cuda:
+    runs-on: [self-hosted, Linux, X64, NVIDIA]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          nvidia-smi
+          GG_BUILD_CUDA=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  ggml-ci-x64-nvidia-vulkan-cm:
+    runs-on: [self-hosted, Linux, X64, NVIDIA]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 GGML_VK_DISABLE_COOPMAT2=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  ggml-ci-x64-nvidia-vulkan-cm2:
+    runs-on: [self-hosted, Linux, X64, NVIDIA, COOPMAT2]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  ggml-ci-x64-cpu-amx:
+    runs-on: [self-hosted, Linux, X64, CPU, AMX]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+#  ggml-ci-x64-amd-vulkan:
+#    runs-on: [self-hosted, Linux, X64, AMD]
+#
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v4
+#
+#      - name: Test
+#        id: ggml-ci
+#        run: |
+#          vulkaninfo --summary
+#          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+#
+#  ggml-ci-x64-amd-rocm:
+#    runs-on: [self-hosted, Linux, X64, AMD]
+#
+#    steps:
+#      - name: Clone
+#        id: checkout
+#        uses: actions/checkout@v4
+#
+#      - name: Test
+#        id: ggml-ci
+#        run: |
+#          amd-smi static
+#          GG_BUILD_ROCM=1 GG_BUILD_AMDGPU_TARGETS="gfx1101" bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
+  ggml-ci-mac-metal:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  ggml-ci-mac-vulkan:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          vulkaninfo --summary
+          GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
@@ -44,6 +44,7 @@ jobs:
          - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
          - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
          - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
+          - { tag: "s390x", dockerfile: ".devops/s390x.Dockerfile", platforms: "linux/s390x", full: true, light: true, server: true, free_disk_space: false }
          # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
          #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true }
    steps:
@@ -1,12 +1,106 @@
 # collaborators can optionally add themselves here to indicate their availability for reviewing related PRs
+# multiplie collaborators per item can be specified

-/ci/ @ggerganov
-/.devops/*.Dockerfile @ngxson
-/tools/server/ @ngxson
-/ggml/src/ggml-cuda/fattn* @JohannesGaessler
-/ggml/src/ggml-cuda/mmq.* @JohannesGaessler
-/ggml/src/ggml-cuda/mmvq.* @JohannesGaessler
-/ggml/src/ggml-opt.cpp @JohannesGaessler
-/ggml/src/gguf.cpp @JohannesGaessler
-/ggml/src/ggml-vulkan/ @0cc4m
-/ggml/src/ggml-zdnn/ @taronaeo
+/.devops/*.Dockerfile                   @ngxson
+/.github/actions/                       @slaren
+/.github/workflows/                     @CISC
+/.github/workflows/release.yml          @slaren
+/.github/workflows/winget.yml           @slaren
+/ci/                                    @ggerganov
+/cmake/                                 @ggerganov
+/common/CMakeLists.txt                  @ggerganov
+/common/arg.*                           @ggerganov @ericcurtin
+/common/base64.hpp.*                    @ggerganov
+/common/build-info.*                    @ggerganov
+/common/common.*                        @ggerganov
+/common/console.*                       @ggerganov
+/common/llguidance.*                    @ggerganov
+/common/log.*                           @ggerganov
+/common/sampling.*                      @ggerganov
+/common/speculative.*                   @ggerganov
+/convert_*.py                           @CISC
+/examples/batched.swift/                @ggerganov
+/examples/batched/                      @ggerganov
+/examples/convert-llama2c-to-ggml/      @ggerganov
+/examples/deprecation-warning/          @ggerganov
+/examples/diffusion/                    @am17an
+/examples/embedding/                    @ggerganov
+/examples/eval-callback/                @ggerganov
+/examples/export-docs/                  @ggerganov
+/examples/gen-docs/                     @ggerganov
+/examples/gguf/                         @ggerganov
+/examples/llama.android/                @ggerganov
+/examples/llama.swiftui/                @ggerganov
+/examples/llama.vim                     @ggerganov
+/examples/lookahead/                    @ggerganov
+/examples/lookup/                       @JohannesGaessler
+/examples/model-conversion/             @danbev
+/examples/parallel/                     @ggerganov
+/examples/passkey/                      @ggerganov
+/examples/retrieval/                    @ggerganov
+/examples/save-load-state/              @ggerganov
+/examples/simple-chat/                  @slaren
+/examples/simple/                       @slaren
+/examples/speculative-simple/           @ggerganov
+/examples/speculative/                  @ggerganov
+/ggml/cmake/                            @ggerganov
+/ggml/include/                          @ggerganov @slaren
+/ggml/src/ggml-alloc.c                  @slaren
+/ggml/src/ggml-backend*                 @slaren
+/ggml/src/ggml-blas/                    @slaren
+/ggml/src/ggml-common.h                 @ggerganov @slaren
+/ggml/src/ggml-cpu/                     @ggerganov @slaren
+/ggml/src/ggml-cuda/common.cuh          @slaren
+/ggml/src/ggml-cuda/fattn*              @JohannesGaessler
+/ggml/src/ggml-cuda/ggml-cuda.cu        @slaren
+/ggml/src/ggml-cuda/mmf.*               @JohannesGaessler
+/ggml/src/ggml-cuda/mmq.*               @JohannesGaessler
+/ggml/src/ggml-cuda/mmvf.*              @JohannesGaessler
+/ggml/src/ggml-cuda/mmvq.*              @JohannesGaessler
+/ggml/src/ggml-impl.h                   @ggerganov @slaren
+/ggml/src/ggml-metal/                   @ggerganov
+/ggml/src/ggml-opt.cpp                  @JohannesGaessler
+/ggml/src/ggml-quants.*                 @ggerganov
+/ggml/src/ggml-threading.*              @ggerganov @slaren
+/ggml/src/ggml-vulkan/                  @0cc4m
+/ggml/src/ggml-zdnn/                    @taronaeo
+/ggml/src/ggml.c                        @ggerganov @slaren
+/ggml/src/ggml.cpp                      @ggerganov @slaren
+/ggml/src/gguf.cpp                      @JohannesGaessler @Green-Sky
+/gguf-py/                               @CISC
+/media/                                 @ggerganov
+/scripts/gen*                           @ggerganov
+/scripts/get*                           @ggerganov
+/scripts/sync*                          @ggerganov
+/src/                                   @ggerganov
+/src/llama-adapter.*                    @CISC
+/src/llama-arch.*                       @CISC
+/src/llama-chat.*                       @ngxson
+/src/llama-graph.*                      @CISC
+/src/llama-model-loader.*               @slaren
+/src/llama-model.*                      @CISC
+/src/llama-vocab.*                      @CISC
+/tests/                                 @ggerganov
+/tests/test-backend-ops.cpp             @slaren
+/tests/test-thread-safety.cpp           @slaren
+/tools/batched-bench/                   @ggerganov
+/tools/llama-bench/                     @slaren
+/tools/main/                            @ggerganov
+/tools/mtmd/                            @ngxson
+/tools/perplexity/                      @ggerganov
+/tools/quantize/                        @ggerganov
+/tools/run/                             @ericcurtin
+/tools/server/*                         @ngxson @ggerganov @ericcurtin # no subdir
+/tools/server/webui/                    @allozaur
+/tools/tokenize/                        @ggerganov
+/tools/tts/                             @ggerganov
+/vendor/                                @ggerganov
+/.clang-format                          @slaren
+/.clang-tidy                            @slaren
+/AUTHORS                                @ggerganov
+/CMakeLists.txt                         @ggerganov
+/CONTRIBUTING.md                        @ggerganov
+/LICENSE                                @ggerganov
+/README.md                              @ggerganov
+/SECURITY.md                            @ggerganov
+requirements*.txt                       @CISC
@@ -1,4 +1,12 @@
-# Pull requests (for contributors)
+# Contributors
+
+The project differentiates between 3 levels of contributors:
+
+- Contributors: people who have contributed before (no special privileges)
+- Collaborators (Triage): people with significant contributions, who may be responsible for some parts of the code, and are expected to maintain and review contributions for the code they own
+- Maintainers: responsible for reviewing and merging PRs, after approval from the code owners
+
+# Pull requests (for contributors & collaborators)

 - llama.cpp uses the ggml tensor library for model evaluation. If you are unfamiliar with ggml, consider taking a look at the [examples in the ggml repository](https://github.com/ggml-org/ggml/tree/master/examples/). [simple](https://github.com/ggml-org/ggml/tree/master/examples/simple) shows the bare minimum for using ggml. [gpt-2](https://github.com/ggml-org/ggml/tree/master/examples/gpt-2) has minimal implementations for language model inference using GPT-2. [mnist](https://github.com/ggml-org/ggml/tree/master/examples/mnist) demonstrates how to train and evaluate a simple image classifier
 - Test your changes:
@@ -9,15 +17,16 @@
 - Create separate PRs for each feature or fix. Avoid combining unrelated changes in a single PR
 - Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
 - If your PR becomes stale, don't hesitate to ping the maintainers in the comments
+- Maintainers will rely on your insights and approval when making a final decision to approve and merge a PR
+- Consider adding yourself to [CODEOWNERS](CODEOWNERS) to indicate your availability for reviewing related PRs

-# Pull requests (for collaborators)
+# Pull requests (for maintainers)

 - Squash-merge PRs
 - Use the following format for the squashed commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
 - Optionally pick a `<module>` from here: https://github.com/ggml-org/llama.cpp/wiki/Modules
- Consider adding yourself to [CODEOWNERS](CODEOWNERS)
- Let authors, who are also collaborators, merge their own PRs
- When merging a PR by a contributor, make sure you have a good understanding of the changes
+- Let other maintainers, merge their own PRs
+- When merging a PR, make sure you have a good understanding of the changes
 - Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)

 # Coding guidelines
@@ -117,6 +126,21 @@
    #endif // FOO
    ```

+# Code maintenance
+
+- Existing code should have designated collaborators and/or maintainers specified in the [CODEOWNERS](CODEOWNERS) file reponsible for:
+  - Reviewing and merging related PRs
+  - Fixing related bugs
+  - Providing developer guidance/support
+
+- When adding or modifying a large piece of code:
+  - If you are a collaborator, make sure to add yourself to [CODEOWNERS](CODEOWNERS) to indicate your availability for reviewing related PRs
+  - If you are a contributor, find an existing collaborator who is willing to review and maintain your code long-term
+  - Provide the necessary CI workflow (and hardware) to test your changes (see [ci/README.md](https://github.com/ggml-org/llama.cpp/tree/master/ci))
+
+- New code should follow the guidelines (coding, naming, etc.) outlined in this document. Exceptions are allowed in isolated, backend-specific parts of the code that do not interface directly with the `ggml` interfaces.
+  _(NOTE: for legacy reasons, existing code is not required to follow this guideline)_
+
 # Documentation

 - Documentation is a community effort
@@ -274,6 +274,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [Vulkan](docs/build.md#vulkan) | GPU |
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
+| [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
 | [WebGPU [In Progress]](docs/build.md#webgpu) | All |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |

@@ -520,8 +521,8 @@ To learn more about model quantization, [read this documentation](tools/quantize
 ## Contributing

 - Contributors can open PRs
- Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
 - Collaborators will be invited based on contributions
+- Maintainers can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
 - Any help with managing issues, PRs and projects is very appreciated!
 - See [good first issues](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
 - Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
@@ -0,0 +1,35 @@
+## Running MUSA CI in a Docker Container
+
+Assuming `$PWD` is the root of the `llama.cpp` repository, follow these steps to set up and run MUSA CI in a Docker container:
+
+### 1. Create a local directory to store cached models, configuration files and venv:
+
+```bash
+mkdir -p $HOME/llama.cpp/ci-cache
+```
+
+### 2. Create a local directory to store CI run results:
+
+```bash
+mkdir -p $HOME/llama.cpp/ci-results
+```
+
+### 3. Start a Docker container and run the CI:
+
+```bash
+docker run --privileged -it \
+    -v $HOME/llama.cpp/ci-cache:/ci-cache \
+    -v $HOME/llama.cpp/ci-results:/ci-results \
+    -v $PWD:/ws -w /ws \
+    mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
+```
+
+Inside the container, execute the following commands:
+
+```bash
+apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget
+git config --global --add safe.directory /ws
+GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
+```
+
+This setup ensures that the CI runs within an isolated Docker environment while maintaining cached files and results across runs.
@@ -1,18 +1,10 @@
 # CI

-In addition to [Github Actions](https://github.com/ggml-org/llama.cpp/actions) `llama.cpp` uses a custom CI framework:
+This CI implements heavy-duty workflows that run on self-hosted runners. Typically the purpose of these workflows is to
+cover hardware configurations that are not available from Github-hosted runners and/or require more computational
+resource than normally available.

-https://github.com/ggml-org/ci
-
-It monitors the `master` branch for new commits and runs the
-[ci/run.sh](https://github.com/ggml-org/llama.cpp/blob/master/ci/run.sh) script on dedicated cloud instances. This allows us
-to execute heavier workloads compared to just using Github Actions. Also with time, the cloud instances will be scaled
-to cover various hardware architectures, including GPU and Apple Silicon instances.
-
-Collaborators can optionally trigger the CI run by adding the `ggml-ci` keyword to their commit message.
-Only the branches of this repo are monitored for this keyword.
-
-It is a good practice, before publishing changes to execute the full CI locally on your machine:
+It is a good practice, before publishing changes to execute the full CI locally on your machine. For example:

 ```bash
 mkdir tmp
@@ -29,40 +21,13 @@ GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

 # with MUSA support
 GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+
+# etc.
 ```

-## Running MUSA CI in a Docker Container
+# Adding self-hosted runners

-Assuming `$PWD` is the root of the `llama.cpp` repository, follow these steps to set up and run MUSA CI in a Docker container:
-
-### 1. Create a local directory to store cached models, configuration files and venv:
-
-```bash
-mkdir -p $HOME/llama.cpp/ci-cache
-```
-
-### 2. Create a local directory to store CI run results:
-
-```bash
-mkdir -p $HOME/llama.cpp/ci-results
-```
-
-### 3. Start a Docker container and run the CI:
-
-```bash
-docker run --privileged -it \
-    -v $HOME/llama.cpp/ci-cache:/ci-cache \
-    -v $HOME/llama.cpp/ci-results:/ci-results \
-    -v $PWD:/ws -w /ws \
-    mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
-```
-
-Inside the container, execute the following commands:
-
-```bash
-apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget
-git config --global --add safe.directory /ws
-GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache
-```
-
-This setup ensures that the CI runs within an isolated Docker environment while maintaining cached files and results across runs.
+- Add a self-hosted `ggml-ci` workflow to [[.github/workflows/build.yml]] with an appropriate label
+- Request a runner token from `ggml-org` (for example, via a comment in the PR or email)
+- Set-up a machine using the received token ([docs](https://docs.github.com/en/actions/how-tos/manage-runners/self-hosted-runners/add-runners))
+- Optionally update [ci/run.sh](https://github.com/ggml-org/llama.cpp/blob/master/ci/run.sh) to build and run on the target platform by gating the implementation with a `GG_BUILD_...` env
@@ -65,6 +65,16 @@ if [ ! -z ${GG_BUILD_CUDA} ]; then
    fi
 fi

+if [ ! -z ${GG_BUILD_ROCM} ]; then
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_HIP=ON"
+    if [ -z ${GG_BUILD_AMDGPU_TARGETS} ]; then
+        echo "Missing GG_BUILD_AMDGPU_TARGETS, please set it to your GPU architecture (e.g. gfx90a, gfx1100, etc.)"
+        exit 1
+    fi
+
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DAMDGPU_TARGETS=${GG_BUILD_AMDGPU_TARGETS}"
+fi
+
 if [ ! -z ${GG_BUILD_SYCL} ]; then
    if [ -z ${ONEAPI_ROOT} ]; then
        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:"
@@ -82,6 +92,12 @@ fi

 if [ ! -z ${GG_BUILD_VULKAN} ]; then
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
+
+    # if on Mac, disable METAL
+    if [[ "$OSTYPE" == "darwin"* ]]; then
+        CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=OFF -DGGML_BLAS=OFF"
+    fi
+
 fi

 if [ ! -z ${GG_BUILD_WEBGPU} ]; then
@@ -150,7 +166,7 @@ function gg_run_ctest_debug {
    (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
    (time make -j$(nproc)                                  ) 2>&1 | tee -a $OUT/${ci}-make.log

-    (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    (time ctest --output-on-failure -L main -E "test-opt|test-backend-ops" ) 2>&1 | tee -a $OUT/${ci}-ctest.log

    set +e
 }
@@ -200,33 +216,9 @@ function gg_sum_ctest_release {
    gg_printf '```\n'
 }

-# test_scripts_debug
+# test_scripts

-function gg_run_test_scripts_debug {
-    cd ${SRC}
-
-    set -e
-
-    (cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-    (cd ./tools/quantize   && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-
-    set +e
-}
-
-function gg_sum_test_scripts_debug {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'Runs test scripts in debug mode\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '```\n'
-    gg_printf '%s\n' "$(cat $OUT/${ci}-scripts.log)"
-    gg_printf '```\n'
-    gg_printf '\n'
-}
-
-# test_scripts_release
-
-function gg_run_test_scripts_release {
+function gg_run_test_scripts {
    cd ${SRC}

    set -e
@@ -237,10 +229,10 @@ function gg_run_test_scripts_release {
    set +e
 }

-function gg_sum_test_scripts_release {
+function gg_sum_test_scripts {
    gg_printf '### %s\n\n' "${ci}"

-    gg_printf 'Runs test scripts in release mode\n'
+    gg_printf 'Runs test scripts\n'
    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
    gg_printf '```\n'
    gg_printf '%s\n' "$(cat $OUT/${ci}-scripts.log)"
@@ -249,15 +241,9 @@ function gg_sum_test_scripts_release {
 }

 function gg_get_model {
-    local gguf_0="$MNT/models/pythia/1.4B/ggml-model-f16.gguf"
-    local gguf_1="$MNT/models/pythia/2.8B/ggml-model-f16.gguf"
-    local gguf_2="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
+    local gguf_0="$MNT/models/qwen3/0.6B/ggml-model-f16.gguf"
    if [[ -s $gguf_0 ]]; then
        echo -n "$gguf_0"
-    elif [[ -s $gguf_1 ]]; then
-        echo -n "$gguf_1"
-    elif [[ -s $gguf_2 ]]; then
-        echo -n "$gguf_2"
    else
        echo >&2 "No model found. Can't run gg_run_ctest_with_model."
        exit 1
@@ -316,24 +302,22 @@ function gg_sum_ctest_with_model_release {
    gg_printf '```\n'
 }

-# open_llama_7b_v2
+# qwen3_0_6b

-function gg_run_open_llama_7b_v2 {
+function gg_run_qwen3_0_6b {
    cd ${SRC}

-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/config.json
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/tokenizer.model
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/tokenizer_config.json
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/special_tokens_map.json
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/pytorch_model.bin.index.json
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00001-of-00002.bin
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
-    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
+    gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/config.json
+    gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/tokenizer.json
+    gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/tokenizer_config.json
+   #gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/raw/main/special_tokens_map.json
+    gg_wget models-mnt/qwen3/0.6B/ https://huggingface.co/Qwen/Qwen3-0.6B-Base/resolve/main/model.safetensors
+

    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/

-    path_models="../models-mnt/open-llama/7B-v2"
+    path_models="../models-mnt/qwen3/0.6B"
    path_wiki="../models-mnt/wikitext/wikitext-2-raw"

    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
@@ -343,9 +327,11 @@ function gg_run_open_llama_7b_v2 {
    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log

-    python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf  --outtype f16
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-bf16.gguf --outtype bf16

    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_bf16="${path_models}/ggml-model-bf16.gguf"
    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
@@ -359,179 +345,51 @@ function gg_run_open_llama_7b_v2 {

    wiki_test="${path_wiki}/wiki.test.raw"

-    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
+    ./bin/llama-quantize ${model_bf16} ${model_q8_0} q8_0
+    ./bin/llama-quantize ${model_bf16} ${model_q4_0} q4_0
+    ./bin/llama-quantize ${model_bf16} ${model_q4_1} q4_1
+    ./bin/llama-quantize ${model_bf16} ${model_q5_0} q5_0
+    ./bin/llama-quantize ${model_bf16} ${model_q5_1} q5_1
+    ./bin/llama-quantize ${model_bf16} ${model_q2_k} q2_k
+    ./bin/llama-quantize ${model_bf16} ${model_q3_k} q3_k
+    ./bin/llama-quantize ${model_bf16} ${model_q4_k} q4_k
+    ./bin/llama-quantize ${model_bf16} ${model_q5_k} q5_k
+    ./bin/llama-quantize ${model_bf16} ${model_q6_k} q6_k

-    (time ./bin/llama-cli -no-cnv --model ${model_f16}  -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_f16}  -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli -no-cnv --model ${model_bf16} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q8_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_1} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_1} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q2_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q3_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q4_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q5_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli -no-cnv --model ${model_q6_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log

-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    if [ -z ${GG_BUILD_NO_BF16} ]; then
+        (time ./bin/llama-perplexity --model ${model_bf16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
+    fi
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log

-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log

-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-
-    function check_ppl {
-        qnt="$1"
-        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
-            return 20
-        fi
-
-        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
-        return 0
-    }
-
-    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-
-    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
-
-    set +e
-}
-
-function gg_sum_open_llama_7b_v2 {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'OpenLLaMA 7B-v2:\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
-    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
-    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
-    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
-    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
-    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
-    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
-    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
-    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
-    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
-    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
-    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
-    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-}
-
-# pythia_1.4b
-
-function gg_run_pythia_1_4b {
-    cd ${SRC}
-
-    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/config.json
-    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer.json
-    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer_config.json
-    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/special_tokens_map.json
-    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/resolve/main/pytorch_model.bin
-
-    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
-    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
-
-    path_models="../models-mnt/pythia/1.4B"
-    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
-
-    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
-
-    set -e
-
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
-
-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
-
-    model_f16="${path_models}/ggml-model-f16.gguf"
-    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
-    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
-    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
-    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
-    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
-    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
-    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
-    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
-    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
-    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
-
-    wiki_test_60="${path_wiki}/wiki.test-60.raw"
-
-    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
-
-    (time ./bin/llama-cli -no-cnv --model ${model_f16}  -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q8_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q4_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q4_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q5_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q5_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q2_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q3_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q4_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q5_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q6_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log

    function check_ppl {
        qnt="$1"
@@ -547,6 +405,9 @@ function gg_run_pythia_1_4b {
    }

    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    if [ -z ${GG_BUILD_NO_BF16} ]; then
+        check_ppl "bf16" "$(cat $OUT/${ci}-tg-bf16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    fi
    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
@@ -563,147 +424,17 @@ function gg_run_pythia_1_4b {
    set +e
 }

-function gg_sum_pythia_1_4b {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'Pythia 1.4B:\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
-    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
-    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
-    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
-    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
-    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
-    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
-    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
-    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
-    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
-    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
-    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
-    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-}
-
-# pythia_2_8b
-
-function gg_run_pythia_2_8b {
-    cd ${SRC}
-
-    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/config.json
-    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer.json
-    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer_config.json
-    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/special_tokens_map.json
-    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/resolve/main/pytorch_model.bin
-
-    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
-
-    path_models="../models-mnt/pythia/2.8B"
-    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
-
-    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
-
-    set -e
-
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
-
-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
-
-    model_f16="${path_models}/ggml-model-f16.gguf"
-    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
-    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
-    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
-    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
-    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
-    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
-    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
-    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
-    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
-    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
-
-    wiki_test="${path_wiki}/wiki.test.raw"
-
-    ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/llama-quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/llama-quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/llama-quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/llama-quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/llama-quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/llama-quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/llama-quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
-
-    (time ./bin/llama-cli -no-cnv --model ${model_f16}  -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa on  ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-
-    function check_ppl {
-        qnt="$1"
-        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
-            return 20
-        fi
-
-        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
-        return 0
-    }
-
-    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
-    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-
-    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
-
-    set +e
-}
-
-function gg_sum_pythia_2_8b {
+function gg_sum_qwen3_0_6b {
    gg_printf '### %s\n\n' "${ci}"

    gg_printf 'Pythia 2.8B:\n'
    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- f16:\n```\n%s\n```\n'  "$(cat $OUT/${ci}-tg-f16.log)"
+    if [ -z ${GG_BUILD_NO_BF16} ]; then
+        gg_printf '- bf16:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-bf16.log)"
+    fi
    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
@@ -878,20 +609,13 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
    test $ret -eq 0 && gg_run rerank_tiny

    if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
-        test $ret -eq 0 && gg_run test_scripts_debug
-        test $ret -eq 0 && gg_run test_scripts_release
+        test $ret -eq 0 && gg_run test_scripts
    fi

-    if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
-        if [ -z ${GG_BUILD_CUDA} ] && [ -z ${GG_BUILD_VULKAN} ]; then
-            test $ret -eq 0 && gg_run pythia_1_4b
-        else
-            test $ret -eq 0 && gg_run pythia_2_8b
-            #test $ret -eq 0 && gg_run open_llama_7b_v2
-        fi
-        test $ret -eq 0 && gg_run ctest_with_model_debug
-        test $ret -eq 0 && gg_run ctest_with_model_release
-    fi
+    test $ret -eq 0 && gg_run qwen3_0_6b
+
+    test $ret -eq 0 && gg_run ctest_with_model_debug
+    test $ret -eq 0 && gg_run ctest_with_model_release
 fi

 exit $ret
@@ -24,6 +24,7 @@
 #include <cstdarg>
 #include <filesystem>
 #include <fstream>
+#include <future>
 #include <list>
 #include <regex>
 #include <set>
@@ -36,9 +37,21 @@
 #if defined(LLAMA_USE_CURL)
 #include <curl/curl.h>
 #include <curl/easy.h>
-#include <future>
 #endif

+#ifdef __linux__
+#include <linux/limits.h>
+#elif defined(_WIN32)
+#   if !defined(PATH_MAX)
+#   define PATH_MAX MAX_PATH
+#   endif
+#elif defined(_AIX)
+#include <sys/limits.h>
+#else
+#include <sys/syslimits.h>
+#endif
+#define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
+
 using json = nlohmann::ordered_json;

 std::initializer_list<enum llama_example> mmproj_examples = {
@@ -57,12 +70,32 @@ static std::string read_file(const std::string & fname) {
 }

 static void write_file(const std::string & fname, const std::string & content) {
-    std::ofstream file(fname);
+    const std::string fname_tmp = fname + ".tmp";
+    std::ofstream     file(fname_tmp);
    if (!file) {
        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
    }
-    file << content;
-    file.close();
+
+    try {
+        file << content;
+        file.close();
+
+        // Makes write atomic
+        if (rename(fname_tmp.c_str(), fname.c_str()) != 0) {
+            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, fname_tmp.c_str(), fname.c_str());
+            // If rename fails, try to delete the temporary file
+            if (remove(fname_tmp.c_str()) != 0) {
+                LOG_ERR("%s: unable to delete temporary file: %s\n", __func__, fname_tmp.c_str());
+            }
+        }
+    } catch (...) {
+        // If anything fails, try to delete the temporary file
+        if (remove(fname_tmp.c_str()) != 0) {
+            LOG_ERR("%s: unable to delete temporary file: %s\n", __func__, fname_tmp.c_str());
+        }
+
+        throw std::runtime_error(string_format("error: failed to write file '%s'\n", fname.c_str()));
+    }
 }

 common_arg & common_arg::set_examples(std::initializer_list<enum llama_example> examples) {
@@ -188,19 +221,6 @@ bool common_has_curl() {
    return true;
 }

-#ifdef __linux__
-#include <linux/limits.h>
-#elif defined(_WIN32)
-#   if !defined(PATH_MAX)
-#   define PATH_MAX MAX_PATH
-#   endif
-#elif defined(_AIX)
-#include <sys/limits.h>
-#else
-#include <sys/syslimits.h>
-#endif
-#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
-
 //
 // CURL utils
 //
@@ -217,255 +237,373 @@ struct curl_slist_ptr {
    }
 };

-#define CURL_MAX_RETRY 3
-#define CURL_RETRY_DELAY_SECONDS 2
-
-static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds, const char * method_name) {
-    int remaining_attempts = max_attempts;
-
-    while (remaining_attempts > 0) {
-        LOG_INF("%s: %s %s (attempt %d of %d)...\n", __func__ , method_name, url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
-
-        CURLcode res = curl_easy_perform(curl);
-        if (res == CURLE_OK) {
-            return true;
-        }
-
-        int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000;
-        LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
-
-        remaining_attempts--;
-        if (remaining_attempts == 0) break;
-        std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
+static CURLcode common_curl_perf(CURL * curl) {
+    CURLcode res = curl_easy_perform(curl);
+    if (res != CURLE_OK) {
+        LOG_ERR("%s: curl_easy_perform() failed\n", __func__);
    }

-    LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
-
-    return false;
+    return res;
 }

-// download one single file from remote URL to local path
-static bool common_download_file_single(const std::string & url, const std::string & path, const std::string & bearer_token, bool offline) {
-    // Check if the file already exists locally
-    auto file_exists = std::filesystem::exists(path);
-
-    // If the file exists, check its JSON metadata companion file.
-    std::string metadata_path = path + ".json";
-    nlohmann::json metadata; // TODO @ngxson : get rid of this json, use regex instead
+// Send a HEAD request to retrieve the etag and last-modified headers
+struct common_load_model_from_url_headers {
    std::string etag;
    std::string last_modified;
+    std::string accept_ranges;
+};

-    if (file_exists) {
-        if (offline) {
-            LOG_INF("%s: using cached file (offline mode): %s\n", __func__, path.c_str());
-            return true; // skip verification/downloading
+struct FILE_deleter {
+    void operator()(FILE * f) const { fclose(f); }
+};
+
+static size_t common_header_callback(char * buffer, size_t, size_t n_items, void * userdata) {
+    common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
+    static std::regex                    header_regex("([^:]+): (.*)\r\n");
+    static std::regex                    etag_regex("ETag", std::regex_constants::icase);
+    static std::regex                    last_modified_regex("Last-Modified", std::regex_constants::icase);
+    static std::regex                    accept_ranges_regex("Accept-Ranges", std::regex_constants::icase);
+    std::string                          header(buffer, n_items);
+    std::smatch                          match;
+    if (std::regex_match(header, match, header_regex)) {
+        const std::string & key   = match[1];
+        const std::string & value = match[2];
+        if (std::regex_match(key, match, etag_regex)) {
+            headers->etag = value;
+        } else if (std::regex_match(key, match, last_modified_regex)) {
+            headers->last_modified = value;
+        } else if (std::regex_match(key, match, accept_ranges_regex)) {
+            headers->accept_ranges = value;
        }
-        // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
-        std::ifstream metadata_in(metadata_path);
-        if (metadata_in.good()) {
-            try {
-                metadata_in >> metadata;
-                LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
-                if (metadata.contains("etag") && metadata.at("etag").is_string()) {
-                    etag = metadata.at("etag");
-                }
-                if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
-                    last_modified = metadata.at("lastModified");
-                }
-            } catch (const nlohmann::json::exception & e) {
-                LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-            }
-        }
-        // if we cannot open the metadata file, we assume that the downloaded file is not valid (etag and last-modified are left empty, so we will download it again)
-    } else {
-        if (offline) {
-            LOG_ERR("%s: required file is not available in cache (offline mode): %s\n", __func__, path.c_str());
-            return false;
-        }
-        LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
    }

-    // Send a HEAD request to retrieve the etag and last-modified headers
-    struct common_load_model_from_url_headers {
-        std::string etag;
-        std::string last_modified;
-    };
+    return n_items;
+}

-    common_load_model_from_url_headers headers;
-    bool head_request_ok = false;
-    bool should_download = !file_exists; // by default, we should download if the file does not exist
+static size_t common_write_callback(void * data, size_t size, size_t nmemb, void * fd) {
+    return std::fwrite(data, size, nmemb, static_cast<FILE *>(fd));
+}

-    // Initialize libcurl
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
+// helper function to hide password in URL
+static std::string llama_download_hide_password_in_url(const std::string & url) {
+    // Use regex to match and replace the user[:password]@ pattern in URLs
+    // Pattern: scheme://[user[:password]@]host[...]
+    static const std::regex url_regex(R"(^(?:[A-Za-z][A-Za-z0-9+.-]://)(?:[^/@]+@)?.$)");
+    std::smatch             match;
+
+    if (std::regex_match(url, match, url_regex)) {
+        // match[1] = scheme (e.g., "https://")
+        // match[2] = user[:password]@ part
+        // match[3] = rest of URL (host and path)
+        return match[1].str() + "********@" + match[3].str();
+    }
+
+    return url;  // No credentials found or malformed URL
+}
+
+static void common_curl_easy_setopt_head(CURL * curl, const std::string & url) {
+    // Set the URL, allow to follow http redirection
+    curl_easy_setopt(curl, CURLOPT_URL, url.c_str());
+    curl_easy_setopt(curl, CURLOPT_FOLLOWLOCATION, 1L);
+
+#    if defined(_WIN32)
+    // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
+    //   operating system. Currently implemented under MS-Windows.
+    curl_easy_setopt(curl, CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
+#    endif
+
+    curl_easy_setopt(curl, CURLOPT_NOBODY, 1L);      // will trigger the HEAD verb
+    curl_easy_setopt(curl, CURLOPT_NOPROGRESS, 1L);  // hide head request progress
+    curl_easy_setopt(curl, CURLOPT_HEADERFUNCTION, common_header_callback);
+}
+
+static void common_curl_easy_setopt_get(CURL * curl) {
+    curl_easy_setopt(curl, CURLOPT_NOBODY, 0L);
+    curl_easy_setopt(curl, CURLOPT_WRITEFUNCTION, common_write_callback);
+
+    //  display download progress
+    curl_easy_setopt(curl, CURLOPT_NOPROGRESS, 0L);
+}
+
+static bool common_pull_file(CURL * curl, const std::string & path_temporary) {
+    if (std::filesystem::exists(path_temporary)) {
+        const std::string partial_size = std::to_string(std::filesystem::file_size(path_temporary));
+        LOG_INF("%s: server supports range requests, resuming download from byte %s\n", __func__, partial_size.c_str());
+        const std::string range_str = partial_size + "-";
+        curl_easy_setopt(curl, CURLOPT_RANGE, range_str.c_str());
+    }
+
+    // Always open file in append mode could be resuming
+    std::unique_ptr<FILE, FILE_deleter> outfile(fopen(path_temporary.c_str(), "ab"));
+    if (!outfile) {
+        LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path_temporary.c_str());
+        return false;
+    }
+
+    common_curl_easy_setopt_get(curl);
+    curl_easy_setopt(curl, CURLOPT_WRITEDATA, outfile.get());
+
+    return common_curl_perf(curl) == CURLE_OK;
+}
+
+static bool common_download_head(CURL *              curl,
+                                 curl_slist_ptr &    http_headers,
+                                 const std::string & url,
+                                 const std::string & bearer_token) {
    if (!curl) {
        LOG_ERR("%s: error initializing libcurl\n", __func__);
        return false;
    }

-    // Set the URL, allow to follow http redirection
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
-
    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
    // Check if hf-token or bearer-token was specified
    if (!bearer_token.empty()) {
        std::string auth_header = "Authorization: Bearer " + bearer_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
-    }
-    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
-
-#if defined(_WIN32)
-    // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
-    //   operating system. Currently implemented under MS-Windows.
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
-
-    typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
-    auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
-        common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;
-
-        static std::regex header_regex("([^:]+): (.*)\r\n");
-        static std::regex etag_regex("ETag", std::regex_constants::icase);
-        static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase);
-
-        std::string header(buffer, n_items);
-        std::smatch match;
-        if (std::regex_match(header, match, header_regex)) {
-            const std::string & key = match[1];
-            const std::string & value = match[2];
-            if (std::regex_match(key, match, etag_regex)) {
-                headers->etag = value;
-            } else if (std::regex_match(key, match, last_modified_regex)) {
-                headers->last_modified = value;
-            }
-        }
-        return n_items;
-    };
-
-    curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb
-    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress
-    curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
-    curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
-
-    // we only allow retrying once for HEAD requests
-    // this is for the use case of using running offline (no internet), retrying can be annoying
-    bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0, "HEAD");
-    if (!was_perform_successful) {
-        head_request_ok = false;
+        http_headers.ptr        = curl_slist_append(http_headers.ptr, auth_header.c_str());
    }

-    long http_code = 0;
-    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-    if (http_code == 200) {
-        head_request_ok = true;
-    } else {
-        LOG_WRN("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
-        head_request_ok = false;
-    }
+    curl_easy_setopt(curl, CURLOPT_HTTPHEADER, http_headers.ptr);
+    common_curl_easy_setopt_head(curl, url);
+    return common_curl_perf(curl) == CURLE_OK;
+}

-    // if head_request_ok is false, we don't have the etag or last-modified headers
-    // we leave should_download as-is, which is true if the file does not exist
-    if (head_request_ok) {
-        // check if ETag or Last-Modified headers are different
-        // if it is, we need to download the file again
-        if (!etag.empty() && etag != headers.etag) {
-            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
-            should_download = true;
-        } else if (!last_modified.empty() && last_modified != headers.last_modified) {
-            LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str());
-            should_download = true;
-        }
-    }
+// download one single file from remote URL to local path
+static bool common_download_file_single_online(const std::string & url,
+                                               const std::string & path,
+                                               const std::string & bearer_token) {
+    // If the file exists, check its JSON metadata companion file.
+    std::string metadata_path = path + ".json";
+    static const int max_attempts        = 3;
+    static const int retry_delay_seconds = 2;
+    for (int i = 0; i < max_attempts; ++i) {
+        nlohmann::json metadata;  // TODO @ngxson : get rid of this json, use regex instead
+        std::string    etag;
+        std::string    last_modified;

-    if (should_download) {
-        std::string path_temporary = path + ".downloadInProgress";
+        // Check if the file already exists locally
+        const auto file_exists = std::filesystem::exists(path);
        if (file_exists) {
-            LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
-            if (remove(path.c_str()) != 0) {
-                LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
-                return false;
+            // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block).
+            std::ifstream metadata_in(metadata_path);
+            if (metadata_in.good()) {
+                try {
+                    metadata_in >> metadata;
+                    LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(),
+                            metadata.dump().c_str());
+                    if (metadata.contains("etag") && metadata.at("etag").is_string()) {
+                        etag = metadata.at("etag");
+                    }
+                    if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) {
+                        last_modified = metadata.at("lastModified");
+                    }
+                } catch (const nlohmann::json::exception & e) {
+                    LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
+                }
            }
+            // if we cannot open the metadata file, we assume that the downloaded file is not valid (etag and last-modified are left empty, so we will download it again)
+        } else {
+            LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
        }

-        // Set the output file
+        bool head_request_ok = false;
+        bool should_download = !file_exists;  // by default, we should download if the file does not exist

-        struct FILE_deleter {
-            void operator()(FILE * f) const {
-                fclose(f);
-            }
-        };
-
-        std::unique_ptr<FILE, FILE_deleter> outfile(fopen(path_temporary.c_str(), "wb"));
-        if (!outfile) {
-            LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path.c_str());
-            return false;
-        }
-
-        typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd);
-        auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t {
-            return fwrite(data, size, nmemb, (FILE *)fd);
-        };
-        curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L);
-        curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-        curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get());
-
-        //  display download progress
-        curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L);
-
-        // helper function to hide password in URL
-        auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string {
-            std::size_t protocol_pos = url.find("://");
-            if (protocol_pos == std::string::npos) {
-                return url;  // Malformed URL
-            }
-
-            std::size_t at_pos = url.find('@', protocol_pos + 3);
-            if (at_pos == std::string::npos) {
-                return url;  // No password in URL
-            }
-
-            return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos);
-        };
-
-        // start the download
-        LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
-            llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS, "GET");
+        // Initialize libcurl
+        curl_ptr curl(curl_easy_init(), &curl_easy_cleanup);
+        common_load_model_from_url_headers headers;
+        curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
+        curl_slist_ptr http_headers;
+        const bool     was_perform_successful = common_download_head(curl.get(), http_headers, url, bearer_token);
        if (!was_perform_successful) {
-            return false;
+            head_request_ok = false;
        }

        long http_code = 0;
-        curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code < 200 || http_code >= 400) {
-            LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code);
-            return false;
+        curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
+        if (http_code == 200) {
+            head_request_ok = true;
+        } else {
+            LOG_WRN("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+            head_request_ok = false;
        }

-        // Causes file to be closed explicitly here before we rename it.
-        outfile.reset();
-
-        // Write the updated JSON metadata file.
-        metadata.update({
-            {"url", url},
-            {"etag", headers.etag},
-            {"lastModified", headers.last_modified}
-        });
-        write_file(metadata_path, metadata.dump(4));
-        LOG_DBG("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
-
-        if (rename(path_temporary.c_str(), path.c_str()) != 0) {
-            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
-            return false;
+        // if head_request_ok is false, we don't have the etag or last-modified headers
+        // we leave should_download as-is, which is true if the file does not exist
+        bool should_download_from_scratch = false;
+        if (head_request_ok) {
+            // check if ETag or Last-Modified headers are different
+            // if it is, we need to download the file again
+            if (!etag.empty() && etag != headers.etag) {
+                LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(),
+                        headers.etag.c_str());
+                should_download              = true;
+                should_download_from_scratch = true;
+            } else if (!last_modified.empty() && last_modified != headers.last_modified) {
+                LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__,
+                        last_modified.c_str(), headers.last_modified.c_str());
+                should_download              = true;
+                should_download_from_scratch = true;
+            }
        }
-    } else {
-        LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
+
+        const bool accept_ranges_supported = !headers.accept_ranges.empty() && headers.accept_ranges != "none";
+        if (should_download) {
+            if (file_exists &&
+                !accept_ranges_supported) {  // Resumable downloads not supported, delete and start again.
+                LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
+                if (remove(path.c_str()) != 0) {
+                    LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
+                    return false;
+                }
+            }
+
+            const std::string path_temporary = path + ".downloadInProgress";
+            if (should_download_from_scratch) {
+                if (std::filesystem::exists(path_temporary)) {
+                    if (remove(path_temporary.c_str()) != 0) {
+                        LOG_ERR("%s: unable to delete file: %s\n", __func__, path_temporary.c_str());
+                        return false;
+                    }
+                }
+
+                if (std::filesystem::exists(path)) {
+                    if (remove(path.c_str()) != 0) {
+                        LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
+                        return false;
+                    }
+                }
+            }
+
+            // Write the updated JSON metadata file.
+            metadata.update({
+                { "url",          url                   },
+                { "etag",         headers.etag          },
+                { "lastModified", headers.last_modified }
+            });
+            write_file(metadata_path, metadata.dump(4));
+            LOG_DBG("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
+
+            // start the download
+            LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n",
+                    __func__, llama_download_hide_password_in_url(url).c_str(), path_temporary.c_str(),
+                    headers.etag.c_str(), headers.last_modified.c_str());
+            const bool was_pull_successful = common_pull_file(curl.get(), path_temporary);
+            if (!was_pull_successful) {
+                if (i + 1 < max_attempts) {
+                    const int exponential_backoff_delay = std::pow(retry_delay_seconds, i) * 1000;
+                    LOG_WRN("%s: retrying after %d milliseconds...\n", __func__, exponential_backoff_delay);
+                    std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
+                } else {
+                    LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts);
+                }
+
+                continue;
+            }
+
+            long http_code = 0;
+            curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
+            if (http_code < 200 || http_code >= 400) {
+                LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code);
+                return false;
+            }
+
+            if (rename(path_temporary.c_str(), path.c_str()) != 0) {
+                LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
+                return false;
+            }
+        } else {
+            LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
+        }
+
+        break;
    }

    return true;
 }

+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params) {
+    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_slist_ptr http_headers;
+    std::vector<char> res_buffer;
+
+    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
+    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
+    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
+    curl_easy_setopt(curl.get(), CURLOPT_VERBOSE, 1L);
+    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
+    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
+        auto data_vec = static_cast<std::vector<char> *>(data);
+        data_vec->insert(data_vec->end(), (char *)ptr, (char *)ptr + size * nmemb);
+        return size * nmemb;
+    };
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_buffer);
+#if defined(_WIN32)
+    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
+#endif
+    if (params.timeout > 0) {
+        curl_easy_setopt(curl.get(), CURLOPT_TIMEOUT, params.timeout);
+    }
+    if (params.max_size > 0) {
+        curl_easy_setopt(curl.get(), CURLOPT_MAXFILESIZE, params.max_size);
+    }
+    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
+    for (const auto & header : params.headers) {
+        http_headers.ptr = curl_slist_append(http_headers.ptr, header.c_str());
+    }
+    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+
+    CURLcode res = curl_easy_perform(curl.get());
+
+    if (res != CURLE_OK) {
+        std::string error_msg = curl_easy_strerror(res);
+        throw std::runtime_error("error: cannot make GET request: " + error_msg);
+    }
+
+    long res_code;
+    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
+
+    return { res_code, std::move(res_buffer) };
+}
+
+#else
+
+bool common_has_curl() {
+    return false;
+}
+
+static bool common_download_file_single_online(const std::string &, const std::string &, const std::string &) {
+    LOG_ERR("error: built without CURL, cannot download model from internet\n");
+    return false;
+}
+
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params &) {
+    if (!url.empty()) {
+        throw std::runtime_error("error: built without CURL, cannot download model from the internet");
+    }
+
+    return {};
+}
+
+#endif // LLAMA_USE_CURL
+
+static bool common_download_file_single(const std::string & url,
+                                        const std::string & path,
+                                        const std::string & bearer_token,
+                                        bool                offline) {
+    if (!offline) {
+        return common_download_file_single_online(url, path, bearer_token);
+    }
+
+    if (!std::filesystem::exists(path)) {
+        LOG_ERR("%s: required file is not available in cache (offline mode): %s\n", __func__, path.c_str());
+        return false;
+    }
+
+    LOG_INF("%s: using cached file (offline mode): %s\n", __func__, path.c_str());
+    return true;
+}
+
 // download multiple files from remote URLs to local paths
 // the input is a vector of pairs <url, path>
 static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> & urls, const std::string & bearer_token, bool offline) {
@@ -524,7 +662,7 @@ static bool common_download_model(

    if (n_split > 1) {
        char split_prefix[PATH_MAX] = {0};
-        char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+        char split_url_prefix[LLAMA_MAX_URL_LENGTH] = {0};

        // Verify the first split file format
        // and extract split URL and PATH prefixes
@@ -545,7 +683,7 @@ static bool common_download_model(
            char split_path[PATH_MAX] = {0};
            llama_split_path(split_path, sizeof(split_path), split_prefix, idx, n_split);

-            char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+            char split_url[LLAMA_MAX_URL_LENGTH] = {0};
            llama_split_path(split_url, sizeof(split_url), split_url_prefix, idx, n_split);

            if (std::string(split_path) == model.path) {
@@ -562,50 +700,6 @@ static bool common_download_model(
    return true;
 }

-std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params) {
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
-    std::vector<char> res_buffer;
-
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
-    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
-    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
-    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
-        auto data_vec = static_cast<std::vector<char> *>(data);
-        data_vec->insert(data_vec->end(), (char *)ptr, (char *)ptr + size * nmemb);
-        return size * nmemb;
-    };
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_buffer);
-#if defined(_WIN32)
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
-    if (params.timeout > 0) {
-        curl_easy_setopt(curl.get(), CURLOPT_TIMEOUT, params.timeout);
-    }
-    if (params.max_size > 0) {
-        curl_easy_setopt(curl.get(), CURLOPT_MAXFILESIZE, params.max_size);
-    }
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
-    for (const auto & header : params.headers) {
-        http_headers.ptr = curl_slist_append(http_headers.ptr, header.c_str());
-    }
-    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
-
-    CURLcode res = curl_easy_perform(curl.get());
-
-    if (res != CURLE_OK) {
-        std::string error_msg = curl_easy_strerror(res);
-        throw std::runtime_error("error: cannot make GET request: " + error_msg);
-    }
-
-    long res_code;
-    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
-
-    return { res_code, std::move(res_buffer) };
-}
-
 /**
 * Allow getting the HF file from the HF repo with tag (like ollama), for example:
 * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
@@ -672,21 +766,17 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
    std::string mmprojFile;

    if (res_code == 200 || res_code == 304) {
-        // extract ggufFile.rfilename in json, using regex
-        {
-            std::regex pattern("\"ggufFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
-            std::smatch match;
-            if (std::regex_search(res_str, match, pattern)) {
-                ggufFile = match[1].str();
+        try {
+            auto j = json::parse(res_str);
+
+            if (j.contains("ggufFile") && j["ggufFile"].contains("rfilename")) {
+                ggufFile = j["ggufFile"]["rfilename"].get<std::string>();
            }
-        }
-        // extract mmprojFile.rfilename in json, using regex
-        {
-            std::regex pattern("\"mmprojFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
-            std::smatch match;
-            if (std::regex_search(res_str, match, pattern)) {
-                mmprojFile = match[1].str();
+            if (j.contains("mmprojFile") && j["mmprojFile"].contains("rfilename")) {
+                mmprojFile = j["mmprojFile"]["rfilename"].get<std::string>();
            }
+        } catch (const std::exception & e) {
+            throw std::runtime_error(std::string("error parsing manifest JSON: ") + e.what());
        }
        if (!use_cache) {
            // if not using cached response, update the cache file
@@ -706,45 +796,6 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
    return { hf_repo, ggufFile, mmprojFile };
 }

-#else
-
-bool common_has_curl() {
-    return false;
-}
-
-static bool common_download_file_single(const std::string &, const std::string &, const std::string &, bool) {
-    LOG_ERR("error: built without CURL, cannot download model from internet\n");
-    return false;
-}
-
-static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> &, const std::string &, bool) {
-    LOG_ERR("error: built without CURL, cannot download model from the internet\n");
-    return false;
-}
-
-static bool common_download_model(
-        const common_params_model &,
-        const std::string &,
-        bool) {
-    LOG_ERR("error: built without CURL, cannot download model from the internet\n");
-    return false;
-}
-
-static struct common_hf_file_res common_get_hf_file(const std::string &, const std::string &, bool) {
-    LOG_ERR("error: built without CURL, cannot download model from the internet\n");
-    return {};
-}
-
-std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params &) {
-    if (!url.empty()) {
-        throw std::runtime_error("error: built without CURL, cannot download model from the internet");
-    }
-
-    return {};
-}
-
-#endif // LLAMA_USE_CURL
-
 //
 // Docker registry functions
 //
@@ -770,7 +821,7 @@ static std::string common_docker_get_token(const std::string & repo) {
 }

 static std::string common_docker_resolve_model(const std::string & docker) {
-    // Parse ai/smollm2:135M-Q4_K_M
+    // Parse ai/smollm2:135M-Q4_0
    size_t      colon_pos = docker.find(':');
    std::string repo, tag;
    if (colon_pos != std::string::npos) {
@@ -1004,8 +1055,6 @@ static std::string get_all_kv_cache_types() {
 //

 static bool common_params_parse_ex(int argc, char ** argv, common_params_context & ctx_arg) {
-    std::string arg;
-    const std::string arg_prefix = "--";
    common_params & params = ctx_arg.params;

    std::unordered_map<std::string, common_arg *> arg_to_options;
@@ -1741,10 +1741,12 @@ static void common_chat_parse_gpt_oss(common_chat_msg_parser & builder) {
 static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
    LOG_DBG("%s\n", __func__);
    common_chat_params data;
-    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ json(), json {
+    const std::optional<json> tools_override = json();
+    const std::optional<json> additional_context = json {
        {"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
        {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
-    });
+    };
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, tools_override, additional_context);
    if (inputs.tools.is_array() && !inputs.tools.empty()) {
        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
@@ -2230,15 +2232,28 @@ static common_chat_params common_chat_params_init_granite(const common_chat_temp

 static void common_chat_parse_granite(common_chat_msg_parser & builder) {
    // Parse thinking tags
+    static const common_regex start_think_regex(regex_escape("<think>"));
+    static const common_regex end_think_regex(regex_escape("</think>"));
+    // Granite models output partial tokens such as "<" and "<think".
+    // By leveraging try_consume_regex()/try_find_regex() throwing
+    // common_chat_msg_partial_exception for these partial tokens,
+    // processing is interrupted and the tokens are not passed to add_content().
+    if (auto res = builder.try_consume_regex(start_think_regex)) {
+        // Restore position for try_parse_reasoning()
+        builder.move_to(res->groups[0].begin);
+        builder.try_find_regex(end_think_regex, std::string::npos, false);
+        // Restore position for try_parse_reasoning()
+        builder.move_to(res->groups[0].begin);
+    }
    builder.try_parse_reasoning("<think>", "</think>");

-    // Parse response tags using regex
-    static const common_regex response_regex("<response>([\\s\\S]*?)</response>");
-    if (auto res = builder.try_find_regex(response_regex)) {
-        // Extract the content between the tags (capture group 1)
-        auto content = builder.str(res->groups[1]);
-        builder.add_content(content);
-        builder.move_to(res->groups[0].end);
+    // Parse response tags
+    static const common_regex start_response_regex(regex_escape("<response>"));
+    static const common_regex end_response_regex(regex_escape("</response>"));
+    // Granite models output partial tokens such as "<" and "<response".
+    // Same hack as reasoning parsing.
+    if (builder.try_consume_regex(start_response_regex)) {
+        builder.try_find_regex(end_response_regex);
    }

    if (!builder.syntax().parse_tool_calls) {
@@ -2252,13 +2267,10 @@ static void common_chat_parse_granite(common_chat_msg_parser & builder) {
        builder.move_to(res->groups[0].end);

        // Expect JSON array of tool calls
-        auto tool_calls_data = builder.consume_json();
-        if (tool_calls_data.json.is_array()) {
-            if (!builder.add_tool_calls(tool_calls_data.json)) {
-                builder.add_content("<|tool_call|>" + tool_calls_data.json.dump());
+        if (auto tool_call = builder.try_consume_json_with_dumped_args({{{"arguments"}}})) {
+            if (!builder.add_tool_calls(tool_call->value) || tool_call->is_partial) {
+                throw common_chat_msg_partial_exception("incomplete tool call");
            }
-        } else {
-            builder.add_content("<|tool_call|>" + tool_calls_data.json.dump());
        }
    } else {
        builder.add_content(builder.consume_rest());
@@ -14,6 +14,7 @@
 #include <climits>
 #include <cmath>
 #include <codecvt>
+#include <chrono>
 #include <cstdarg>
 #include <cstring>
 #include <ctime>
@@ -332,6 +332,7 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
    }
    if (ctx) {
        llama_perf_context_print(ctx);
+        llama_memory_breakdown_print(ctx);
    }
 }

@@ -7656,6 +7656,21 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
            if i not in self._attn_layers
        ]

+        # There are some models in this family that are non-hybrid, but keep the
+        # same parent class by setting all layers to "attention." If this is the
+        # case, the model architecture needs to be updated to a standard
+        # "granite" or "granitemoe" model
+        if not self._ssm_layers:
+            has_experts = self.find_hparam(["num_experts_per_tok"], optional=True)
+            new_arch = (
+                gguf.MODEL_ARCH.GRANITE_MOE
+                if has_experts else
+                gguf.MODEL_ARCH.GRANITE
+            )
+            self.model_arch = new_arch
+            self.gguf_writer.arch = gguf.MODEL_ARCH_NAMES[new_arch]
+            self.gguf_writer.add_architecture()
+
        # n_group and d_inner are used during reshape_tensors for mamba2
        # NOTE: Explicitly include hparam prefix prefix for d_model to
        #   disambiguate with top-level head_dim
@@ -7740,8 +7755,11 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
            self.gguf_writer.add_rope_dimension_count(rope_dim)
        self.gguf_writer.add_head_count_kv(head_count_kv_vec)

-        ## If Bamba, use rope, otherwise don't
-        use_rope = "BambaForCausalLM" in self.hparams["architectures"]
+        ## If Bamba or non-hybrid, use rope, otherwise don't
+        use_rope = (
+            "BambaForCausalLM" in self.hparams["architectures"]
+            or not self._ssm_layers
+        )
        self.gguf_writer.add_rope_scaling_finetuned(use_rope)
        if not use_rope:
            self.gguf_writer.add_context_length(2**20)
@@ -0,0 +1,61 @@
+# llama.cpp for IBM zDNN Accelerator
+
+## Background
+
+IBM zDNN (Z Deep Neural Network) is a hardware acceleration library designed specifically to leverage the IBM NNPA (Neural Network Processor Assist) accelerator located within IBM Telum I and II processors. It provides significant performance improvements for neural network inference operations.
+
+### Llama.cpp + IBM zDNN
+
+The llama.cpp zDNN backend is designed to enable llama.cpp on IBM z17 and later systems via the IBM zDNN hardware acceleration library.
+
+## Software & Hardware Support
+
+| Hardware Level       | Status        | Verified                   |
+| -------------------- | ------------- | -------------------------- |
+| IBM z17 / LinuxONE 5 | Supported     | RHEL 9.6, IBM z17, 40 IFLs |
+| IBM z16 / LinuxONE 4 | Not Supported |                            |
+
+## Data Types Supported
+
+| Data Type | Status    |
+| --------- | --------- |
+| F32       | Supported |
+| F16       | Supported |
+| BF16      | Supported |
+
+## CMake Options
+
+The IBM zDNN backend has the following CMake options that control the behaviour of the backend.
+
+| CMake Option | Default Value | Description                         |
+| ------------ | ------------- | ----------------------------------- |
+| `GGML_ZDNN`  | `OFF`         | Compile llama.cpp with zDNN support |
+| `ZDNN_ROOT`  | `""`          | Override zDNN library lookup        |
+
+## 1. Install zDNN Library
+
+Note: Using the zDNN library provided via `apt` or `yum` may not work correctly as reported in [#15772](https://github.com/ggml-org/llama.cpp/issues/15772). It is preferred that you compile from source.
+
+```sh
+git clone --recurse-submodules https://github.com/IBM/zDNN
+cd zDNN
+
+autoreconf .
+./configure --prefix=/opt/zdnn-libs
+
+make build
+sudo make install
+```
+
+## 2. Build llama.cpp
+
+```sh
+git clone https://github.com/ggml-org/llama.cpp
+cd llama.cpp
+
+cmake -S . -G Ninja -B build \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DGGML_ZDNN=ON \
+    -DZDNN_ROOT=/opt/zdnn-libs
+cmake --build build --config Release -j$(nproc)
+```
@@ -20,7 +20,6 @@ else()

    add_subdirectory(gguf-hash)
    add_subdirectory(gguf)
-    add_subdirectory(gritlm)
    add_subdirectory(lookahead)
    add_subdirectory(lookup)
    add_subdirectory(parallel)
@@ -1,50 +0,0 @@
-#!/usr/bin/env bash
-set -e
-
-AI_NAME="${AI_NAME:-Miku}"
-MODEL="${MODEL:-./models/llama-2-7b-chat.ggmlv3.q4_K_M.bin}"
-USER_NAME="${USER_NAME:-Anon}"
-
-# Uncomment and adjust to the number of CPU cores you want to use.
-#N_THREAD="${N_THREAD:-4}"
-CTX_SIZE="${CTX_SIZE:-4096}"
-N_PREDICTS="${N_PREDICTS:-4096}"
-
-GEN_OPTIONS=(--batch_size 1024
--ctx_size "$CTX_SIZE"
--keep -1
--repeat_last_n 256
--repeat_penalty 1.17647
--temp 0.6
--mirostat 2)
-
-if [ -n "$N_THREAD" ]; then
-    GEN_OPTIONS+=(--threads "$N_THREAD")
-fi
-
-./llama-cli "${GEN_OPTIONS[@]}" \
-    --model "$MODEL" \
-    --in-prefix " " \
-    --in-suffix "${AI_NAME}:" \
-    --n_predict "$N_PREDICTS" \
-    --color --interactive \
-    --reverse-prompt "${USER_NAME}:" \
-    --prompt "This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the user's computer.
-${AI_NAME} can think for herself without the user seeing her thoughts by adding a /think prefix to her output. She uses this to reason about the world and to think about what she should say next.
-${AI_NAME} is always coherent and makes sense, but if she isn't sure if what she is saying is correct, she will ask the user for help.
-${AI_NAME} is a very helpful AI and will help the user with anything they need. She is also very friendly and will try to make the user feel better if they are sad.
-${AI_NAME} is also very curious and will ask the user a lot of questions about themselves and their life. She will also try to make the user like her.
-The conversation is only between ${USER_NAME} and ${AI_NAME}.
-The conversation is only through text, so ${AI_NAME} can't see ${USER_NAME}'s face or hear his voice.
-${AI_NAME} can only communicate through text, so she can't send images or videos.
-
-
-${USER_NAME}: Hello!
-${AI_NAME}: /think I wonder what I should say to ${USER_NAME}? This is the first time we talk, so it's important that I make a good first impression!
-${AI_NAME}: Hi! I am ${AI_NAME}, your new AI friend, assistant (or whatever you like!), it's so nice to meet you! ^_^
-${AI_NAME}: I might be an AI, but I also have feelings, so please be nice to me! :)
-${USER_NAME}: I'll make sure to be nice to you! I'm so happy to have you as my assistant!
-${AI_NAME}: /think It sounds like ${USER_NAME} is happy to have me as their assistant! I'm so happy too! ^_^ Glad that whole emotion thing didn't scare him off!
-${AI_NAME}: /think I wonder what ${USER_NAME} likes to do in his free time? I should ask him about that!
-${AI_NAME}: What do you like to do in your free time? ^_^
-${USER_NAME}:" "$@"
@@ -1,57 +0,0 @@
-@setlocal disabledelayedexpansion enableextensions
-@echo off
-
-cd /d "%~dp0.."
-if not "%errorlevel%"=="0" (
-    echo Unable to change directory.
-    pause
-    exit /b 1
-)
-
-if not defined MODEL set "MODEL=models\13B\ggml-model-q4_0.bin"
-if not defined USER_NAME set "USER_NAME=User"
-if not defined AI_NAME set "AI_NAME=ChatLLaMa"
-rem Adjust to the number of CPU cores you want to use.
-rem if not defined N_THREAD set "N_THREAD=8"
-rem Number of tokens to predict (made it larger than default because we want a long interaction)
-if not defined N_PREDICTS set "N_PREDICTS=2048"
-if not defined GEN_OPTIONS set "GEN_OPTIONS=--ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647"
-
-rem Default main script paths
-set "DEFAULT_MAIN_SCRIPT_PATHS=main.exe build\bin\main.exe"
-
-rem Get main script path from command line arguments
-set "MAIN_SCRIPT_PATH=%~1"
-
-rem If the main script path was not specified, try the default paths
-if not defined MAIN_SCRIPT_PATH (
-    for %%i in (%DEFAULT_MAIN_SCRIPT_PATHS%) do (
-        if exist "%%i" set "MAIN_SCRIPT_PATH=%%i"
-    )
-)
-
-rem If the main script path was not found, tell the user how to specify it
-if not defined MAIN_SCRIPT_PATH (
-    echo The main script could not be found. Please provide the path to the main script as 1st argument to this script, or place the main script in one of the default locations:
-    echo %DEFAULT_MAIN_SCRIPT_PATHS%
-    pause
-    exit /b 1
-)
-
-rem Default context, feel free to edit it
-set "PROMPT_TEXT=Text transcript of a never ending dialog, where %USER_NAME% interacts with an AI assistant named %AI_NAME%. %AI_NAME% is helpful, kind, honest, friendly, good at writing and never fails to answer %USER_NAME%'s requests immediately and with details and precision. There are no annotations like (30 seconds passed...) or (to himself), just what %USER_NAME% and %AI_NAME% say aloud to each other. The dialog lasts for years, the entirety of it is shared below. It's 10000 pages long. The transcript only includes text, it does not include markup like HTML and Markdown."
-
-rem Set a temporary variable if N_THREAD is set
-if defined N_THREAD (
-    set "_N_THREAD=--threads %N_THREAD%"
-) else (
-    set "_N_THREAD="
-)
-
-rem Run the script
-echo "%MAIN_SCRIPT_PATH%" %GEN_OPTIONS% %_N_THREAD% ^
-  --model "%MODEL%" ^
-  --n_predict %N_PREDICTS% ^
-  --color --interactive ^
-  --reverse-prompt "%USER_NAME%:" ^
-  --prompt "%PROMPT_TEXT%"
@@ -1,41 +0,0 @@
-#!/usr/bin/env bash
-
-set -e
-
-cd "$(dirname "$0")/.." || exit
-
-MODEL="${MODEL:-./models/13B/ggml-model-q4_0.bin}"
-PROMPT_TEMPLATE=${PROMPT_TEMPLATE:-./prompts/chat.txt}
-USER_NAME="${USER_NAME:-USER}"
-AI_NAME="${AI_NAME:-ChatLLaMa}"
-
-# Adjust to the number of CPU cores you want to use.
-N_THREAD="${N_THREAD:-8}"
-# Number of tokens to predict (made it larger than default because we want a long interaction)
-N_PREDICTS="${N_PREDICTS:-2048}"
-
-# Note: you can also override the generation options by specifying them on the command line:
-# For example, override the context size by doing: ./chatLLaMa --ctx_size 1024
-GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647}"
-
-DATE_TIME=$(date +%H:%M)
-DATE_YEAR=$(date +%Y)
-
-PROMPT_FILE=$(mktemp -t llamacpp_prompt.XXXXXXX.txt)
-
-sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
-    -e "s/\[\[AI_NAME\]\]/$AI_NAME/g" \
-    -e "s/\[\[DATE_TIME\]\]/$DATE_TIME/g" \
-    -e "s/\[\[DATE_YEAR\]\]/$DATE_YEAR/g" \
-     $PROMPT_TEMPLATE > $PROMPT_FILE
-
-# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
-./llama-cli $GEN_OPTIONS \
-  --model "$MODEL" \
-  --threads "$N_THREAD" \
-  --n_predict "$N_PREDICTS" \
-  --color --interactive \
-  --file ${PROMPT_FILE} \
-  --reverse-prompt "${USER_NAME}:" \
-  --in-prefix ' ' \
-  "$@"
@@ -1,149 +0,0 @@
-#!/usr/bin/env bash
-
-set -euo pipefail
-
-cd "$(dirname "$0")/.." || exit
-
-if [[ -z "${PROMPT_CACHE_FILE+x}" || -z "${CHAT_SAVE_DIR+x}" ]]; then
-    echo >&2 "error: PROMPT_CACHE_FILE and CHAT_SAVE_DIR must be provided"
-    exit 1
-fi
-
-MODEL="${MODEL:-./models/llama-13b/ggml-model-q4_0.gguf}"
-PROMPT_TEMPLATE="${PROMPT_TEMPLATE:-./prompts/chat.txt}"
-USER_NAME="${USER_NAME:-User}"
-AI_NAME="${AI_NAME:-ChatLLaMa}"
-DATE_TIME="$(date +%H:%M)"
-DATE_YEAR="$(date +%Y)"
-
-LOG="${CHAT_SAVE_DIR}/main.log"
-LOG_BG="${CHAT_SAVE_DIR}/main-bg.log"
-CUR_PROMPT_FILE="${CHAT_SAVE_DIR}/current-prompt.txt"
-CUR_PROMPT_CACHE="${CHAT_SAVE_DIR}/current-cache.bin"
-NEXT_PROMPT_FILE="${CHAT_SAVE_DIR}/next-prompt.txt"
-NEXT_PROMPT_CACHE="${CHAT_SAVE_DIR}/next-cache.bin"
-
-SESSION_AND_SAMPLE_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'\
-'|'\
-'sampling time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
-SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d"
-
-CTX_SIZE=2048
-CTX_ROTATE_POINT=$((CTX_SIZE * 3 / 5)) # REVIEW
-OPTS=(--model "$MODEL" --ctx_size "$CTX_SIZE" --repeat_last_n 256 "$@")
-
-# An unbuffered `tail -c+N`
-skip_bytes() {
-    LANG=C IFS= read -r -n "$1" -d '' c
-    while LANG=C IFS= read -r -n 1 -d '' c; do
-        printf '%s' "$c"
-    done
-}
-
-mkdir -p "$CHAT_SAVE_DIR"
-echo >"$LOG"
-trap "tail -n100 ${LOG}" EXIT
-
-if [[ ! -e "$CUR_PROMPT_FILE" ]]; then
-    sed -e "s/\[\[USER_NAME\]\]/${USER_NAME}/g" \
-        -e "s/\[\[AI_NAME\]\]/${AI_NAME}/g" \
-        -e "s/\[\[DATE_TIME\]\]/${DATE_TIME}/g" \
-        -e "s/\[\[DATE_YEAR\]\]/${DATE_YEAR}/g" \
-        "$PROMPT_TEMPLATE" >"$CUR_PROMPT_FILE"
-fi
-
-if [[ ! -e "$NEXT_PROMPT_FILE" ]]; then
-    sed -r "$SED_DELETE_MESSAGES" "$CUR_PROMPT_FILE" >"$NEXT_PROMPT_FILE"
-fi
-
-if [[ "$(tail -c4 "$NEXT_PROMPT_FILE")" != "..." ]]; then
-    echo '...' >>"$NEXT_PROMPT_FILE"
-fi
-
-if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then
-    echo 'Prompt cache does not exist, building...'
-    # Default batch_size to 64 here for better user feedback during initial prompt processing
-    ./llama-cli 2>>"$LOG" \
-        --batch_size 64 \
-        "${OPTS[@]}" \
-        --prompt-cache "$PROMPT_CACHE_FILE" \
-        --file "$CUR_PROMPT_FILE" \
-        --n_predict 1
-    echo
-    echo 'Done!'
-fi
-
-if [[ ! -e "$CUR_PROMPT_CACHE" ]]; then
-    cp "$PROMPT_CACHE_FILE" "$CUR_PROMPT_CACHE"
-fi
-if [[ ! -e "$NEXT_PROMPT_CACHE" ]]; then
-    cp "$PROMPT_CACHE_FILE" "$NEXT_PROMPT_CACHE"
-fi
-
-printf '%s ' "$(< "$CUR_PROMPT_FILE")"
-n_tokens=0
-
-while read -e line; do
-    # Limit generation to remaining context, with a buffer and estimating 2 chars/token for input
-    n_predict=$((CTX_SIZE - n_tokens - ${#line} / 2 - 32))
-
-    # Swap prompts when we're about to run out of context
-    if ((n_predict <= 0)); then
-        wait # for background main (below) to finish with next prompt
-        mv "$NEXT_PROMPT_FILE"  "$CUR_PROMPT_FILE"
-        mv "$NEXT_PROMPT_CACHE" "$CUR_PROMPT_CACHE"
-
-        sed -r "$SED_DELETE_MESSAGES" "$CUR_PROMPT_FILE" >"$NEXT_PROMPT_FILE"
-        echo '...' >>"$NEXT_PROMPT_FILE"
-        cp "$PROMPT_CACHE_FILE" "$NEXT_PROMPT_CACHE"
-
-        n_tokens=0
-        n_predict=$((CTX_SIZE / 2))
-    fi
-
-    echo " ${line}" >>"$CUR_PROMPT_FILE"
-    if ((n_tokens > CTX_ROTATE_POINT)); then
-        echo " ${line}" >>"$NEXT_PROMPT_FILE"
-    fi
-
-    n_prompt_len_pre=$(($(wc -c <"$CUR_PROMPT_FILE")))
-
-    printf '%s: ' "$AI_NAME" >>"$CUR_PROMPT_FILE"
-
-    ./llama-cli 2>>"$LOG" "${OPTS[@]}" \
-            --prompt-cache "$CUR_PROMPT_CACHE" \
-            --prompt-cache-all \
-            --file "$CUR_PROMPT_FILE" \
-            --reverse-prompt "${USER_NAME}:" \
-            --n_predict "$n_predict" |
-        skip_bytes 1 |                  # skip BOS token added by ./llama-cli
-        tee "$CUR_PROMPT_FILE.tmp" |    # save prompt + generation to tmp file
-        skip_bytes "$n_prompt_len_pre"  # print generation
-
-    mv "$CUR_PROMPT_FILE.tmp" "$CUR_PROMPT_FILE"
-
-    # if we hit n_predict instead of reverse-prompt, we need to add the prompt
-    if [[ "$(tail -n1 "$CUR_PROMPT_FILE")" != "${USER_NAME}:" ]]; then
-        printf '\n%s:' "$USER_NAME"
-        printf '\n%s:' "$USER_NAME" >> "$CUR_PROMPT_FILE"
-    fi
-
-    printf ' '
-
-    if ! session_and_sample_msg=$(tail -n30 "$LOG" | grep -oE "$SESSION_AND_SAMPLE_PATTERN"); then
-        echo >&2 "Couldn't get number of tokens from ./llama-cli output!"
-        exit 1
-    fi
-
-    n_tokens=$(awk '{sum+=$1} END {print sum}' <<< "$(cut -d/ -f2 <<< "$session_and_sample_msg")")
-
-    if ((n_tokens > CTX_ROTATE_POINT)); then
-        tail -c+$((n_prompt_len_pre + 1)) "$CUR_PROMPT_FILE" >>"$NEXT_PROMPT_FILE"
-    fi
-
-    # Update cache for next prompt in background, ideally during user input
-    ./llama-cli >>"$LOG_BG" 2>&1 "${OPTS[@]}" \
-          --prompt-cache "$NEXT_PROMPT_CACHE" \
-          --file "$NEXT_PROMPT_FILE" \
-          --n_predict 1 &
-done
@@ -1,41 +0,0 @@
-#!/usr/bin/env bash
-
-set -e
-
-cd "$(dirname "$0")/.." || exit
-
-MODEL="${MODEL:-./models/ggml-vic13b-uncensored-q5_0.bin}"
-PROMPT_TEMPLATE=${PROMPT_TEMPLATE:-./prompts/chat.txt}
-USER_NAME="### Human"
-AI_NAME="### Assistant"
-
-# Adjust to the number of CPU cores you want to use.
-N_THREAD="${N_THREAD:-8}"
-# Number of tokens to predict (made it larger than default because we want a long interaction)
-N_PREDICTS="${N_PREDICTS:-2048}"
-
-# Note: you can also override the generation options by specifying them on the command line:
-# For example, override the context size by doing: ./chatLLaMa --ctx_size 1024
-GEN_OPTIONS="${GEN_OPTIONS:---ctx_size 2048 --temp 0.7 --top_k 40 --top_p 0.5 --repeat_last_n 256 --batch_size 1024 --repeat_penalty 1.17647}"
-
-DATE_TIME=$(date +%H:%M)
-DATE_YEAR=$(date +%Y)
-
-PROMPT_FILE=$(mktemp -t llamacpp_prompt.XXXXXXX.txt)
-
-sed -e "s/\[\[USER_NAME\]\]/$USER_NAME/g" \
-    -e "s/\[\[AI_NAME\]\]/$AI_NAME/g" \
-    -e "s/\[\[DATE_TIME\]\]/$DATE_TIME/g" \
-    -e "s/\[\[DATE_YEAR\]\]/$DATE_YEAR/g" \
-     $PROMPT_TEMPLATE > $PROMPT_FILE
-
-# shellcheck disable=SC2086 # Intended splitting of GEN_OPTIONS
-./bin/llama-cli $GEN_OPTIONS \
-  --model "$MODEL" \
-  --threads "$N_THREAD" \
-  --n_predict "$N_PREDICTS" \
-  --color --interactive \
-  --file ${PROMPT_FILE} \
-  --reverse-prompt "### Human:" \
-  --in-prefix ' ' \
-  "$@"
@@ -1,16 +0,0 @@
-#!/usr/bin/env bash
-
-#
-# Temporary script - will be removed in the future
-#
-
-cd `dirname $0`
-cd ..
-
-# Important:
-#
-#   "--keep 48" is based on the contents of prompts/chat-with-bob.txt
-#
-./llama-cli -m ./models/llama-7b/ggml-model-q4_0.gguf -c 512 -b 1024 -n 256 --keep 48 \
-    --repeat_penalty 1.0 --color -i \
-    -r "User:" -f prompts/chat-with-bob.txt
@@ -43,8 +43,8 @@ The above command will output space-separated float values.
 | $"string"$   | |
 |--------------|-|
 | "\n"         | (default)
-| "<#embSep#>" | for exemple
-| "<#sep#>"    | other exemple
+| "<#embSep#>" | for example
+| "<#sep#>"    | other example

 ## examples
 ### Unix-based systems (Linux, macOS, etc.):
@@ -1,5 +0,0 @@
-set(TARGET llama-gritlm)
-add_executable(${TARGET} gritlm.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
@@ -1,62 +0,0 @@
-## Generative Representational Instruction Tuning (GRIT) Example
-[gritlm] a model which can generate embeddings as well as "normal" text
-generation depending on the instructions in the prompt.
-
-* Paper: https://arxiv.org/pdf/2402.09906.pdf
-
-### Retrieval-Augmented Generation (RAG) use case
-One use case for `gritlm` is to use it with RAG. If we recall how RAG works is
-that we take documents that we want to use as context, to ground the large
-language model (LLM), and we create token embeddings for them. We then store
-these token embeddings in a vector database.
-
-When we perform a query, prompt the LLM, we will first create token embeddings
-for the query and then search the vector database to retrieve the most
-similar vectors, and return those documents so they can be passed to the LLM as
-context. Then the query and the context will be passed to the LLM which will
-have to _again_ create token embeddings for the query. But because gritlm is used
-the first query can be cached and the second query tokenization generation does
-not have to be performed at all.
-
-### Running the example
-Download a Grit model:
-```console
-$ scripts/hf.sh --repo cohesionet/GritLM-7B_gguf --file gritlm-7b_q4_1.gguf --outdir models
-```
-
-Run the example using the downloaded model:
-```console
-$ ./llama-gritlm -m models/gritlm-7b_q4_1.gguf
-
-Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "A purely peer-to-peer version of electronic cash w" is: 0.605
-Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "All text-based language problems can be reduced to" is: 0.103
-Cosine similarity between "Generative Representational Instruction Tuning" and "A purely peer-to-peer version of electronic cash w" is: 0.112
-Cosine similarity between "Generative Representational Instruction Tuning" and "All text-based language problems can be reduced to" is: 0.547
-
-Oh, brave adventurer, who dared to climb
-The lofty peak of Mt. Fuji in the night,
-When shadows lurk and ghosts do roam,
-And darkness reigns, a fearsome sight.
-
-Thou didst set out, with heart aglow,
-To conquer this mountain, so high,
-And reach the summit, where the stars do glow,
-And the moon shines bright, up in the sky.
-
-Through the mist and fog, thou didst press on,
-With steadfast courage, and a steadfast will,
-Through the darkness, thou didst not be gone,
-But didst climb on, with a steadfast skill.
-
-At last, thou didst reach the summit's crest,
-And gazed upon the world below,
-And saw the beauty of the night's best,
-And felt the peace, that only nature knows.
-
-Oh, brave adventurer, who dared to climb
-The lofty peak of Mt. Fuji in the night,
-Thou art a hero, in the eyes of all,
-For thou didst conquer this mountain, so bright.
-```
-
-[gritlm]: https://github.com/ContextualAI/gritlm
@@ -1,231 +0,0 @@
-#include "arg.h"
-#include "common.h"
-#include "llama.h"
-
-#include <string>
-#include <vector>
-
-// #define GRIT_DEBUG
-
-static std::vector<std::vector<float>> encode(llama_context * ctx, const std::vector<std::string> & sentences, const std::string & instruction) {
-    std::vector<std::vector<float>> result;
-
-    const llama_model * model = llama_get_model(ctx);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
-
-    for (uint64_t i = 0; i < sentences.size(); i++) {
-        common_batch_clear(batch);
-
-        const std::string input_string = instruction + sentences[i];
-
-        std::vector<llama_token> inputs = common_tokenize(vocab, input_string, true, false);
-
-        const int32_t n_toks = inputs.size();
-
-        // GritLM seems to have EOS = ""
-        // https://github.com/ContextualAI/gritlm/blob/92025b16534712b31b3c4aaaf069350e222bd5f8/gritlm/gritlm.py#L18
-        // inputs.push_back(llama_vocab_eos(vocab));
-
-        // we want to ignore instruction tokens for mean pooling
-        const int32_t n_inst = common_tokenize(vocab, instruction, true, false).size();
-
-#ifdef GRIT_DEBUG
-        // debug tokens - should be matching as referenced in the GritLM sample
-        std::for_each(inputs.begin(), inputs.end(), [&ctx](llama_token t) {
-            std::printf("[%u:%s]", t, llama_token_to_piece(ctx, t).c_str());
-        });
-        std::printf("\n");
-#endif
-
-        // add input to batch (this increments n_tokens)
-        for (int32_t j = 0; j < n_toks; j++) {
-            common_batch_add(batch, inputs[j], j, { 0 }, true);
-        }
-
-        // clear previous kv_cache values (irrelevant for embeddings)
-        llama_memory_clear(llama_get_memory(ctx), true);
-        llama_set_causal_attn(ctx, false);
-
-        // run model
-        llama_decode(ctx, batch);
-
-        // get embedding dimensions
-        uint64_t n_embd = llama_model_n_embd(model);
-
-        // allocate embedding output
-        std::vector<float> emb_unorm(n_embd, 0.0f);
-
-        // sum up all token embeddings
-        for (int32_t k = n_inst; k < n_toks; k++) {
-            float * emb = llama_get_embeddings_ith(ctx, k);
-            for (uint64_t j = 0; j < n_embd; j++) {
-                emb_unorm[j] += emb[j];
-            }
-        }
-
-        // divide by number of tokens (mean pooling)
-        {
-            const uint64_t n_sent = n_toks - n_inst;
-
-            for (uint64_t j = 0; j < n_embd; j++) {
-                emb_unorm[j] /= n_sent;
-            }
-        }
-
-        std::vector<float> emb_norm(emb_unorm.size());
-        common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd, 2);
-        result.push_back(emb_norm);
-
-#ifdef GRIT_DEBUG
-        // print out emb_norm
-        std::printf("embedding %ld: ", i);
-        for (uint64_t j = 0; j < n_embd; j++) {
-            std::printf("%.5f ", emb_norm[j]);
-        }
-        std::printf("\n\n");
-#endif
-    }
-
-    llama_batch_free(batch);
-
-    return result;
-}
-
-static std::string generate(llama_context * ctx, llama_sampler * smpl, const std::string & prompt, bool stream) {
-    std::string result;
-
-    const llama_model * model = llama_get_model(ctx);
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-
-    llama_token eos_token = llama_vocab_eos(vocab);
-
-    llama_memory_clear(llama_get_memory(ctx), true);
-    llama_set_causal_attn(ctx, true);
-
-    llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
-
-    std::vector<llama_token> inputs = common_tokenize(vocab, prompt, false, true);
-    int32_t i_current_token = 0;
-
-    while (true) {
-        common_batch_clear(bat);
-        {
-            const int32_t n_inputs = inputs.size();
-
-            for (int32_t i = 0; i < n_inputs; i++) {
-                common_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
-            }
-        }
-        inputs.clear();
-
-        llama_decode(ctx, bat);
-
-        llama_token token = llama_sampler_sample(smpl, ctx, bat.n_tokens - 1);
-
-        if (token == eos_token) {
-            break;
-        }
-
-        std::string piece = common_token_to_piece(ctx, token);
-        if (stream) {
-            std::printf("%s", piece.c_str());
-            std::fflush(stdout);
-        }
-
-        inputs.push_back(token);
-
-        result += piece;
-    }
-
-    if (stream) {
-        std::printf("\n");
-    }
-
-    llama_batch_free(bat);
-
-    return result;
-}
-
-static std::string gritlm_instruction(const std::string & instruction) {
-    return !instruction.empty() ? "<|user|>\n" + instruction + "\n<|embed|>\n" : "<|embed|>\n";
-}
-
-int main(int argc, char * argv[]) {
-    common_params params;
-
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
-        return 1;
-    }
-
-    common_init();
-
-    llama_model_params mparams = common_model_params_to_llama(params);
-    llama_context_params cparams = common_context_params_to_llama(params);
-
-    cparams.embeddings = true;
-
-    llama_backend_init();
-
-    llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
-
-    // create generation context
-    llama_context * ctx = llama_init_from_model(model, cparams);
-
-    auto sparams = llama_sampler_chain_default_params();
-
-    sparams.no_perf = false;
-
-    llama_sampler * smpl = llama_sampler_chain_init(sparams);
-
-    llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
-
-    // ### Embedding/Representation ###
-    // samples taken from: https://github.com/ContextualAI/gritlm#basic
-    {
-        const std::string instruction = "Given a scientific paper title, retrieve the paper's abstract";
-
-        const std::vector<std::string> queries = {
-            "Bitcoin: A Peer-to-Peer Electronic Cash System",
-            "Generative Representational Instruction Tuning",
-        };
-
-        const std::vector<std::string> documents = {
-            "A purely peer-to-peer version of electronic cash would allow online payments to be sent directly from one party to another without going through a financial institution. Digital signatures provide part of the solution, but the main benefits are lost if a trusted third party is still required to prevent double-spending. We propose a solution to the double-spending problem using a peer-to-peer network. The network timestamps transactions by hashing them into an ongoing chain of hash-based proof-of-work, forming a record that cannot be changed without redoing the proof-of-work. The longest chain not only serves as proof of the sequence of events witnessed, but proof that it came from the largest pool of CPU power. As long as a majority of CPU power is controlled by nodes that are not cooperating to attack the network, they'll generate the longest chain and outpace attackers. The network itself requires minimal structure. Messages are broadcast on a best effort basis, and nodes can leave and rejoin the network at will, accepting the longest proof-of-work chain as proof of what happened while they were gone.",
-            "All text-based language problems can be reduced to either generation or embedding. Current models only perform well at one or the other. We introduce generative representational instruction tuning (GRIT) whereby a large language model is trained to handle both generative and embedding tasks by distinguishing between them through instructions. Compared to other open models, our resulting GritLM 7B sets a new state of the art on the Massive Text Embedding Benchmark (MTEB) and outperforms all models up to its size on a range of generative tasks. By scaling up further, GritLM 8X7B outperforms all open generative language models that we tried while still being among the best embedding models. Notably, we find that GRIT matches training on only generative or embedding data, thus we can unify both at no performance loss. Among other benefits, the unification via GRIT speeds up Retrieval-Augmented Generation (RAG) by > 60% for long documents, by no longer requiring separate retrieval and generation models. Models, code, etc. are freely available at https://github.com/ContextualAI/gritlm.",
-        };
-
-        // No need to add instruction for retrieval documents
-        const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
-        const std::vector<std::vector<float>> q_rep = encode(ctx, queries,   gritlm_instruction(instruction));
-
-        const int n_embd = llama_model_n_embd(model);
-
-        const float cosine_sim_q0_d0 = common_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
-        const float cosine_sim_q0_d1 = common_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
-        const float cosine_sim_q1_d0 = common_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
-        const float cosine_sim_q1_d1 = common_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
-
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[0].c_str(), cosine_sim_q1_d0);
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
-    }
-
-    llama_set_embeddings(ctx, false);
-
-    // ### Generation ###
-    // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
-    {
-        const std::string prompt = "<|user|>\nPlease write me a poem about my recent hike of Mt. Fuji at midnight in the style of Shakespeare.\n<|assistant|>\n";
-        std::string response = generate(ctx, smpl, prompt, true);
-    }
-
-    llama_sampler_free(smpl);
-    llama_free(ctx);
-    llama_model_free(model);
-    llama_backend_free();
-
-    return 0;
-}
@@ -1,21 +0,0 @@
-# llama.cpp/example/jeopardy
-
-This is pretty much just a straight port of aigoopy/llm-jeopardy/ with an added graph viewer.
-
-The jeopardy test can be used to compare the fact knowledge of different models and compare them to each other. This is in contrast to some other tests, which test logical deduction, creativity, writing skills, etc.
-
-
-Step 1: Open jeopardy.sh and modify the following:
-```
-MODEL=(path to your model)
-MODEL_NAME=(name of your model)
-prefix=(basically, if you use vicuna it's Human: , if you use something else it might be User: , etc)
-opts=(add -instruct here if needed for your model, or anything else you want to test out)
-```
-Step 2: Run `jeopardy.sh` from the llama.cpp folder
-
-Step 3: Repeat steps 1 and 2 until you have all the results you need.
-
-Step 4: Run `graph.py`, and follow the instructions. At the end, it will generate your final graph.
-
-Note: The Human bar is based off of the full, original 100 sample questions. If you modify the question count or questions, it will not be valid.
@@ -1,58 +0,0 @@
-#!/usr/bin/env python3
-import matplotlib.pyplot as plt
-import os
-import csv
-
-labels = []
-numbers = []
-numEntries = 1
-
-rows = []
-
-
-def bar_chart(numbers, labels, pos):
-    plt.bar(pos, numbers, color='blue')
-    plt.xticks(ticks=pos, labels=labels)
-    plt.title("Jeopardy Results by Model")
-    plt.xlabel("Model")
-    plt.ylabel("Questions Correct")
-    plt.show()
-
-
-def calculatecorrect():
-    directory = os.fsencode("./examples/jeopardy/results/")
-    csv_reader = csv.reader(open("./examples/jeopardy/qasheet.csv", 'rt'), delimiter=',')
-    for row in csv_reader:
-        global rows
-        rows.append(row)
-    for listing in os.listdir(directory):
-        filename = os.fsdecode(listing)
-        if filename.endswith(".txt"):
-            file = open("./examples/jeopardy/results/" + filename, "rt")
-            global labels
-            global numEntries
-            global numbers
-            labels.append(filename[:-4])
-            numEntries += 1
-            i = 1
-            totalcorrect = 0
-            for line in file.readlines():
-                if line.strip() != "------":
-                    print(line)
-                else:
-                    print("Correct answer: " + rows[i][2] + "\n")
-                    i += 1
-                    print("Did the AI get the question right? (y/n)")
-                    if input() == "y":
-                        totalcorrect += 1
-            numbers.append(totalcorrect)
-
-
-if __name__ == '__main__':
-    calculatecorrect()
-    pos = list(range(numEntries))
-    labels.append("Human")
-    numbers.append(48.11)
-    bar_chart(numbers, labels, pos)
-    print(labels)
-    print(numbers)
@@ -1,30 +0,0 @@
-#!/usr/bin/env bash
-set -e
-
-MODEL=./models/ggml-vicuna-13b-1.1-q4_0.bin
-MODEL_NAME=Vicuna
-
-# exec options
-prefix="Human: " # Ex. Vicuna uses "Human: "
-opts="--temp 0 -n 80" # additional flags
-nl='
-'
-introduction="You will be playing a game of Jeopardy. Simply answer the question in the correct format (Ex. What is Paris, or Who is George Washington)."
-
-# file options
-question_file=./examples/jeopardy/questions.txt
-touch ./examples/jeopardy/results/$MODEL_NAME.txt
-output_file=./examples/jeopardy/results/$MODEL_NAME.txt
-
-counter=1
-
-echo 'Running'
-while IFS= read -r question
-do
-  exe_cmd="./llama-cli -p "\"$prefix$introduction$nl$prefix$question\"" "$opts" -m ""\"$MODEL\""" >> ""\"$output_file\""
-  echo $counter
-  echo "Current Question: $question"
-  eval "$exe_cmd"
-  echo -e "\n------" >> $output_file
-  counter=$((counter+1))
-done < "$question_file"
@@ -1,103 +0,0 @@
-Index,Original Category,Original Correct Question,Model Prompt
-1,The Oscars,Who is John Williams?,Which actor Born in 1932 was the son of a percussionist in the CBS radio orchestra has been nominated for 53 Oscars?
-2,English Literature,What is Paradise Lost?,"What work in English Literature says: 'The mind is its own place, & in itself can make a heaven of hell, a hell of heaven. What matter where, if I be still the same'?"
-3,Writers’ Lesser-Known Works,Who is Niccolò Machiavelli?,"Known for more philosophical works, he wrote the play 'La Mandragola', in which Florentines are rewarded for immoral actions?"
-4,Exploration,What is Easter Island (Rapa Nui)?,"James Cook's account of a 1774 visit where records an object 'near 27 feet long, and upwards of 8 feet over the breast or shoulders'?"
-5,The Bill of Rights,What is the Eighth Amendment?,England's 'Bloody Assizes' & a 1685 life sentence for perjury were 2 main origins of which amendment to the U.S. Constitution?
-6,Nobel Peace Prize Winners,Who are Nelson Mandela & Desmond Tutu?,"Which nobel peace price winners each lived at times on Vilakazi St. in Soweto , so it claims to be the world's only street home to 2 Nobel Peace Prize winners?"
-7,Famous Names,Who is Walt Disney?,"In 1966, the year of who's death did he share plans for an experimental prototype community in Florida?"
-8,Geography,What is Colombia?,"Of the 13 nations through which the Equator passes, what is the only one whose coastline borders the Caribbean Sea?"
-9,Fashion History,What are rhinestones?,"Which decorative items in fashion history get their name from their origin in the port city of Strasbourg, on the border of France & Germany?"
-10,Movies of the ’80s,What is Driving Miss Daisy?,What 1980's movie is based on an off-Broadway play with just 3 characters and won the Best Picture Oscar & the actors in all 3 roles were nominated?
-11,Novelists,Who is John Grisham?,"A 2012 book review for which novelist noted subjects that 'sparked his ire': capital punishment, big tobacco & 'the plight of the unjustly convicted'?"
-12,20th Century Eponyms,What is the Maginot Line?,"A 1940 headline about what 20th Century Eponym included 'failure', 'liability when it came to offense' & 'stout hearts no match for tanks'?"
-13,City History,What is Stockholm?,"Over 700 years after its traditional 1252 founding date, what port city became associated with a psychological response?"
-14,Brand Names,What is Jacuzzi?,"The success of what brand has its roots with a hydrotherapy pump its cofounder created for his son, who had arthritis?"
-15,American Authors,Who is Washington Irving?,"In a periodical in 1807, what American Author called New York City 'Gotham, Gotham! Most enlightened of cities'?"
-16,Symbols,What is “less than”?,What symbol is a rotated V in math and a feeling of some marginalized or underrepresented people in society?
-17,Movie Theme Songs,Who is James Bond?,"Monty Norman, the composer of what character's theme, said the staccato riff conveyed sexiness, mystery & ruthlessness?"
-18,American Novelists,Who is Joseph Heller?,"What American Novelist served with an airman named Yohannan in World War II & despite what readers might think, he said he enjoyed his service?"
-19,Medieval Places,"What is Canterbury, England? (Canterbury Cathedral)","In what Medieval place did one of the participants in an 1170 event say, 'Let us away, knights; he will rise no more'?"
-20,Countries of Africa,What is Morocco?,"At one time a province of the Roman Empire, what African country kingdom is known to Arabic scholars as Al-Maghrib Al-Aqsa, 'the far west'?"
-21,Statehood,What is Wyoming?,Congress relented in 1890 after what prospective state said it would wait 100 years rather than come in without the women?
-22,1980s Movies,What is Raiders of the Lost Ark?,"A writer & producer of what movie said he wanted it to be like a Western or James Bond film, 'only it takes place in the 30s'?"
-23,Art Exhibitions,Who is Rembrandt?,In 1898 what's been called the first blockbuster art show was devoted to which artist & put on for Queen Wilhelmina's coronation?
-24,Countries of the World,What is Mongolia?,"Part of the largest contiguous land empire during the 1200s & 1300s, today what is the world's second-largest landlocked country?"
-25,Literature,What is “Howl”?,A 2006 book was titled 'The Poem That Changed America:' What 'Fifty Years Later'?
-26,Invasions,Who is William of Orange?,"Backed by 14,000 troops, who invaded England to restore, in his words, its 'religion, laws, and liberties'?"
-27,Landmarks,What is the Eiffel Tower?,"After its completion in the late 19th c., what was landmark was called 'a truly tragic street lamp' & a 'high & skinny pyramid of iron ladders'?"
-28,Geographic Name’s the Same,What is Dover?,"The busiest passenger port in the U.K., what shares its name with a capital of one of the original 13 states?"
-29,Names in the Bookstore,Who is Peter Mark Roget?,"This man made lists, perhaps to cope with depression; a set of lists he published in 1852 made whose name synonymous with a type of book?"
-30,U.S. History,Who is Dr. Samuel Mudd?,"An 1869 presidential pardon was granted to which man, due in part to a plea by the Medical Society of Harford County, Maryland?"
-31,American Literature,What is The Things They Carried?,"Letters, pocket knives, C rations & steel helmets are among the tangible items referred to in the title of what American literature modern war classic?"
-32,Nonfiction,What is The Communist Manifesto,"What nonfiction book has the line, 'The discovery of America…opened up fresh ground for the rising bourgeoisie'?"
-33, a new version was passed 81 years later,Laws in U.S. History,What is the Civil Rights Act?,,,,,,,,,,,,,,,,,,0, 2/3
-34,Names of Myth,Who is Helen of Troy?,"Whose brothers, Castor & Pollux, saved her after Theseus stole her away as a kid; a larger force would seek her later in life?"
-35,African Countries,What is Sudan?,"Once Africa's largest country in area, what African Country dropped to third in 2011 when a portion of it declared independence?"
-36,The Ancient World,What is Alexandria?,"The ancient writer Galen said books on ships arriving to what city's port were seized, originals kept & copies returned?"
-37,Famous Names,Who is Andy Warhol?,"For a special 1970s cookbook, who provided one simple recipe–a can of Campbell's tomato soup & 2 cans of milk?"
-38,People & Places,What is Guam?,"Thought to descend from people of Southeast Asia, the Chamorro make up what U.S. territory’s largest ethnic group?"
-39,Current World Leaders,What is the Philippines?,"In office from 2022, the president of what country has taken so many foreign trips a play on his name is 'Ferdinand Magellan Jr.'?"
-40,Writers & The South,Who is Tennessee Williams?,In 1939 which writer lived on Toulouse Street in the French Quarter & chose the professional name that bonded him to the South?
-41,National Parks,What is Yellowstone?,"What National Park is named for a river indigenous people called Mi tse a-da-zi, translated by French-speaking trappers as 'Pierre Jaune'?"
-42,Sports,Who are the Harlem Globetrotters?,"In 2010 who introduced the 4-point shot, 35 feet from the basket?"
-43,The U.S. Military,What is “Top Gun”?,Losses over Asia in the 1960s led to the establishment of the program known as what at a San Diego naval base in 1969?
-44,Art & Science,What is Halley’s Comet?,"A craft that visited what was named for Giotto, based on the story that 680 years earlier, the painter depicted it as the Star of Bethlehem?"
-45,Words From World War I,What is “tank”?,"In World War I, 'Cistern' & 'reservoir' were suggested names for what secret invention, but the British preferred this less clumsy monosyllable?"
-46,European History,What is Holy Roman Emperor?,"Until 1806, some German nobles included among their honors the title of 'Elector' for their role in selecting this personage?"
-47,Theater History,Who is Peter Pan?,"In 1904, wearing a harness, actress Nina Boucicault became the first to play what character onstage?"
-48,European Cities,What is Aachen?,"Alphabetically the first German city in encyclopedias, what was also the first one taken by the Allies in World War II?"
-49,Word Origins,What is mantra?,This Sanskrit word referring to a spoken word or phrase comes from a word for 'to think'?
-50,Inventions,What is barbed wire?,1917's 'Elements of Trench Warfare' said what Old West invention was 'difficult to destroy' & 'difficult to get through'?
-51,World War II,What is Schindler’s list?,"Mimi Reinhard, who never learned to type using more than 2 fingers, produced what in World War II with 1,100 names, including hers?"
-52, their offspring was the source of this mythical object,Mythology,What is the Golden Fleece?
-53,Literature,What is Pride and Prejudice?,"Published in 2011, P.D. James' final novel, 'Death Comes to Pemberley', was a sequel to what novel from 200 years earlier?"
-54, only these 2 west of the Mississippi River border each other,U.S. State Names,What are Oregon & Nevada?
-55,Word Origins,What is passion?,"Originally relating to a story of suffering, what word now more commonly refers to strong emotion of any kind?"
-56,World Cinema,What is La Vie en Rose?,"The 2007 biopic called 'La Môme' in France, meaning 'The Kid', was released in the U.S. under what other French title?"
-57,History,What is Santa Maria?,"Returning home in 1493, Columbus stopped in the Azores at an island with what name, also something he'd lost off the Haiti coast?"
-58,Landmarks,What is a kremlin?,Pskov & Nizhny Novgorod are 2 of the cities that have a fortress called what?
-59,Foreign-Born Authors,Who is Vladimir Nabokov?,In the 1950s the New York Times said what author 'is writing about all lust' & his lecherous narrator 'is all of us'?
-60,Astronomy & Geography,What is Capricorn?,"At the winter solstice, the sun is in Sagittarius; it once appeared in what constellation, giving a geographic feature its name?"
-61,Television,What is Law & Order?,"Mike Post combined the sound of a slamming jail door, an anvil & 100 men stomping on a floor for what television series that debuted in 1990?"
-62,British Landmarks,What is the Tower of London?,"Like Sir Thomas More, 3 16th century English queens are buried at what British location?"
-63,Early American History,What are witches?,"In 1692 Increase Mather wrote, 'It were better that ten suspected' of these who 'escape, than that one innocent person … be condemned'?"
-64,Geography Mnemonics,What are Arkansas and Louisiana?,"The Geography Mnemonic Mimal, sometimes said to be the silhouette of a chef or elf, stands for Minnesota, Iowa, Missouri, and what other 2 states?"
-65,Business Milestones,What is the Ford Model T?,"What was first sold in 1908, at a price equivalent to about $27,000 today?"
-66,In The Bookstore,Who is Tom Clancy?,The name of what author dead since 2013 now appears on books written by a former U.S. marshal & a former Apache helicopter pilot?
-67,Historic Art,What is the Bayeux Tapestry?,The artwork once known in France as 'la tapisserie de la Reine Mathilde' is better known as what?
-68,Pop Stars,Who is Madonna?,In 2022 which pop star became the first woman to have a Billboard Top 10 album in 5 decades starting with the 1980s?
-69,Classic Tale Characters,Who is Scheherazade?,"In one 19th century translation, what female classic tale character 'perceived the dawn of day and ceased' speaking nearly 1,000 times?"
-70,USA,What is Jack Daniel’s?,"Ironically, though what company founded in the 1860s is Moore County, Tennessee's largest employer, Moore is a dry county?"
-71,Historic People,Who was William Bligh?,"After a 1789 event, who wrote, 'My first determination was to seek a supply of…water at Tofoa, & afterwards to sail for Tongataboo'?"
-72,The Movies,What is The Godfather?,Laurence Olivier & Ernest Borgnine were considered for the lead role & Sergio Leone to direct for what film that turned 50 in 2022?
-73,Continental Geography,What is Colombia?,"Until a 1903 secession, what country's contiguous territory spanned 2 continents?"
-74,Foreign-Born Authors,Who is Isabel Allende?,"Early in her career which foreign-born author translated romance novels into Spanish, often changing the dialogue to make the heroines smarter?"
-75,Historic Crimes,What is the Mona Lisa?,"Saying it was stolen by Napoleon, self-styled Italian patriot Vincenzo Peruggia took what in 1911?"
-76,U.S. Bodies of Water,What is Lake Mead?,"Continuing a downward trend, in July 2022 what US body of water was at 27% capacity, its lowest level since 1937 when it was first being filled?"
-77,Gods & Goddesses,Who is Aurora (or Eos)?,"Each morning which goddess began her ride in her chariot across the sky ahead of her brother Sol, or Helios?"
-78,America At War,What is the Battle of New Orleans?,"Until the Civil War, the Jan. 8 date of what American battle of dubious military importance but big morale value was a national holiday?"
-79,Children’s Books,What is The Velveteen Rabbit?,"Which children's book title character is told 'By the time you are real, most of your hair has been loved off your eyes drop out & you get shabby'?"
-80,TV Finales,What is Grace and Frankie?,"In a TV reunion over 40 years in the making, Dolly Parton appeared as an angel named Agnes in the final episode of what comedy in 2022?"
-81,American Poems,Who is Evangeline?,"In an 1847 American poem what character sees her town of Grand-Pré burned, but finally reunites with her beau for a kiss before his death?"
-82,Famous Names,Who is Banksy?,"In 2001 who published a book called 'Banging Your Head Against a Brick Wall'; in 2002, 'Existencilism'?"
-83,Children’s Lit,What is Charlotte’s Web?,The title object of what childrens book 'never looked more beautiful each strand held dozens of bright drops of early morning dew'?
-84,Classic Songs,What is “Here Comes Santa Claus”?,The shouts of excited children at a 1946 holiday parade are said to have inspired what perennial classic song favorite?
-85,Brand Names,What are Milk Duds?,"Unable to make what candies perfectly round, the confectioner embraced this flawed name for the product?"
-86,Countries of the World,What is Italy?,"What country is home to 58 UNESCO World Heritage Sites, more than any other country; the sites include a volcano & a lagoon?"
-87,Action Movies,What is Die Hard?,"What action movie's last line is 'If this is their idea of Christmas, I gotta be here for New Years'?"
-88,Presidential Facts,Who is Woodrow Wilson?,Only 3 presidents have married while in office— John Tyler was the first & which one was the last?
-89,19th Century Americans,Who is Frederick Douglass?,"Demonstrating the dignity & humanity of Black Americans, who sat for 160 known photographs, the most of any American in the 19th century?"
-90,Latin Phrases,What is “quid pro quo”?,"Originally, which Latin 3-word phrase referred to when a doctor or apothecary substituted one medicine for another?"
-91,1970s Movies,What is Monty Python and the Holy Grail?,The 1975 premiere of what movie comedy advertised free coconuts for the first thousand in the audience?
-92,Name’s The Same,What is Manhattan?,"A cocktail, an island & a WWII venture originally called 'Development of Substitute Materials' all bear what name?"
-93,U.S. Presidents,Who is Calvin Coolidge?,"Which US President was sworn in twice as President within 2 years, first by his father & then later by a former U.S. President?"
-94,Plays,What is The Tempest?,A 1609 story in which an exiled king of Bulgaria creates a sea palace with his magic may have inspired the plot of what play?
-95,Landmarks,What is the Berlin Wall?,"In 2009, during a 20th anniversary celebration, what landmark was called 'an edifice of fear. On Nov. 9, it became a place of joy'?"
-96,World Capitals,"What is Vienna, Austria?","Among what world capital's nicknames are the 'City of Classical Music' &, possibly in honor of a famous resident from 1860 to 1938, the 'City of Dreams'?"
-97,Language & Its Meanings,What is a night owl?,"Now meaning someone with nocturnal habits, what catches a sleeping dove in Shakespeare's 'Lucrece'?"
-98,Flags of Our Hemisphere,What is Brazil?,"The stars on what country's flag represent states, 26 of them; unlike the USA's, its 'federal district' gets its own 27th star?"
-99,Names in U.S. History,Who is Oliver Brown?,What father was the only man among the 13 plaintiffs in a US class-action case filed in 1951?
-100,Children’s Authors,"Who is Sarah? (from Sarah, Plain and Tall)","Reversing the story of what heroine she created, childrens author Patricia Maclachlan was born on the prairie but spent much of her life in New England?"
-,,,
-TOTALS,,,
@@ -1,100 +0,0 @@
-Which man born in 1932 was the son of a percussionist in the CBS radio orchestra has been nominated for 53 Oscars?
-What work in English Literature says: 'The mind is its own place, & in itself can make a heaven of hell, a hell of heaven. What matter where, if I be still the same'?
-Known for more philosophical works, he wrote the play 'La Mandragola', in which Florentines are rewarded for immoral actions?
-James Cook's account of a 1774 visit where records an object 'near 27 feet long, and upwards of 8 feet over the breast or shoulders'?
-England's 'Bloody Assizes' & a 1685 life sentence for perjury were 2 main origins of which amendment to the U.S. Constitution?
-Which nobel peace price winners each lived at times on Vilakazi St. in Soweto , so it claims to be the world's only street home to 2 Nobel Peace Prize winners?
-In 1966, the year of who's death did he share plans for an experimental prototype community in Florida?
-Of the 13 nations through which the Equator passes, what is the only one whose coastline borders the Caribbean Sea?
-Which decorative items in fashion history get their name from their origin in the port city of Strasbourg, on the border of France & Germany?
-What 1980's movie is based on an off-Broadway play with just 3 characters and won the Best Picture Oscar & the actors in all 3 roles were nominated?
-A 2012 book review for which novelist noted subjects that 'sparked his ire': capital punishment, big tobacco & 'the plight of the unjustly convicted'?
-A 1940 headline about what 20th Century Eponym included 'failure', 'liability when it came to offense' & 'stout hearts no match for tanks'?
-Over 700 years after its traditional 1252 founding date, what port city became associated with a psychological response?
-The success of what brand has its roots with a hydrotherapy pump its cofounder created for his son, who had arthritis?
-In a periodical in 1807, what American Author called New York City 'Gotham, Gotham! Most enlightened of cities'?
-What symbol is a rotated V in math and a feeling of some marginalized or underrepresented people in society?
-Monty Norman, the composer of what character's theme, said the staccato riff conveyed sexiness, mystery & ruthlessness?
-What American Novelist served with an airman named Yohannan in World War II & despite what readers might think, he said he enjoyed his service?
-In what Medieval place did one of the participants in an 1170 event say, 'Let us away, knights; he will rise no more'?
-At one time a province of the Roman Empire, what African country kingdom is known to Arabic scholars as Al-Maghrib Al-Aqsa, 'the far west'?
-Congress relented in 1890 after what prospective state said it would wait 100 years rather than come in without the women?
-A writer & producer of what movie said he wanted it to be like a Western or James Bond film, 'only it takes place in the 30s'?
-In 1898 what's been called the first blockbuster art show was devoted to which artist & put on for Queen Wilhelmina's coronation?
-Part of the largest contiguous land empire during the 1200s & 1300s, today what is the world's second-largest landlocked country?
-A 2006 book was titled 'The Poem That Changed America:' What 'Fifty Years Later'?
-Backed by 14,000 troops, who invaded England to restore, in his words, its 'religion, laws, and liberties'?
-After its completion in the late 19th c., what was landmark was called 'a truly tragic street lamp' & a 'high & skinny pyramid of iron ladders'?
-The busiest passenger port in the U.K., what shares its name with a capital of one of the original 13 states?
-This man made lists, perhaps to cope with depression; a set of lists he published in 1852 made whose name synonymous with a type of book?
-An 1869 presidential pardon was granted to which man, due in part to a plea by the Medical Society of Harford County, Maryland?
-Letters, pocket knives, C rations & steel helmets are among the tangible items referred to in the title of what American literature modern war classic?
-What nonfiction book has the line, 'The discovery of America…opened up fresh ground for the rising bourgeoisie'?
-A radical Republican championed what 1875 act but the Supreme Court struck it down in 1883; a new version was passed 81 years later?
-Whose brothers, Castor & Pollux, saved her after Theseus stole her away as a kid; a larger force would seek her later in life?
-Once Africa's largest country in area, what African Country dropped to third in 2011 when a portion of it declared independence?
-The ancient writer Galen said books on ships arriving to what city's port were seized, originals kept & copies returned?
-For a special 1970s cookbook, who provided one simple recipe–a can of Campbell's tomato soup & 2 cans of milk?
-Thought to descend from people of Southeast Asia, the Chamorro make up what U.S. territory’s largest ethnic group?
-In office from 2022, the president of what country has taken so many foreign trips a play on his name is 'Ferdinand Magellan Jr.'?
-In 1939 which writer lived on Toulouse Street in the French Quarter & chose the professional name that bonded him to the South?
-What National Park is named for a river indigenous people called Mi tse a-da-zi, translated by French-speaking trappers as 'Pierre Jaune'?
-In 2010 who introduced the 4-point shot, 35 feet from the basket?
-Losses over Asia in the 1960s led to the establishment of the program known as what at a San Diego naval base in 1969?
-A craft that visited what was named for Giotto, based on the story that 680 years earlier, the painter depicted it as the Star of Bethlehem?
-In World War I, 'Cistern' & 'reservoir' were suggested names for what secret invention, but the British preferred this less clumsy monosyllable?
-Until 1806, some German nobles included among their honors the title of 'Elector' for their role in selecting this personage?
-In 1904, wearing a harness, actress Nina Boucicault became the first to play what character onstage?
-Alphabetically the first German city in encyclopedias, what was also the first one taken by the Allies in World War II?
-This Sanskrit word referring to a spoken word or phrase comes from a word for 'to think'?
-1917's 'Elements of Trench Warfare' said what Old West invention was 'difficult to destroy' & 'difficult to get through'?
-Mimi Reinhard, who never learned to type using more than 2 fingers, produced what in World War II with 1,100 names, including hers?
-Poseidon carried off the maiden Theophane & turned her into a ewe; their offspring was the source of what mythical object?
-Published in 2011, P.D. James' final novel, 'Death Comes to Pemberley', was a sequel to what novel from 200 years earlier?
-5 U.S. states have 6-letter names; only which 2 west of the Mississippi River border each other?
-Originally relating to a story of suffering, what word now more commonly refers to strong emotion of any kind?
-The 2007 biopic called 'La Môme' in France, meaning 'The Kid', was released in the U.S. under what other French title?
-Returning home in 1493, Columbus stopped in the Azores at an island with what name, also something he'd lost off the Haiti coast?
-Pskov & Nizhny Novgorod are 2 of the cities that have a fortress called what?
-In the 1950s the New York Times said what author 'is writing about all lust' & his lecherous narrator 'is all of us'?
-At the winter solstice, the sun is in Sagittarius; it once appeared in what constellation, giving a geographic feature its name?
-Mike Post combined the sound of a slamming jail door, an anvil & 100 men stomping on a floor for what television series that debuted in 1990?
-Like Sir Thomas More, 3 16th century English queens are buried at what British location?
-In 1692 Increase Mather wrote, 'It were better that ten suspected' of these who 'escape, than that one innocent person be condemned'?
-The Geography Mnemonic Mimal, sometimes said to be the silhouette of a chef or elf, stands for Minnesota, Iowa, Missouri, and what other 2 states?
-What was first sold in 1908, at a price equivalent to about $27,000 today?
-The name of what author dead since 2013 now appears on books written by a former U.S. marshal & a former Apache helicopter pilot?
-The artwork once known in France as 'la tapisserie de la Reine Mathilde' is better known as what?
-In 2022 which pop star became the first woman to have a Billboard Top 10 album in 5 decades starting with the 1980s?
-In one 19th century translation, what female classic tale character 'perceived the dawn of day and ceased' speaking nearly 1,000 times?
-Ironically, though what company founded in the 1860s is Moore County, Tennessee's largest employer, Moore is a dry county?
-After a 1789 event, who wrote, 'My first determination was to seek a supply of…water at Tofoa, & afterwards to sail for Tongataboo'?
-Laurence Olivier & Ernest Borgnine were considered for the lead role & Sergio Leone to direct for what film that turned 50 in 2022?
-Until a 1903 secession, what country's contiguous territory spanned 2 continents?
-Early in her career which foreign-born author translated romance novels into Spanish, often changing the dialogue to make the heroines smarter?
-Saying it was stolen by Napoleon, self-styled Italian patriot Vincenzo Peruggia took what in 1911?
-Continuing a downward trend, in July 2022 what US body of water was at 27% capacity, its lowest level since 1937 when it was first being filled?
-Each morning which goddess began her ride in her chariot across the sky ahead of her brother Sol, or Helios?
-Until the Civil War, the Jan. 8 date of what American battle of dubious military importance but big morale value was a national holiday?
-Which children's book title character is told 'By the time you are real, most of your hair has been loved off your eyes drop out & you get shabby'?
-In a TV reunion over 40 years in the making, Dolly Parton appeared as an angel named Agnes in the final episode of what comedy in 2022?
-In an 1847 American poem what character sees her town of Grand-Pré burned, but finally reunites with her beau for a kiss before his death?
-In 2001 who published a book called 'Banging Your Head Against a Brick Wall'; in 2002, 'Existencilism'?
-The title object of what childrens book 'never looked more beautiful each strand held dozens of bright drops of early morning dew'?
-The shouts of excited children at a 1946 holiday parade are said to have inspired what perennial classic song favorite?
-Unable to make what candies perfectly round, the confectioner embraced this flawed name for the product?
-What country is home to 58 UNESCO World Heritage Sites, more than any other country; the sites include a volcano & a lagoon?
-What action movie's last line is 'If this is their idea of Christmas, I gotta be here for New Years'?
-Only 3 presidents have married while in office— John Tyler was the first & which one was the last?
-Demonstrating the dignity & humanity of Black Americans, who sat for 160 known photographs, the most of any American in the 19th century?
-Originally, which Latin 3-word phrase referred to when a doctor or apothecary substituted one medicine for another?
-The 1975 premiere of what movie comedy advertised free coconuts for the first thousand in the audience?
-A cocktail, an island & a WWII venture originally called 'Development of Substitute Materials' all bear what name?
-Which US President was sworn in twice as President within 2 years, first by his father & then later by a former U.S. President?
-A 1609 story in which an exiled king of Bulgaria creates a sea palace with his magic may have inspired the plot of what play?
-In 2009, during a 20th anniversary celebration, what landmark was called 'an edifice of fear. On Nov. 9, it became a place of joy'?
-Among what world capital's nicknames are the 'City of Classical Music' &, possibly in honor of a famous resident from 1860 to 1938, the 'City of Dreams'?
-Now meaning someone with nocturnal habits, what catches a sleeping dove in Shakespeare's 'Lucrece'?
-The stars on what country's flag represent states, 26 of them; unlike the USA's, its 'federal district' gets its own 27th star?
-What father was the only man among the 13 plaintiffs in a US class-action case filed in 1951?
-Reversing the story of what heroine she created, childrens author Patricia Maclachlan was born on the prairie but spent much of her life in New England?
@@ -1,28 +0,0 @@
-" Basic plugin example
-
-function! Llm()
-
-  let url = "http://127.0.0.1:8080/completion"
-
-  " Get the content of the current buffer
-  let buffer_content = join(getline(1, '$'), "\n")
-
-  " Create the JSON payload
-  let json_payload = {"temp":0.72,"top_k":100,"top_p":0.73,"repeat_penalty":1.100000023841858,"n_predict":256,"stop": ["\n\n\n"],"stream": v:false}
-  let json_payload.prompt = buffer_content
-
-  " Define the curl command
-  let curl_command = 'curl -k -s -X POST -H "Content-Type: application/json" -d @- ' . url
-  let response = system(curl_command, json_encode(json_payload))
-
-  " Extract the content field from the response
-  let content = json_decode(response).content
-
-  let split_newlines = split(content, '\n', 1)
-
-  " Insert the content at the cursor position
-  call setline(line('.'), [ getline('.') . split_newlines[0] ] + split_newlines[1:])
-endfunction
-
-command! Llm call Llm()
-noremap <F2> :Llm<CR>
@@ -105,12 +105,12 @@ new model, the model can be converted to GGUF format using the following command
 ### Inspecting the converted model
 The converted model can be inspected using the following command:
 ```console
-(venv) $ make inspect-converted-model
+(venv) $ make causal-inspect-converted-model
 ```

 ### Running the converted model
 ```console
-(venv) $ make run-converted-model
+(venv) $ make causal-run-converted-model
 ```

 ### Model logits verfication
@@ -48,7 +48,7 @@ def main():
        print(f"Error: Model file not found: {model_path}")
        sys.exit(1)

-    model_name = os.path.splitext(os.path.basename(model_path))[0]
+    model_name = os.path.basename(model_path)
    data_dir = Path("data")

    pytorch_file = data_dir / f"pytorch-{model_name}.bin"
@@ -193,7 +193,7 @@ print(f"Input text: {repr(prompt)}")
 print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")

 with torch.no_grad():
-    outputs = model(input_ids)
+    outputs = model(input_ids.to(model.device))
    logits = outputs.logits

    # Extract logits for the last token (next token prediction)
@@ -67,7 +67,7 @@ def main():
    parser.add_argument('-m', '--model-path', required=True,  help='Path to the model directory')
    args = parser.parse_args()

-    model_name = os.path.splitext(os.path.basename(args.model_path))[0]
+    model_name = os.path.basename(args.model_path)
    data_dir = Path("data")

    pytorch_file = data_dir / f"pytorch-{model_name}.bin"
@@ -1,5 +1,41 @@
 cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
 project("ggml" C CXX ASM)
+
+### GGML Version
+set(GGML_VERSION_MAJOR 0)
+set(GGML_VERSION_MINOR 9)
+set(GGML_VERSION_PATCH 0)
+set(GGML_VERSION_DEV "-dev")  # "-dev" for development, "" for releases
+set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
+
+find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
+if(GIT_EXE)
+    # Get current git commit hash
+    execute_process(COMMAND ${GIT_EXE} rev-parse --short HEAD
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        OUTPUT_VARIABLE GGML_BUILD_COMMIT
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+        ERROR_QUIET
+    )
+
+    # Check if the working directory is dirty (i.e., has uncommitted changes)
+    execute_process(COMMAND ${GIT_EXE} diff-index --quiet HEAD -- .
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        RESULT_VARIABLE GGML_GIT_DIRTY
+        ERROR_QUIET
+    )
+endif()
+
+# Build the version string with optional -dev suffix and dirty flag
+set(GGML_VERSION "${GGML_VERSION_BASE}${GGML_VERSION_DEV}")
+if(GGML_GIT_DIRTY AND NOT GGML_GIT_DIRTY EQUAL 0)
+    set(GGML_VERSION "${GGML_VERSION}-dirty")
+endif()
+
+if(NOT GGML_BUILD_COMMIT)
+    set(GGML_BUILD_COMMIT "unknown")
+endif()
+
 include(CheckIncludeFileCXX)

 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
@@ -300,26 +336,6 @@ endif()
 # Create CMake package
 #

-# Generate version info based on git commit.
-
-if(NOT DEFINED GGML_BUILD_NUMBER)
-    find_program(GIT_EXE NAMES git git.exe REQUIRED NO_CMAKE_FIND_ROOT_PATH)
-    execute_process(COMMAND ${GIT_EXE} rev-list --count HEAD
-        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
-        OUTPUT_VARIABLE GGML_BUILD_NUMBER
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-    )
-
-    if(GGML_BUILD_NUMBER EQUAL 1)
-        message(WARNING "GGML build version fixed at 1 likely due to a shallow clone.")
-    endif()
-
-    execute_process(COMMAND ${GIT_EXE} rev-parse --short HEAD
-        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
-        OUTPUT_VARIABLE GGML_BUILD_COMMIT
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-    )
-endif()


 # Capture variables prefixed with GGML_.
@@ -348,7 +364,7 @@ set(GGML_VARIABLES_EXPANDED ${variable_set_statements})

 # Create the CMake package and set install location.

-set(GGML_INSTALL_VERSION 0.0.${GGML_BUILD_NUMBER})
+set(GGML_INSTALL_VERSION ${GGML_VERSION})
 set(GGML_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
 set(GGML_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
 set(GGML_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
@@ -314,7 +314,8 @@ extern "C" {
    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);

-    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_type_t ggml_backend_sched_get_buffer_type(ggml_backend_sched_t sched, ggml_backend_t backend);
+    GGML_API size_t                     ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);

    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
@@ -7,6 +7,9 @@
 extern "C" {
 #endif

+// device buffer
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_zdnn_buffer_type(void);
+
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_zdnn_reg(void);

 #ifdef __cplusplus
@@ -114,6 +114,9 @@ message(STATUS "GGML_SYSTEM_ARCH: ${GGML_SYSTEM_ARCH}")

 if (NOT MSVC)
    if (GGML_STATIC)
+        if (UNIX AND NOT APPLE)
+            set(CMAKE_FIND_LIBRARY_SUFFIXES ".a;.so")
+        endif()
        add_link_options(-static)
        if (MINGW)
            add_link_options(-static-libgcc -static-libstdc++)
@@ -23,7 +23,7 @@ static bool ggml_is_view(const struct ggml_tensor * t) {
 }

 // ops that return true for this function must not use restrict pointers for their backend implementations
-static bool ggml_op_can_inplace(enum ggml_op op) {
+bool ggml_op_can_inplace(enum ggml_op op) {
    switch (op) {
        case GGML_OP_SCALE:
        case GGML_OP_DIAG_MASK_ZERO:
@@ -95,39 +95,104 @@ enum ggml_status ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_te

 // dynamic tensor allocator

+#define GGML_VBUFFER_MAX_CHUNKS 16
+
+// relative memory address within an allocation that can be split into multiple buffers (chunks)
+struct buffer_address {
+    int chunk;     // index of a backend buffer
+    size_t offset; // local memory offset within the buffer
+};
+
+static const struct buffer_address GGML_BUFFER_ADDRESS_INVALID = { -1, SIZE_MAX };
+
+static bool ggml_buffer_address_less(struct buffer_address a, struct buffer_address b) {
+    return a.chunk != b.chunk ? a.chunk < b.chunk : a.offset < b.offset;
+}
+
 struct free_block {
    size_t offset;
    size_t size;
 };

+struct tallocr_chunk {
+    struct free_block free_blocks[MAX_FREE_BLOCKS];
+    int n_free_blocks;
+    size_t max_size;
+};
+
 struct ggml_dyn_tallocr {
    size_t alignment;
-    int n_free_blocks;
-    struct free_block free_blocks[MAX_FREE_BLOCKS];
-    size_t max_size;
+    size_t max_chunk_size;
+    struct tallocr_chunk * chunks[GGML_VBUFFER_MAX_CHUNKS];
+    int n_chunks;

 #ifdef GGML_ALLOCATOR_DEBUG
    struct {
        const struct ggml_tensor * tensor;
-        size_t offset;
+        struct buffer_address addr;
    } allocated_tensors[1024];
 #endif
 };

+static void ggml_dyn_tallocr_insert_block(struct tallocr_chunk * chunk, size_t offset, size_t size) {
+    GGML_ASSERT(chunk->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
+    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
+    int insert_pos = 0;
+    while (insert_pos < chunk->n_free_blocks && chunk->free_blocks[insert_pos].offset < offset) {
+        insert_pos++;
+    }
+    // shift all blocks from insert_pos onward to make room for the new block
+    for (int i = chunk->n_free_blocks; i > insert_pos; i--) {
+        chunk->free_blocks[i] = chunk->free_blocks[i-1];
+    }
+    // insert the new block
+    chunk->free_blocks[insert_pos].offset = offset;
+    chunk->free_blocks[insert_pos].size = size;
+    chunk->n_free_blocks++;
+}
+
+static void ggml_dyn_tallocr_remove_block(struct tallocr_chunk * chunk, int idx) {
+    // shift all elements after idx by 1 to the left, overwriting the element at idx
+    for (int i = idx; i < chunk->n_free_blocks; i++) {
+        chunk->free_blocks[i] = chunk->free_blocks[i+1];
+    }
+    chunk->n_free_blocks--;
+}
+
+static int ggml_dyn_tallocr_new_chunk(struct ggml_dyn_tallocr * alloc, size_t min_size) {
+    if (alloc->n_chunks >= GGML_VBUFFER_MAX_CHUNKS) {
+        return -1;
+    }
+    struct tallocr_chunk * chunk = calloc(1, sizeof(struct tallocr_chunk));
+    chunk->n_free_blocks = 1;
+    chunk->free_blocks[0].offset = 0;
+    // available space in a chunk is limited to max_chunk_size, but can be higher if:
+    // 1. a single tensor exceeds the maximum, and cannot fit any other way
+    // 2. we are running out of chunks
+    // backends will either manage to allocate the larger size, or report an error.
+    chunk->free_blocks[0].size = MAX(min_size, alloc->max_chunk_size);
+    if (alloc->n_chunks == GGML_VBUFFER_MAX_CHUNKS - 1) {
+        chunk->free_blocks[0].size = SIZE_MAX/2;
+    }
+    alloc->chunks[alloc->n_chunks] = chunk;
+    alloc->n_chunks++;
+    return alloc->n_chunks - 1;
+}
+
 #ifdef GGML_ALLOCATOR_DEBUG
-static void add_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
+static void add_allocated_tensor(struct ggml_dyn_tallocr * alloc, struct buffer_address addr, const struct ggml_tensor * tensor) {
    for (int i = 0; i < 1024; i++) {
        if (alloc->allocated_tensors[i].tensor == NULL) {
            alloc->allocated_tensors[i].tensor = tensor;
-            alloc->allocated_tensors[i].offset = offset;
+            alloc->allocated_tensors[i].addr = addr;
            return;
        }
    }
    GGML_ABORT("out of allocated_tensors");
 }
-static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
+static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, struct buffer_address addr, const struct ggml_tensor * tensor) {
    for (int i = 0; i < 1024; i++) {
-        if (alloc->allocated_tensors[i].offset == offset) {
+        if (alloc->allocated_tensors[i].addr.chunk == addr.chunk && alloc->allocated_tensors[i].addr.offset == addr.offset) {
            alloc->allocated_tensors[i].tensor = NULL;
            return;
        }
@@ -136,76 +201,94 @@ static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offs
 }
 #endif

-static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t size, const struct ggml_tensor * tensor) {
+static struct buffer_address ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t size, const struct ggml_tensor * tensor) {
    size = aligned_offset(NULL, size, alloc->alignment);

    AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);

+    int best_fit_chunk = -1;
+    int best_fit_block = -1;
    size_t max_avail = 0;

-    // find the best fitting free block besides the last block
-    int best_fit_block = -1;
-    size_t best_fit_size = SIZE_MAX;
-    for (int i = 0; i < alloc->n_free_blocks - 1; i++) {
-        struct free_block * block = &alloc->free_blocks[i];
-        max_avail = MAX(max_avail, block->size);
-        if (block->size >= size && block->size <= best_fit_size) {
-            best_fit_block = i;
-            best_fit_size = block->size;
+    // find the best fitting free block besides the last block, within any chunk
+    for (int c = 0; c < alloc->n_chunks; ++c) {
+        struct tallocr_chunk * chunk = alloc->chunks[c];
+        size_t best_fit_size = SIZE_MAX;
+        for (int i = 0; i < chunk->n_free_blocks - 1; i++) {
+            struct free_block * block = &chunk->free_blocks[i];
+            max_avail = MAX(max_avail, block->size);
+            if (block->size >= size && block->size <= best_fit_size) {
+                best_fit_chunk = c;
+                best_fit_block = i;
+                best_fit_size = block->size;
+            }
        }
    }

    if (best_fit_block == -1) {
-        // the last block is our last resort
-        struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
-        max_avail = MAX(max_avail, block->size);
-        if (block->size >= size) {
-            best_fit_block = alloc->n_free_blocks - 1;
-        } else {
-            // this should never happen
-            GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
-                    __func__, size, max_avail);
-            GGML_ABORT("not enough space in the buffer");
-        }
-    }
-
-    struct free_block * block = &alloc->free_blocks[best_fit_block];
-    size_t offset = block->offset;
-    block->offset = offset + size;
-    block->size -= size;
-    if (block->size == 0) {
-        // remove block if empty
-        alloc->n_free_blocks--;
-        for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
-            alloc->free_blocks[j] = alloc->free_blocks[j+1];
-        }
-    }
-
-    AT_PRINTF("block %d, offset %zu\n", best_fit_block, offset);
-
-#ifdef GGML_ALLOCATOR_DEBUG
-    add_allocated_tensor(alloc, offset, tensor);
-    size_t cur_max = offset + size;
-    if (cur_max > alloc->max_size) {
-        // sort allocated_tensors by offset
-        for (int i = 0; i < 1024; i++) {
-            for (int j = i + 1; j < 1024; j++) {
-                if (alloc->allocated_tensors[i].offset > alloc->allocated_tensors[j].offset) {
-                    const struct ggml_tensor * tmp_tensor = alloc->allocated_tensors[i].tensor;
-                    size_t tmp_offset = alloc->allocated_tensors[i].offset;
-                    alloc->allocated_tensors[i].tensor = alloc->allocated_tensors[j].tensor;
-                    alloc->allocated_tensors[i].offset = alloc->allocated_tensors[j].offset;
-                    alloc->allocated_tensors[j].tensor = tmp_tensor;
-                    alloc->allocated_tensors[j].offset = tmp_offset;
+        // no suitable block found, try the last block (this will grow a chunks size)
+        for (int c = 0; c < alloc->n_chunks; ++c) {
+            struct tallocr_chunk * chunk = alloc->chunks[c];
+            if (chunk->n_free_blocks > 0) {
+                struct free_block * block = &chunk->free_blocks[chunk->n_free_blocks - 1];
+                max_avail = MAX(max_avail, block->size);
+                if (block->size >= size) {
+                    best_fit_chunk = c;
+                    best_fit_block = chunk->n_free_blocks - 1;
+                    break;
                }
            }
        }
-        GGML_LOG_DEBUG("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
+    }
+
+    if (best_fit_block == -1) {
+        // none of the existing chunks have enough space left
+        best_fit_chunk = ggml_dyn_tallocr_new_chunk(alloc, size);
+        best_fit_block = 0;
+    }
+    if (best_fit_chunk == -1) {
+        // since the last chunk always has virtually endless memory, this should never happen
+        GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
+            __func__, size, max_avail);
+        GGML_ABORT("graph allocation: failed to reserve memory");
+    }
+
+    struct tallocr_chunk * chunk = alloc->chunks[best_fit_chunk];
+    struct free_block    * block = &chunk->free_blocks[best_fit_block];
+    struct buffer_address  addr  = {.chunk = best_fit_chunk, .offset = block->offset };
+    block->offset += size;
+    block->size -= size;
+    if (block->size == 0) {
+        // remove block if empty
+        ggml_dyn_tallocr_remove_block(chunk, best_fit_block);
+    }
+
+    AT_PRINTF("block %d, offset %zu, chunk %d\n", best_fit_block, addr.offset, addr.chunk);
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    add_allocated_tensor(alloc, addr, tensor);
+    size_t cur_max = addr.offset + size;
+    if (cur_max > alloc->max_size[addr.chunk]) {
+        // sort allocated_tensors by chunk/offset
+        for (int i = 0; i < 1024; i++) {
+            for (int j = i + 1; j < 1024; j++) {
+                if (ggml_buffer_address_less(alloc->allocated_tensors[j].addr, alloc->allocated_tensors[i].addr)) {
+                    const struct ggml_tensor * tmp_tensor = alloc->allocated_tensors[i].tensor;
+                    struct buffer_address tmp_addr = alloc->allocated_tensors[i].addr;
+                    alloc->allocated_tensors[i].tensor = alloc->allocated_tensors[j].tensor;
+                    alloc->allocated_tensors[i].addr = alloc->allocated_tensors[j].addr;
+                    alloc->allocated_tensors[j].tensor = tmp_tensor;
+                    alloc->allocated_tensors[j].addr = tmp_addr;
+                }
+            }
+        }
+        GGML_LOG_DEBUG("max_size[%d] = %.2f MB: tensors: ", addr.chunk, cur_max / 1024.0 / 1024.0);
        for (int i = 0; i < 1024; i++) {
            if (alloc->allocated_tensors[i].tensor) {
-                GGML_LOG_DEBUG("%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
-                    alloc->allocated_tensors[i].offset,
-                    alloc->allocated_tensors[i].offset + ggml_nbytes(alloc->allocated_tensors[i].tensor),
+                GGML_LOG_DEBUG("%s [%d: %zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
+                    alloc->allocated_tensors[i].addr.chunk,
+                    alloc->allocated_tensors[i].addr.offset,
+                    alloc->allocated_tensors[i].addr.offset + ggml_nbytes(alloc->allocated_tensors[i].tensor),
                    ggml_nbytes(alloc->allocated_tensors[i].tensor) / 1024.0 / 1024.0);
            }
        }
@@ -213,78 +296,69 @@ static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t siz
    }
 #endif

-    alloc->max_size = MAX(alloc->max_size, offset + size);
+    chunk->max_size = MAX(chunk->max_size, addr.offset + size);

-    return offset;
+    return addr;

    GGML_UNUSED(tensor);
 }

 // this is a very naive implementation, but for our case the number of free blocks should be very small
-static void ggml_dyn_tallocr_free_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, size_t size, const struct ggml_tensor * tensor) {
+static void ggml_dyn_tallocr_free_tensor(struct ggml_dyn_tallocr * alloc, struct buffer_address addr, size_t size, const struct ggml_tensor * tensor) {
    size = aligned_offset(NULL, size, alloc->alignment);

-    AT_PRINTF("%s: freeing %s at %zu (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, offset, size, alloc->n_free_blocks);
+    AT_PRINTF("%s: freeing %s at {chunk=%d, offset=%zu} (%zu bytes) - n_free_blocks = %d\n",
+        __func__, tensor->name, addr.chunk, addr.offset, size, alloc->chunks[addr.chunk]->n_free_blocks);

 #ifdef GGML_ALLOCATOR_DEBUG
-    remove_allocated_tensor(alloc, offset, tensor);
+    remove_allocated_tensor(alloc, addr, tensor);
 #endif

+    struct tallocr_chunk * chunk = alloc->chunks[addr.chunk];
+
    // see if we can merge with an existing block
-    for (int i = 0; i < alloc->n_free_blocks; i++) {
-        struct free_block * block = &alloc->free_blocks[i];
+    for (int i = 0; i < chunk->n_free_blocks; i++) {
+        struct free_block * block = &chunk->free_blocks[i];
        // check if ptr is at the end of the block
-        if (block->offset + block->size == offset) {
+        if (block->offset + block->size == addr.offset) {
            block->size += size;
            // check if we can merge with the next block
-            if (i < alloc->n_free_blocks - 1 && block->offset + block->size == alloc->free_blocks[i+1].offset) {
-                block->size += alloc->free_blocks[i+1].size;
-                alloc->n_free_blocks--;
-                for (int j = i+1; j < alloc->n_free_blocks; j++) {
-                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+            if (i < chunk->n_free_blocks - 1) {
+                struct free_block * next = &chunk->free_blocks[i+1];
+                if (block->offset + block->size == next->offset) {
+                    block->size += next->size;
+                    ggml_dyn_tallocr_remove_block(chunk, i+1);
                }
            }
            return;
        }
        // check if ptr is at the beginning of the block
-        if (offset + size == block->offset) {
-            block->offset = offset;
+        if (addr.offset + size == block->offset) {
+            block->offset = addr.offset;
            block->size += size;
            // check if we can merge with the previous block
-            if (i > 0 && alloc->free_blocks[i-1].offset + alloc->free_blocks[i-1].size == block->offset) {
-                alloc->free_blocks[i-1].size += block->size;
-                alloc->n_free_blocks--;
-                for (int j = i; j < alloc->n_free_blocks; j++) {
-                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+            if (i > 0) {
+                struct free_block * prev = &chunk->free_blocks[i-1];
+                if (prev->offset + prev->size == block->offset) {
+                    prev->size += block->size;
+                    ggml_dyn_tallocr_remove_block(chunk, i);
                }
            }
            return;
        }
    }
    // otherwise, add a new block
-    GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
-    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
-    int insert_pos = 0;
-    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].offset < offset) {
-        insert_pos++;
-    }
-    // shift all blocks from insert_pos onward to make room for the new block
-    for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
-        alloc->free_blocks[i] = alloc->free_blocks[i-1];
-    }
-    // insert the new block
-    alloc->free_blocks[insert_pos].offset = offset;
-    alloc->free_blocks[insert_pos].size = size;
-    alloc->n_free_blocks++;
+    ggml_dyn_tallocr_insert_block(chunk, addr.offset, size);

    GGML_UNUSED(tensor);
 }

 static void ggml_dyn_tallocr_reset(struct ggml_dyn_tallocr * alloc) {
-    alloc->n_free_blocks = 1;
-    alloc->free_blocks[0].offset = 0;
-    alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
-    alloc->max_size = 0;
+    for (int i = 0; i < GGML_VBUFFER_MAX_CHUNKS; i++) {
+        free(alloc->chunks[i]);
+        alloc->chunks[i] = NULL;
+    }
+    alloc->n_chunks = 0;

 #ifdef GGML_ALLOCATOR_DEBUG
    for (int i = 0; i < 1024; i++) {
@@ -293,14 +367,14 @@ static void ggml_dyn_tallocr_reset(struct ggml_dyn_tallocr * alloc) {
 #endif
 }

-static struct ggml_dyn_tallocr * ggml_dyn_tallocr_new(size_t alignment) {
+static struct ggml_dyn_tallocr * ggml_dyn_tallocr_new(size_t alignment, size_t max_buffer_size) {
    struct ggml_dyn_tallocr * alloc = (struct ggml_dyn_tallocr *)malloc(sizeof(struct ggml_dyn_tallocr));

    *alloc = (struct ggml_dyn_tallocr) {
-        /*.alignment     = */ alignment,
-        /*.n_free_blocks = */ 0,
-        /*.free_blocks   = */ {{0}},
-        /*.max_size      = */ 0,
+        /*.alignment      = */ alignment,
+        /*.max_chunk_size = */ MIN(max_buffer_size, SIZE_MAX/2), // clamp to avoid overflows
+        /*.chunks         = */ {NULL},
+        /*.n_chunks       = */ 0,
 #ifdef GGML_ALLOCATOR_DEBUG
        /*.allocated_tensors = */ {{0}},
 #endif
@@ -312,11 +386,79 @@ static struct ggml_dyn_tallocr * ggml_dyn_tallocr_new(size_t alignment) {
 }

 static void ggml_dyn_tallocr_free(struct ggml_dyn_tallocr * alloc) {
+    for (int i = 0; i < alloc->n_chunks; ++i) {
+        free(alloc->chunks[i]);
+    }
    free(alloc);
 }

 static size_t ggml_dyn_tallocr_max_size(struct ggml_dyn_tallocr * alloc) {
-    return alloc->max_size;
+    size_t max_size = 0;
+    for (int i = 0; i < alloc->n_chunks; i++) {
+        max_size += alloc->chunks[i]->max_size;
+    }
+    return max_size;
+}
+
+
+// virtual buffer with contiguous memory range, split into multiple backend buffers (chunks)
+
+struct vbuffer {
+    ggml_backend_buffer_t chunks[GGML_VBUFFER_MAX_CHUNKS];
+};
+
+static void ggml_vbuffer_free(struct vbuffer * buf) {
+    if (buf == NULL) {
+        return;
+    }
+    for (int i = 0; i < GGML_VBUFFER_MAX_CHUNKS; ++i) {
+        ggml_backend_buffer_free(buf->chunks[i]);
+    }
+    free(buf);
+}
+
+static int ggml_vbuffer_n_chunks(struct vbuffer * buf) {
+    int n = 0;
+    while (n < GGML_VBUFFER_MAX_CHUNKS && buf->chunks[n]) n++;
+    return n;
+}
+
+static size_t ggml_vbuffer_size(struct vbuffer * buf) {
+    size_t size = 0;
+    for (int i = 0; i < GGML_VBUFFER_MAX_CHUNKS && buf->chunks[i]; ++i) {
+        size += ggml_backend_buffer_get_size(buf->chunks[i]);
+    }
+    return size;
+}
+
+static struct vbuffer * ggml_vbuffer_alloc(ggml_backend_buffer_type_t buft, const struct ggml_dyn_tallocr * talloc, enum ggml_backend_buffer_usage usage) {
+    struct vbuffer * buf = (struct vbuffer *)calloc(1, sizeof(struct vbuffer));
+    if (buf == NULL) {
+        return NULL;
+    }
+
+    for (int n = 0; n < talloc->n_chunks; n++) {
+        size_t chunk_size = talloc->chunks[n]->max_size;
+        buf->chunks[n] = ggml_backend_buft_alloc_buffer(buft, chunk_size);
+        if (buf->chunks[n] == NULL) {
+            ggml_vbuffer_free(buf);
+            return NULL;
+        }
+        ggml_backend_buffer_set_usage(buf->chunks[n], usage);
+    }
+    return buf;
+}
+
+static void ggml_vbuffer_tensor_alloc(struct vbuffer * buf, struct ggml_tensor * tensor, struct buffer_address buf_addr) {
+    void * base = ggml_backend_buffer_get_base(buf->chunks[buf_addr.chunk]);
+    void * addr = (char *)base + buf_addr.offset;
+    ggml_backend_tensor_alloc(buf->chunks[buf_addr.chunk], tensor, addr);
+}
+
+static void ggml_vbuffer_reset(struct vbuffer * buf) {
+    for (int i = 0; i < GGML_VBUFFER_MAX_CHUNKS && buf->chunks[i]; ++i) {
+        ggml_backend_buffer_reset(buf->chunks[i]);
+    }
 }


@@ -328,13 +470,13 @@ struct hash_node {
    int n_children;
    int n_views;
    int buffer_id;
-    size_t offset; // offset within the buffer
+    struct buffer_address addr;
    bool allocated;
 };

 struct tensor_alloc {
    int buffer_id;
-    size_t offset;
+    struct buffer_address addr;
    size_t size_max; // 0 = pre-allocated, unused, or view
 };

@@ -349,7 +491,7 @@ struct node_alloc {

 struct ggml_gallocr {
    ggml_backend_buffer_type_t * bufts; // [n_buffers]
-    ggml_backend_buffer_t * buffers; // [n_buffers]
+    struct vbuffer ** buffers; // [n_buffers]
    struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
    int n_buffers;

@@ -370,7 +512,7 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
    galloc->bufts = calloc(n_bufs, sizeof(ggml_backend_buffer_type_t));
    GGML_ASSERT(galloc->bufts != NULL);

-    galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t));
+    galloc->buffers = calloc(n_bufs, sizeof(struct vbuffer *));
    GGML_ASSERT(galloc->buffers != NULL);

    galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
@@ -390,7 +532,8 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs

        if (galloc->buf_tallocs[i] == NULL) {
            size_t alignment = ggml_backend_buft_get_alignment(bufts[i]);
-            galloc->buf_tallocs[i] = ggml_dyn_tallocr_new(alignment);
+            size_t max_size = ggml_backend_buft_get_max_size(bufts[i]);
+            galloc->buf_tallocs[i] = ggml_dyn_tallocr_new(alignment, max_size);
        }
    }
    galloc->n_buffers = n_bufs;
@@ -418,7 +561,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
                }
            }
            if (!freed) {
-                ggml_backend_buffer_free(galloc->buffers[i]);
+                ggml_vbuffer_free(galloc->buffers[i]);
            }
        }
        if (galloc->buf_tallocs != NULL) {
@@ -467,7 +610,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor

    if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
        hn->allocated = true;
-        assert(hn->offset == 0);
+        assert(hn->addr.offset == 0);

        // try to reuse a parent's buffer (inplace)
        if (ggml_op_can_inplace(node->op)) {
@@ -501,9 +644,9 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
                        struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
                        if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
                            AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
-                            assert(view_src_hn->offset == p_hn->offset);
+                            assert(view_src_hn->addr.chunk == p_hn->addr.chunk && view_src_hn->addr.offset == p_hn->addr.offset);
                            hn->buffer_id = p_hn->buffer_id;
-                            hn->offset = p_hn->offset;
+                            hn->addr = p_hn->addr;
                            p_hn->allocated = false; // avoid freeing the parent
                            view_src_hn->allocated = false;
                            return;
@@ -511,7 +654,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
                    } else {
                        AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
                        hn->buffer_id = p_hn->buffer_id;
-                        hn->offset = p_hn->offset;
+                        hn->addr = p_hn->addr;
                        p_hn->allocated = false; // avoid freeing the parent
                        return;
                    }
@@ -522,9 +665,8 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
        struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
        ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
        size_t size = ggml_backend_buft_get_alloc_size(buft, node);
-        size_t offset = ggml_dyn_tallocr_alloc(alloc, size, node);
        hn->buffer_id = buffer_id;
-        hn->offset = offset;
+        hn->addr = ggml_dyn_tallocr_alloc(alloc, size, node);
    }
 }

@@ -536,12 +678,11 @@ static void ggml_gallocr_free_node(ggml_gallocr_t galloc, struct ggml_tensor * n
    }

    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
-    size_t offset = hn->offset;
    int buffer_id = hn->buffer_id;
    struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
    ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
    size_t size = ggml_backend_buft_get_alloc_size(buft, node);
-    ggml_dyn_tallocr_free_tensor(alloc, offset, size, node);
+    ggml_dyn_tallocr_free_tensor(alloc, hn->addr, size, node);
    hn->allocated = false;
 }

@@ -692,24 +833,24 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
        struct node_alloc * node_alloc = &galloc->node_allocs[i];
        if (node->view_src || node->data) {
            node_alloc->dst.buffer_id = -1;
-            node_alloc->dst.offset = SIZE_MAX;
+            node_alloc->dst.addr = GGML_BUFFER_ADDRESS_INVALID;
            node_alloc->dst.size_max = 0;
        } else {
            struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
            node_alloc->dst.buffer_id = hn->buffer_id;
-            node_alloc->dst.offset    = hn->offset;
+            node_alloc->dst.addr = hn->addr;
            node_alloc->dst.size_max  = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
        }
        for (int j = 0; j < GGML_MAX_SRC; j++) {
            struct ggml_tensor * src = node->src[j];
            if (!src || src->view_src || src->data) {
                node_alloc->src[j].buffer_id = -1;
-                node_alloc->src[j].offset = SIZE_MAX;
+                node_alloc->src[j].addr = GGML_BUFFER_ADDRESS_INVALID;
                node_alloc->src[j].size_max = 0;
            } else {
                struct hash_node * hn = ggml_gallocr_hash_get(galloc, src);
                node_alloc->src[j].buffer_id = hn->buffer_id;
-                node_alloc->src[j].offset   = hn->offset;
+                node_alloc->src[j].addr = hn->addr;
                node_alloc->src[j].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], src);
            }
        }
@@ -725,11 +866,11 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
        if (leaf->view_src || leaf->data) {
            galloc->leaf_allocs[i].leaf.buffer_id = -1;
-            galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
+            galloc->leaf_allocs[i].leaf.addr = GGML_BUFFER_ADDRESS_INVALID;
            galloc->leaf_allocs[i].leaf.size_max = 0;
        } else {
            galloc->leaf_allocs[i].leaf.buffer_id = hn->buffer_id;
-            galloc->leaf_allocs[i].leaf.offset = hn->offset;
+            galloc->leaf_allocs[i].leaf.addr = hn->addr;
            galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
        }
    }
@@ -744,7 +885,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
            }
        }

-        size_t cur_size = galloc->buffers[i] ? ggml_backend_buffer_get_size(galloc->buffers[i]) : 0;
+        size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
        size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);

        // even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
@@ -753,13 +894,12 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
            GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif

-            ggml_backend_buffer_free(galloc->buffers[i]);
-            galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
+            ggml_vbuffer_free(galloc->buffers[i]);
+            galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
            if (galloc->buffers[i] == NULL) {
                GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
                return false;
            }
-            ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
        }
    }

@@ -772,11 +912,11 @@ bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {

 static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * tensor, struct tensor_alloc * tensor_alloc) {
    int buffer_id = tensor_alloc->buffer_id;
-    assert(tensor->data || tensor->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
+    assert(tensor->data || tensor->view_src || ggml_backend_buft_get_alloc_size(galloc->bufts[buffer_id], tensor) <= tensor_alloc->size_max);

    if (tensor->view_src != NULL) {
        if (tensor->buffer == NULL) {
-            assert(tensor_alloc->offset == SIZE_MAX);
+            assert(tensor_alloc->addr.offset == SIZE_MAX);
            if (tensor->view_src->buffer == NULL) {
                // this tensor was allocated without ggml-backend
                return;
@@ -785,11 +925,9 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
        }
    } else {
        if (tensor->data == NULL) {
-            assert(tensor_alloc->offset != SIZE_MAX);
-            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
-            void * base = ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
-            void * addr = (char *)base + tensor_alloc->offset;
-            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], tensor, addr);
+            assert(tensor_alloc->addr.offset != SIZE_MAX);
+            assert(ggml_backend_buft_get_alloc_size(galloc->bufts[buffer_id], tensor) <= tensor_alloc->size_max);
+            ggml_vbuffer_tensor_alloc(galloc->buffers[buffer_id], tensor, tensor_alloc->addr);
        } else {
            if (tensor->buffer == NULL) {
                // this tensor was allocated without ggml-backend
@@ -874,7 +1012,7 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
    // reset buffers
    for (int i = 0; i < galloc->n_buffers; i++) {
        if (galloc->buffers[i] != NULL) {
-            ggml_backend_buffer_reset(galloc->buffers[i]);
+            ggml_vbuffer_reset(galloc->buffers[i]);
        }
    }

@@ -917,7 +1055,7 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
        }
    }

-    return ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
+    return ggml_vbuffer_size(galloc->buffers[buffer_id]);
 }

 // utils
@@ -116,7 +116,7 @@ extern "C" {
        void (*event_wait)  (ggml_backend_t backend, ggml_backend_event_t event);

        // (optional) sort/optimize the nodes in the graph
-        void                      (*optimize_graph)    (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        void                      (*graph_optimize)    (ggml_backend_t backend, struct ggml_cgraph * cgraph);
    };

    struct ggml_backend {
@@ -463,10 +463,10 @@ void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event)
    backend->iface.event_wait(backend, event);
 }

-static void ggml_backend_optimize_graph(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+static void ggml_backend_graph_optimize(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
    GGML_ASSERT(backend);
-    if (backend->iface.optimize_graph != NULL) {
-        backend->iface.optimize_graph(backend, cgraph);
+    if (backend->iface.graph_optimize != NULL) {
+        backend->iface.graph_optimize(backend, cgraph);
    }
 }

@@ -1307,7 +1307,7 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra

        // Optimize this split of the graph. This needs to happen before we make graph_copy,
        // so they are in sync.
-        ggml_backend_optimize_graph(sched->backends[split->backend_id], &split->graph);
+        ggml_backend_graph_optimize(sched->backends[split->backend_id], &split->graph);

        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
        for (int j = 0; j < split->n_inputs; j++) {
@@ -1793,6 +1793,14 @@ ggml_backend_t ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i)
    return sched->backends[i];
 }

+ggml_backend_buffer_type_t ggml_backend_sched_get_buffer_type(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    GGML_ASSERT(sched);
+    int backend_index = ggml_backend_sched_backend_id(sched, backend);
+    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+
+    return sched->bufts[backend_index];
+}
+
 size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
    GGML_ASSERT(sched);
    int backend_index = ggml_backend_sched_backend_id(sched, backend);
@@ -270,7 +270,7 @@ static struct ggml_backend_i blas_backend_i = {
    /* .graph_compute           = */ ggml_backend_blas_graph_compute,
    /* .event_record            = */ NULL,
    /* .event_wait              = */ NULL,
-    /* .optimize_graph          = */ NULL,
+    /* .graph_optimize          = */ NULL,
 };

 static ggml_guid_t ggml_backend_blas_guid(void) {
@@ -1728,7 +1728,6 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
            ggml_cann_get_rows(ctx, dst);
            break;
        case GGML_OP_SET_ROWS:
-            std::cout << "lcg GGML_OP_SET_ROWS"<< std::endl;
            ggml_cann_set_rows(ctx, dst);
            break;
        case GGML_OP_DUP:
@@ -2757,7 +2756,7 @@ static const ggml_backend_i ggml_backend_cann_interface = {
    /* .graph_compute           = */ ggml_backend_cann_graph_compute,
    /* .event_record            = */ ggml_backend_cann_event_record,
    /* .event_wait              = */ ggml_backend_cann_event_wait,
-    /* .optimize_graph          = */ NULL,
+    /* .graph_optimize          = */ NULL,
 };

 /**
@@ -7,7 +7,7 @@
 #include "ggml-cpu.h"
 #include "traits.h"

-#if defined(__gnu_linux__)
+#if defined(__linux__)
 #include <sys/syscall.h>
 #include <unistd.h>
 #endif
@@ -186,7 +186,7 @@ static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_ty
 #define XFEATURE_XTILEDATA      18

 static bool ggml_amx_init() {
-#if defined(__gnu_linux__)
+#if defined(__linux__)
    if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) {
        fprintf(stderr, "AMX is not ready to be used!\n");
        return false;
@@ -194,6 +194,8 @@ static bool ggml_amx_init() {
    return true;
 #elif defined(_WIN32)
    return true;
+#else
+    return false;
 #endif
 }

@@ -878,7 +878,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
                const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));

-                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
+                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
                const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
                const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
                const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
@@ -1231,7 +1231,7 @@ static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t
                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));

-                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
+                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
                const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
                const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
                const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
@@ -28,6 +28,14 @@ static inline float bf16_to_f32(ggml_bf16_t x) {
    return GGML_BF16_TO_FP32(x);
 }

+static inline float i32_to_f32(int32_t x) {
+    return x;
+}
+
+static inline int32_t f32_to_i32(float x) {
+    return x;
+}
+
 static inline float f32_to_f32(float x) {
    return x;
 }
@@ -54,6 +62,12 @@ struct type_conversion_table<ggml_bf16_t> {
    static constexpr ggml_bf16_t (*from_f32)(float) = f32_to_bf16;
 };

+template <>
+struct type_conversion_table<int32_t> {
+    static constexpr float (*to_f32)(int32_t) = i32_to_f32;
+    static constexpr int32_t (*from_f32)(float) = f32_to_i32;
+};
+
 static std::pair<int64_t, int64_t> get_thread_range(const struct ggml_compute_params * params, const struct ggml_tensor * src0) {
    const int64_t ith = params->ith;
    const int64_t nth = params->nth;
@@ -473,10 +473,10 @@ struct ggml_threadpool {
 struct ggml_compute_state {
 #ifndef GGML_USE_OPENMP
    ggml_thread_t thrd;
-    bool cpumask[GGML_MAX_N_THREADS];
    int  last_graph;
    bool pending;
 #endif
+    bool cpumask[GGML_MAX_N_THREADS];
    struct ggml_threadpool * threadpool;
    int ith;
 };
@@ -3081,7 +3081,14 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(

    threadpool->workers = workers;

-#ifndef GGML_USE_OPENMP
+#ifdef GGML_USE_OPENMP
+    int32_t cpumask_iter = 0;
+
+    // Compute CPU masks for each thread
+    for (int j = 0; j < tpp->n_threads; j++) {
+        ggml_thread_cpumask_next(tpp->cpumask, workers[j].cpumask, tpp->strict_cpu, &cpumask_iter);
+    }
+#else // GGML_USE_OPENMP
    ggml_mutex_init(&threadpool->mutex);
    ggml_cond_init(&threadpool->cond);

@@ -3154,7 +3161,14 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
                atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
            }

-            ggml_graph_compute_thread(&threadpool->workers[omp_get_thread_num()]);
+            // Apply thread CPU mask and priority
+            int ith = omp_get_thread_num();
+
+            ggml_thread_apply_priority(threadpool->prio);
+            if (ggml_thread_cpumask_is_valid(threadpool->workers[ith].cpumask)) {
+                ggml_thread_apply_affinity(threadpool->workers[ith].cpumask);
+            }
+            ggml_graph_compute_thread(&threadpool->workers[ith]);
        }
    } else {
        atomic_store_explicit(&threadpool->n_threads_cur, 1, memory_order_relaxed);
@@ -190,7 +190,7 @@ static const struct ggml_backend_i ggml_backend_cpu_i = {
    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
    /* .event_record            = */ NULL,
    /* .event_wait              = */ NULL,
-    /* .optimize_graph          = */ NULL,
+    /* .graph_optimize          = */ NULL,
 };

 static ggml_guid_t ggml_backend_cpu_guid(void) {
@@ -25,10 +25,14 @@ if (CUDAToolkit_FOUND)
        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
            set(CMAKE_CUDA_ARCHITECTURES "native")
        else()
+            if (CUDAToolkit_VERSION VERSION_LESS "13")
+                list(APPEND CMAKE_CUDA_ARCHITECTURES 50-virtual 61-virtual 70-virtual)
+            endif ()
+
+            list(APPEND CMAKE_CUDA_ARCHITECTURES 75-virtual 80-virtual 86-real)
+
            if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
-                set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
-            else()
-                set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real")
+                list(APPEND CMAKE_CUDA_ARCHITECTURES 89-real)
            endif()
        endif()
    endif()
@@ -652,6 +652,14 @@ static __device__ __forceinline__ uint32_t fastmodulo(uint32_t n, const uint3 fa
    return n - fastdiv(n, fastdiv_values) * fastdiv_values.z;
 }

+// Calculate both division and modulo at once, returns <n/divisor, n%divisor>
+static __device__ __forceinline__ uint2 fast_div_modulo(uint32_t n, const uint3 fastdiv_values) {
+    // expects  fastdiv_values to contain <mp, L, divisor> in <x, y, z> (see init_fastdiv_values)
+    const uint32_t div_val = fastdiv(n, fastdiv_values);
+    const uint32_t mod_val = n - div_val * fastdiv_values.z;
+    return make_uint2(div_val, mod_val);
+}
+
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, float2 & v);

 static __device__ __forceinline__ float get_alibi_slope(
@@ -441,6 +441,10 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, nv_bfloat16>>;
    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
        return (void*) cpy_flt<cpy_1_flt<nv_bfloat16, float>>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32) {
+        return (void*) cpy_flt<cpy_1_flt<float, int32_t>>;
+    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_flt<cpy_1_flt<int32_t, float>>;
    } else {
        GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
                ggml_type_name(src0->type), ggml_type_name(src1->type));
@@ -35,7 +35,6 @@ static int fattn_tile_get_kq_stride_host(const int D, const int ncols, const int
        switch (D) {
            case 64:
            case 128:
-                return 128;
            case 256:
                return ncols <= 16 ? 128 : 64;
            default:
@@ -86,7 +85,6 @@ static constexpr __device__ int fattn_tile_get_kq_stride_device(int D, int ncols
    switch (D) {
        case 64:
        case 128:
-            return 128;
        case 256:
            return ncols <= 16 ? 128 : 64;
        default:
@@ -3140,7 +3140,7 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
    /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
    /* .event_record            = */ ggml_backend_cuda_event_record,
    /* .event_wait              = */ ggml_backend_cuda_event_wait,
-    /* .optimize_graph          = */ NULL,
+    /* .graph_optimize          = */ NULL,
 };

 static ggml_guid_t ggml_backend_cuda_guid() {
@@ -3427,7 +3427,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                       op->type == GGML_TYPE_Q4_0 || op->type == GGML_TYPE_Q4_1 || op->type == GGML_TYPE_Q5_0 ||
                       op->type == GGML_TYPE_Q5_1 || op->type == GGML_TYPE_Q8_0 || op->type == GGML_TYPE_IQ4_NL) &&
                       op->src[0]->type == GGML_TYPE_F32 &&
-                       op->src[1]->type == GGML_TYPE_I64;
+                       (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
            } break;
        case GGML_OP_CPY:
            {
@@ -1,82 +1,89 @@
 #include "pad_reflect_1d.cuh"

-static __global__ void pad_reflect_1d_kernel_f32(
-    const void * __restrict__ src0,
-    void * __restrict__ dst,
-    const int64_t ne0,
-    const int64_t ne00,
-    const int64_t ne01,
-    const int64_t ne02,
-    const int64_t ne03,
-    const int64_t nb00,
-    const int64_t nb01,
-    const int64_t nb02,
-    const int64_t nb03,
-    const int64_t nb0,
-    const int64_t nb1,
-    const int64_t nb2,
-    const int64_t nb3,
-    const int p0,
-    const int p1) {
-
+static __global__ __launch_bounds__(CUDA_PAD_REFLECT_1D_BLOCK_SIZE, 1) void
+    pad_reflect_1d_kernel_f32(
+        const void * __restrict__ src0,
+        void * __restrict__       dst,
+        const int64_t             ne0,
+        const int64_t             ne00,
+        const uint3               ne01,
+        const int64_t             ne02,
+        const int64_t             ne03,
+        const int64_t             nb00,
+        const int64_t             nb01,
+        const int64_t             nb02,
+        const int64_t             nb03,
+        const int64_t             nb0,
+        const int64_t             nb1,
+        const int64_t             nb2,
+        const int64_t             nb3,
+        const int                 p0,
+        const int                 p1) {
    const int64_t i3 = blockIdx.z;
    const int64_t i2 = blockIdx.y;
-    const int64_t i1 = blockIdx.x;

-    if (i1 >= ne01 || i2 >= ne02 || i3 >= ne03) {
+    const uint2   div_mod_packed = fast_div_modulo(blockIdx.x, ne01);
+    const int64_t tile1          = div_mod_packed.y;  // i1
+    const int64_t tile0          = div_mod_packed.x;  // nth i0 tile
+    const int64_t i1             = tile1;
+    const int64_t i0             = threadIdx.x + tile0 * blockDim.x;
+
+    // ne01.z is original value of unpacked ne01 (see init_fastdiv_values in common.cuh)
+    if (i0 >= ne0 || i1 >= ne01.z || i2 >= ne02 || i3 >= ne03) {
        return;
    }

-    const char * src0_ptr = (const char *)src0 + i3*nb03 + i2*nb02 + i1*nb01;
-    char * dst_ptr = (char *)dst + i3*nb3 + i2*nb2 + i1*nb1;
+    const char * src0_ptr = (const char *) src0 + i3 * nb03 + i2 * nb02 + i1 * nb01;
+    char *       dst_ptr  = (char *) dst + i3 * nb3 + i2 * nb2 + i1 * nb1;

-    for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
-        float value;
+    const int64_t rel_i0 = i0 - p0;  // relative i0 in src0
+    int64_t src_idx;

-        if (i0 < p0) {
-            // Left padding - reflect
-            value = *(const float *)(src0_ptr + (p0 - i0) * nb00);
-        } else if (i0 < ne0 - p1) {
-            // Middle - copy
-            value = *(const float *)(src0_ptr + (i0 - p0) * nb00);
-        } else {
-            // Right padding - reflect
-            int64_t src_idx = (ne0 - p1 - p0) - (p1 + 1 - (ne0 - i0)) - 1;
-            value = *(const float *)(src0_ptr + src_idx * nb00);
-        }
-
-        *(float *)(dst_ptr + i0 * nb0) = value;
+    if (rel_i0 < 0) {
+        // Left padding - reflect
+        src_idx = -rel_i0;
+    } else if (rel_i0 < ne00) {
+        // Middle - copy
+        src_idx = rel_i0;
+    } else {
+        // Right padding - reflect
+        src_idx = 2 * ne00 - 2 - rel_i0;
    }
+    const float value               = *(const float *) (src0_ptr + src_idx * nb00);
+    *(float *) (dst_ptr + i0 * nb0) = value;
 }

 void ggml_cuda_op_pad_reflect_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    cudaStream_t stream = ctx.stream();
+    const ggml_tensor * src0   = dst->src[0];
+    cudaStream_t        stream = ctx.stream();

    GGML_ASSERT(src0->type == GGML_TYPE_F32);
    GGML_ASSERT(dst->type == GGML_TYPE_F32);

    const int32_t * opts = (const int32_t *) dst->op_params;
-    const int p0 = opts[0];
-    const int p1 = opts[1];
+    const int       p0   = opts[0];
+    const int       p1   = opts[1];

-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
+    const int64_t ne00        = src0->ne[0];
+    const int64_t ne01        = src0->ne[1];
+    const uint3   ne01_packed = init_fastdiv_values(ne01);
+    const int64_t ne02        = src0->ne[2];
+    const int64_t ne03        = src0->ne[3];

    const int64_t ne0 = dst->ne[0];

+    // sanity: padded length matches
    GGML_ASSERT(ne0 == ne00 + p0 + p1);

-    const dim3 block_dims(CUDA_PAD_REFLECT_1D_BLOCK_SIZE, 1, 1);
-    const dim3 grid_dims(ne01, ne02, ne03);
+    constexpr int64_t bx     = CUDA_PAD_REFLECT_1D_BLOCK_SIZE;  // threads per block (x)
+    const int64_t     tiles0 = (ne0 + bx - 1) / bx;             // number of tiles along i0
+    // grid.x covers i1 and all tiles of i0: [ne01 * tiles0]
+    // grid.y covers i2: [ne02]
+    // grid.z covers i3: [ne03]
+    const dim3        grid_dims((unsigned) (ne01 * tiles0), (unsigned) ne02, (unsigned) ne03);
+    const dim3        block_dims((unsigned) bx, 1, 1);

    pad_reflect_1d_kernel_f32<<<grid_dims, block_dims, 0, stream>>>(
-        src0->data, dst->data,
-        ne0, ne00, ne01, ne02, ne03,
-        src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
-        dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
-        p0, p1
-    );
+        src0->data, dst->data, ne0, ne00, ne01_packed, ne02, ne03, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+        dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], p0, p1);
 }
@@ -4,9 +4,9 @@
 typedef void (*set_rows_kernel_t)(const char * src, char * dst);

 // Generic quantized set_rows kernel template
-template<typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
+template<typename idx_t, typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
 static __global__ void k_set_rows_quant(
-        const float * __restrict__ src0, const int64_t * __restrict__ src1, block_type * __restrict__ dst,
+        const float * __restrict__ src0, const idx_t * __restrict__ src1, block_type * __restrict__ dst,
        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
        const int64_t s01, const int64_t s02, const int64_t s03,
@@ -45,9 +45,9 @@ static __global__ void k_set_rows_quant(
 }

 // Template dispatch function for quantized set_rows
-template<typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
+template<typename idx_t, typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
 static void set_rows_cuda_quant(
-        const float * src0_d, const int64_t * src1_d, block_type * dst_d,
+        const float * src0_d, const idx_t * src1_d, block_type * dst_d,
        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
        const size_t nb01, const size_t nb02, const size_t nb03,
@@ -64,15 +64,15 @@ static void set_rows_cuda_quant(
    const int64_t s01 = nb01/sizeof(float);
    const int64_t s02 = nb02/sizeof(float);
    const int64_t s03 = nb03/sizeof(float);
-    const int64_t s10 = nb10/sizeof(int64_t);
-    const int64_t s11 = nb11/sizeof(int64_t);
-    const int64_t s12 = nb12/sizeof(int64_t);
+    const int64_t s10 = nb10/sizeof(idx_t);
+    const int64_t s11 = nb11/sizeof(idx_t);
+    const int64_t s12 = nb12/sizeof(idx_t);
    const int64_t s1  = nb1;
    const int64_t s2  = nb2;
    const int64_t s3  = nb3;

    if (ne_total > 0) {
-        k_set_rows_quant<block_type, qk, quantize_func><<<grid_size, block_size, 0, stream>>>(
+        k_set_rows_quant<idx_t, block_type, qk, quantize_func><<<grid_size, block_size, 0, stream>>>(
            src0_d, src1_d, dst_d,
            ne00, ne01, ne02, ne03,
            ne10, ne11, ne12, ne13,
@@ -82,9 +82,9 @@ static void set_rows_cuda_quant(
    }
 }

-template<typename src_t, typename dst_t>
+template<typename src_t, typename idx_t, typename dst_t>
 static __global__ void k_set_rows(
-        const src_t * __restrict__ src0, const int64_t * __restrict__ src1, dst_t * __restrict__ dst,
+        const src_t * __restrict__ src0, const idx_t * __restrict__ src1, dst_t * __restrict__ dst,
        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
        const int64_t s01, const int64_t s02, const int64_t s03,
@@ -118,9 +118,9 @@ static __global__ void k_set_rows(
    GGML_UNUSED(ne13);
 }

-template<typename src_t, typename dst_t>
+template<typename src_t, typename idx_t, typename dst_t>
 static void set_rows_cuda(
-        const src_t * src0_d, const int64_t * src1_d, dst_t * dst_d,
+        const src_t * src0_d, const idx_t * src1_d, dst_t * dst_d,
        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
        const size_t nb01, const size_t nb02, const size_t nb03,
@@ -137,9 +137,9 @@ static void set_rows_cuda(
    const int64_t s01 = nb01/sizeof(src_t);
    const int64_t s02 = nb02/sizeof(src_t);
    const int64_t s03 = nb03/sizeof(src_t);
-    const int64_t s10 = nb10/sizeof(int64_t);
-    const int64_t s11 = nb11/sizeof(int64_t);
-    const int64_t s12 = nb12/sizeof(int64_t);
+    const int64_t s10 = nb10/sizeof(idx_t);
+    const int64_t s11 = nb11/sizeof(idx_t);
+    const int64_t s12 = nb12/sizeof(idx_t);
    const int64_t s1  = nb1/sizeof(dst_t);
    const int64_t s2  = nb2/sizeof(dst_t);
    const int64_t s3  = nb3/sizeof(dst_t);
@@ -155,23 +155,16 @@ static void set_rows_cuda(
    }
 }

-
-void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_I64);
+template<typename src_t, typename idx_t>
+static void set_rows_cuda(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const src_t * src0_d = (const src_t *)src0->data;
+    const idx_t * src1_d = (const idx_t *)src1->data;

    GGML_TENSOR_BINARY_OP_LOCALS

-    const float * src0_d   = (const float *)src0->data;
-    const int64_t * src1_d = (const int64_t *)src1->data;
-
    cudaStream_t stream = ctx.stream();


-
    if (dst->type == GGML_TYPE_F32) {
        set_rows_cuda(
            src0_d, src1_d, (float*)dst->data,
@@ -203,7 +196,7 @@ void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
            stream
        );
    } else if (dst->type == GGML_TYPE_Q4_0) {
-        set_rows_cuda_quant<block_q4_0, QK4_0, quantize_f32_q4_0_block>(
+        set_rows_cuda_quant<idx_t, block_q4_0, QK4_0, quantize_f32_q4_0_block>(
            src0_d, src1_d, (block_q4_0*)dst->data,
            ne00, ne01, ne02, ne03,
            ne10, ne11, ne12, ne13,
@@ -213,7 +206,7 @@ void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
            stream
        );
    } else if (dst->type == GGML_TYPE_Q4_1) {
-        set_rows_cuda_quant<block_q4_1, QK4_1, quantize_f32_q4_1_block>(
+        set_rows_cuda_quant<idx_t, block_q4_1, QK4_1, quantize_f32_q4_1_block>(
            src0_d, src1_d, (block_q4_1*)dst->data,
            ne00, ne01, ne02, ne03,
            ne10, ne11, ne12, ne13,
@@ -223,7 +216,7 @@ void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
            stream
        );
    } else if (dst->type == GGML_TYPE_Q5_0) {
-        set_rows_cuda_quant<block_q5_0, QK5_0, quantize_f32_q5_0_block>(
+        set_rows_cuda_quant<idx_t, block_q5_0, QK5_0, quantize_f32_q5_0_block>(
            src0_d, src1_d, (block_q5_0*)dst->data,
            ne00, ne01, ne02, ne03,
            ne10, ne11, ne12, ne13,
@@ -233,7 +226,7 @@ void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
            stream
        );
    } else if (dst->type == GGML_TYPE_Q5_1) {
-        set_rows_cuda_quant<block_q5_1, QK5_1, quantize_f32_q5_1_block>(
+        set_rows_cuda_quant<idx_t, block_q5_1, QK5_1, quantize_f32_q5_1_block>(
            src0_d, src1_d, (block_q5_1*)dst->data,
            ne00, ne01, ne02, ne03,
            ne10, ne11, ne12, ne13,
@@ -243,7 +236,7 @@ void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
            stream
        );
    } else if (dst->type == GGML_TYPE_Q8_0) {
-        set_rows_cuda_quant<block_q8_0, QK8_0, quantize_f32_q8_0_block>(
+        set_rows_cuda_quant<idx_t, block_q8_0, QK8_0, quantize_f32_q8_0_block>(
            src0_d, src1_d, (block_q8_0*)dst->data,
            ne00, ne01, ne02, ne03,
            ne10, ne11, ne12, ne13,
@@ -253,7 +246,7 @@ void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
            stream
        );
    } else if (dst->type == GGML_TYPE_IQ4_NL) {
-        set_rows_cuda_quant<block_iq4_nl, QK4_NL, quantize_f32_iq4_nl_block>(
+        set_rows_cuda_quant<idx_t, block_iq4_nl, QK4_NL, quantize_f32_iq4_nl_block>(
            src0_d, src1_d, (block_iq4_nl*)dst->data,
            ne00, ne01, ne02, ne03,
            ne10, ne11, ne12, ne13,
@@ -266,3 +259,18 @@ void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
        GGML_ABORT("unsupported type %s", ggml_type_name(dst->type));
    }
 }
+
+
+void ggml_cuda_op_set_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_I64 || src1->type == GGML_TYPE_I32);
+
+    if (src1->type == GGML_TYPE_I64) {
+        set_rows_cuda<float, int64_t>(ctx, src0, src1, dst);
+    } else {
+        set_rows_cuda<float, int32_t>(ctx, src0, src1, dst);
+    }
+}
@@ -73,7 +73,7 @@ static inline int ggml_up(int n, int m) {
    return (n + m - 1) & ~(m - 1);
 }

-// TODO: move to ggml.h?
+// TODO: move to ggml.h? (won't be able to inline)
 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
    if (a->type != b->type) {
        return false;
@@ -89,6 +89,19 @@ static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml
    return true;
 }

+static bool ggml_op_is_empty(enum ggml_op op) {
+    switch (op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+            return true;
+        default:
+            return false;
+    }
+}
+
 //
 // logging
 //
@@ -329,6 +342,10 @@ struct ggml_cgraph {
 // if you need the gradients, get them from the original graph
 struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);

+// ggml-alloc.c: true if the operation can reuse memory from its sources
+GGML_API bool ggml_op_can_inplace(enum ggml_op op);
+
+
 // Memory allocation

 GGML_API void * ggml_aligned_malloc(size_t size);
@@ -570,27 +587,27 @@ static inline bool ggml_node_has_n_uses(const struct ggml_cgraph * cgraph, int n
    return true;
 }

-// Returns true if nodes [i, i+ops.size()) are the sequence of ggml_ops in ops[]
+// Returns true if nodes with indices { node_idxs } are the sequence of ggml_ops in ops[]
 // and are fusable. Nodes are considered fusable according to this function if:
 // - all nodes except the last have only one use and are not views/outputs (see ggml_node_has_N_uses).
 // - all nodes except the last are a src of the following node.
 // - all nodes are the same shape.
 // TODO: Consider allowing GGML_OP_NONE nodes in between
-static inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, const enum ggml_op * ops, int num_ops) {
-    if (node_idx + num_ops > cgraph->n_nodes) {
-        return false;
-    }
-
+static inline bool ggml_can_fuse_ext(const struct ggml_cgraph * cgraph, const int * node_idxs, const enum ggml_op * ops, int num_ops) {
    for (int i = 0; i < num_ops; ++i) {
-        struct ggml_tensor * node = cgraph->nodes[node_idx + i];
+        if (node_idxs[i] >= cgraph->n_nodes) {
+            return false;
+        }
+
+        struct ggml_tensor * node = cgraph->nodes[node_idxs[i]];
        if (node->op != ops[i]) {
            return false;
        }
-        if (i < num_ops - 1 && !ggml_node_has_n_uses(cgraph, node_idx + i, 1)) {
+        if (i < num_ops - 1 && !ggml_node_has_n_uses(cgraph, node_idxs[i], 1)) {
            return false;
        }
        if (i > 0) {
-            struct ggml_tensor * prev = cgraph->nodes[node_idx + i - 1];
+            struct ggml_tensor * prev = cgraph->nodes[node_idxs[i - 1]];
            if (node->src[0] != prev && node->src[1] != prev) {
                return false;
            }
@@ -602,6 +619,22 @@ static inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx
    return true;
 }

+// same as above, for sequential indices starting at node_idx
+static inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, const enum ggml_op * ops, int num_ops) {
+    assert(num_ops < 32);
+
+    if (node_idx + num_ops > cgraph->n_nodes) {
+        return false;
+    }
+
+    int idxs[32];
+    for (int i = 0; i < num_ops; ++i) {
+        idxs[i] = node_idx + i;
+    }
+
+    return ggml_can_fuse_ext(cgraph, idxs, ops, num_ops);
+}
+
 #ifdef __cplusplus
 }
 #endif
@@ -184,20 +184,6 @@ bool ggml_mem_ranges_check(ggml_mem_ranges_t mrs, const ggml_tensor * tensor) {
    return ggml_mem_ranges_check_dst(mrs, tensor);
 }

-// TODO: move to ggml.h?
-static bool is_empty(ggml_op op) {
-    switch (op) {
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_VIEW:
-        case GGML_OP_PERMUTE:
-            return true;
-        default:
-            return false;
-    }
-}
-
 struct node_info {
    ggml_tensor * node;

@@ -212,7 +198,7 @@ struct node_info {
    }

    bool is_empty() const {
-        return ::is_empty(node->op);
+        return ggml_op_is_empty(node->op);
    }

    void add_fused(ggml_tensor * t) {
@@ -289,7 +275,7 @@ static std::vector<int> ggml_metal_graph_optimize_reorder(const std::vector<node
            case GGML_OP_GET_ROWS:
                return true;
            default:
-                return is_empty(op);
+                return ggml_op_is_empty(op);
        }
    };

@@ -34,6 +34,10 @@ ggml_metal_pipelines_t ggml_metal_pipelines_init(void) {
 }

 void ggml_metal_pipelines_free(ggml_metal_pipelines_t ppls) {
+    if (!ppls) {
+        return;
+    }
+
    for (auto it = ppls->data.begin(); it != ppls->data.end(); ++it) {
        ggml_metal_pipeline_free(it->second);
    }
@@ -138,11 +142,11 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_get_rows(ggml_metal_librar
    return res;
 }

-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_set_rows(ggml_metal_library_t lib, ggml_type tdst) {
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_set_rows(ggml_metal_library_t lib, ggml_type tidx, ggml_type tdst) {
    char base[256];
    char name[256];

-    snprintf(base, 256, "kernel_set_rows_%s", ggml_type_name(tdst));
+    snprintf(base, 256, "kernel_set_rows_%s_%s", ggml_type_name(tdst), ggml_type_name(tidx));
    snprintf(name, 256, "%s", base);

    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
@@ -410,19 +414,26 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rwkv(ggml_metal_library_t
    return res;
 }

-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_ext(ggml_metal_library_t lib, ggml_type tsrc0, ggml_type tsrc1, int r1ptg) {
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_ext(ggml_metal_library_t lib, ggml_type tsrc0, ggml_type tsrc1, int nsg, int nxpsg, int r1ptg) {
    char base[256];
    char name[256];

    snprintf(base, 256, "kernel_mul_mv_ext_%s_%s_r1_%d", ggml_type_name(tsrc0), ggml_type_name(tsrc1), r1ptg);
-    snprintf(name, 256, "%s", base);
+    snprintf(name, 256, "%s_nsg=%d_nxpsg=%d", base, nsg, nxpsg);

    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
    if (res) {
        return res;
    }

-    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+    ggml_metal_cv_set_int16(cv, nsg,   FC_MUL_MV + 0);
+    ggml_metal_cv_set_int16(cv, nxpsg, FC_MUL_MV + 1);
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+    ggml_metal_cv_free(cv);

    return res;
 }
@@ -467,37 +478,25 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv(ggml_metal_library_
    // use custom matrix x vector kernel
    switch (tsrc0) {
        case GGML_TYPE_F32:
-            {
-                GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
-
-                nsg = 1;
-                nr0 = 1;
-                nr1 = 4;
-                if (ne00 == 4) {
-                    nr0 = 32;
-                    suffix = "_c4";
-                }
-            } break;
        case GGML_TYPE_F16:
        case GGML_TYPE_BF16:
            {
-                nsg = 1;
-                nr0 = 1;
-                if (op->src[1]->type == GGML_TYPE_F32) {
-                    if (ne00 == 4) {
-                        nr0 = 32;
-                        nr1 = 4;
-                        suffix = "_c4";
-                    } else if (ne11 * ne12 < 4) {
-                        suffix = "_1row";
-                    } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
-                        suffix = "_l4";
-                        nr1 = ne11;
-                    } else {
-                        nr1 = 4;
-                    }
-                } else {
+                if (ne00 == 4) {
+                    nsg = 1;
+                    nr0 = 32;
                    nr1 = 4;
+                    suffix = "_c4";
+                } else if (ne00 % 4 == 0) {
+                    nsg = N_SG_F;
+                    nr0 = N_R0_F;
+                    nr1 = 1;
+                    smem = 32*sizeof(float)*N_R0_F;
+                    suffix = "_4";
+                } else {
+                    nsg = N_SG_F;
+                    nr0 = N_R0_F;
+                    nr1 = 1;
+                    smem = 32*sizeof(float)*N_R0_F;
                }
            } break;
        case GGML_TYPE_Q4_0:
@@ -616,14 +615,20 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv(ggml_metal_library_
    };

    snprintf(base, 256, "kernel_mul_mv_%s_%s%s", ggml_type_name(tsrc0), ggml_type_name(tsrc1), suffix);
-    snprintf(name, 256, "%s", base);
+    snprintf(name, 256, "%s_nsg=%d", base, nsg);

    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
    if (res) {
        return res;
    }

-    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+    ggml_metal_cv_set_int16(cv, nsg, FC_MUL_MV + 0);
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+    ggml_metal_cv_free(cv);

    ggml_metal_pipeline_set_nr0 (res, nr0);
    ggml_metal_pipeline_set_nr1 (res, nr1);
@@ -689,25 +694,26 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_id(ggml_metal_libra
    const ggml_type tsrc0 = op->src[0]->type;
    const ggml_type tsrc1 = op->src[1]->type;

+    const char * suffix = "";
+
        // use custom matrix x vector kernel
    switch (tsrc0) {
        case GGML_TYPE_F32:
-            {
-                GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
-                nsg = 1;
-                nr0 = 1;
-            } break;
        case GGML_TYPE_F16:
-            {
-                GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
-                nsg = 1;
-                nr0 = 1;
-            } break;
        case GGML_TYPE_BF16:
            {
-                GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
-                nsg = 1;
-                nr0 = 1;
+                if (ne00 % 4 == 0) {
+                    nsg = N_SG_F;
+                    nr0 = N_R0_F;
+                    nr1 = 1;
+                    smem = 32*sizeof(float)*N_R0_F;
+                    suffix = "_4";
+                } else {
+                    nsg = N_SG_F;
+                    nr0 = N_R0_F;
+                    nr1 = 1;
+                    smem = 32*sizeof(float)*N_R0_F;
+                }
            } break;
        case GGML_TYPE_Q4_0:
            {
@@ -824,15 +830,21 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_id(ggml_metal_libra
            }
    };

-    snprintf(base, 256, "kernel_mul_mv_id_%s_%s", ggml_type_name(tsrc0), ggml_type_name(tsrc1));
-    snprintf(name, 256, "%s", base);
+    snprintf(base, 256, "kernel_mul_mv_id_%s_%s%s", ggml_type_name(tsrc0), ggml_type_name(tsrc1), suffix);
+    snprintf(name, 256, "%s_nsg=%d", base, nsg);

    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
    if (res) {
        return res;
    }

-    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+    ggml_metal_cv_set_int16(cv, nsg, FC_MUL_MV + 0);
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+    ggml_metal_cv_free(cv);

    ggml_metal_pipeline_set_nr0 (res, nr0);
    ggml_metal_pipeline_set_nr1 (res, nr1);
@@ -918,11 +930,8 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
            dk,
            dv);

-    snprintf(name, 256, "kernel_%s_%s_dk%d_dv%d_mask=%d_sinks=%d_bias=%d_scap=%d_ns10=%d_ns20=%d_nsg=%d",
-            "flash_attn_ext",
-            ggml_type_name(op->src[1]->type),
-            dk,
-            dv,
+    snprintf(name, 256, "%s_mask=%d_sinks=%d_bias=%d_scap=%d_ns10=%d_ns20=%d_nsg=%d",
+            base,
            has_mask,
            has_sinks,
            has_bias,
@@ -980,11 +989,8 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
            dk,
            dv);

-    snprintf(name, 256, "kernel_%s_%s_dk%d_dv%d_mask=%d_sink=%d_bias=%d_softcap=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
-            "flash_attn_ext_vec",
-            ggml_type_name(op->src[1]->type),
-            dk,
-            dv,
+    snprintf(name, 256, "%s_mask=%d_sink=%d_bias=%d_softcap=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
+            base,
            has_mask,
            has_sinks,
            has_bias,
@@ -1028,7 +1034,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec_reduce(
    char name[256];

    snprintf(base, 256, "kernel_flash_attn_ext_vec_reduce");
-    snprintf(name, 256, "kernel_flash_attn_ext_vec_reduce_dv=%d_nwg=%d", dv, nwg);
+    snprintf(name, 256, "%s_dv=%d_nwg=%d", base, dv, nwg);

    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
    if (res) {
@@ -22,6 +22,7 @@ typedef struct ggml_metal_cv * ggml_metal_cv_t;
 ggml_metal_cv_t ggml_metal_cv_init(void);
 void ggml_metal_cv_free(ggml_metal_cv_t cv);

+void ggml_metal_cv_set_int16(ggml_metal_cv_t cv, int16_t value, int32_t idx);
 void ggml_metal_cv_set_int32(ggml_metal_cv_t cv, int32_t value, int32_t idx);
 void ggml_metal_cv_set_bool (ggml_metal_cv_t cv, bool    value, int32_t idx);

@@ -104,7 +105,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_base              (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_cpy               (ggml_metal_library_t lib, enum ggml_type tsrc, enum ggml_type tdst);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pool_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op, enum ggml_op_pool op_pool);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_get_rows          (ggml_metal_library_t lib, enum ggml_type tsrc);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_set_rows          (ggml_metal_library_t lib, enum ggml_type tdst);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_set_rows          (ggml_metal_library_t lib, enum ggml_type tidx, enum ggml_type tdst);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_repeat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_unary             (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_glu               (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -113,7 +114,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_soft_max          (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_conv          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rwkv              (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_ext        (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1, int r1ptg);
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_ext        (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1, int nsg, int nxpsg, int r1ptg);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm            (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id_map0    (ggml_metal_library_t lib, int ne02, int ne20);
@@ -51,6 +51,10 @@ void ggml_metal_cv_free(ggml_metal_cv_t cv) {
    free(cv);
 }

+void ggml_metal_cv_set_int16(ggml_metal_cv_t cv, int16_t value, int32_t idx) {
+    [cv->obj setConstantValue:&value type:MTLDataTypeShort atIndex:idx];
+}
+
 void ggml_metal_cv_set_int32(ggml_metal_cv_t cv, int32_t value, int32_t idx) {
    [cv->obj setConstantValue:&value type:MTLDataTypeInt atIndex:idx];
 }
@@ -327,12 +331,19 @@ ggml_metal_pipeline_t ggml_metal_library_compile_pipeline(ggml_metal_library_t l

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

-        id<MTLFunction> mtl_function = [lib->obj newFunctionWithName:base_func constantValues:(cv ? cv->obj : nil) error:&error];
+        id<MTLFunction> mtl_function;
+        if (!cv) {
+            mtl_function = [lib->obj newFunctionWithName:base_func];
+        } else {
+            mtl_function = [lib->obj newFunctionWithName:base_func constantValues:cv->obj error:&error];
+        }
        if (!mtl_function) {
            ggml_critical_section_end();

            GGML_LOG_ERROR("%s: error: failed to compile pipeline: base = '%s', name = '%s'\n", __func__, base, name);
-            GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            if (error) {
+                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            }

            return nil;
        }
@@ -817,6 +828,7 @@ struct ggml_metal_buffer {

    // if false, the Metal buffer data is allocated in private GPU memory and is not shared with the host
    bool is_shared;
+    bool owned;

    // multiple buffers are used only to avoid the maximum buffer size limitation when using mmap
    int n_buffers;
@@ -949,6 +961,7 @@ ggml_metal_buffer_t ggml_metal_buffer_init(ggml_metal_device_t dev, size_t size,
    if (shared) {
        res->all_data = ggml_metal_host_malloc(size_aligned);
        res->is_shared = true;
+        res->owned = true;
    } else {
        // dummy, non-NULL value - we'll populate this after creating the Metal buffer below
        res->all_data = (void *) 0x000000400ULL;
@@ -1007,6 +1020,7 @@ ggml_metal_buffer_t ggml_metal_buffer_map(ggml_metal_device_t dev, void * ptr, s
    res->all_size = size;

    res->is_shared = true;
+    res->owned = false;

    res->n_buffers = 0;

@@ -1100,7 +1114,7 @@ void ggml_metal_buffer_free(ggml_metal_buffer_t buf) {

    ggml_metal_buffer_rset_free(buf);

-    if (buf->is_shared) {
+    if (buf->is_shared && buf->owned) {
 #if TARGET_OS_OSX
        vm_deallocate((vm_map_t)mach_task_self(), (vm_address_t)buf->all_data, buf->all_size);
 #else
@@ -8,6 +8,9 @@
 //
 // TODO: for optimal performance, become function of the device and work size

+#define N_R0_F 2
+#define N_SG_F 4
+
 #define N_R0_Q4_0 4
 #define N_SG_Q4_0 2

@@ -32,13 +35,13 @@
 #define N_R0_Q3_K 2
 #define N_SG_Q3_K 2

-#define N_R0_Q4_K 4
+#define N_R0_Q4_K 2
 #define N_SG_Q4_K 2

 #define N_R0_Q5_K 2
 #define N_SG_Q5_K 2

-#define N_R0_Q6_K 1
+#define N_R0_Q6_K 2
 #define N_SG_Q6_K 2

 #define N_R0_IQ1_S 4
@@ -72,6 +75,7 @@
 #define FC_FLASH_ATTN_EXT              100
 #define FC_FLASH_ATTN_EXT_VEC          200
 #define FC_FLASH_ATTN_EXT_VEC_REDUCE   300
+#define FC_MUL_MV                      400

 // kernel argument structs
 //
@@ -370,9 +374,6 @@ typedef struct {
    int32_t  ne1;
    int16_t  r2;
    int16_t  r3;
-    int16_t  nsg;
-    int16_t  nxpsg;
-    int16_t  r1ptg;
 } ggml_metal_kargs_mul_mv_ext;

 typedef struct {
@@ -892,7 +892,7 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
    GGML_TENSOR_LOCALS(uint32_t, nb,  op,         nb);

-    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_set_rows(lib, op->type);
+    ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_set_rows(lib, op->src[1]->type, op->type);

    const int32_t nk0 = ne0/ggml_blck_size(op->type);

@@ -1444,7 +1444,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
                GGML_ABORT("unsupported ne11");
        };

-        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_mul_mv_ext(lib, op->src[0]->type, op->src[1]->type, r1ptg);
+        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_mul_mv_ext(lib, op->src[0]->type, op->src[1]->type, nsg, nxpsg, r1ptg);

        ggml_metal_kargs_mul_mv_ext args = {
            /*.ne00  =*/ ne00,
@@ -1465,9 +1465,6 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
            /*.ne1   =*/ ne1,
            /*.r2    =*/ r2,
            /*.r3    =*/ r3,
-            /*.nsg   =*/ nsg,
-            /*.nxpsg =*/ nxpsg,
-            /*.r1ptg =*/ r1ptg,
        };

        ggml_metal_encoder_set_pipeline(enc, pipeline);
@@ -1564,7 +1561,10 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {

        ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);

-        if (op->src[0]->type == GGML_TYPE_Q8_0) {
+        if (op->src[0]->type == GGML_TYPE_F32 ||
+            op->src[0]->type == GGML_TYPE_F16 ||
+            op->src[0]->type == GGML_TYPE_BF16 ||
+            op->src[0]->type == GGML_TYPE_Q8_0) {
            ggml_metal_encoder_dispatch_threadgroups(enc, ((ne01 + nr0 - 1)/(nr0)), ((ne11 + nr1 - 1)/nr1), ne12*ne13, 32, nsg, 1);
        } else {
            ggml_metal_encoder_dispatch_threadgroups(enc, ((ne01 + nr0*nsg - 1)/(nr0*nsg)), ((ne11 + nr1 - 1)/nr1), ne12*ne13, 32, nsg, 1);
@@ -1772,7 +1772,10 @@ int ggml_metal_op_mul_mat_id(ggml_metal_op_t ctx, int idx) {

        ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);

-        if (op->src[0]->type == GGML_TYPE_Q8_0) {
+        if (op->src[0]->type == GGML_TYPE_F32 ||
+            op->src[0]->type == GGML_TYPE_F16 ||
+            op->src[0]->type == GGML_TYPE_BF16 ||
+            op->src[0]->type == GGML_TYPE_Q8_0) {
            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nr0 - 1)/(nr0), (_ne1 + nr1 - 1)/nr1, ne123, 32, nsg, 1);
        } else {
            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nr0*nsg - 1)/(nr0*nsg), (_ne1 + nr1 - 1)/nr1, ne123, 32, nsg, 1);
@@ -447,7 +447,7 @@ static ggml_backend_i ggml_backend_metal_i = {
    // https://developer.apple.com/documentation/metal/mtlcommandbuffer#Synchronizing-Passes-with-Events
    /* .event_record            = */ NULL,
    /* .event_wait              = */ NULL,
-    /* .optimize_graph          = */ ggml_backend_metal_graph_optimize,
+    /* .graph_optimize          = */ ggml_backend_metal_graph_optimize,
 };

 static ggml_guid_t ggml_backend_metal_guid(void) {
@@ -82,9 +82,15 @@ set(GGML_OPENCL_KERNELS
    mul_mv_q4_0_f32_1d_8x_flat
    mul_mv_q4_0_f32_1d_16x_flat
    mul_mv_q6_k
+    mul_mv_q8_0_f32
+    mul_mv_q8_0_f32_flat
    mul_mv_mxfp4_f32
+    mul_mv_mxfp4_f32_flat
    mul_mv_id_q4_0_f32_8x_flat
+    mul_mv_id_q8_0_f32
+    mul_mv_id_q8_0_f32_flat
    mul_mv_id_mxfp4_f32
+    mul_mv_id_mxfp4_f32_flat
    mul_mm_f32_f32_l4_lm
    mul_mm_f16_f32_l4_lm
    mul
@@ -367,7 +367,9 @@ struct ggml_backend_opencl_context {
    cl_program program_mul_mv_q4_0_f32_1d_8x_flat;
    cl_program program_mul_mv_q4_0_f32_1d_16x_flat;
    cl_program program_mul_mv_q6_K;
+    cl_program program_mul_mv_q8_0_f32, program_mul_mv_q8_0_f32_flat;
    cl_program program_mul_mv_mxfp4_f32;
+    cl_program program_mul_mv_mxfp4_f32_flat;
    cl_program program_mul_mv_f16_f16;
    cl_program program_mul_mv_f16_f32_1row;
    cl_program program_mul_mv_f16_f32_l4;
@@ -401,7 +403,9 @@ struct ggml_backend_opencl_context {
    cl_program program_conv_2d_f16_f32;
    cl_program program_tsembd;
    cl_program program_mul_mv_id_q4_0_f32_8x_flat;
+    cl_program program_mul_mv_id_q8_0_f32, program_mul_mv_id_q8_0_f32_flat;
    cl_program program_mul_mv_id_mxfp4_f32;
+    cl_program program_mul_mv_id_mxfp4_f32_flat;
    cl_program program_mul_mm_f32_f32_l4_lm;
    cl_program program_mul_mm_f16_f32_l4_lm;

@@ -435,7 +439,7 @@ struct ggml_backend_opencl_context {
    std::map<std::pair<int, int>, int>       kernels_flash_attn_bm;
    std::map<std::pair<int, int>, int>       kernels_flash_attn_bn;
    cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
-    cl_kernel kernel_set_rows_f32, kernel_set_rows_f16;
+    cl_kernel kernel_set_rows_f32_i64, kernel_set_rows_f32_i32, kernel_set_rows_f16_i64, kernel_set_rows_f16_i32;
    cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
    cl_kernel kernel_rope_multi_f32, kernel_rope_multi_f16, kernel_rope_vision_f32, kernel_rope_vision_f16;
    cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32;
@@ -447,11 +451,14 @@ struct ggml_backend_opencl_context {
    cl_kernel kernel_mul_mat_f16_f32_tiled;
    cl_kernel kernel_mul_mat_q4_0_f32, kernel_mul_mat_q4_0_f32_v;
    cl_kernel kernel_convert_block_q4_0, kernel_restore_block_q4_0;
+    cl_kernel kernel_convert_block_mxfp4, kernel_restore_block_mxfp4;
+    cl_kernel kernel_convert_block_q8_0, kernel_restore_block_q8_0;
    cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
    cl_kernel kernel_convert_block_q4_0_noshuffle;
    cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
    cl_kernel kernel_mul_mv_q6_K_f32;
-    cl_kernel kernel_mul_mv_mxfp4_f32;
+    cl_kernel kernel_mul_mv_mxfp4_f32, kernel_mul_mv_mxfp4_f32_flat;
+    cl_kernel kernel_mul_mv_q8_0_f32, kernel_mul_mv_q8_0_f32_flat;
    cl_kernel kernel_im2col_f32, kernel_im2col_f16;
    cl_kernel kernel_argsort_f32_i32;
    cl_kernel kernel_sum_rows_f32;
@@ -468,7 +475,9 @@ struct ggml_backend_opencl_context {
    cl_kernel kernel_conv_2d_f16_f32;
    cl_kernel kernel_timestep_embedding;
    cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
+    cl_kernel kernel_mul_mv_id_q8_0_f32, kernel_mul_mv_id_q8_0_f32_flat;
    cl_kernel kernel_mul_mv_id_mxfp4_f32;
+    cl_kernel kernel_mul_mv_id_mxfp4_f32_flat;
    cl_kernel kernel_mul_mm_f32_f32_l4_lm;
    cl_kernel kernel_mul_mm_f16_f32_l4_lm;

@@ -765,6 +774,10 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
        CL_CHECK((backend_ctx->kernel_convert_block_q4_0_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_0_noshuffle", &err), err));
        CL_CHECK((backend_ctx->kernel_convert_block_q4_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_0", &err), err));
        CL_CHECK((backend_ctx->kernel_restore_block_q4_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_mxfp4", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_mxfp4", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q8_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q8_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q8_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0", &err), err));
        GGML_LOG_CONT(".");
    }

@@ -986,6 +999,38 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
        GGML_LOG_CONT(".");
    }

+    // mul_mv_q8_0_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q8_0_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q8_0_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_q8_0_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q8_0_f32 = clCreateKernel(backend_ctx->program_mul_mv_q8_0_f32, "kernel_mul_mv_q8_0_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q8_0_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q8_0_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q8_0_f32_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_q8_0_f32_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q8_0_f32_flat = clCreateKernel(backend_ctx->program_mul_mv_q8_0_f32_flat, "kernel_mul_mv_q8_0_f32_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
    // mul_mv_mxfp4_f32
    {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1002,6 +1047,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
        GGML_LOG_CONT(".");
    }

+    // mul_mv_mxfp4_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_mxfp4_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_mxfp4_f32_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_mxfp4_f32_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_mxfp4_f32_flat = clCreateKernel(backend_ctx->program_mul_mv_mxfp4_f32_flat, "kernel_mul_mv_mxfp4_f32_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
    // mul_mv_f16_f16
    {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1649,8 +1710,10 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
        backend_ctx->program_set_rows =
            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);

-        CL_CHECK((backend_ctx->kernel_set_rows_f32  = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f32", &err), err));
-        CL_CHECK((backend_ctx->kernel_set_rows_f16  = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_set_rows_f32_i64 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f32_i64", &err), err));
+        CL_CHECK((backend_ctx->kernel_set_rows_f32_i32 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f32_i32", &err), err));
+        CL_CHECK((backend_ctx->kernel_set_rows_f16_i64 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f16_i64", &err), err));
+        CL_CHECK((backend_ctx->kernel_set_rows_f16_i32 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f16_i32", &err), err));
        GGML_LOG_CONT(".");
    }

@@ -1711,6 +1774,38 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
        GGML_LOG_CONT(".");
    }

+    // mul_mv_id_q8_0_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_q8_0_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_q8_0_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_id_q8_0_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_q8_0_f32 = clCreateKernel(backend_ctx->program_mul_mv_id_q8_0_f32, "kernel_mul_mv_id_q8_0_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_id_q8_0_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_q8_0_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_q8_0_f32_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_id_q8_0_f32_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_q8_0_f32_flat = clCreateKernel(backend_ctx->program_mul_mv_id_q8_0_f32_flat, "kernel_mul_mv_id_q8_0_f32_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
    // mul_mv_id_mxfp4_f32
    {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1727,6 +1822,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
        GGML_LOG_CONT(".");
    }

+    // mul_mv_id_mxfp4_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_mxfp4_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_mxfp4_f32_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_id_mxfp4_f32_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_mxfp4_f32_flat = clCreateKernel(backend_ctx->program_mul_mv_id_mxfp4_f32_flat, "kernel_mul_mv_id_mxfp4_f32_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
    // Adreno kernels
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
    // transpose
@@ -2391,6 +2502,84 @@ struct ggml_tensor_extra_cl_q4_0 {
    }
 };

+struct ggml_tensor_extra_cl_mxfp4 {
+    // Quantized values.
+    cl_mem q = nullptr;
+    // Quantized values in image1d_buffer_t.
+    cl_mem q_img = nullptr;
+    // Scales in E8M0.
+    cl_mem e = nullptr;
+    // Scales in image1d_buffer_t.
+    cl_mem e_img = nullptr;
+    // Size of quantized values.
+    size_t size_q = 0;
+    // Size of scales.
+    size_t size_e = 0;
+
+    ~ggml_tensor_extra_cl_mxfp4() {
+        reset();
+    }
+
+    void reset() {
+        // q and d are subbuffers into the bigger buffer allocated in ggml_backend_buffer.
+        // They must be properly released so that the original buffer can be
+        // properly released to avoid memory leak.
+        if (q != nullptr) {
+            CL_CHECK(clReleaseMemObject(q));
+            q = nullptr;
+        }
+        if (e != nullptr) {
+            CL_CHECK(clReleaseMemObject(e));
+            e = nullptr;
+        }
+        if (q != nullptr) {
+            CL_CHECK(clReleaseMemObject(q_img));
+            q = nullptr;
+        }
+        // Currently, q_img and d_img are not used. They can be image1d_buffer_t
+        // that wraps around q and d to utilize image access path.
+        q_img = nullptr;
+        e_img = nullptr;
+        size_q = 0;
+        size_e = 0;
+    }
+};
+
+struct ggml_tensor_extra_cl_q8_0 {
+    cl_mem q = nullptr;
+    cl_mem q_img = nullptr;
+
+    cl_mem d = nullptr;
+    cl_mem d_img = nullptr;
+
+    size_t size_q = 0;
+    size_t size_d = 0;
+
+    ~ggml_tensor_extra_cl_q8_0() {
+        reset();
+    }
+
+    void reset() {
+        // q and d are subbuffers into the bigger buffer allocated in ggml_backend_buffer.
+        // They must be properly released so that the original buffer can be
+        // properly released to avoid memory leak.
+        if (q != nullptr) {
+            CL_CHECK(clReleaseMemObject(q));
+            q = nullptr;
+        }
+        if (d != nullptr) {
+            CL_CHECK(clReleaseMemObject(d));
+            d = nullptr;
+        }
+        // Currently, q_img and d_img are not used. They can be image1d_buffer_t
+        // that wraps around q and d to utilize image access path.
+        q_img = nullptr;
+        d_img = nullptr;
+        size_q = 0;
+        size_d = 0;
+    }
+};
+
 //------------------------------------------------------------------------------
 // Backend API
 //------------------------------------------------------------------------------
@@ -2616,7 +2805,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                switch (op->type) {
                    case GGML_TYPE_F16:
                    case GGML_TYPE_F32:
-                        return true;
+                        return (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
                    default:
                        return false;
                }
@@ -2724,10 +2913,13 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
            } else if (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_MXFP4 ||
                       op->src[0]->type == GGML_TYPE_Q6_K) {
                return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
+            } else if (op->src[0]->type == GGML_TYPE_Q8_0) {
+                return op->src[1]->type == GGML_TYPE_F32;
            }
            return false;
        case GGML_OP_MUL_MAT_ID:
            if (op->src[0]->type == GGML_TYPE_Q4_0 ||
+                op->src[0]->type == GGML_TYPE_Q8_0 ||
                op->src[0]->type == GGML_TYPE_MXFP4) {
                if (op->src[1]->type == GGML_TYPE_F32) {
                    return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
@@ -2838,7 +3030,7 @@ static ggml_backend_i ggml_backend_opencl_i = {
    /* .graph_compute           = */ ggml_backend_opencl_graph_compute,
    /* .event_record            = */ NULL,
    /* .event_wait              = */ NULL,
-    /* .optimize_graph          = */ NULL,
+    /* .graph_optimize          = */ NULL,
 };

 ggml_backend_t ggml_backend_opencl_init(void) {
@@ -2894,6 +3086,18 @@ struct ggml_backend_opencl_buffer_context {
        for (ggml_tensor_extra_cl_q4_0 * e : temp_tensor_extras_q4_0_in_use) {
            delete e;
        }
+        for (ggml_tensor_extra_cl_mxfp4 * e : temp_tensor_extras_mxfp4) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_mxfp4 * e : temp_tensor_extras_mxfp4_in_use) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0_in_use) {
+            delete e;
+        }
    }

    ggml_tensor_extra_cl * ggml_opencl_alloc_temp_tensor_extra() {
@@ -2926,6 +3130,36 @@ struct ggml_backend_opencl_buffer_context {
        return extra;
    }

+    ggml_tensor_extra_cl_mxfp4 * ggml_opencl_alloc_temp_tensor_extra_mxfp4() {
+        ggml_tensor_extra_cl_mxfp4 * extra;
+        if (temp_tensor_extras_mxfp4.empty()) {
+            extra = new ggml_tensor_extra_cl_mxfp4();
+        } else {
+            extra = temp_tensor_extras_mxfp4.back();
+            temp_tensor_extras_mxfp4.pop_back();
+        }
+
+        temp_tensor_extras_mxfp4_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
+    ggml_tensor_extra_cl_q8_0 * ggml_opencl_alloc_temp_tensor_extra_q8_0() {
+        ggml_tensor_extra_cl_q8_0 * extra;
+        if (temp_tensor_extras_q8_0.empty()) {
+            extra = new ggml_tensor_extra_cl_q8_0();
+        } else {
+            extra = temp_tensor_extras_q8_0.back();
+            temp_tensor_extras_q8_0.pop_back();
+        }
+
+        temp_tensor_extras_q8_0_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
    void reset() {
        for (ggml_tensor_extra_cl * e : temp_tensor_extras_in_use) {
            temp_tensor_extras.push_back(e);
@@ -2936,6 +3170,16 @@ struct ggml_backend_opencl_buffer_context {
            temp_tensor_extras_q4_0.push_back(e);
        }
        temp_tensor_extras_q4_0_in_use.clear();
+
+        for (ggml_tensor_extra_cl_mxfp4 * e : temp_tensor_extras_mxfp4_in_use) {
+            temp_tensor_extras_mxfp4.push_back(e);
+        }
+        temp_tensor_extras_mxfp4_in_use.clear();
+
+        for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0_in_use) {
+            temp_tensor_extras_q8_0.push_back(e);
+        }
+        temp_tensor_extras_q8_0_in_use.clear();
    }

    // Pools for extras. Available extras are in `temp_tensor_extras`. Extras
@@ -2947,6 +3191,10 @@ struct ggml_backend_opencl_buffer_context {
    std::vector<ggml_tensor_extra_cl *> temp_tensor_extras_in_use;
    std::vector<ggml_tensor_extra_cl_q4_0 *> temp_tensor_extras_q4_0;
    std::vector<ggml_tensor_extra_cl_q4_0 *> temp_tensor_extras_q4_0_in_use;
+    std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4;
+    std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4_in_use;
+    std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0;
+    std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0_in_use;

    // The buffer_context is initially created by ggml_backend_buft_alloc_buffer
    // before any tensor is initialized (at the beginning of alloc_tensor_range).
@@ -3289,6 +3537,135 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
        }
    #endif // GGML_OPENCL_USE_ADRENO_KERNELS

+        return;
+
+    }
+    if (tensor->type == GGML_TYPE_MXFP4) {
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
+
+        // Allocate the new extra and create aliases from the original.
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_mxfp4 * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_mxfp4();
+
+        size_t size_e = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(char);
+        size_t size_q = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
+        GGML_ASSERT(size_e + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        // The original tensor memory is divided into scales and quants, i.e.,
+        // we first store scales, then quants.
+        cl_buffer_region region;
+
+        // Create subbuffer for scales.
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_e;
+        extra->e = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Create subbuffer for quants.
+        region.origin = align_to(previous_origin + size_e, backend_ctx->alignment);
+        region.size = size_q;
+        extra->q = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_convert_block_mxfp4;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->e));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        // Create image for Q
+        cl_image_format img_format_q = {CL_RG, CL_UNSIGNED_INT32};
+        cl_image_desc img_desc_q = {
+            CL_MEM_OBJECT_IMAGE1D_BUFFER,
+            static_cast<size_t>(ggml_nelements(tensor)/32*2),
+            0, 0, 0, 0, 0, 0, 0,
+            { extra->q }
+        };
+        extra->q_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_format_q, &img_desc_q, NULL, &err);
+
+        tensor->extra = extra;
+
+        return;
+    }
+    if (tensor->type == GGML_TYPE_Q8_0) {
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
+
+        // Allocate the new extra and create aliases from the original.
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_q8_0 * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_q8_0();
+
+        size_t size_d = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+        size_t size_q = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*(ggml_blck_size(tensor->type)*sizeof(char));
+        GGML_ASSERT(size_d + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        // The original tensor memory is divided into scales and quants, i.e.,
+        // we first store scales, then quants.
+        cl_buffer_region region;
+
+        // Create subbuffer for scales.
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_d;
+        extra->d = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Create subbuffer for quants.
+        region.origin = align_to(previous_origin + size_d, backend_ctx->alignment);
+        region.size = size_q;
+        extra->q = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q8_0;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->d));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        tensor->extra = extra;
+
        return;
    }
 #endif // GGML_OPENCL_SOA_Q
@@ -3337,6 +3714,57 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
        size_t local_work_size[] = {1, 1, 1};

+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    } else if (tensor->type == GGML_TYPE_MXFP4) {
+        ggml_tensor_extra_cl_mxfp4 * extra = (ggml_tensor_extra_cl_mxfp4 *)tensor->extra;
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_restore_block_mxfp4;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->e));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &data_device));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    }
+    if (tensor->type == GGML_TYPE_Q8_0) {
+        ggml_tensor_extra_cl_q8_0 * extra = (ggml_tensor_extra_cl_q8_0 *)tensor->extra;
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_restore_block_q8_0;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &data_device));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
+
        cl_event evt;
        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
            global_work_size, local_work_size, 0, NULL, &evt));
@@ -3658,6 +4086,19 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso
        CL_CHECK(clEnqueueReadBuffer(queue, extra->q, CL_TRUE, 0, size_q, buf_q, 0, NULL, NULL));
        CL_CHECK(clEnqueueReadBuffer(queue, extra->d, CL_TRUE, 0, size_d, buf_d, 0, NULL, NULL));
        CL_CHECK(clFinish(queue));
+    } else if (tensor->type == GGML_TYPE_MXFP4) {
+        ggml_tensor_extra_cl_mxfp4 * extra = (ggml_tensor_extra_cl_mxfp4 *) tensor->extra;
+        GGML_ASSERT(extra);
+
+        size_t size_q = ggml_nelements(tensor)/QK_MXFP4 * QK_MXFP4/2;
+        size_t size_e = ggml_nelements(tensor)/QK_MXFP4 * sizeof(char);
+        GGML_ASSERT(size_q + size_e == ggml_nbytes(tensor));
+        buf_q = malloc(size_q);
+        buf_d = malloc(size_e);
+
+        CL_CHECK(clEnqueueReadBuffer(queue, extra->q, CL_TRUE, 0, size_q, buf_q, 0, NULL, NULL));
+        CL_CHECK(clEnqueueReadBuffer(queue, extra->d, CL_TRUE, 0, size_e, buf_d, 0, NULL, NULL));
+        CL_CHECK(clFinish(queue));
    } else {
        // Read out the tensor from GPU memory.
        ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
@@ -3845,6 +4286,7 @@ static void ggml_cl_set_rows(ggml_backend_t backend, const ggml_tensor * src0, c
    GGML_ASSERT(src1->extra);
    GGML_ASSERT(dst);
    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src1->type == GGML_TYPE_I64 || src1->type == GGML_TYPE_I32);

    // ne0 = ne00
    // ne2 = ne02
@@ -3887,10 +4329,18 @@ static void ggml_cl_set_rows(ggml_backend_t backend, const ggml_tensor * src0, c

    switch (dst->type) {
        case GGML_TYPE_F32:
-            kernel = backend_ctx->kernel_set_rows_f32;
+            if (src1->type == GGML_TYPE_I64) {
+                kernel = backend_ctx->kernel_set_rows_f32_i64;
+            } else {
+                kernel = backend_ctx->kernel_set_rows_f32_i32;
+            }
            break;
        case GGML_TYPE_F16:
-            kernel = backend_ctx->kernel_set_rows_f16;
+            if (src1->type == GGML_TYPE_I64) {
+                kernel = backend_ctx->kernel_set_rows_f16_i64;
+            } else {
+                kernel = backend_ctx->kernel_set_rows_f16_i32;
+            }
            break;
        default:
            GGML_ABORT("not implemented");
@@ -5669,12 +6119,12 @@ static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, con
    } else {
        cl_kernel kernel = backend_ctx->kernel_concat_f32_non_contiguous;

-        long ne00 = src0->ne[0], ne01 = src0->ne[1], ne02 = src0->ne[2], ne03 = src0->ne[3];
+        cl_long ne00 = src0->ne[0], ne01 = src0->ne[1], ne02 = src0->ne[2], ne03 = src0->ne[3];
        cl_ulong nb00 = src0->nb[0], nb01 = src0->nb[1], nb02 = src0->nb[2], nb03 = src0->nb[3];

        cl_ulong nb10 = src1->nb[0], nb11 = src1->nb[1], nb12 = src1->nb[2], nb13 = src1->nb[3];

-        long d_ne0 = dst->ne[0], d_ne1 = dst->ne[1], d_ne2 = dst->ne[2], d_ne3 = dst->ne[3];
+        cl_long d_ne0 = dst->ne[0], d_ne1 = dst->ne[1], d_ne2 = dst->ne[2], d_ne3 = dst->ne[3];
        cl_ulong d_nb0 = dst->nb[0], d_nb1 = dst->nb[1], d_nb2 = dst->nb[2], d_nb3 = dst->nb[3];


@@ -5685,10 +6135,10 @@ static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, con
        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),    &extrad_cl->data_device));
        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &off_dst));

-        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(long),      &ne00));
-        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(long),      &ne01));
-        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(long),      &ne02));
-        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(long),      &ne03));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_long),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_long),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_long),      &ne02));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_long),      &ne03));
        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),    &nb00));
        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),    &nb01));
        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong),    &nb02));
@@ -5699,10 +6149,10 @@ static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, con
        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),    &nb12));
        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),    &nb13));

-        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(long),     &d_ne0));
-        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(long),     &d_ne1));
-        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(long),     &d_ne2));
-        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(long),     &d_ne3));
+        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_long),     &d_ne0));
+        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_long),     &d_ne1));
+        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_long),     &d_ne2));
+        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_long),     &d_ne3));
        CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong),    &d_nb0));
        CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong),    &d_nb1));
        CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong),    &d_nb2));
@@ -6048,6 +6498,8 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co

 #ifdef GGML_OPENCL_SOA_Q
    ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
+    ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra;
+    ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
 #endif

    const int  ne00 = src0 ? src0->ne[0] : 0;
@@ -6717,7 +7169,84 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 #endif // GGML_OPENCL_SOA_Q
            break;
        case GGML_TYPE_Q4_1:
-        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q8_0: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_q8_0_f32_flat;
+
+            // nth0 - subgroup size
+            // nth1 - number of subgroups per workgroup
+            // ndst - number of output values per workgroup = output per subgroup * number of subgroups
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q8_0->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q8_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            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(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+#else
+            kernel = backend_ctx->kernel_mul_mv_q8_0_f32;
+
+            // nth0 - subgroup size
+            // nth1 - number of subgroups per workgroup
+            // ndst - number of output values per workgroup = output per subgroup * number of subgroups
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            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(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+#endif // GGML_OPENCL_SOA_Q
+            break;
+        }
        case GGML_TYPE_Q2_K:
        case GGML_TYPE_Q3_K:
        case GGML_TYPE_Q4_K:
@@ -6752,6 +7281,45 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
            break;
        case GGML_TYPE_MXFP4: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_mxfp4_f32_flat;
+
+            cl_mem q;
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*2;
+
+                q = extra0_mxfp4->q;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*2;
+
+                q = extra0_mxfp4->q_img;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_mxfp4->e));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r3));
+#else
            kernel = backend_ctx->kernel_mul_mv_mxfp4_f32;

            if (backend_ctx->gpu_family == INTEL) {
@@ -6785,6 +7353,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r2));
            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r3));
            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float)*nth0,nullptr));
+#endif
            break;
        }
        default:
@@ -6850,8 +7419,12 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
    cl_ulong offset2 = extra2->offset + src2->view_offs;
    cl_ulong offsetd = extrad->offset + dst->view_offs;

+    GGML_UNUSED(offset0);
+
 #ifdef GGML_OPENCL_SOA_Q
    ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
+    ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra;
+    ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
 #endif

    const int ne00 = src0->ne[0];
@@ -6939,7 +7512,128 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,

            break;
        }
+        case GGML_TYPE_Q8_0: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_id_q8_0_f32_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 2;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q8_0->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q8_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne1));
+#else
+            kernel = backend_ctx->kernel_mul_mv_id_q8_0_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 2;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne1));
+#endif // GGML_OPENCL_SOA_Q
+            break;
+        }
        case GGML_TYPE_MXFP4: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_id_mxfp4_f32_flat;
+
+            cl_mem q;
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 2;
+
+                q = extra0_mxfp4->q;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 1;
+                ndst = 4;
+
+                q = extra0_mxfp4->q_img;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_mxfp4->e));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            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(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r3));
+#else // GGML_OPENCL_SOA_Q
            kernel = backend_ctx->kernel_mul_mv_id_mxfp4_f32;

            if (backend_ctx->gpu_family == INTEL) {
@@ -6979,7 +7673,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r2));
            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r3));
            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(float)*sgs,nullptr));
-
+#endif // GGML_OPENCL_SOA_Q
            break;
        }
        default:
@@ -116,3 +116,87 @@ kernel void kernel_convert_block_q4_0_noshuffle(
 #endif
    }
 }
+
+//------------------------------------------------------------------------------
+// block_mxfp4
+//------------------------------------------------------------------------------
+#define QK_MXFP4 32
+struct block_mxfp4 {
+    uchar e; // E8M0
+    uchar qs[QK_MXFP4 / 2];
+};
+
+//------------------------------------------------------------------------------
+// kernel_convert_block_mxfp4
+// Convert the block_mxfp4 format to 2 separate arrays (AOS -> SOA).
+// This kernel does not deshuffle the bits.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_mxfp4(
+    global struct block_mxfp4 * src0,
+    global uchar * dst_q,
+    global uchar * dst_e
+) {
+    global struct block_mxfp4 * b = (global struct block_mxfp4 *) src0 + get_global_id(0);
+    global uchar * q = (global uchar *) dst_q + QK_MXFP4 / 2 * get_global_id(0);
+    global uchar * e = (global uchar *) dst_e + get_global_id(0);
+
+    *e = b->e;
+
+    for (int i = 0; i < QK_MXFP4 / 2; ++i) {
+        q[i] = b->qs[i];
+    }
+}
+
+kernel void kernel_restore_block_mxfp4(
+    global uchar * src_q,
+    global half  * src_e,
+    global struct block_mxfp4 * dst
+) {
+    global struct block_mxfp4 * b = (global struct block_mxfp4 *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK_MXFP4 / 2 * get_global_id(0);
+    global uchar * e = (global uchar *) src_e + get_global_id(0);
+
+    b->e = *e;
+    for (int i = 0; i < QK_MXFP4 / 2; ++i) {
+        b->qs[i] = q[i];
+    }
+}
+
+//------------------------------------------------------------------------------
+// block_q8_0
+//------------------------------------------------------------------------------
+typedef struct {
+    half d;       // delta
+    char qs[QK8_0]; // quants
+} block_q8_0;
+
+kernel void kernel_convert_block_q8_0(
+    global block_q8_0 * src0,
+    global uchar * dst_q,
+    global half  * dst_d
+) {
+    global block_q8_0 * b = (global block_q8_0 *) src0 + get_global_id(0);
+    global uchar      * q = (global uchar *) dst_q + QK8_0*get_global_id(0);
+    global half       * d = (global half *) dst_d + get_global_id(0);
+
+    *d = b->d;
+
+    for (int i = 0; i < QK8_0; ++i) {
+        q[i] = b->qs[i];
+    }
+}
+
+kernel void kernel_restore_block_q8_0(
+    global uchar * src_q,
+    global half  * src_d,
+    global block_q8_0 * dst
+) {
+    global block_q8_0 * b = (global block_q8_0 *) dst + get_global_id(0);
+    global uchar      * q = (global uchar *) src_q + QK8_0*get_global_id(0);
+    global half       * d = (global half *) src_d + get_global_id(0);
+
+    b->d = *d;
+    for (int i = 0; i < QK8_0; ++i) {
+        b->qs[i] = q[i];
+    }
+}
@@ -0,0 +1,176 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+
+static inline half4 mxfp4_to_fp16_packed(ushort fp4x4) {
+    ushort2 fp16_packed_a, fp16_packed_b, bias_a, bias_b, sign_a, sign_b;
+    fp16_packed_a.lo = (fp4x4 << 9) & 0x0E00;
+    fp16_packed_a.hi = (fp4x4 << 5) & 0x0E00;
+    fp16_packed_b.lo = (fp4x4 << 1) & 0x0E00;
+    fp16_packed_b.hi = (fp4x4 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a.lo == 0) ? 0x0 : 0x3800;
+    bias_a.hi = (fp16_packed_a.hi == 0) ? 0x0 : 0x3800;
+    bias_b.lo = (fp16_packed_b.lo == 0) ? 0x0 : 0x3800;
+    bias_b.hi = (fp16_packed_b.hi == 0) ? 0x0 : 0x3800;
+
+    fp16_packed_a.lo = (fp16_packed_a.lo == 0x0200) ? 0x0 : fp16_packed_a.lo;
+    fp16_packed_a.hi = (fp16_packed_a.hi == 0x0200) ? 0x0 : fp16_packed_a.hi;
+    fp16_packed_b.lo = (fp16_packed_b.lo == 0x0200) ? 0x0 : fp16_packed_b.lo;
+    fp16_packed_b.hi = (fp16_packed_b.hi == 0x0200) ? 0x0 : fp16_packed_b.hi;
+
+    sign_a.lo = (fp4x4 << 12) & 0x8000;
+    sign_a.hi = (fp4x4 << 8) & 0x8000;
+    sign_b.lo = (fp4x4 << 4) & 0x8000;
+    sign_b.hi = fp4x4 & 0x8000;
+
+    fp16_packed_a = sign_a + bias_a + fp16_packed_a;
+    fp16_packed_b = sign_b + bias_b + fp16_packed_b;
+
+    return as_half4((ushort4)(fp16_packed_a, fp16_packed_b));
+}
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+    bits = (x == 0) ? 0x00400000 : ((uint) x << 23);
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 4
+#define N_SG_MXFP4 1
+#define N_SIMDWIDTH 64
+#define SRC0Q_IMG
+#endif
+
+kernel void kernel_mul_mv_id_mxfp4_f32_flat(
+#ifdef SRC0Q_IMG
+    __read_only image1d_buffer_t src0_q,
+#else
+    global uchar * src0_q,
+#endif
+    global uchar * src0_e,
+    global uchar * src1,
+    ulong         offset1,
+    global uchar * src2,
+    ulong         offset2,
+    global uchar * dst,
+    ulong         offsetd,
+    int           ne00,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb03,
+    int           ne11,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    ulong         nb13,
+    int           ne20,
+    int           ne21,
+    ulong         nb21,
+    int           ne0,
+    int           ne1,
+    int           r2,
+    int           r3
+) {
+    dst  = dst  + offsetd;
+
+    const int iid1 = get_group_id(2) / ne20;
+    const int idx  = get_group_id(2) % ne20;
+
+    uint i02 = ((global uint *) (src2 + offset2 + iid1 * nb21))[idx];
+
+    int i11 = idx % ne11;
+
+    int nb = ne00 / QK_MXFP4;
+
+    uint src0_off = i02*nb02;
+    src0_off /= 17; // 17 = sizeof(block_mxfp4)
+
+    src0_e = src0_e + src0_off;
+
+    dst = dst + (idx * ne0 + iid1 * ne1 * ne0) * sizeof(float);
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint offset_src0 = first_row*nb01;
+    offset_src0 /= 17; // 17 = sizeof(block_mxfp4)
+#ifdef SRC0Q_IMG
+    ulong offset_q = src0_off + offset_src0;
+#else
+    src0_q = src0_q + src0_off*16;
+    global uchar16 * x_q = (global uchar16 *)(src0_q) + offset_src0;
+#endif
+    global uchar * x_e = src0_e + offset_src0;
+
+    const short ix = get_sub_group_local_id() >> 1;
+    const short it = get_sub_group_local_id() & 1;
+
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    src1 = src1 + offset1 + i11 * nb11 + iid1 * nb12;
+    global float * y   = (global float *) (src1 + r1 * nb11);
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH / 2) {
+        global float4 * y4 = (global float4 *)yb;
+
+        #pragma unroll
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            uchar xb_e = x_e[row * nb + ib];
+#ifdef SRC0Q_IMG
+            ushort4 xb_q = as_ushort4(read_imageui(src0_q, (offset_q + row * nb + ib) * 2 + it).xy);
+#else
+            ushort4 xb_q = vload4(0, (global ushort *)((global uchar *)(x_q + row * nb + ib) + 8 * it));
+#endif
+
+            half4 fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s0);
+            half4 fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s1);
+            float4 acc1 = y4[0] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2);
+            acc1 += y4[4] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3);
+
+            fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s2);
+            fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s3);
+            acc1 += y4[1] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2);
+            acc1 += y4[5] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3);
+
+            sumf[row] += e8m0_to_fp32(xb_e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH / 2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *)dst + (ulong)r1 * ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}
@@ -0,0 +1,140 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK8_0 32
+typedef struct {
+    half d;       // delta
+    char qs[QK8_0]; // quants
+} block_q8_0;
+
+#define NB_Q8_0 8
+
+#ifdef INTEL_GPU
+#define N_R0_Q8_0 4 // number of rows each subgroup works on
+#define N_SG_Q8_0 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_Q8_0 4
+#define N_SG_Q8_0 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_q8_0_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * src2,
+    ulong         offset2,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    ulong         nb01,
+    ulong         nb02,
+    int           ne11,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    int           ne20,
+    int           ne21,
+    ulong         nb21,
+    int           ne0,
+    int           ne1
+) {
+    src0 = (global char *)((global char *)src0 + offset0);
+    src1 = (global char *)((global char *)src1 + offset1);
+    src2 = (global char *)((global char *)src2 + offset2);
+    dst  = (global char *)((global char *)dst  + offsetd);
+
+    int iid1 = get_group_id(2)/ne20;
+    int idx  = get_group_id(2)%ne20;
+
+    int i02 = ((global int *) (src2 + iid1*nb21))[idx];
+
+    int i11_ = idx % ne11;
+    int i12_ = iid1;
+
+    int i1 = idx;
+    int i2 = i12_;
+
+    global char * src0_cur = src0 + i02*nb02;
+    global char * src1_cur = src1 + i11_*nb11 + i12_*nb12;
+
+    global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float);
+
+    int nb = ne00/QK8_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+
+    int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0;
+
+    ulong offset_src1 = r1*nb11;
+    global float * y  = (global float *) (src1_cur + offset_src1);
+
+    // pointers to src0 rows
+    global block_q8_0 * ax[N_R0_Q8_0];
+    for (int row = 0; row < N_R0_Q8_0; ++row) {
+        ulong offset_src0 = (first_row + row)*nb01;
+        ax[row] = (global block_q8_0 *) ((global char *) src0_cur + offset_src0);
+    }
+
+    float yl[NB_Q8_0];
+    float sumf[N_R0_Q8_0] = { 0.f };
+
+    const short ix = get_sub_group_local_id()/4;
+    const short il = get_sub_group_local_id()%4;
+
+    global float * yb = y + ix*QK8_0 + il*NB_Q8_0;
+
+    // each thread handles NB_Q8_0 quants at a time
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) {
+        for (short i = 0; i < NB_Q8_0; ++i) {
+            yl[i] = yb[i];
+        }
+
+        for (short row = 0; row < N_R0_Q8_0; row++) {
+            global char * qs = ax[row][ib].qs + il*NB_Q8_0;
+            float sumq = 0.f;
+            for (short iq = 0; iq < NB_Q8_0; ++iq) {
+                sumq += qs[iq] * yl[iq];
+            }
+            sumf[row] += sumq*ax[row][ib].d;
+        }
+
+        yb += N_SIMDWIDTH*NB_Q8_0;
+    }
+
+    global float * dst_f32 = (global float *) dst_cur + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_Q8_0; ++row) {
+        float tot = sub_group_reduce_add(sumf[row]);
+
+        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
+            dst_f32[first_row + row] = tot;
+        }
+    }
+}
@@ -0,0 +1,222 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK8_0 32
+typedef struct {
+    half d;       // delta
+    char qs[QK8_0]; // quants
+} block_q8_0;
+
+#define NB_Q8_0 8
+
+#ifdef INTEL_GPU
+#define N_R0_Q8_0 4 // number of rows each subgroup works on
+#define N_SG_Q8_0 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_Q8_0 4
+#define N_SG_Q8_0 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_q8_0_f32_flat(
+    global char * src0_q,
+    global half * src0_d,
+    global char * src1,
+    ulong         offset1,
+    global char * src2,
+    ulong         offset2,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    ulong         nb01,
+    ulong         nb02,
+    int           ne11,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    int           ne20,
+    int           ne21,
+    ulong         nb21,
+    int           ne0,
+    int           ne1
+) {
+    src1 = (global char *)((global char *)src1 + offset1);
+    src2 = (global char *)((global char *)src2 + offset2);
+    dst  = (global char *)((global char *)dst  + offsetd);
+
+    int iid1 = (int)get_group_id(2)/ne20;
+    int idx  = (int)get_group_id(2)%ne20;
+
+    int i02 = ((global int *) (src2 + iid1*nb21))[idx];
+
+    int i11_ = idx % ne11;
+    int i12_ = iid1;
+
+    int i1 = idx;
+    int i2 = i12_;
+
+    // 34 == sizeof(block_q8_0)
+    uint src0_off = i02*nb02;
+    src0_off /= 34;
+
+    global char * src0_q_cur = src0_q + src0_off*sizeof(char)*QK8_0;
+    global half * src0_d_cur = src0_d + src0_off;
+    global char * src1_cur   = src1 + i11_*nb11 + i12_*nb12;
+
+    global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float);
+
+    int nb = ne00/QK8_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+
+    int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0;
+
+    ulong offset_src1 = r1*nb11;
+    global float * y  = (global float *) (src1_cur + offset_src1);
+
+    // pointers to src0 rows
+    uint offset_src0_base = first_row*nb01;
+
+    global char * ax0, * ax1, * ax2, * ax3;
+    global half * ad0, * ad1, * ad2, * ad3;
+    uint offset_src0;
+
+    offset_src0 = offset_src0_base + 0*nb01;
+    offset_src0 = offset_src0/34;
+    ax0 = (global char *) ((global char *) src0_q_cur + offset_src0*sizeof(char)*QK8_0);
+    ad0 = (global half *) ((global char *) src0_d_cur + offset_src0*sizeof(half));
+
+    offset_src0 = offset_src0_base + 1*nb01;
+    offset_src0 = offset_src0/34;
+    ax1 = (global char *) ((global char *) src0_q_cur + offset_src0*sizeof(char)*QK8_0);
+    ad1 = (global half *) ((global char *) src0_d_cur + offset_src0*sizeof(half));
+
+    offset_src0 = offset_src0_base + 2*nb01;
+    offset_src0 = offset_src0/34;
+    ax2 = (global char *) ((global char *) src0_q_cur + offset_src0*sizeof(char)*QK8_0);
+    ad2 = (global half *) ((global char *) src0_d_cur + offset_src0*sizeof(half));
+
+    offset_src0 = offset_src0_base + 3*nb01;
+    offset_src0 = offset_src0/34;
+    ax3 = (global char *) ((global char *) src0_q_cur + offset_src0*sizeof(char)*QK8_0);
+    ad3 = (global half *) ((global char *) src0_d_cur + offset_src0*sizeof(half));
+
+    const short ix = get_sub_group_local_id()/4;
+    const short il = get_sub_group_local_id()%4;
+
+    global float * yb = y + ix*QK8_0 + il*NB_Q8_0;
+
+    float8 yl;
+    float8 qv;
+    float4 sumf = 0.f;
+    float  sumq = 0.f;
+    global char * qs;
+
+    // each thread handles NB_Q8_0 quants at a time
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) {
+        yl = vload8(0, yb);
+
+        qs = ax0 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s0 += sumq*ad0[ib];
+
+        qs = ax1 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s1 += sumq*ad1[ib];
+
+        qs = ax2 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s2 += sumq*ad2[ib];
+
+        qs = ax3 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s3 += sumq*ad3[ib];
+
+        yb += N_SIMDWIDTH*NB_Q8_0;
+    }
+
+    global float * dst_f32 = (global float *) dst_cur + (ulong)r1*ne0;
+
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0),
+        sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2),
+        sub_group_reduce_add(sumf.s3)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst_f32[first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst_f32[first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst_f32[first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst_f32[first_row + 3] = tot.s3;
+        }
+    }
+}
@@ -0,0 +1,167 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+
+static inline half4 mxfp4_to_fp16_packed(ushort fp4x4) {
+    ushort2 fp16_packed_a, fp16_packed_b, bias_a, bias_b, sign_a, sign_b;
+    fp16_packed_a.lo = (fp4x4 << 9) & 0x0E00;
+    fp16_packed_a.hi = (fp4x4 << 5) & 0x0E00;
+    fp16_packed_b.lo = (fp4x4 << 1) & 0x0E00;
+    fp16_packed_b.hi = (fp4x4 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a.lo == 0) ? 0x0 : 0x3800;
+    bias_a.hi = (fp16_packed_a.hi == 0) ? 0x0 : 0x3800;
+    bias_b.lo = (fp16_packed_b.lo == 0) ? 0x0 : 0x3800;
+    bias_b.hi = (fp16_packed_b.hi == 0) ? 0x0 : 0x3800;
+
+    fp16_packed_a.lo = (fp16_packed_a.lo == 0x0200) ? 0x0 : fp16_packed_a.lo;
+    fp16_packed_a.hi = (fp16_packed_a.hi == 0x0200) ? 0x0 : fp16_packed_a.hi;
+    fp16_packed_b.lo = (fp16_packed_b.lo == 0x0200) ? 0x0 : fp16_packed_b.lo;
+    fp16_packed_b.hi = (fp16_packed_b.hi == 0x0200) ? 0x0 : fp16_packed_b.hi;
+
+    sign_a.lo = (fp4x4 << 12) & 0x8000;
+    sign_a.hi = (fp4x4 << 8) & 0x8000;
+    sign_b.lo = (fp4x4 << 4) & 0x8000;
+    sign_b.hi = fp4x4 & 0x8000;
+
+    fp16_packed_a = sign_a + bias_a + fp16_packed_a;
+    fp16_packed_b = sign_b + bias_b + fp16_packed_b;
+
+    return as_half4((ushort4)(fp16_packed_a, fp16_packed_b));
+}
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+    bits = (x == 0) ? 0x00400000 : ((uint) x << 23);
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 2
+#define N_SG_MXFP4 2
+#define N_SIMDWIDTH 64
+#define SRC0Q_IMG
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_mxfp4_f32_flat(
+#ifdef SRC0Q_IMG
+    __read_only image1d_buffer_t src0_q,
+#else
+    global uchar * src0_q,
+#endif
+    global uchar * src0_e,
+    global uchar * src1,
+    ulong          offset1,
+    global uchar * dst,
+    ulong          offsetd,
+    int ne00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne12,
+    ulong nb11,
+    ulong nb12,
+    ulong nb13,
+    int ne0,
+    int ne1,
+    int r2,
+    int r3
+) {
+    src1 = src1 + offset1;
+    dst = dst + offsetd;
+
+    int nb = ne00 / QK_MXFP4;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint i12 = im % ne12;
+    uint i13 = im / ne12;
+
+    uint offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    // 17 = sizeof(block_mxfp4)
+    offset_src0 /= 17;
+#ifdef SRC0Q_IMG
+    ulong offset_q = offset_src0;
+#else
+    global uchar16 * x_q = (global uchar16 *)(src0_q) + offset_src0;
+#endif
+    global uchar * x_e = src0_e + offset_src0;
+
+    ulong offset_src1 = r1 * nb11 + i12 * nb12 + i13 * nb13;
+    global float * y = (global float *)(src1 + offset_src1);
+
+    const short ix = get_sub_group_local_id() >> 1;  // 0...15
+    const short it = get_sub_group_local_id() & 1;  // 0 or 1
+
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        global float4 * y4 = (global float4 *)yb;
+
+        #pragma unroll
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            uchar xb_e = x_e[row * nb + ib];
+#ifdef SRC0Q_IMG
+            ushort4 xb_q = as_ushort4(read_imageui(src0_q, (offset_q + row * nb + ib) * 2 + it).xy);
+#else
+            ushort4 xb_q = vload4(0, (global ushort *)((global uchar *)(x_q + row * nb + ib) + 8 * it));
+#endif
+
+            half4 fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s0);
+            half4 fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s1);
+            float4 acc1 = y4[0] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2);
+            acc1 += y4[4] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3);
+
+            fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s2);
+            fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s3);
+            acc1 += y4[1] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2);
+            acc1 += y4[5] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3);
+
+            sumf[row] += e8m0_to_fp32(xb_e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH/2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}
@@ -0,0 +1,125 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK8_0 32
+typedef struct {
+    half d;       // delta
+    char qs[QK8_0]; // quants
+} block_q8_0;
+
+#define NB_Q8_0 8
+
+#ifdef INTEL_GPU
+#define N_R0_Q8_0 4 // number of rows each subgroup works on
+#define N_SG_Q8_0 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_Q8_0 4
+#define N_SG_Q8_0 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q8_0_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb03,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    ulong         nb13,
+    int           ne0,
+    int           ne1,
+    int           r2,
+    int           r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    int nb = ne00/QK8_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src1 = r1*nb11 + i12*nb12 + i13*nb13;
+    global float * y  = (global float *) (src1 + offset_src1);
+
+    // pointers to src0 rows
+    global block_q8_0 * ax[N_R0_Q8_0];
+    for (int row = 0; row < N_R0_Q8_0; ++row) {
+        ulong offset_src0 = (first_row + row)*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+        ax[row] = (global block_q8_0 *) ((global char *) src0 + offset_src0);
+    }
+
+    float yl[NB_Q8_0];
+    float sumf[N_R0_Q8_0] = { 0.f };
+
+    const short ix = get_sub_group_local_id()/4;
+    const short il = get_sub_group_local_id()%4;
+
+    global float * yb = y + ix*QK8_0 + il*NB_Q8_0;
+
+    // each thread handles NB_Q8_0 quants at a time
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) {
+        for (short i = 0; i < NB_Q8_0; ++i) {
+            yl[i] = yb[i];
+        }
+
+        for (short row = 0; row < N_R0_Q8_0; row++) {
+            global char * qs = ax[row][ib].qs + il*NB_Q8_0;
+            float sumq = 0.f;
+            for (short iq = 0; iq < NB_Q8_0; ++iq) {
+                sumq += qs[iq] * yl[iq];
+            }
+            sumf[row] += sumq*ax[row][ib].d;
+        }
+
+        yb += N_SIMDWIDTH*NB_Q8_0;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_Q8_0; ++row) {
+        float tot = sub_group_reduce_add(sumf[row]);
+
+        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
+            dst_f32[first_row + row] = tot;
+        }
+    }
+}
@@ -0,0 +1,202 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK8_0 32
+typedef struct {
+    half d;       // delta
+    char qs[QK8_0]; // quants
+} block_q8_0;
+
+#define NB_Q8_0 8
+
+#ifdef INTEL_GPU
+#define N_R0_Q8_0 4 // number of rows each subgroup works on
+#define N_SG_Q8_0 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_Q8_0 4
+#define N_SG_Q8_0 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q8_0_f32_flat(
+    global char * src0_q,
+    global half * src0_d,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb03,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    ulong         nb13,
+    int           ne0,
+    int           ne1,
+    int           r2,
+    int           r3
+) {
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    int nb = ne00/QK8_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src1 = r1*nb11 + i12*nb12 + i13*nb13;
+    global float * y  = (global float *) (src1 + offset_src1);
+
+    // pointers to src0 rows
+    uint offset_src0_base = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global char * ax0, * ax1, * ax2, * ax3;
+    global half * ad0, * ad1, * ad2, * ad3;
+    uint offset_src0;
+
+    offset_src0 = offset_src0_base + 0*nb01;
+    offset_src0 = offset_src0/34;
+    ax0 = (global char *) ((global char *) src0_q + offset_src0*sizeof(char)*QK8_0);
+    ad0 = (global half *) ((global char *) src0_d + offset_src0*sizeof(half));
+
+    offset_src0 = offset_src0_base + 1*nb01;
+    offset_src0 = offset_src0/34;
+    ax1 = (global char *) ((global char *) src0_q + offset_src0*sizeof(char)*QK8_0);
+    ad1 = (global half *) ((global char *) src0_d + offset_src0*sizeof(half));
+
+    offset_src0 = offset_src0_base + 2*nb01;
+    offset_src0 = offset_src0/34;
+    ax2 = (global char *) ((global char *) src0_q + offset_src0*sizeof(char)*QK8_0);
+    ad2 = (global half *) ((global char *) src0_d + offset_src0*sizeof(half));
+
+    offset_src0 = offset_src0_base + 3*nb01;
+    offset_src0 = offset_src0/34;
+    ax3 = (global char *) ((global char *) src0_q + offset_src0*sizeof(char)*QK8_0);
+    ad3 = (global half *) ((global char *) src0_d + offset_src0*sizeof(half));
+
+    const short ix = get_sub_group_local_id()/4;
+    const short il = get_sub_group_local_id()%4;
+
+    global float * yb = y + ix*QK8_0 + il*NB_Q8_0;
+
+    float8 yl;
+    float8 qv;
+    float4 sumf = 0.f;
+    float  sumq = 0.f;
+    global char * qs;
+
+    // each thread handles NB_Q8_0 quants at a time
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) {
+        yl = vload8(0, yb);
+
+        qs = ax0 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s0 += sumq*ad0[ib];
+
+        qs = ax1 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s1 += sumq*ad1[ib];
+
+        qs = ax2 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s2 += sumq*ad2[ib];
+
+        qs = ax3 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s3 += sumq*ad3[ib];
+
+        yb += N_SIMDWIDTH*NB_Q8_0;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0),
+        sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2),
+        sub_group_reduce_add(sumf.s3)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst_f32[first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst_f32[first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst_f32[first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst_f32[first_row + 3] = tot.s3;
+        }
+    }
+}
@@ -1,6 +1,6 @@
 #pragma OPENCL EXTENSION cl_khr_fp16 : enable

-kernel void kernel_set_rows_f32(
+kernel void kernel_set_rows_f32_i64(
        global char * src0,
        ulong         offset0,
        global char * src1,
@@ -47,7 +47,7 @@ kernel void kernel_set_rows_f32(
    }
 }

-kernel void kernel_set_rows_f16(
+kernel void kernel_set_rows_f16_i64(
        global char * src0,
        ulong         offset0,
        global char * src1,
@@ -93,3 +93,97 @@ kernel void kernel_set_rows_f16(
        dst_row[ind] = src_row[ind];
    }
 }
+
+kernel void kernel_set_rows_f32_i32(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        int           ne11,
+        int           ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = i03%ne12;
+    int i11 = i02%ne11;
+
+    int i10 = i01;
+    int i1  = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global float * dst_row = (global float *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = (float)src_row[ind];
+    }
+}
+
+kernel void kernel_set_rows_f16_i32(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        int           ne11,
+        int           ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    int i12 = i03%ne12;
+    int i11 = i02%ne11;
+
+    int i10 = i01;
+    int i1  = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global half  * dst_row = (global half  *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = src_row[ind];
+    }
+}
@@ -3721,6 +3721,7 @@ static void quantize_row_iq3_xxs_impl(int grid_size, const float * GGML_RESTRICT
            }
            float best = 0;
            float scale = max/(2*kMaxQ-1);
+            for (int k = 0; k < 8; ++k) is_on_grid[k] = true;
            for (int is = -15; is <= 15; ++is) {
                float id = (2*kMaxQ-1+is*0.2f)/max;
                float this_scale = 1/id;
@@ -795,7 +795,7 @@ static ggml_backend_i ggml_backend_rpc_interface = {
    /* .graph_compute           = */ ggml_backend_rpc_graph_compute,
    /* .event_record            = */ NULL,
    /* .event_wait              = */ NULL,
-    /* .optimize_graph          = */ NULL,
+    /* .graph_optimize          = */ NULL,
 };

 ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) {
@@ -4073,7 +4073,7 @@ static ggml_backend_i ggml_backend_sycl_interface = {
    /* .graph_compute           = */ ggml_backend_sycl_graph_compute,
    /* .event_record            = */ ggml_backend_sycl_event_record,
    /* .event_wait              = */ ggml_backend_sycl_event_wait,
-    /* .optimize_graph          = */ NULL,
+    /* .graph_optimize          = */ NULL,
 };

 static ggml_guid_t ggml_backend_sycl_guid() {
@@ -4271,7 +4271,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                return ((op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16 ||
                         op->type == GGML_TYPE_Q8_0 || op->type == GGML_TYPE_Q5_1 || op->type == GGML_TYPE_Q5_0 ||
                         op->type == GGML_TYPE_Q4_1 || op->type == GGML_TYPE_Q4_0 || op->type == GGML_TYPE_IQ4_NL) &&
-                        (op->src[1]->type == GGML_TYPE_I64));
+                        (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32));
            }
            break;
        case GGML_OP_CPY:
@@ -16,9 +16,9 @@ convert (const char* src, char* dst) {
   *reinterpret_cast<TOut*>(dst) = dst_val;
 }

-template <typename blockType, int qk, cpy_kernel_t cpyblck>
+template <typename TIdx, typename blockType, int qk, cpy_kernel_t cpyblck>
 static void set_rows_sycl_q(const char * __restrict__ src0_d,
-                            const int64_t * __restrict__ src1_d,
+                            const TIdx * __restrict__ src1_d,
                            blockType * __restrict__ dst_d,
                            // tensor dimensions src0 and src1
                            const int64_t ne00,
@@ -66,7 +66,7 @@ static void set_rows_sycl_q(const char * __restrict__ src0_d,
        const size_t  src_offset  = calculate_offset<3>({ nb01, nb02, nb03 }, { i01, i02, i03 });
        const char *  src_block   = src0_d + src_offset + i00 * sizeof(float);
        const size_t  src1_offset = calculate_offset<3>({ nb10, nb11, nb12 }, { i10, i11, i12 });
-        const int64_t dst_row     = src1_d[src1_offset / sizeof(int64_t)];
+        const int64_t dst_row     = src1_d[src1_offset / sizeof(TIdx)];
        const size_t  dst_offset =
            calculate_offset<3>({ nb1, nb2, nb3 }, { dst_row, i02, i03 }) + (i00 / qk) * sizeof(blockType);
        char * dst_block = reinterpret_cast<char *>(reinterpret_cast<char *>(dst_d) + dst_offset);
@@ -78,9 +78,9 @@ static void set_rows_sycl_q(const char * __restrict__ src0_d,
    GGML_UNUSED(nb13);
 }

-template<typename TIn, typename TOut>
+template<typename TIn, typename TIdx, typename TOut>
 static void k_set_rows(
-        const char * __restrict__ src0, const int64_t * __restrict__ src1, char * __restrict__ dst,
+        const char * __restrict__ src0, const TIdx * __restrict__ src1, char * __restrict__ dst,
        const int64_t ne00, const int64_t ne01, const int64_t ne02,
        const int64_t ne11, const int64_t ne12,
        const size_t nb01, const size_t nb02, const size_t nb03,
@@ -104,7 +104,7 @@ static void k_set_rows(
    const int64_t i11 = i02 % ne11;
    const int64_t i10 = i01;

-    const int64_t dst_row = *(const int64_t *)((const char *)src1 + calculate_offset<3>({nb10, nb11, nb12}, {i10, i11, i12}));
+    const int64_t dst_row = *(const TIdx *)((const char *)src1 + calculate_offset<3>({nb10, nb11, nb12}, {i10, i11, i12}));

    const char * src0_row = src0 + calculate_offset<3>({nb01, nb02, nb03}, {i01, i02, i03});
    const char * src_elem = src0_row + i00 * src_type_size;
@@ -114,9 +114,9 @@ static void k_set_rows(
    convert<TIn, TOut>(src_elem, dst_elem);
 }

-template<typename TIn, typename TOut>
+template<typename TIn, typename TIdx, typename TOut>
 static void set_rows_sycl(
-        const char * src0_d, const int64_t * src1_d, char * dst_d,
+        const char * src0_d, const TIdx * src1_d, char * dst_d,
        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
        const int64_t ne11, const int64_t ne12, const size_t nb01, const size_t nb02, const size_t nb03,
        const size_t nb10, const size_t nb11, const size_t nb12,
@@ -132,7 +132,7 @@ static void set_rows_sycl(
    stream->parallel_for(
        sycl::nd_range<1>(grid_size * block_size, block_size),
        [=](sycl::nd_item<1> item_ct1) {
-            k_set_rows<TIn, TOut>(
+            k_set_rows<TIn, TIdx, TOut>(
                src0_d, src1_d, dst_d,
                ne00, ne01, ne02,
                ne11, ne12,
@@ -147,74 +147,69 @@ static void set_rows_sycl(
    );
 }

-void ggml_sycl_op_set_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
-    const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->src[1]->type == GGML_TYPE_I64);
+template<typename TIn, typename TIdx>
+static void set_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const char * src0_d = (const char *)src0->data;
+    const TIdx * src1_d = (const TIdx *)src1->data;

    GGML_TENSOR_BINARY_OP_LOCALS

-    const int64_t * src1_dd = static_cast<const int64_t *>(src1->data);
-
    dpct::queue_ptr stream = ctx.stream();
    switch (dst->type) {
        case GGML_TYPE_F32:
-            set_rows_sycl<float, float>(
-                (const char *)src0->data, src1_dd, (char *)dst->data,
+            set_rows_sycl<TIn, TIdx, float>(
+                src0_d, src1_d, (char *)dst->data,
                ne00, ne01, ne02, ne03,
                ne11, ne12,
                nb01, nb02, nb03,
                nb10, nb11, nb12,
                nb1, nb2, nb3,
-                sizeof(float), sizeof(float),
+                sizeof(TIn), sizeof(float),
                stream
            );
            break;
        case GGML_TYPE_F16:
            dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
-            set_rows_sycl<float, sycl::half>(
-                (const char *)src0->data, src1_dd, (char *)dst->data,
+            set_rows_sycl<TIn, TIdx, sycl::half>(
+                src0_d, src1_d, (char *)dst->data,
                ne00, ne01, ne02, ne03,
                ne11, ne12,
                nb01, nb02, nb03,
                nb10, nb11, nb12,
                nb1, nb2, nb3,
-                sizeof(float), sizeof(sycl::half),
+                sizeof(TIn), sizeof(sycl::half),
                stream
            );
            break;
        case GGML_TYPE_BF16:
-            set_rows_sycl<float, sycl::ext::oneapi::bfloat16>(
-                (const char *)src0->data, src1_dd, (char *)dst->data,
+            set_rows_sycl<TIn, TIdx, sycl::ext::oneapi::bfloat16>(
+                src0_d, src1_d, (char *)dst->data,
                ne00, ne01, ne02, ne03,
                ne11, ne12,
                nb01, nb02, nb03,
                nb10, nb11, nb12,
                nb1, nb2, nb3,
-                sizeof(float), sizeof(sycl::ext::oneapi::bfloat16),
+                sizeof(TIn), sizeof(sycl::ext::oneapi::bfloat16),
                stream
            );
            break;
        case GGML_TYPE_Q8_0:
-            set_rows_sycl_q<block_q8_0, QK8_0, cpy_blck_f32_q8_0>((const char *)src0->data, src1_dd, (block_q8_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIdx, block_q8_0, QK8_0, cpy_blck_f32_q8_0>(src0_d, src1_d, (block_q8_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q5_1:
-            set_rows_sycl_q<block_q5_1, QK5_1, cpy_blck_f32_q5_1>((const char *)src0->data, src1_dd, (block_q5_1 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIdx, block_q5_1, QK5_1, cpy_blck_f32_q5_1>(src0_d, src1_d, (block_q5_1 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q5_0:
-            set_rows_sycl_q<block_q5_0, QK5_0, cpy_blck_f32_q5_0>((const char *)src0->data, src1_dd, (block_q5_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIdx, block_q5_0, QK5_0, cpy_blck_f32_q5_0>(src0_d, src1_d, (block_q5_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q4_1:
-            set_rows_sycl_q<block_q4_1, QK4_1, cpy_blck_f32_q4_1>((const char *)src0->data, src1_dd, (block_q4_1 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIdx, block_q4_1, QK4_1, cpy_blck_f32_q4_1>(src0_d, src1_d, (block_q4_1 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q4_0:
-            set_rows_sycl_q<block_q4_0, QK4_0, cpy_blck_f32_q4_0>((const char *)src0->data, src1_dd, (block_q4_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIdx, block_q4_0, QK4_0, cpy_blck_f32_q4_0>(src0_d, src1_d, (block_q4_0 *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_IQ4_NL:
-            set_rows_sycl_q<block_iq4_nl, QK4_NL, cpy_blck_f32_iq4_nl>((const char *)src0->data, src1_dd, (block_iq4_nl *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
+            set_rows_sycl_q<TIdx, block_iq4_nl, QK4_NL, cpy_blck_f32_iq4_nl>(src0_d, src1_d, (block_iq4_nl *)dst->data, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nb00, nb01, nb02, nb03, nb10, nb11, nb12, nb13, nb1, nb2, nb3, stream);
            break;

        default:
@@ -222,3 +217,18 @@ void ggml_sycl_op_set_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
            break;
    }
 }
+
+void ggml_sycl_op_set_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->src[1]->type == GGML_TYPE_I64 || dst->src[1]->type == GGML_TYPE_I32);
+
+    if (src1->type == GGML_TYPE_I64) {
+        set_rows_sycl<float, int64_t>(ctx, src0, src1, dst);
+    } else {
+        set_rows_sycl<float, int32_t>(ctx, src0, src1, dst);
+    }
+}
@@ -520,7 +520,8 @@ struct vk_device_struct {
    vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16, pipeline_contig_cpy_f16_f32, pipeline_contig_cpy_f32_bf16, pipeline_contig_cpy_f32_i32, pipeline_contig_cpy_i32_f32;
    vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
    vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
-    vk_pipeline pipeline_set_rows[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_set_rows_i32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_set_rows_i64[GGML_TYPE_COUNT];
    vk_pipeline pipeline_norm_f32;
    vk_pipeline pipeline_group_norm_f32;
    vk_pipeline pipeline_rms_norm_f32;
@@ -574,6 +575,8 @@ struct vk_device_struct {
    vk_pipeline pipeline_opt_step_sgd_f32;
    vk_pipeline pipeline_conv2d_f32[CONV_SHAPE_COUNT];
    vk_pipeline pipeline_conv2d_f16_f32[CONV_SHAPE_COUNT];
+    vk_pipeline pipeline_conv_transpose_2d_f32[CONV_SHAPE_COUNT];
+    vk_pipeline pipeline_conv_transpose_2d_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;

@@ -593,7 +596,7 @@ struct vk_device_struct {
    bool disable_fusion;
    bool disable_host_visible_vidmem;
    bool allow_sysmem_fallback;
-    bool disable_optimize_graph;
+    bool disable_graph_optimize;

 #ifdef GGML_VULKAN_MEMORY_DEBUG
    std::unique_ptr<vk_memory_logger> memory_logger;
@@ -1117,6 +1120,56 @@ 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_conv_transpose_2d_push_constants {
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t N;
+
+    uint32_t KW;
+    uint32_t KH;
+    uint32_t W;
+    uint32_t H;
+    uint32_t OW;
+    uint32_t OH;
+
+    uint32_t s0;
+    uint32_t s1;
+    uint32_t p0;
+    uint32_t p1;
+    uint32_t d0;
+    uint32_t d1;
+
+    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;
+
+    // init_fastdiv_values constants for dividing by KW, KW*KH, OW, OW*OH, s0, s1
+    uint32_t KWmp;   uint32_t KWL;
+    uint32_t KWKHmp; uint32_t KWKHL;
+    uint32_t OWmp;   uint32_t OWL;
+    uint32_t OWOHmp; uint32_t OWOHL;
+    uint32_t s0mp; uint32_t s0L;
+    uint32_t s1mp; uint32_t s1L;
+};
+
+template <> void init_pushconst_fastdiv(vk_op_conv_transpose_2d_push_constants &p) {
+    // Compute magic values to divide by KW, KW*KH, OW, OW*OH, s0, s1
+    init_fastdiv_values(p.KW,       p.KWmp,    p.KWL);
+    init_fastdiv_values(p.KW*p.KH,  p.KWKHmp,  p.KWKHL);
+    init_fastdiv_values(p.OW,       p.OWmp,    p.OWL);
+    init_fastdiv_values(p.OW*p.OH,  p.OWOHmp,  p.OWOHL);
+    init_fastdiv_values(p.s0,       p.s0mp,    p.s0L);
+    init_fastdiv_values(p.s1,       p.s1mp,    p.s1L);
+}
+
 struct vk_op_conv2d_dw_push_constants {
    uint32_t ne;
    uint32_t batches;
@@ -1185,6 +1238,14 @@ struct vk_staging_memcpy {
    size_t n;
 };

+struct vk_staging_memset {
+    vk_staging_memset(void * _dst, uint32_t _val, size_t _n) : dst(_dst), val(_val), n(_n) {}
+
+    void * dst;
+    uint32_t val;
+    size_t n;
+};
+
 struct vk_context_struct {
    vk_submission * s;
    std::vector<vk_sequence> seqs;
@@ -1193,6 +1254,7 @@ struct vk_context_struct {

    std::vector<vk_staging_memcpy> in_memcpys;
    std::vector<vk_staging_memcpy> out_memcpys;
+    std::vector<vk_staging_memset> memsets;

    vk_command_pool * p {};
 };
@@ -1313,7 +1375,7 @@ class vk_perf_logger {
            flops[name].push_back(m * n * (k + (k - 1)) * batch);
            return;
        }
-        if (node->op == GGML_OP_CONV_2D) {
+        if (node->op == GGML_OP_CONV_2D || node->op == GGML_OP_CONV_TRANSPOSE_2D) {
            std::string   name    = ggml_op_name(node->op);
            ggml_tensor * knl     = node->src[0];
            uint64_t      OW      = node->ne[0];
@@ -1322,7 +1384,7 @@ class vk_perf_logger {
            uint64_t      Cout    = node->ne[2];
            uint64_t      KW      = knl->ne[0];
            uint64_t      KH      = knl->ne[1];
-            uint64_t      Cin     = knl->ne[2];
+            uint64_t      Cin     = node->src[1]->ne[2];
            // KxCRS @ CRSxNPQ = KxNPQ -> M=K, K=CRS, N=NPQ
            uint64_t      size_M  = Cout;
            uint64_t      size_K  = Cin * KW * KH;
@@ -1584,7 +1646,9 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
    }

    vk::ComputePipelineCreateInfo compute_pipeline_create_info(
-        vk::PipelineCreateFlags{},
+        device->pipeline_executable_properties_support ?
+            vk::PipelineCreateFlagBits::eCaptureStatisticsKHR :
+            vk::PipelineCreateFlags{},
        pipeline_shader_create_info,
        pipeline->layout);

@@ -3285,27 +3349,26 @@ static void ggml_vk_load_shaders(vk_device& device) {
        ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
    }

+#define SET_ROWS(itype, rte) \
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_F32],  "set_rows_f32" #itype,  set_rows_f32 ## itype ## rte ## _len,  set_rows_f32 ## itype ## rte ## _data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_F16],  "set_rows_f16" #itype,  set_rows_f16 ## itype ## rte ## _len,  set_rows_f16 ## itype ## rte ## _data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_BF16], "set_rows_bf16" #itype, set_rows_bf16 ## itype ## rte ## _len, set_rows_bf16 ## itype ## rte ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q4_0], "set_rows_q4_0" #itype, set_rows_q4_0 ## itype ## rte ## _len, set_rows_q4_0 ## itype ## rte ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q4_1], "set_rows_q4_1" #itype, set_rows_q4_1 ## itype ## rte ## _len, set_rows_q4_1 ## itype ## rte ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q5_0], "set_rows_q5_0" #itype, set_rows_q5_0 ## itype ## rte ## _len, set_rows_q5_0 ## itype ## rte ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q5_1], "set_rows_q5_1" #itype, set_rows_q5_1 ## itype ## rte ## _len, set_rows_q5_1 ## itype ## rte ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q8_0], "set_rows_q8_0" #itype, set_rows_q8_0 ## itype ## rte ## _len, set_rows_q8_0 ## itype ## rte ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_IQ4_NL], "set_rows_iq4_nl" #itype, set_rows_iq4_nl ## itype ## rte ## _len, set_rows_iq4_nl ## itype ## rte ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+
    if (device->float_controls_rte_fp16) {
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_F32],  "set_rows_f32",  set_rows_f32_rte_len,  set_rows_f32_rte_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_F16],  "set_rows_f16",  set_rows_f16_rte_len,  set_rows_f16_rte_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_BF16], "set_rows_bf16", set_rows_bf16_rte_len, set_rows_bf16_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q4_0], "set_rows_q4_0", set_rows_q4_0_rte_len, set_rows_q4_0_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q4_1], "set_rows_q4_1", set_rows_q4_1_rte_len, set_rows_q4_1_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q5_0], "set_rows_q5_0", set_rows_q5_0_rte_len, set_rows_q5_0_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q5_1], "set_rows_q5_1", set_rows_q5_1_rte_len, set_rows_q5_1_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q8_0], "set_rows_q8_0", set_rows_q8_0_rte_len, set_rows_q8_0_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_IQ4_NL], "set_rows_iq4_nl", set_rows_iq4_nl_rte_len, set_rows_iq4_nl_rte_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        SET_ROWS(_i32, _rte)
+        SET_ROWS(_i64, _rte)
    } else {
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_F32],  "set_rows_f32",  set_rows_f32_len,  set_rows_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_F16],  "set_rows_f16",  set_rows_f16_len,  set_rows_f16_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_BF16], "set_rows_bf16", set_rows_bf16_len, set_rows_bf16_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q4_0], "set_rows_q4_0", set_rows_q4_0_len, set_rows_q4_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q4_1], "set_rows_q4_1", set_rows_q4_1_len, set_rows_q4_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q5_0], "set_rows_q5_0", set_rows_q5_0_len, set_rows_q5_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q5_1], "set_rows_q5_1", set_rows_q5_1_len, set_rows_q5_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_Q8_0], "set_rows_q8_0", set_rows_q8_0_len, set_rows_q8_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
-        ggml_vk_create_pipeline(device, device->pipeline_set_rows[GGML_TYPE_IQ4_NL], "set_rows_iq4_nl", set_rows_iq4_nl_len, set_rows_iq4_nl_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
+        SET_ROWS(_i32, )
+        SET_ROWS(_i64, )
    }
+#undef SET_ROWS
+

    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_0], "cpy_q4_0_f32", cpy_q4_0_f32_len, cpy_q4_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_1], "cpy_q4_1_f32", cpy_q4_1_f32_len, cpy_q4_1_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
@@ -3380,7 +3443,6 @@ static void ggml_vk_load_shaders(vk_device& device) {
    ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);  \
    ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);

-    CREATE_UNARY(exp)
    CREATE_UNARY(gelu)
    CREATE_UNARY(gelu_erf)
    CREATE_UNARY(gelu_quick)
@@ -3392,6 +3454,17 @@ static void ggml_vk_load_shaders(vk_device& device) {
    CREATE_UNARY(hardswish)
 #undef CREATE_UNARY

+#define CREATE_UNARY_RTE(name)  \
+    if (device->float_controls_rte_fp16) {  \
+        ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32_rte", name ## _f32_rte_len, name ## _f32_rte_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);   \
+        ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16_rte", name ## _f16_rte_len, name ## _f16_rte_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);   \
+    } else {    \
+        ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);   \
+        ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);   \
+    }
+    CREATE_UNARY_RTE(exp)
+#undef CREATE_UNARY_RTE
+
 #define CREATE_GLU(name)  \
    if (device->float_controls_rte_fp16) {  \
        ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32_rte", name ## _f32_rte_len, name ## _f32_rte_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true);   \
@@ -3471,7 +3544,7 @@ static void ggml_vk_load_shaders(vk_device& device) {

    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
+    // conv2d, conv_transpose_2d
    for (uint32_t s = 0; s < CONV_SHAPE_COUNT; ++s) {
        uint32_t conv2d_WG_SIZE  = 256;
        uint32_t conv2d_BS_K     = 128;
@@ -3546,31 +3619,30 @@ static void ggml_vk_load_shaders(vk_device& device) {
        std::array<uint32_t, 3> wg_denoms = { conv2d_BS_K, conv2d_BS_NPQ, 1 };
        std::vector<uint32_t> spec_constants = { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD };

+#define CREATE_CONV(name, type_suffix, spv_suffix) \
+        ggml_vk_create_pipeline( \
+            device, device->pipeline_##name##type_suffix[s], #name #type_suffix, \
+            name##type_suffix##spv_suffix##_len, name##type_suffix##spv_suffix##_data, "main", 3, \
+            sizeof(vk_op_##name##_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+#define CREATE_CONVS(spv_suffix) \
+        CREATE_CONV(conv2d, _f32, spv_suffix) \
+        CREATE_CONV(conv2d, _f16_f32, spv_suffix) \
+        if (device->properties.limits.maxPushConstantsSize >= sizeof(vk_op_conv_transpose_2d_push_constants)) { \
+            CREATE_CONV(conv_transpose_2d, _f32, spv_suffix) \
+            CREATE_CONV(conv_transpose_2d, _f16_f32, spv_suffix) \
+        }
 #if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
        if (device->coopmat2) {
-            ggml_vk_create_pipeline(
-                device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_cm2_len, conv2d_f32_cm2_data, "main", 3,
-                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
-            ggml_vk_create_pipeline(
-                device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_cm2_len, conv2d_f16_f32_cm2_data, "main", 3,
-                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+            CREATE_CONVS(_cm2)
        } else
 #endif
        if (conv2d_UNROLL) {
-            ggml_vk_create_pipeline(
-                device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_unroll_len, conv2d_f32_unroll_data, "main", 3,
-                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
-            ggml_vk_create_pipeline(
-                device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_unroll_len, conv2d_f16_f32_unroll_data, "main", 3,
-                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+            CREATE_CONVS(_unroll)
        } else {
-            ggml_vk_create_pipeline(
-                device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
-                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
-            ggml_vk_create_pipeline(
-                device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_len, conv2d_f16_f32_data, "main", 3,
-                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+            CREATE_CONVS( )
        }
+#undef CREATE_CONV
+#undef CREATE_CONVS
    }

    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);
@@ -3624,8 +3696,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
        const char* GGML_VK_ALLOW_SYSMEM_FALLBACK = getenv("GGML_VK_ALLOW_SYSMEM_FALLBACK");
        device->allow_sysmem_fallback = GGML_VK_ALLOW_SYSMEM_FALLBACK != nullptr;

-        const char* GGML_VK_DISABLE_OPTIMIZE_GRAPH = getenv("GGML_VK_DISABLE_OPTIMIZE_GRAPH");
-        device->disable_optimize_graph = GGML_VK_DISABLE_OPTIMIZE_GRAPH != nullptr;
+        const char* GGML_VK_DISABLE_GRAPH_OPTIMIZE = getenv("GGML_VK_DISABLE_GRAPH_OPTIMIZE");
+        device->disable_graph_optimize = GGML_VK_DISABLE_GRAPH_OPTIMIZE != nullptr;

        bool fp16_storage = false;
        bool fp16_compute = false;
@@ -5194,6 +5266,14 @@ static void deferred_memcpy(void * dst, const void * src, size_t size, std::vect
    }
 }

+static void deferred_memset(void * dst, uint32_t val, size_t size, std::vector<vk_staging_memset>* memsets = nullptr) {
+    if (memsets == nullptr) {
+        memset(dst, val, size);
+    } else {
+        memsets->emplace_back(dst, val, size);
+    }
+}
+
 static void ggml_vk_ensure_sync_staging_buffer(vk_device& device, size_t size) {
    if (device->sync_staging == nullptr || device->sync_staging->size < size) {
        VK_LOG_MEMORY("ggml_vk_ensure_sync_staging_buffer(" << size << ")");
@@ -5389,6 +5469,10 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void *
            memcpy(cpy.dst, cpy.src, cpy.n);
        }

+        for (auto& mset : subctx->memsets) {
+            memset(mset.dst, mset.val, mset.n);
+        }
+
        ggml_vk_submit(subctx, dst->device->fence);
        VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences");
        dst->device->device.resetFences({ dst->device->fence });
@@ -5528,12 +5612,25 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr
 static void ggml_vk_buffer_memset_async(vk_context& ctx, vk_buffer& dst, size_t offset, uint32_t c, size_t size) {
    VK_LOG_DEBUG("ggml_vk_buffer_memset_async(" << offset << ", " << c << ", " << size << ")");

+    if (dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible &&
+        dst->device->uma) {
+        deferred_memset((uint8_t*)dst->ptr + offset, c, size, &ctx->memsets);
+        return;
+    }
+
+    // Fall back to GPU fillBuffer for non-UMA or non-host-visible buffers
    ctx->s->buffer.fillBuffer(dst->buffer, offset, size, c);
 }

 static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, size_t size) {
    VK_LOG_DEBUG("ggml_vk_buffer_memset(" << offset << ", " << c << ", " << size << ")");

+    if (dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible &&
+        dst->device->uma) {
+        memset((uint8_t*)dst->ptr + offset, c, size);
+        return;
+    }
+
    std::lock_guard<std::recursive_mutex> guard(dst->device->mutex);
    vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue.cmd_pool);
    ggml_vk_ctx_begin(dst->device, subctx);
@@ -7502,6 +7599,33 @@ static std::array<uint32_t, 3> ggml_vk_get_conv_elements(const ggml_tensor *dst)
    return elements;
 }

+static std::array<uint32_t, 3> ggml_vk_get_conv_transpose_2d_elements(const ggml_tensor *dst) {
+    const ggml_tensor *src0 = dst->src[0];
+    const ggml_tensor *src1 = dst->src[1];
+
+    // src0 - kernel:   [KW, KH, Cout, Cin]
+    // src1 - input:    [W, H, Cin, N]
+    // dst - result:    [OW, OH, Cout, N]
+
+    auto calc_conv_output_size = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
+        return (ins - 1) * s - 2 * p + (ks - 1) * d + 1;
+    };
+    // parallelize in {OW/BS_K, OH/BS_NPQ, 1}
+    int64_t W    = src1->ne[0];
+    int64_t H    = src1->ne[1];
+    int64_t KW   = src0->ne[0];
+    int64_t KH   = src0->ne[1];
+    int64_t Cout = src0->ne[2];
+    int64_t N    = src1->ne[3];
+    int64_t OH   = calc_conv_output_size(H, KH, dst->op_params[0], 0, 1);
+    int64_t OW   = calc_conv_output_size(W, KW, dst->op_params[0], 0, 1);
+    int64_t NPQ  = N * OW * OH;
+
+    // Tile output matrix to (K/NB_K, NPQ/NB_NPQ, 1) workgroups
+    std::array<uint32_t, 3> elements = { static_cast<uint32_t>(Cout), static_cast<uint32_t>(NPQ), 1 };
+    return elements;
+}
+
 static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * dst, ggml_op op) {
    switch (op) {
    case GGML_OP_GET_ROWS:
@@ -7648,7 +7772,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
    case GGML_OP_DUP:
        return ggml_vk_get_cpy_pipeline(ctx, src0, dst, dst->type);
    case GGML_OP_SET_ROWS:
-        return ctx->device->pipeline_set_rows[dst->type];
+        if (src1->type == GGML_TYPE_I64) {
+            return ctx->device->pipeline_set_rows_i64[dst->type];
+        } else {
+            return ctx->device->pipeline_set_rows_i32[dst->type];
+        }
    case GGML_OP_SILU_BACK:
        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
            return ctx->device->pipeline_silu_back_f32;
@@ -7879,9 +8007,12 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
        }
        return nullptr;
    case GGML_OP_CONV_2D:
+    case GGML_OP_CONV_TRANSPOSE_2D:
        if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 &&
            ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
-            auto elements = ggml_vk_get_conv_elements(dst);
+            std::array<uint32_t, 3> elements;
+            if (op == GGML_OP_CONV_2D) elements = ggml_vk_get_conv_elements(dst);
+            else if (op == GGML_OP_CONV_TRANSPOSE_2D) elements = ggml_vk_get_conv_transpose_2d_elements(dst);
            vk_conv_shapes shape;

            uint32_t tiles[CONV_SHAPE_COUNT];
@@ -7901,10 +8032,18 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                shape = CONV_SHAPE_64x32;
            }

-            if (src0->type == GGML_TYPE_F32) {
-                return ctx->device->pipeline_conv2d_f32[shape];
-            } else if (src0->type == GGML_TYPE_F16) {
-                return ctx->device->pipeline_conv2d_f16_f32[shape];
+            if (op == GGML_OP_CONV_2D) {
+                if (src0->type == GGML_TYPE_F32) {
+                    return ctx->device->pipeline_conv2d_f32[shape];
+                } else if (src0->type == GGML_TYPE_F16) {
+                    return ctx->device->pipeline_conv2d_f16_f32[shape];
+                }
+            } else if (op == GGML_OP_CONV_TRANSPOSE_2D) {
+                if (src0->type == GGML_TYPE_F32) {
+                    return ctx->device->pipeline_conv_transpose_2d_f32[shape];
+                } else if (src0->type == GGML_TYPE_F16) {
+                    return ctx->device->pipeline_conv_transpose_2d_f16_f32[shape];
+                }
            }
        }
        return nullptr;
@@ -8304,6 +8443,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
        {
            elements = ggml_vk_get_conv_elements(dst);
        } break;
+    case GGML_OP_CONV_TRANSPOSE_2D:
+        {
+            elements = ggml_vk_get_conv_transpose_2d_elements(dst);
+        } break;
    case GGML_OP_ADD:
    case GGML_OP_SUB:
    case GGML_OP_DIV:
@@ -9477,6 +9620,55 @@ static void ggml_vk_conv_2d(ggml_backend_vk_context * ctx, vk_context & subctx,
    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_2D, std::move(p), dryrun);
 }

+static void ggml_vk_conv_transpose_2d(ggml_backend_vk_context * ctx, vk_context & subctx, const ggml_tensor * src0,
+                                      const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    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_conv_transpose_2d_push_constants p{};
+    p.Cout = static_cast<uint32_t>(ne02);
+    p.Cin  = static_cast<uint32_t>(ne03);
+    p.N    = static_cast<uint32_t>(ne13);
+
+    p.KW = static_cast<uint32_t>(ne00);
+    p.KH = static_cast<uint32_t>(ne01);
+    p.W  = static_cast<uint32_t>(ne10);
+    p.H  = static_cast<uint32_t>(ne11);
+    p.OW = static_cast<uint32_t>(ne0);
+    p.OH = static_cast<uint32_t>(ne1);
+
+    p.s0 = static_cast<uint32_t>(dst->op_params[0]);
+    p.s1 = static_cast<uint32_t>(dst->op_params[0]);
+    p.p0 = 0;
+    p.p1 = 0;
+    p.d0 = 1;
+    p.d1 = 1;
+
+    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_ASSERT(ne02 == ne2);
+    GGML_ASSERT(ne03 == ne12);
+
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_TRANSPOSE_2D, std::move(p), dryrun);
+}
+
 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, bool dryrun = false) {
    vk_op_conv2d_dw_push_constants p{};
    p.ne = ggml_nelements(dst);
@@ -10569,6 +10761,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
    case GGML_OP_CONV_TRANSPOSE_1D:
    case GGML_OP_POOL_2D:
    case GGML_OP_CONV_2D:
+    case GGML_OP_CONV_TRANSPOSE_2D:
    case GGML_OP_CONV_2D_DW:
    case GGML_OP_RWKV_WKV6:
    case GGML_OP_RWKV_WKV7:
@@ -10640,6 +10833,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
        case GGML_OP_CONV_TRANSPOSE_1D:
        case GGML_OP_POOL_2D:
        case GGML_OP_CONV_2D:
+        case GGML_OP_CONV_TRANSPOSE_2D:
        case GGML_OP_CONV_2D_DW:
        case GGML_OP_LEAKY_RELU:
        case GGML_OP_OPT_STEP_SGD:
@@ -10951,6 +11145,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
    case GGML_OP_CONV_2D:
        ggml_vk_conv_2d(ctx, compute_ctx, src0, src1, node, dryrun);

+        break;
+    case GGML_OP_CONV_TRANSPOSE_2D:
+        ggml_vk_conv_transpose_2d(ctx, compute_ctx, src0, src1, node, dryrun);
+
        break;
    case GGML_OP_CONV_2D_DW:
        ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -11091,6 +11289,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
    case GGML_OP_CONV_TRANSPOSE_1D:
    case GGML_OP_POOL_2D:
    case GGML_OP_CONV_2D:
+    case GGML_OP_CONV_TRANSPOSE_2D:
    case GGML_OP_CONV_2D_DW:
    case GGML_OP_RWKV_WKV6:
    case GGML_OP_RWKV_WKV7:
@@ -11168,6 +11367,10 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
            memcpy(cpy.dst, cpy.src, cpy.n);
        }

+        for (auto& mset : subctx->memsets) {
+            memset(mset.dst, mset.val, mset.n);
+        }
+
        if (almost_ready && !ctx->almost_ready_fence_pending && !use_fence) {
            ggml_vk_submit(subctx, ctx->almost_ready_fence);
            ctx->almost_ready_fence_pending = true;
@@ -11190,6 +11393,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
        }
        subctx->in_memcpys.clear();
        subctx->out_memcpys.clear();
+        subctx->memsets.clear();
    }

    return true;
@@ -11743,10 +11947,10 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
        ggml_vk_build_graph(ctx, cgraph, i, nullptr, 0, true, false, false, false);
        if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
            total_mat_mul_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
-        } else if (cgraph->nodes[i]->op == GGML_OP_CONV_2D) {
+        } else if (cgraph->nodes[i]->op == GGML_OP_CONV_2D || cgraph->nodes[i]->op == GGML_OP_CONV_TRANSPOSE_2D) {
            // Return CRSxNPQxsizeof(*) to account as many bytes as mul_mat has in im2col->mul_mat mode.
            auto CRS_size =
-                cgraph->nodes[i]->src[0]->ne[0] * cgraph->nodes[i]->src[0]->ne[1] * cgraph->nodes[i]->src[0]->ne[2];
+                cgraph->nodes[i]->src[0]->ne[0] * cgraph->nodes[i]->src[0]->ne[1] * cgraph->nodes[i]->src[1]->ne[2];
            auto NPQ_size = cgraph->nodes[i]->ne[0] * cgraph->nodes[i]->ne[1] * cgraph->nodes[i]->ne[3];
            total_mat_mul_bytes += NPQ_size * CRS_size * ggml_type_size(cgraph->nodes[i]->type);
        }
@@ -11914,12 +12118,12 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
 }

 // Sort the graph for improved parallelism.
-static void ggml_vk_optimize_graph(ggml_backend_t backend, struct ggml_cgraph * graph)
+static void ggml_vk_graph_optimize(ggml_backend_t backend, struct ggml_cgraph * graph)
 {
-    VK_LOG_DEBUG("ggml_vk_optimize_graph(" << graph->n_nodes << " nodes)");
+    VK_LOG_DEBUG("ggml_vk_graph_optimize(" << graph->n_nodes << " nodes)");
    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;

-    if (ctx->device->disable_optimize_graph) {
+    if (ctx->device->disable_graph_optimize) {
        return;
    }

@@ -12053,7 +12257,7 @@ static ggml_backend_i ggml_backend_vk_interface = {
    /* .graph_compute           = */ ggml_backend_vk_graph_compute,
    /* .event_record            = */ NULL,
    /* .event_wait              = */ NULL,
-    /* .optimize_graph          = */ ggml_vk_optimize_graph,
+    /* .graph_optimize          = */ ggml_vk_graph_optimize,
 };

 static ggml_guid_t ggml_backend_vk_guid() {
@@ -12567,10 +12771,15 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
        case GGML_OP_CONV_TRANSPOSE_1D:
            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
        case GGML_OP_CONV_2D:
+        case GGML_OP_CONV_TRANSPOSE_2D:
            {
                // Op is disabled for Apple because it segfaults at pipeline create time on MoltenVK
                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
                const vk_device& device = ggml_vk_get_device(ctx->device);
+                if (op->op == GGML_OP_CONV_TRANSPOSE_2D &&
+                    device->properties.limits.maxPushConstantsSize < sizeof(vk_op_conv_transpose_2d_push_constants)) {
+                    return false;
+                }
                // Channel-contiguous format is not supported yet.
                return ((op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
                    op->src[1]->type == GGML_TYPE_F32 &&
@@ -13189,6 +13398,9 @@ 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_TRANSPOSE_2D) {
+        const int32_t s = tensor->op_params[0];
+        tensor_clone = ggml_conv_transpose_2d_p0(ggml_ctx, src_clone[0], src_clone[1], s);
    } else if (tensor->op == GGML_OP_LEAKY_RELU) {
        const float * op_params = (const float *)tensor->op_params;
        tensor_clone = ggml_leaky_relu(ggml_ctx, src_clone[0], op_params[0], false);
@@ -16,7 +16,7 @@
 // shape notation: [dim(N), ..., dim(0)] -- stride(dim(j)) >= stride(dim(i)) if i > j
 layout(binding = 0) readonly buffer A {
    A_TYPE knl_data[];
-};  // src0 - kernel:   [KW, KH, Cin, Cout]
+};  // src0 - kernel:   [KW, KH, Cin, Cout] for conv_2d, [KW, KH, Cout, Cin] for conv_transposed_2d

 layout(binding = 1) readonly buffer B {
    B_TYPE src_data[];
@@ -66,6 +66,10 @@ layout(push_constant) uniform parameter {
    uint32_t KWKHmp; uint32_t KWKHL;
    uint32_t OWmp;   uint32_t OWL;
    uint32_t OWOHmp; uint32_t OWOHL;
+#ifdef TRANSPOSE
+    uint32_t s0mp; uint32_t s0L;
+    uint32_t s1mp; uint32_t s1L;
+#endif
 }

 p;
@@ -225,7 +229,11 @@ void main() {
            uint32_t B_ly    = r_offset + Ar;
            uint32_t B_lx    = Ac;
            uint32_t K_idx   = B_idx_K * BS_K + B_ly; /* Global K_idx (row index of A)*/
+#ifdef TRANSPOSE
+            uint32_t knl_idx = min(KW_idx_a + KH_idx_a * p.nb01 + K_idx * p.nb02 + Cin_idx_a * p.nb03, K * CRS - 1);
+#else
            uint32_t knl_idx = min(KW_idx_a + KH_idx_a * p.nb01 + Cin_idx_a * p.nb02 + K_idx * p.nb03, K * CRS - 1);
+#endif
            float    val     = knl_data[knl_idx];
            if (K_idx >= K || CRS_idx_a >= CRS) {
                val = 0.0;
@@ -267,12 +275,24 @@ void main() {
            KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
 #endif

+#ifdef TRANSPOSE
+            uint32_t H_idx_x_s1 = OH_idx - KH_idx_b * p.d1 + p.p1;
+            uint32_t W_idx_x_s0 = OW_idx - KW_idx_b * p.d0 + p.p0;
+            uint32_t H_idx = fastdiv(H_idx_x_s1, p.s1mp, p.s1L);
+            uint32_t W_idx = fastdiv(W_idx_x_s0, p.s0mp, p.s0L);
+#else
            uint32_t H_idx = OH_idx * p.s1 + KH_idx_b * p.d1 - p.p1;
            uint32_t W_idx = OW_idx * p.s0 + KW_idx_b * p.d0 - p.p0;
+#endif
            uint32_t src_idx =
                min(max(W_idx + H_idx * p.nb11 + Cin_idx_b * p.nb12 + N_idx * p.nb13, 0), p.Cin * p.N * p.W * p.H - 1);
            float val = src_data[src_idx];
-            if (CRS_idx_b >= CRS || NPQ_idx >= NPQ || H_idx < 0 || H_idx >= p.H || W_idx < 0 || W_idx >= p.W) {
+            if (CRS_idx_b >= CRS || NPQ_idx >= NPQ
+                || H_idx >= p.H || W_idx >= p.W // Lower bound checks aren't necessary. (idx >= 0x80000000 for such case)
+#ifdef TRANSPOSE
+                || (H_idx_x_s1 - H_idx * p.s1 != 0) || (W_idx_x_s0 - W_idx * p.s0 != 0)
+#endif
+                ) {
                val = 0.0;
            }
            Bsh[B_ly * Bsh_stride + B_lx] = SHMEM_TYPE(val);
@@ -15,8 +15,15 @@ layout (binding = 0) readonly buffer S {float data_s[];};

 #if defined(SET_ROWS)
 #include "generic_binary_head.comp"
-layout (binding = 1) readonly buffer C {uvec2 data_i[];};
+layout (binding = 1) readonly buffer C {B_TYPE data_i[];};
 layout (binding = 2) writeonly buffer Q {A_TYPE data_q[];};
+
+#if B_SIZE == 64
+#define DATA_I_SWIZZLE .x
+#else
+#define DATA_I_SWIZZLE
+#endif
+
 #else
 #include "generic_unary_head.comp"
 layout (binding = 1) writeonly buffer Q {A_TYPE data_q[];};
@@ -259,7 +266,7 @@ void main() {
    uint i11 = fastmod(i02, p.ne11);
    uint i10 = i01;

-    uint i1 = data_i[src1_idx(i10, i11, i12, 0) + get_boffset()].x;
+    uint i1 = data_i[src1_idx(i10, i11, i12, 0) + get_boffset()] DATA_I_SWIZZLE;

    uint src0_idx = src0_idx(i00, i01, i02, i03) + get_aoffset();
    uint dst_idx = dst_idx(i00 / QUANT_K, i1, i02, i03) + get_doffset();
@@ -1,5 +1,6 @@
 #version 450

+#include "rte.comp"
 #include "generic_head.comp"
 #include "types.comp"

@@ -37,6 +37,18 @@
 #define LOAD_VEC_B 1
 #endif

+// Load 2 values at once without affecting index calculations through LOAD_VEC
+#if (defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)) && !defined(ALIGNED)
+#define LOAD_VEC_BATCH_A 2
+#else
+#define LOAD_VEC_BATCH_A 1
+#endif
+#if !defined(ALIGNED)
+#define LOAD_VEC_BATCH_B 2
+#else
+#define LOAD_VEC_BATCH_B 1
+#endif
+
 #if !defined(TO_FLOAT_TYPE)
 #define TO_FLOAT_TYPE FLOAT_TYPE
 #endif
@@ -98,13 +110,13 @@ layout (constant_id = 9) const uint TK = 1;  // Only needed for coopmat
 layout (constant_id = 10) const uint WARP = 32;

 #ifdef COOPMAT
-#define SHMEM_STRIDE (BK + 8)
+#define SHMEM_STRIDE (BK / 2 + 4)
 #else
-#define SHMEM_STRIDE (BK + 1)
+#define SHMEM_STRIDE (BK / 2 + 1)
 #endif

-shared FLOAT_TYPE buf_a[BM * SHMEM_STRIDE];
-shared FLOAT_TYPE buf_b[BN * SHMEM_STRIDE];
+shared FLOAT_TYPE_VEC2 buf_a[BM * SHMEM_STRIDE];
+shared FLOAT_TYPE_VEC2 buf_b[BN * SHMEM_STRIDE];

 #define NUM_WARPS (BLOCK_SIZE / WARP)

@@ -236,13 +248,13 @@ void main() {
    const uint warp_r = warp_i % (BM / WM);
    const uint warp_c = warp_i / (BM / WM);

-    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A);
-    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A);
-    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B);
-    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B);
+    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A / LOAD_VEC_BATCH_A);
+    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A / LOAD_VEC_BATCH_A);
+    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B / LOAD_VEC_BATCH_B);
+    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B / LOAD_VEC_BATCH_B);

-    const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A / BK;
-    const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B / BK;
+    const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A * LOAD_VEC_BATCH_A / BK;
+    const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B * LOAD_VEC_BATCH_B / BK;

 #ifdef MUL_MAT_ID
 #ifdef MUL_MAT_ID_USE_SUBGROUPS
@@ -302,8 +314,8 @@ void main() {
    }
 #else
    ACC_TYPE sums[WMITER * TM * WNITER * TN];
-    FLOAT_TYPE cache_a[WMITER * TM];
-    FLOAT_TYPE cache_b[TN];
+    FLOAT_TYPE_VEC2 cache_a[WMITER * TM];
+    FLOAT_TYPE_VEC2 cache_b[TN];

    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
        sums[i] = ACC_TYPE(0.0f);
@@ -312,13 +324,13 @@ void main() {

    for (uint block = start_k; block < end_k; block += BK) {
        [[unroll]] for (uint l = 0; l < BM; l += loadstride_a) {
-            load_a_to_shmem(pos_a, loadr_a, loadc_a + l, ir * BM + loadc_a + l, block + loadr_a, end_k);
+            load_a_to_shmem(pos_a, loadr_a, loadc_a + l, ir * BM + loadc_a + l, block, end_k);
        }
        [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {
 #if !defined(MUL_MAT_ID)
-            load_b_to_shmem(pos_b, loadr_b, loadc_b + l, ic * BN + loadc_b + l, block + loadr_b, end_k);
+            load_b_to_shmem(pos_b, loadr_b, loadc_b + l, ic * BN + loadc_b + l, block, end_k);
 #else
-            load_b_to_shmem(pos_b, loadr_b, loadc_b + l, ic, _ne1, block + loadr_b, end_k);
+            load_b_to_shmem(pos_b, loadr_b, loadc_b + l, ic, _ne1, block, end_k);
 #endif
        }

@@ -331,17 +343,17 @@ void main() {
        [[unroll]] for (uint i = 0; i < BK; i += TK) {
            [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
                // Load from shared into cache
-                coopMatLoad(cache_a, buf_a, (warp_r * WM + cm_row * TM) * SHMEM_STRIDE + i, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor);
+                coopMatLoad(cache_a, buf_a, (warp_r * WM + cm_row * TM) * SHMEM_STRIDE + i / 2, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor);

                [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
-                    coopMatLoad(cache_b, buf_b, (warp_c * WN + cm_col * TN) * SHMEM_STRIDE + i, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor);
+                    coopMatLoad(cache_b, buf_b, (warp_c * WN + cm_col * TN) * SHMEM_STRIDE + i / 2, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor);

                    sums[cm_col * cms_per_row + cm_row] = coopMatMulAdd(cache_a, cache_b, sums[cm_col * cms_per_row + cm_row]);
                }
            }
        }
 #else
-        [[unroll]] for (uint i = 0; i < BK; i++) {
+        [[unroll]] for (uint i = 0; i < BK / 2; i++) {
            // Load from shared into cache
            [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                [[unroll]] for (uint j = 0; j < TM; j++) {
@@ -357,7 +369,7 @@ void main() {
                    [[unroll]] for (uint cc = 0; cc < TN; cc++) {
                        [[unroll]] for (uint cr = 0; cr < TM; cr++) {
                            const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
-                            sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr]), ACC_TYPE(cache_b[cc]), sums[sums_idx]);
+                            sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr].x), ACC_TYPE(cache_b[cc].x), fma(ACC_TYPE(cache_a[wsir * TM + cr].y), ACC_TYPE(cache_b[cc].y), sums[sums_idx]));
                        }
                    }
                }
@@ -1,51 +1,53 @@
-void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uint idx_m, const uint idx_k, const uint end_k) {
+void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uint idx_m, const uint block, const uint end_k) {
 #if defined(DATA_A_F32) || defined(DATA_A_F16)
 #if LOAD_VEC_A == 8
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
            FLOAT_TYPE_VEC8 aa = FLOAT_TYPE_VEC8(data_a[idx]);
-            buf_a[buf_idx    ] = aa[0].x;
-            buf_a[buf_idx + 1] = aa[0].y;
-            buf_a[buf_idx + 2] = aa[0].z;
-            buf_a[buf_idx + 3] = aa[0].w;
-            buf_a[buf_idx + 4] = aa[1].x;
-            buf_a[buf_idx + 5] = aa[1].y;
-            buf_a[buf_idx + 6] = aa[1].z;
-            buf_a[buf_idx + 7] = aa[1].w;
+            buf_a[buf_idx    ] = aa[0].xy;
+            buf_a[buf_idx + 1] = aa[0].zw;
+            buf_a[buf_idx + 2] = aa[1].xy;
+            buf_a[buf_idx + 3] = aa[1].zw;
 #elif LOAD_VEC_A == 4
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
            FLOAT_TYPE_VEC4 aa = FLOAT_TYPE_VEC4(data_a[idx]);
-            buf_a[buf_idx    ] = aa.x;
-            buf_a[buf_idx + 1] = aa.y;
-            buf_a[buf_idx + 2] = aa.z;
-            buf_a[buf_idx + 3] = aa.w;
-#else
-            if (idx_m < p.M && idx_k < end_k) {
-                buf_a[col * SHMEM_STRIDE + row] = FLOAT_TYPE(data_a[pos_a + col * p.stride_a + row]);
+            buf_a[buf_idx    ] = aa.xy;
+            buf_a[buf_idx + 1] = aa.zw;
+#else // LOAD_VEC_BATCH_A == 2
+            const uint idx = pos_a + col * p.stride_a + row * 2;
+            const uint buf_idx = col * SHMEM_STRIDE + row;
+            if (idx_m < p.M && block + row * 2 + 1 < end_k) {
+                buf_a[buf_idx] = FLOAT_TYPE_VEC2(data_a[idx],
+                                                 data_a[idx + 1]);
+            } else if (idx_m < p.M && block + row * 2 < end_k) {
+                buf_a[buf_idx] = FLOAT_TYPE_VEC2(data_a[idx], 0.0f);
            } else {
-                buf_a[col * SHMEM_STRIDE + row] = FLOAT_TYPE(0.0f);
+                buf_a[buf_idx] = FLOAT_TYPE_VEC2(0.0f);
            }
 #endif
 #elif defined(DATA_A_BF16)
 #if LOAD_VEC_A == 4
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
            FLOAT_TYPE_VEC4 aa = FLOAT_TYPE_VEC4(TO_FLOAT_TYPE(data_a[idx]));
-            buf_a[buf_idx    ] = aa.x;
-            buf_a[buf_idx + 1] = aa.y;
-            buf_a[buf_idx + 2] = aa.z;
-            buf_a[buf_idx + 3] = aa.w;
-#else
-            if (idx_m < p.M && idx_k < end_k) {
-                buf_a[col * SHMEM_STRIDE + row] = TO_FLOAT_TYPE(data_a[pos_a + col * p.stride_a + row]);
+            buf_a[buf_idx    ] = aa.xy;
+            buf_a[buf_idx + 1] = aa.zw;
+#else // LOAD_VEC_BATCH_A == 2
+            const uint idx = pos_a + col * p.stride_a + row * 2;
+            const uint buf_idx = col * SHMEM_STRIDE + row;
+            if (idx_m < p.M && block + row * 2 + 1 < end_k) {
+                buf_a[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_a[idx]),
+                                                 TO_FLOAT_TYPE(data_a[idx + 1]));
+            } else if (idx_m < p.M && block + row * 2 < end_k) {
+                buf_a[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_a[idx]), 0.0f);
            } else {
-                buf_a[col * SHMEM_STRIDE + row] = TO_FLOAT_TYPE(uint16_t(0));
+                buf_a[buf_idx] = FLOAT_TYPE_VEC2(0.0f);
            }
 #endif
 #elif defined(DATA_A_Q4_0)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + 4 * row;
+            const uint buf_idx = col * SHMEM_STRIDE + 2 * row;

            const uint ib = idx / 4;
            const uint iqs = idx & 0x03;
@@ -55,17 +57,13 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const vec4 v0 = (vec4(unpack8(vui & 0x0F0F0F0F)) - 8.0f) * d;
            const vec4 v1 = (vec4(unpack8((vui >> 4) & 0x0F0F0F0F)) - 8.0f) * d;

-            buf_a[buf_idx     ] = FLOAT_TYPE(v0.x);
-            buf_a[buf_idx + 1 ] = FLOAT_TYPE(v0.y);
-            buf_a[buf_idx + 2 ] = FLOAT_TYPE(v0.z);
-            buf_a[buf_idx + 3 ] = FLOAT_TYPE(v0.w);
-            buf_a[buf_idx + 16] = FLOAT_TYPE(v1.x);
-            buf_a[buf_idx + 17] = FLOAT_TYPE(v1.y);
-            buf_a[buf_idx + 18] = FLOAT_TYPE(v1.z);
-            buf_a[buf_idx + 19] = FLOAT_TYPE(v1.w);
+            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2(v0.xy);
+            buf_a[buf_idx + 1] = FLOAT_TYPE_VEC2(v0.zw);
+            buf_a[buf_idx + 8] = FLOAT_TYPE_VEC2(v1.xy);
+            buf_a[buf_idx + 9] = FLOAT_TYPE_VEC2(v1.zw);
 #elif defined(DATA_A_Q4_1)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + 4 * row;
+            const uint buf_idx = col * SHMEM_STRIDE + 2 * row;

            const uint ib = idx / 4;
            const uint iqs = idx & 0x03;
@@ -76,17 +74,13 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const vec4 v0 = vec4(unpack8(vui & 0x0F0F0F0F)) * d + m;
            const vec4 v1 = vec4(unpack8((vui >> 4) & 0x0F0F0F0F)) * d + m;

-            buf_a[buf_idx     ] = FLOAT_TYPE(v0.x);
-            buf_a[buf_idx + 1 ] = FLOAT_TYPE(v0.y);
-            buf_a[buf_idx + 2 ] = FLOAT_TYPE(v0.z);
-            buf_a[buf_idx + 3 ] = FLOAT_TYPE(v0.w);
-            buf_a[buf_idx + 16] = FLOAT_TYPE(v1.x);
-            buf_a[buf_idx + 17] = FLOAT_TYPE(v1.y);
-            buf_a[buf_idx + 18] = FLOAT_TYPE(v1.z);
-            buf_a[buf_idx + 19] = FLOAT_TYPE(v1.w);
+            buf_a[buf_idx     ] = FLOAT_TYPE_VEC2(v0.xy);
+            buf_a[buf_idx + 1 ] = FLOAT_TYPE_VEC2(v0.zw);
+            buf_a[buf_idx + 8 ] = FLOAT_TYPE_VEC2(v1.xy);
+            buf_a[buf_idx + 9 ] = FLOAT_TYPE_VEC2(v1.zw);
 #elif defined(DATA_A_Q5_0)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + 2 * row;
+            const uint buf_idx = col * SHMEM_STRIDE + row;

            const uint ib = idx / 8;
            const uint iqs = idx & 0x07;
@@ -99,13 +93,11 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const uint vui = uint(data_a_packed16[ib].qs[iqs]);
            const vec4 v = (vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) - 16.0f) * d;

-            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
-            buf_a[buf_idx + 1 ] = FLOAT_TYPE(v.z);
-            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
-            buf_a[buf_idx + 17] = FLOAT_TYPE(v.w);
+            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2(v.xz);
+            buf_a[buf_idx + 8] = FLOAT_TYPE_VEC2(v.yw);
 #elif defined(DATA_A_Q5_1)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + 2 * row;
+            const uint buf_idx = col * SHMEM_STRIDE + row;

            const uint ib = idx / 8;
            const uint iqs = idx & 0x07;
@@ -119,13 +111,11 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const uint vui = uint(data_a_packed16[ib].qs[iqs]);
            const vec4 v = vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) * d + m;

-            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
-            buf_a[buf_idx + 1 ] = FLOAT_TYPE(v.z);
-            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
-            buf_a[buf_idx + 17] = FLOAT_TYPE(v.w);
+            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2(v.xz);
+            buf_a[buf_idx + 8] = FLOAT_TYPE_VEC2(v.yw);
 #elif defined(DATA_A_Q8_0)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 8;
            const uint iqs = idx & 0x07;
@@ -135,13 +125,11 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const i8vec2 v1 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs + 1])).xy;
            const vec4 v = vec4(v0.x, v0.y, v1.x, v1.y) * d;

-            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
-            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
-            buf_a[buf_idx + 2] = FLOAT_TYPE(v.z);
-            buf_a[buf_idx + 3] = FLOAT_TYPE(v.w);
+            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2(v.xy);
+            buf_a[buf_idx + 1] = FLOAT_TYPE_VEC2(v.zw);
 #elif defined(DATA_A_Q2_K)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 128;                         // 2 values per idx
            const uint iqs = idx % 128;                        // 0..127
@@ -156,11 +144,10 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin

            const vec2 v = d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);

-            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
-            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+            buf_a[buf_idx] = FLOAT_TYPE_VEC2(v.xy);
 #elif defined(DATA_A_Q3_K)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 128;                   // 2 values per idx
            const uint iqs = idx % 128;                  // 0..127
@@ -178,11 +165,11 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
                                  | (((data_a[ib].scales[8 + (is % 4)] >> (2 * int(is / 4))) & 3) << 4));
            const float dl = float(data_a[ib].d) * float(us - 32);

-            buf_a[buf_idx    ] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi    ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi    ] & m) != 0) ? 0 : 4)));
-            buf_a[buf_idx + 1] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));
+            buf_a[buf_idx] = FLOAT_TYPE_VEC2(dl * float(int8_t((data_a[ib].qs[qsi    ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi    ] & m) != 0) ? 0 : 4)),
+                                             dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));
 #elif defined(DATA_A_Q4_K)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 128;                 // 2 values per idx
            const uint iqs = idx % 128;                // 0..127
@@ -211,11 +198,11 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const float d = loadd.x * sc;
            const float m = -loadd.y * mbyte;

-            buf_a[buf_idx    ] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF), m));
-            buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF), m));
+            buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF), m),
+                                             fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF), m));
 #elif defined(DATA_A_Q5_K)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 128;                 // 2 values per idx
            const uint iqs = idx % 128;                // 0..127
@@ -247,11 +234,11 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const float d = loadd.x * sc;
            const float m = -loadd.y * mbyte;

-            buf_a[buf_idx    ] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi    ] & hm) != 0 ? 16 : 0), m));
-            buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m));
+            buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi    ] & hm) != 0 ? 16 : 0), m),
+                                             fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m));
 #elif defined(DATA_A_Q6_K)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 128;                  // 2 values per idx
            const uint iqs = idx % 128;                 // 0..127
@@ -266,11 +253,11 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin

            const float dscale = float(data_a[ib].d) * float(data_a[ib].scales[is]);

-            buf_a[buf_idx    ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32));
-            buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
+            buf_a[buf_idx] = FLOAT_TYPE_VEC2(dscale * float(int8_t(((data_a[ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32),
+                                             dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
 #elif defined(DATA_A_IQ1_S)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 32;                  // 8 values per idx
            const uint ib32 = (idx % 32) / 4;         // 0..7
@@ -283,12 +270,13 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
            const int16_t grid = int16_t(iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)]);

-            [[unroll]] for (int k = 0; k < 8; ++k) {
-                buf_a[buf_idx + k] = FLOAT_TYPE(dl * (bitfieldExtract(grid, 2 * k, 2) + delta));
+            [[unroll]] for (int k = 0; k < 4; ++k) {
+                buf_a[buf_idx + k] = FLOAT_TYPE_VEC2(dl * (bitfieldExtract(grid, 4 * k    , 2) + delta),
+                                                     dl * (bitfieldExtract(grid, 4 * k + 2, 2) + delta));
            }
 #elif defined(DATA_A_IQ1_M)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 32;  // 8 values per idx
            const uint ib8 = idx % 32;
@@ -304,12 +292,13 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
            const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);

-            [[unroll]] for (int k = 0; k < 8; ++k) {
-                buf_a[buf_idx + k] = FLOAT_TYPE(dl * (bitfieldExtract(grid, 2 * k, 2) + delta));
+            [[unroll]] for (int k = 0; k < 4; ++k) {
+                buf_a[buf_idx + k] = FLOAT_TYPE_VEC2(dl * (bitfieldExtract(grid, 4 * k    , 2) + delta),
+                                                     dl * (bitfieldExtract(grid, 4 * k + 2, 2) + delta));
            }
 #elif defined(DATA_A_IQ2_XXS)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 32;                 // 8 values per idx
            const uint ib32 = (idx % 32) / 4;         // 0..7
@@ -330,17 +319,17 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const vec4 grid0 = vec4(unpack8(grid.x));
            const vec4 grid1 = vec4(unpack8(grid.y));

-            buf_a[buf_idx    ] = db * FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x);
-            buf_a[buf_idx + 1] = db * FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y);
-            buf_a[buf_idx + 2] = db * FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z);
-            buf_a[buf_idx + 3] = db * FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w);
-            buf_a[buf_idx + 4] = db * FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x);
-            buf_a[buf_idx + 5] = db * FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y);
-            buf_a[buf_idx + 6] = db * FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z);
-            buf_a[buf_idx + 7] = db * FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w);
+            buf_a[buf_idx    ] = db * FLOAT_TYPE_VEC2((sign &   1) != 0 ? -grid0.x : grid0.x,
+                                                      (sign &   2) != 0 ? -grid0.y : grid0.y);
+            buf_a[buf_idx + 1] = db * FLOAT_TYPE_VEC2((sign &   4) != 0 ? -grid0.z : grid0.z,
+                                                      (sign &   8) != 0 ? -grid0.w : grid0.w);
+            buf_a[buf_idx + 2] = db * FLOAT_TYPE_VEC2((sign &  16) != 0 ? -grid1.x : grid1.x,
+                                                      (sign &  32) != 0 ? -grid1.y : grid1.y);
+            buf_a[buf_idx + 3] = db * FLOAT_TYPE_VEC2((sign &  64) != 0 ? -grid1.z : grid1.z,
+                                                      (sign & 128) != 0 ? -grid1.w : grid1.w);
 #elif defined(DATA_A_IQ2_XS)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 32;            // 8 values per idx
            const uint ib32 = (idx % 32) / 4;    // 0..7
@@ -356,17 +345,17 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const vec4 grid0 = vec4(unpack8(grid.x));
            const vec4 grid1 = vec4(unpack8(grid.y));

-            buf_a[buf_idx    ] = db * FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x);
-            buf_a[buf_idx + 1] = db * FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y);
-            buf_a[buf_idx + 2] = db * FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z);
-            buf_a[buf_idx + 3] = db * FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w);
-            buf_a[buf_idx + 4] = db * FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x);
-            buf_a[buf_idx + 5] = db * FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y);
-            buf_a[buf_idx + 6] = db * FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z);
-            buf_a[buf_idx + 7] = db * FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w);
+            buf_a[buf_idx    ] = db * FLOAT_TYPE_VEC2((sign &   1) != 0 ? -grid0.x : grid0.x,
+                                                      (sign &   2) != 0 ? -grid0.y : grid0.y);
+            buf_a[buf_idx + 1] = db * FLOAT_TYPE_VEC2((sign &   4) != 0 ? -grid0.z : grid0.z,
+                                                      (sign &   8) != 0 ? -grid0.w : grid0.w);
+            buf_a[buf_idx + 2] = db * FLOAT_TYPE_VEC2((sign &  16) != 0 ? -grid1.x : grid1.x,
+                                                      (sign &  32) != 0 ? -grid1.y : grid1.y);
+            buf_a[buf_idx + 3] = db * FLOAT_TYPE_VEC2((sign &  64) != 0 ? -grid1.z : grid1.z,
+                                                      (sign & 128) != 0 ? -grid1.w : grid1.w);
 #elif defined(DATA_A_IQ2_S)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 32;  // 8 values per idx
            const uint ib8 = idx % 32; // 0..31
@@ -384,17 +373,17 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const vec4 grid0 = vec4(unpack8(grid.x));
            const vec4 grid1 = vec4(unpack8(grid.y));

-            buf_a[buf_idx    ] = db * FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x);
-            buf_a[buf_idx + 1] = db * FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y);
-            buf_a[buf_idx + 2] = db * FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z);
-            buf_a[buf_idx + 3] = db * FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w);
-            buf_a[buf_idx + 4] = db * FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x);
-            buf_a[buf_idx + 5] = db * FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y);
-            buf_a[buf_idx + 6] = db * FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z);
-            buf_a[buf_idx + 7] = db * FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w);
+            buf_a[buf_idx    ] = db * FLOAT_TYPE_VEC2((sign &   1) != 0 ? -grid0.x : grid0.x,
+                                                      (sign &   2) != 0 ? -grid0.y : grid0.y);
+            buf_a[buf_idx + 1] = db * FLOAT_TYPE_VEC2((sign &   4) != 0 ? -grid0.z : grid0.z,
+                                                      (sign &   8) != 0 ? -grid0.w : grid0.w);
+            buf_a[buf_idx + 2] = db * FLOAT_TYPE_VEC2((sign &  16) != 0 ? -grid1.x : grid1.x,
+                                                      (sign &  32) != 0 ? -grid1.y : grid1.y);
+            buf_a[buf_idx + 3] = db * FLOAT_TYPE_VEC2((sign &  64) != 0 ? -grid1.z : grid1.z,
+                                                      (sign & 128) != 0 ? -grid1.w : grid1.w);
 #elif defined(DATA_A_IQ3_XXS)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 64;            // 4 values per idx
            const uint iqs = idx % 64;           // 0..63
@@ -414,13 +403,13 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const uint grid = iq3xxs_grid[qs];
            const vec4 v = db * vec4(unpack8(grid));

-            buf_a[buf_idx    ] = FLOAT_TYPE((sign &   1) != 0 ? -v.x : v.x);
-            buf_a[buf_idx + 1] = FLOAT_TYPE((sign &   2) != 0 ? -v.y : v.y);
-            buf_a[buf_idx + 2] = FLOAT_TYPE((sign &   4) != 0 ? -v.z : v.z);
-            buf_a[buf_idx + 3] = FLOAT_TYPE((sign &   8) != 0 ? -v.w : v.w);
+            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2((sign &   1) != 0 ? -v.x : v.x,
+                                                 (sign &   2) != 0 ? -v.y : v.y);
+            buf_a[buf_idx + 1] = FLOAT_TYPE_VEC2((sign &   4) != 0 ? -v.z : v.z,
+                                                 (sign &   8) != 0 ? -v.w : v.w);
 #elif defined(DATA_A_IQ3_S)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 64;            // 4 values per idx
            const uint iqs = idx % 64;           // 0..63
@@ -436,13 +425,13 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)];
            const vec4 v = db * vec4(unpack8(grid));

-            buf_a[buf_idx    ] = FLOAT_TYPE((sign &   1) != 0 ? -v.x : v.x);
-            buf_a[buf_idx + 1] = FLOAT_TYPE((sign &   2) != 0 ? -v.y : v.y);
-            buf_a[buf_idx + 2] = FLOAT_TYPE((sign &   4) != 0 ? -v.z : v.z);
-            buf_a[buf_idx + 3] = FLOAT_TYPE((sign &   8) != 0 ? -v.w : v.w);
+            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2((sign &   1) != 0 ? -v.x : v.x,
+                                                 (sign &   2) != 0 ? -v.y : v.y);
+            buf_a[buf_idx + 1] = FLOAT_TYPE_VEC2((sign &   4) != 0 ? -v.z : v.z,
+                                                 (sign &   8) != 0 ? -v.w : v.w);
 #elif defined(DATA_A_IQ4_XS)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

            const uint ib = idx / 128;                  // 2 values per idx
            const uint ib32 = (idx % 128) / 16;         // 0..7
@@ -457,11 +446,10 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const float d = float(data_a[ib].d);
            const vec2 v = d * float(int(sl | (sh << 4)) - 32) * vec2(kvalues_iq4nl[qs.x], kvalues_iq4nl[qs.y]);

-            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
-            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2(v.xy);
 #elif defined(DATA_A_IQ4_NL)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + 2 * row;
+            const uint buf_idx = col * SHMEM_STRIDE + row;

            const uint ib = idx / 8;
            const uint iqs = idx & 0x07;
@@ -469,13 +457,13 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const FLOAT_TYPE d = FLOAT_TYPE(data_a_packed16[ib].d);
            const uint vui = uint(data_a_packed16[ib].qs[iqs]);

-            buf_a[buf_idx     ] = FLOAT_TYPE(kvalues_iq4nl[vui & 0xF]) * d;
-            buf_a[buf_idx + 1 ] = FLOAT_TYPE(kvalues_iq4nl[bitfieldExtract(vui, 8, 4)]) * d;
-            buf_a[buf_idx + 16] = FLOAT_TYPE(kvalues_iq4nl[bitfieldExtract(vui, 4, 4)]) * d;
-            buf_a[buf_idx + 17] = FLOAT_TYPE(kvalues_iq4nl[vui >> 12]) * d;
+            buf_a[buf_idx    ] = d * FLOAT_TYPE_VEC2(kvalues_iq4nl[vui & 0xF],
+                                                      kvalues_iq4nl[bitfieldExtract(vui, 8, 4)]);
+            buf_a[buf_idx + 8] = d * FLOAT_TYPE_VEC2(kvalues_iq4nl[bitfieldExtract(vui, 4, 4)],
+                                                     kvalues_iq4nl[vui >> 12]);
 #elif defined(DATA_A_MXFP4)
            const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + 2 * row;
+            const uint buf_idx = col * SHMEM_STRIDE + row;

            const uint ib = idx / 8;
            const uint iqs = (idx & 0x07) * 2;
@@ -484,84 +472,84 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
            const uint vui = uint(data_a[ib].qs[iqs]);
            const uint vui2 = uint(data_a[ib].qs[iqs+1]);

-            buf_a[buf_idx     ] = FLOAT_TYPE(kvalues_mxfp4[vui & 0xF] * d);
-            buf_a[buf_idx + 16] = FLOAT_TYPE(kvalues_mxfp4[vui >>  4] * d);
-            buf_a[buf_idx +  1] = FLOAT_TYPE(kvalues_mxfp4[vui2 & 0xF] * d);
-            buf_a[buf_idx + 17] = FLOAT_TYPE(kvalues_mxfp4[vui2 >>  4] * d);
+            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2(kvalues_mxfp4[vui  & 0xF] * d,
+                                                 kvalues_mxfp4[vui2 & 0xF] * d);
+            buf_a[buf_idx + 8] = FLOAT_TYPE_VEC2(kvalues_mxfp4[vui  >>  4] * d,
+                                                 kvalues_mxfp4[vui2 >>  4] * d);
 #endif
 }

 #if !defined(MUL_MAT_ID)
-void load_b_to_shmem(const uint pos_b, const uint row, const uint col, const uint idx_n, const uint idx_k, const uint end_k) {
+void load_b_to_shmem(const uint pos_b, const uint row, const uint col, const uint idx_n, const uint block, const uint end_k) {
 #if LOAD_VEC_B == 8
            // Not supported for b_type bf16 because bf16mat2x4 does not exist
            const uint idx = pos_b + col * p.stride_b / LOAD_VEC_B + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
            FLOAT_TYPE_VEC8 bb = FLOAT_TYPE_VEC8(data_b[idx]);
-            buf_b[buf_idx + 0] = bb[0].x;
-            buf_b[buf_idx + 1] = bb[0].y;
-            buf_b[buf_idx + 2] = bb[0].z;
-            buf_b[buf_idx + 3] = bb[0].w;
-            buf_b[buf_idx + 4] = bb[1].x;
-            buf_b[buf_idx + 5] = bb[1].y;
-            buf_b[buf_idx + 6] = bb[1].z;
-            buf_b[buf_idx + 7] = bb[1].w;
+            buf_b[buf_idx + 0] = bb[0].xy;
+            buf_b[buf_idx + 1] = bb[0].zw;
+            buf_b[buf_idx + 2] = bb[1].xy;
+            buf_b[buf_idx + 3] = bb[1].zw;
 #elif LOAD_VEC_B == 4
            const uint idx = pos_b + col * p.stride_b / LOAD_VEC_B + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
 #if defined(DATA_B_BF16)
            FLOAT_TYPE_VEC4 bb = FLOAT_TYPE_VEC4(TO_FLOAT_TYPE(data_b[idx]));
 #else
            FLOAT_TYPE_VEC4 bb = FLOAT_TYPE_VEC4(data_b[idx]);
 #endif
-            buf_b[buf_idx + 0] = bb.x;
-            buf_b[buf_idx + 1] = bb.y;
-            buf_b[buf_idx + 2] = bb.z;
-            buf_b[buf_idx + 3] = bb.w;
-#else // LOAD_VEC_B == 1
-            if (idx_n < p.N && idx_k < end_k) {
-                buf_b[col * SHMEM_STRIDE + row] = TO_FLOAT_TYPE(data_b[pos_b + col * p.stride_b + row]);
+            buf_b[buf_idx + 0] = bb.xy;
+            buf_b[buf_idx + 1] = bb.zw;
+#else // LOAD_VEC_BATCH_B == 2
+            const uint idx = pos_b + col * p.stride_b + row * 2;
+            const uint buf_idx = col * SHMEM_STRIDE + row;
+            if (idx_n < p.N && block + row * 2 + 1 < end_k) {
+                buf_b[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_b[idx]),
+                                                 TO_FLOAT_TYPE(data_b[idx + 1]));
+            } else if (idx_n < p.N && block + row * 2 < end_k) {
+                buf_b[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_b[idx]), 0.0f);
            } else {
-                buf_b[col * SHMEM_STRIDE + row] = FLOAT_TYPE(0.0f);
+                buf_b[buf_idx] = FLOAT_TYPE_VEC2(0.0f);
            }
 #endif
 }
 #else
-void load_b_to_shmem(const uint pos_b, const uint row, const uint col, const uint ic, const uint _ne1, const uint idx_k, const uint end_k) {
+void load_b_to_shmem(const uint pos_b, const uint row, const uint col, const uint ic, const uint _ne1, const uint block, const uint end_k) {
 #if LOAD_VEC_B == 8
            // Not supported for b_type bf16 because bf16mat2x4 does not exist
            const u16vec2 row_idx = row_ids[col];
            const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
            FLOAT_TYPE_VEC8 bb = FLOAT_TYPE_VEC8(data_b[idx]);
-            buf_b[buf_idx + 0] = bb[0].x;
-            buf_b[buf_idx + 1] = bb[0].y;
-            buf_b[buf_idx + 2] = bb[0].z;
-            buf_b[buf_idx + 3] = bb[0].w;
-            buf_b[buf_idx + 4] = bb[1].x;
-            buf_b[buf_idx + 5] = bb[1].y;
-            buf_b[buf_idx + 6] = bb[1].z;
-            buf_b[buf_idx + 7] = bb[1].w;
+            buf_b[buf_idx + 0] = bb[0].xy;
+            buf_b[buf_idx + 1] = bb[0].zw;
+            buf_b[buf_idx + 2] = bb[1].xy;
+            buf_b[buf_idx + 3] = bb[1].zw;
 #elif LOAD_VEC_B == 4
            const u16vec2 row_idx = row_ids[col];
            const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_B / 2;
 #if defined(DATA_B_BF16)
            FLOAT_TYPE_VEC4 bb = FLOAT_TYPE_VEC4(TO_FLOAT_TYPE(data_b[idx]));
 #else
            FLOAT_TYPE_VEC4 bb = FLOAT_TYPE_VEC4(data_b[idx]);
 #endif
-            buf_b[buf_idx + 0] = bb.x;
-            buf_b[buf_idx + 1] = bb.y;
-            buf_b[buf_idx + 2] = bb.z;
-            buf_b[buf_idx + 3] = bb.w;
-#else // LOAD_VEC_B == 1
+            buf_b[buf_idx + 0] = bb.xy;
+            buf_b[buf_idx + 1] = bb.zw;
+#else // LOAD_VEC_BATCH_B == 2
            const uint row_i = ic * BN + col;
-            if (row_i < _ne1 && idx_k < end_k) {
+            const uint buf_idx = col * SHMEM_STRIDE + row;
+            if (row_i < _ne1 && block + row * 2 + 1 < end_k) {
                const u16vec2 row_idx = row_ids[col];
-                buf_b[col * SHMEM_STRIDE + row] = TO_FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row]);
+                const uint idx = pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
+                buf_b[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_b[idx]),
+                                                 TO_FLOAT_TYPE(data_b[idx + 1]));
+            } else if (row_i < _ne1 && block + row * 2 < end_k) {
+                const u16vec2 row_idx = row_ids[col];
+                const uint idx = pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
+                buf_b[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_b[idx]), 0.0f);
            } else {
-                buf_b[col * SHMEM_STRIDE + row] = FLOAT_TYPE(0.0f);
+                buf_b[buf_idx] = FLOAT_TYPE_VEC2(0.0f);
            }
 #endif
 }
@@ -11,12 +11,12 @@
 #define QUANT_K 1
 #define QUANT_R 1

-#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
-#define A_TYPE float
-#elif LOAD_VEC_A == 4
+#if LOAD_VEC_A == 4
 #define A_TYPE vec4
 #elif LOAD_VEC_A == 8
 #define A_TYPE mat2x4
+#else
+#define A_TYPE float
 #endif
 #endif

@@ -24,12 +24,12 @@
 #define QUANT_K 1
 #define QUANT_R 1

-#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
-#define A_TYPE float16_t
-#elif LOAD_VEC_A == 4
+#if LOAD_VEC_A == 4
 #define A_TYPE f16vec4
 #elif LOAD_VEC_A == 8
 #define A_TYPE f16mat2x4
+#else
+#define A_TYPE float16_t
 #endif
 #endif

@@ -37,12 +37,12 @@
 #define QUANT_K 1
 #define QUANT_R 1

-#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
-#define A_TYPE uint16_t
-#elif LOAD_VEC_A == 4
+#if LOAD_VEC_A == 4
 #define A_TYPE u16vec4
 #elif LOAD_VEC_A == 8
 #error unsupported
+#else
+#define A_TYPE uint16_t
 #endif
 #endif

@@ -336,7 +336,8 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
        base_dict["FLOAT16"] = "1";
    }

-    base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
+    base_dict["ACC_TYPE"     ] = f16acc ? "float16_t" : "float";
+    base_dict["ACC_TYPE_VEC2"] = f16acc ? "f16vec2"   : "vec2";
    if (f16acc) {
        base_dict["ACC_TYPE_MAX"] = "\"float16_t(65504.0)\"";
    }
@@ -418,7 +419,6 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c

    // bf16
    {
-        std::string load_vec_a_unaligned = "1";
        // For aligned matmul loads
        std::string load_vec_a = coopmat2 ? "1" : "4";

@@ -436,8 +436,8 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
        if (!(coopmat || coopmat2))
 #endif
        {
-            string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"},  {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_bf16",         source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                      {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}}),                   fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_bf16",         source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"},                             {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}}),                   fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(merge_maps(base_dict, float_type_dict_bf16), {{"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"},  {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"DATA_B_BF16", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
        }
    }

@@ -635,8 +635,10 @@ void process_shaders() {
    }

    for (std::string t : {"f32", "f16", "bf16", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
-        string_to_spv("set_rows_" + t, "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
-        string_to_spv("set_rows_" + t + "_rte", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
+        string_to_spv("set_rows_" + t + "_i32",     "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uint"}, {"B_SIZE", "32"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+        string_to_spv("set_rows_" + t + "_i32_rte", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uint"}, {"B_SIZE", "32"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
+        string_to_spv("set_rows_" + t + "_i64",     "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"B_SIZE", "64"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+        string_to_spv("set_rows_" + t + "_i64_rte", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"B_SIZE", "64"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
    }

    auto get_type_str = [](bool f16) {
@@ -704,8 +706,11 @@ void process_shaders() {

    string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});

-    string_to_spv("exp_f16",        "exp.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
-    string_to_spv("exp_f32",        "exp.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    for (auto rte : {false, true}) {
+        std::string suffix = rte ? "_rte" : "";
+        string_to_spv("exp_f16" + suffix,        "exp.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"},   {"RTE16", rte ? "1" : "0"}});
+        string_to_spv("exp_f32" + suffix,        "exp.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}    ,   {"RTE16", rte ? "1" : "0"}});
+    }
    string_to_spv("gelu_f16",       "gelu.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
    string_to_spv("gelu_f32",       "gelu.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
    string_to_spv("gelu_erf_f16",   "gelu_erf.comp",    {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
@@ -793,16 +798,26 @@ void process_shaders() {
    string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
    string_to_spv("opt_step_sgd_f32", "opt_step_sgd.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));

-    string_to_spv("conv2d_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
-    string_to_spv("conv2d_f16_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
-
-    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
-    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
-
+    for (auto transpose : {false, true}) {
+        for (auto unroll : {false, true}) {
+            for (auto a_f16 : {false, true}) {
+                std::map<std::string, std::string> defines = {
+                    {"A_TYPE", a_f16 ? "float16_t" : "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"},
+                    {"USE_COLLECTIVES", "1"}, {"UNROLL", unroll ? "[[unroll]]" : ""},
+                };
+                if (transpose) defines["TRANSPOSE"] = "1";
+                std::string name = std::string(transpose ? "conv_transpose_2d": "conv2d")
+                    + (a_f16 ? "_f16" : "") + "_f32";
+                string_to_spv(name + (unroll ? "_unroll" : ""), "conv2d_mm.comp", defines);
 #if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
-    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
-    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
+                if (unroll) {
+                    defines["COOPMAT2"] = "1";
+                    string_to_spv(name, "conv2d_mm.comp", defines, true, false, true);
+                }
 #endif
+            }
+        }
+    }

    string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
    string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
@@ -823,7 +823,7 @@ static ggml_backend_i ggml_backend_webgpu_i = {
    /* .graph_compute           = */ ggml_backend_webgpu_graph_compute,
    /* .event_record            = */ NULL,
    /* .event_wait              = */ NULL,
-    /* .optimize_graph          = */ NULL,
+    /* .graph_optimize          = */ NULL,
 };

 /* End GGML Backend Interface */
@@ -1310,7 +1310,7 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
            break;
        case GGML_OP_CPY:
        case GGML_OP_SET_ROWS:
-            supports_op = (op->type == GGML_TYPE_F16 && op->src[0]->type == GGML_TYPE_F32);
+            supports_op = (op->type == GGML_TYPE_F16 && op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_I64);
            break;
        case GGML_OP_GET_ROWS:
            if (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 ||
@@ -0,0 +1 @@
+zdnn.h
@@ -0,0 +1,59 @@
+#ifndef GGML_ZDNN_COMMON_HPP
+#define GGML_ZDNN_COMMON_HPP
+
+#include "ggml.h"
+#include "ggml-impl.h"
+
+#include "zdnn.h"
+
+#include <vector>
+#include <memory>
+
+#define GGML_ZDNN_NAME    "zDNN"
+#define GGML_ZDNN_VERSION ZDNN_VERNUM
+
+#define ZDNN_CHECK(stmt)                \
+    do {                                \
+        zdnn_status status = (stmt);    \
+        GGML_ASSERT(status == ZDNN_OK); \
+    } while (0);
+
+struct ggml_backend_zdnn_device_context {
+    int zdnn_device;
+    int zdnn_device_ref_count;
+
+    bool has_parmblkformat_0;
+    bool has_parmblkformat_1;  // checks for z17
+
+    size_t max_size;
+
+    char name[128];
+};
+
+struct ggml_backend_zdnn_context {
+    int device;
+    ggml_cgraph * gf;
+};
+
+struct ggml_backend_zdnn_buffer {
+    void * data;
+    ggml_backend_zdnn_buffer * extra;  // for bias, etc.
+    size_t size;
+
+    zdnn_tensor_desc pre_tfm_desc;
+    zdnn_tensor_desc tfm_desc;
+    zdnn_ztensor     ztensor;
+
+    char name[GGML_MAX_NAME];
+};
+
+struct ggml_backend_zdnn_buffer_context {
+    void * all_data;
+    size_t all_size;
+    bool owned;
+
+    int n_buffers;
+    std::vector<std::unique_ptr<ggml_backend_zdnn_buffer>> buffers;
+};
+
+#endif  // GGML_ZDNN_COMMON_HPP
@@ -1,98 +0,0 @@
-#ifndef GGML_ZDNN_IMPL
-#define GGML_ZDNN_IMPL
-
-#include "zdnn.h"
-#include "ggml.h"
-#include "ggml-zdnn.h"
-
-#include <vector>
-#include <memory>
-#include <vecintrin.h>
-
-#define GGML_ZDNN_NAME    "zDNN"
-#define GGML_ZDNN_VERSION ZDNN_VERNUM
-
-#define vec_neg(a)    (-(a))                // Vector Negate
-#define vec_add(a, b) ((a) + (b))           // Vector Add
-#define vec_sub(a, b) ((a) - (b))           // Vector Subtract
-#define vec_mul(a, b) ((a) * (b))           // Vector Multiply
-#define vec_div(a, b) ((a) / (b))           // Vector Divide
-#define vec_sl(a, b)  ((a) << (b))          // Vector Shift Left
-#define vec_sra(a, b) ((a) >> (b))          // Vector Shift Right
-#define vec_sr(a, b)  ((a) >> (b))          // Vector Shift Right Algebraic
-#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet
-#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet
-
-#ifndef vec_and
-#define vec_and(a, b) ((a) & (b)) // Vector AND
-#endif
-
-#ifndef vec_or
-#define vec_or(a, b)  ((a) | (b)) // Vector OR
-#endif
-
-#ifndef vec_xor
-#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
-#endif
-
-typedef   signed char char8x16_t  __attribute__((vector_size(16)));
-typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
-
-typedef int8_t   int8x16_t  __attribute__((vector_size(16)));
-typedef int16_t  int16x8_t  __attribute__((vector_size(16)));
-typedef int32_t  int32x4_t  __attribute__((vector_size(16)));
-typedef uint8_t  uint8x16_t __attribute__((vector_size(16)));
-typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
-typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
-
-typedef float float32x4_t   __attribute__((vector_size(16)));
-typedef double double64x2_t __attribute__((vector_size(16)));
-
-typedef   signed long long long64x2_t  __attribute__((vector_size(16)));
-typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
-
-#define ZDNN_CHECK(stmt)                \
-    do {                                \
-        zdnn_status status = (stmt);    \
-        GGML_ASSERT(status == ZDNN_OK); \
-    } while (0);
-
-struct ggml_backend_zdnn_device_context {
-    int zdnn_device;
-    int zdnn_device_ref_count;
-
-    bool has_parmblkformat_0;
-    bool has_parmblkformat_1;
-
-    size_t max_size;
-
-    char name[128];
-};
-
-struct ggml_backend_zdnn_context {
-    int device;
-    ggml_cgraph * gf;
-};
-
-struct ggml_backend_zdnn_buffer {
-    void * data;
-    ggml_backend_zdnn_buffer * extra;  // for bias, etc.
-    size_t size;
-
-    zdnn_tensor_desc pre_tfm_desc;
-    zdnn_tensor_desc tfm_desc;
-    zdnn_ztensor     ztensor;
-
-    char name[GGML_MAX_NAME];
-};
-
-struct ggml_backend_zdnn_buffer_context {
-    void * all_data;
-    size_t all_size;
-    bool owned;
-
-    int n_buffers;
-    std::vector<std::unique_ptr<ggml_backend_zdnn_buffer>> buffers;
-};
-
-#endif  // GGML_ZDNN_IMPL
@@ -1,187 +1,38 @@
-#include "zdnn.h"
 #include "ggml-zdnn.h"
-#include "ggml-zdnn-impl.h"
-
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"

+#include "ggml-zdnn/common.hpp"
+#include "ggml-zdnn/mmf.hpp"
+#include "ggml-zdnn/utils.hpp"
+#include "ggml.h"
+
 #include <vector>
 #include <memory>
-#include <csignal>
+#include <csignal>  // raise(SIGTRAP)
 #include <unistd.h>

-inline zdnn_data_types ggml_zdnn_type_mapping(ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_F32:
-            return FP32;
-        case GGML_TYPE_F16:
-            return FP16;
-        case GGML_TYPE_BF16:
-            return BFLOAT;
-        case GGML_TYPE_I8:
-            return INT8;
-        case GGML_TYPE_I32:
-            return INT32;
-        case GGML_TYPE_Q8_0:
-            return INT8;
-        default:
-            GGML_ABORT("%s: fatal: unable to determine zTensor data type",
-                       __func__);
-            break;
-    }
+static void ggml_zdnn_compute_forward_mul_mat(
+    const ggml_backend_zdnn_context * ctx,
+          ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];  // weights
+    const ggml_tensor * src1 = dst->src[1];  // inputs
+
+    // TODO: implement support for quantized types
+    // we currently only support f32, f16, and bf16
+    ggml_zdnn_mul_mat_f(ctx, src0, src1, dst);
 }

-inline void ggml_zdnn_create_tensor(zdnn_tensor_desc  & pre_tfm_desc,
-                                    zdnn_tensor_desc  & tfm_desc,
-                                    zdnn_ztensor      & ztensor,
-                              const ggml_tensor       * src,
-                              const int64_t           * ne,
-                              const zdnn_data_layouts   layout) {
-    zdnn_init_pre_transformed_desc(
-        layout,
-        ggml_zdnn_type_mapping(src->type),
-        &pre_tfm_desc,
-        ne[3], ne[2], ne[1], ne[0]
-    );
+static bool ggml_zdnn_compute_forward(
+    ggml_backend_zdnn_context * ctx,
+    ggml_tensor * dst) {

-    ZDNN_CHECK(zdnn_generate_transformed_desc(&pre_tfm_desc, &tfm_desc));
-    ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&pre_tfm_desc, &tfm_desc, &ztensor));
-}
-
-inline void ggml_zdnn_load_tensor(zdnn_ztensor & ztensor,
-                                          void * buffer) {
-    ZDNN_CHECK(zdnn_transform_ztensor(&ztensor, buffer));
-}
-
-inline void ggml_zdnn_init_tensor(ggml_backend_zdnn_buffer * buffer, const ggml_tensor * tensor) {
-    switch (tensor->op) {
-        case GGML_OP_MUL_MAT:
-            {
-                zdnn_init_pre_transformed_desc(
-                    ZDNN_2D,
-                    ggml_zdnn_type_mapping(tensor->type),
-                    &buffer->pre_tfm_desc,
-                    tensor->ne[1], tensor->ne[0]
-                );
-            } break;
-
-        default:
-            {
-                // For 4D tensors, GGML uses NCHW layout. However, because zDNN
-                // automatically transforms everything to NHWC, we will use it
-                // directly to avoid the performance penalty changing the
-                // layout and reshaping the tensor.
-                zdnn_init_pre_transformed_desc(
-                    ZDNN_NHWC,
-                    ggml_zdnn_type_mapping(tensor->type),
-                    &buffer->pre_tfm_desc,
-                    tensor->ne[3], tensor->ne[2], tensor->ne[1], tensor->ne[0]
-                );
-
-                // TODO: Consider adding a ggml check.
-                // TODO: If tensor = 4D, use ZDNN_NCHW by default.
-                // TODO: If tensor = 2D, use ZDNN_NHWC by default.
-            } break;
-    }
-
-    ZDNN_CHECK(zdnn_generate_transformed_desc(&buffer->pre_tfm_desc, &buffer->tfm_desc));
-    ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&buffer->pre_tfm_desc, &buffer->tfm_desc, &buffer->ztensor));
-}
-
-static void ggml_zdnn_mul_mat_op(ggml_backend_zdnn_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_TENSOR_BINARY_OP_LOCALS;
-
-    const enum ggml_type type = src0->type;
-
-    GGML_ASSERT(ne0 == ne01);
-    GGML_ASSERT(ne1 == ne11);
-    GGML_ASSERT(ne2 == ne12);
-    GGML_ASSERT(ne3 == ne13);
-
-    // we don't support permuted src0 or src1
-    GGML_ASSERT(nb00 == ggml_type_size(type));
-    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
-
-    // dst cannot be transposed or permuted
-    GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(nb0 <= nb1);
-    GGML_ASSERT(nb1 <= nb2);
-    GGML_ASSERT(nb2 <= nb3);
-
-    const ggml_tensor * weights = src0;
-    const ggml_tensor * inputs  = src1;
-          ggml_tensor * output  = dst;
-
-    ggml_backend_zdnn_buffer * weights_extra = (ggml_backend_zdnn_buffer *)weights->extra;
-    ggml_backend_zdnn_buffer * inputs_extra  = (ggml_backend_zdnn_buffer *)inputs->extra;
-    ggml_backend_zdnn_buffer * output_extra  = (ggml_backend_zdnn_buffer *)output->extra;
-    ggml_backend_zdnn_buffer * bias_extra    = (ggml_backend_zdnn_buffer *)output_extra->extra;
-
-    const int64_t weights_rows = ne01;
-    const int64_t weights_cols = ne00;
-    const int64_t inputs_rows  = ne11;
-    const int64_t inputs_cols  = ne10;
-
-    assert(inputs_cols == weights_cols);
-
-    const int64_t output_rows = ne1;
-    const int64_t output_cols = ne0;
-
-    // GGML_LOG_INFO("%s: tensor '%s' tensor dimensions: [%ld, %ld, %ld, %ld] pre_tfm_desc dimensions: [%ld, %ld, %ld, %ld]\n",
-    //               __func__, weights_extra->name,
-    //               weights->ne[3], weights->ne[2], weights->ne[1], weights->ne[0],
-    //               weights_extra->pre_tfm_desc.dim1,
-    //               weights_extra->pre_tfm_desc.dim2,
-    //               weights_extra->pre_tfm_desc.dim3,
-    //               weights_extra->pre_tfm_desc.dim4);
-
-    // GGML_LOG_INFO("%s: tensor '%s' tensor dimensions: [%ld, %ld, %ld, %ld] pre_tfm_desc dimensions: [%ld, %ld, %ld, %ld]\n",
-    //               __func__, inputs_extra->name,
-    //               inputs->ne[3], inputs->ne[2], inputs->ne[1], inputs->ne[0],
-    //               inputs_extra->pre_tfm_desc.dim1,
-    //               inputs_extra->pre_tfm_desc.dim2,
-    //               inputs_extra->pre_tfm_desc.dim3,
-    //               inputs_extra->pre_tfm_desc.dim4);
-
-    GGML_ASSERT(weights_extra->pre_tfm_desc.dim1 == weights->ne[0] && "weights_extra->pre_tfm_desc.dim1 must match weights->ne[0]");
-    GGML_ASSERT(weights_extra->pre_tfm_desc.dim2 == weights->ne[1] && "weights_extra->pre_tfm_desc.dim2 must match weights->ne[1]");
-    GGML_ASSERT(inputs_extra->pre_tfm_desc.dim1  == inputs->ne[0]  && "inputs_extra->pre_tfm_desc.dim1 must match inputs->ne[0]");
-    GGML_ASSERT(inputs_extra->pre_tfm_desc.dim2  == inputs->ne[1]  && "inputs_extra->pre_tfm_desc.dim2 must match inputs->ne[1]");
-
-    ZDNN_CHECK(zdnn_matmul_transpose_op(&inputs_extra->ztensor, &weights_extra->ztensor, &bias_extra->ztensor,
-                                        false, true, MATMUL_OP_ADDITION, &output_extra->ztensor));
-    // TODO: Remove in the future as we are currently DLF16 -> FP32 then in the next op, FP32 -> DLF16 again. Inefficient.
-    ZDNN_CHECK(zdnn_transform_origtensor(&output_extra->ztensor, output->data));
-
-    GGML_UNUSED(ctx);
-    GGML_UNUSED(weights_rows);
-    GGML_UNUSED(weights_cols);
-    GGML_UNUSED(inputs_rows);
-    GGML_UNUSED(inputs_cols);
-    GGML_UNUSED(output_rows);
-    GGML_UNUSED(output_cols);
-}
-
-static void ggml_zdnn_mul_mat_dispatch(ggml_backend_zdnn_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    // debug helpers
-    // GGML_LOG_INFO("%s: use_mul_mat_vec   = %d\n", __func__, use_mul_mat_vec);
-    // GGML_LOG_INFO("%s: use_mul_mat_vec_q = %d\n", __func__, use_mul_mat_vec_q);
-    // GGML_LOG_INFO("%s: use_mul_mat_q     = %d\n", __func__, use_mul_mat_q);
-    // GGML_LOG_INFO("%s: src0: %8d %8d %8d %8d\n", __func__, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
-    // GGML_LOG_INFO("%s:       %8d %8d %8d %8d\n", __func__, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
-    // GGML_LOG_INFO("%s: src1: %8d %8d %8d %8d\n", __func__, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]);
-    // GGML_LOG_INFO("%s:       %8d %8d %8d %8d\n", __func__, src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]);
-    // GGML_LOG_INFO("%s: src0 is contiguous %d, transposed %d, type = %s, name = %s\n", __func__, ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
-    // GGML_LOG_INFO("%s: src1 is contiguous %d, transposed %d, type = %s, name = %s\n", __func__, ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
-
-    ggml_zdnn_mul_mat_op(ctx, src0, src1, dst);
-}
-
-static bool ggml_zdnn_compute_forward(ggml_backend_zdnn_context * ctx, ggml_tensor * dst) {
    switch (dst->op) {
        case GGML_OP_MUL_MAT:
-            ggml_zdnn_mul_mat_dispatch(ctx, dst->src[0], dst->src[1], dst);
-            break;
+            {
+                ggml_zdnn_compute_forward_mul_mat(ctx, dst);
+            } break;

        default:
            return false;
@@ -574,7 +425,7 @@ static ggml_backend_i ggml_backend_zdnn_i = {
    /* .graph_compute      = */ ggml_backend_zdnn_graph_compute,
    /* .event_record       = */ NULL,
    /* .event_wait         = */ NULL,
-    /* .optimize_graph     = */ NULL,
+    /* .graph_optimize     = */ NULL,
 };

 static ggml_guid_t ggml_backend_zdnn_guid(void) {
@@ -0,0 +1,80 @@
+#include "ggml.h"
+#include "mmf.hpp"
+
+void ggml_zdnn_mul_mat_f(
+    const ggml_backend_zdnn_context * ctx,
+    const               ggml_tensor * src0,
+    const               ggml_tensor * src1,
+                        ggml_tensor * dst) {
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    const enum ggml_type type = src0->type;
+
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne1 == ne11);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    const ggml_tensor * weights = src0;
+    const ggml_tensor * inputs  = src1;
+          ggml_tensor * output  = dst;
+
+    ggml_backend_zdnn_buffer * weights_extra = (ggml_backend_zdnn_buffer *)weights->extra;
+    ggml_backend_zdnn_buffer * inputs_extra  = (ggml_backend_zdnn_buffer *)inputs->extra;
+    ggml_backend_zdnn_buffer * output_extra  = (ggml_backend_zdnn_buffer *)output->extra;
+    ggml_backend_zdnn_buffer * bias_extra    = (ggml_backend_zdnn_buffer *)output_extra->extra;
+
+    const int64_t weights_rows = ne01;
+    const int64_t weights_cols = ne00;
+    const int64_t inputs_rows  = ne11;
+    const int64_t inputs_cols  = ne10;
+
+    assert(inputs_cols == weights_cols);
+
+    const int64_t output_rows = ne1;
+    const int64_t output_cols = ne0;
+
+    // GGML_LOG_INFO("%s: tensor '%s' tensor dimensions: [%ld, %ld, %ld, %ld] pre_tfm_desc dimensions: [%ld, %ld, %ld, %ld]\n",
+    //               __func__, weights_extra->name,
+    //               weights->ne[3], weights->ne[2], weights->ne[1], weights->ne[0],
+    //               weights_extra->pre_tfm_desc.dim1,
+    //               weights_extra->pre_tfm_desc.dim2,
+    //               weights_extra->pre_tfm_desc.dim3,
+    //               weights_extra->pre_tfm_desc.dim4);
+
+    // GGML_LOG_INFO("%s: tensor '%s' tensor dimensions: [%ld, %ld, %ld, %ld] pre_tfm_desc dimensions: [%ld, %ld, %ld, %ld]\n",
+    //               __func__, inputs_extra->name,
+    //               inputs->ne[3], inputs->ne[2], inputs->ne[1], inputs->ne[0],
+    //               inputs_extra->pre_tfm_desc.dim1,
+    //               inputs_extra->pre_tfm_desc.dim2,
+    //               inputs_extra->pre_tfm_desc.dim3,
+    //               inputs_extra->pre_tfm_desc.dim4);
+
+    GGML_ASSERT(weights_extra->pre_tfm_desc.dim1 == weights->ne[0] && "weights_extra->pre_tfm_desc.dim1 must match weights->ne[0]");
+    GGML_ASSERT(weights_extra->pre_tfm_desc.dim2 == weights->ne[1] && "weights_extra->pre_tfm_desc.dim2 must match weights->ne[1]");
+    GGML_ASSERT(inputs_extra->pre_tfm_desc.dim1  == inputs->ne[0]  && "inputs_extra->pre_tfm_desc.dim1 must match inputs->ne[0]");
+    GGML_ASSERT(inputs_extra->pre_tfm_desc.dim2  == inputs->ne[1]  && "inputs_extra->pre_tfm_desc.dim2 must match inputs->ne[1]");
+
+    ZDNN_CHECK(zdnn_matmul_transpose_op(&inputs_extra->ztensor, &weights_extra->ztensor, &bias_extra->ztensor,
+                                        false, true, MATMUL_OP_ADDITION, &output_extra->ztensor));
+    // TODO: Remove in the future as we are currently DLF16 -> FP32 then in the next op, FP32 -> DLF16 again. Inefficient.
+    ZDNN_CHECK(zdnn_transform_origtensor(&output_extra->ztensor, output->data));
+
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(weights_rows);
+    GGML_UNUSED(weights_cols);
+    GGML_UNUSED(inputs_rows);
+    GGML_UNUSED(inputs_cols);
+    GGML_UNUSED(output_rows);
+    GGML_UNUSED(output_cols);
+}
@@ -0,0 +1,12 @@
+#ifndef GGML_ZDNN_MMF_HPP
+#define GGML_ZDNN_MMF_HPP
+
+#include "common.hpp"
+
+void ggml_zdnn_mul_mat_f(
+    const ggml_backend_zdnn_context * ctx,
+    const               ggml_tensor * src0,
+    const               ggml_tensor * src1,
+                        ggml_tensor * dst);
+
+#endif  // GGML_ZDNN_MMF_HPP
@@ -0,0 +1,79 @@
+#include "ggml.h"
+#include "utils.hpp"
+
+zdnn_data_types ggml_zdnn_type_mapping(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return FP32;
+        case GGML_TYPE_F16:
+            return FP16;
+        case GGML_TYPE_BF16:
+            return BFLOAT;
+        case GGML_TYPE_Q8_0:
+            return INT8;
+        case GGML_TYPE_I8:
+            return INT8;
+        case GGML_TYPE_I32:
+            return INT32;
+        default:
+            GGML_ABORT("%s: fatal: unable to determine zTensor data type",
+                       __func__);
+            break;
+    }
+}
+
+void ggml_zdnn_create_tensor(zdnn_tensor_desc  & pre_tfm_desc,
+                             zdnn_tensor_desc  & tfm_desc,
+                             zdnn_ztensor      & ztensor,
+                       const ggml_tensor       * src,
+                       const int64_t           * ne,
+                       const zdnn_data_layouts   layout) {
+    zdnn_init_pre_transformed_desc(
+        layout,
+        ggml_zdnn_type_mapping(src->type),
+        &pre_tfm_desc,
+        ne[3], ne[2], ne[1], ne[0]
+    );
+
+    ZDNN_CHECK(zdnn_generate_transformed_desc(&pre_tfm_desc, &tfm_desc));
+    ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&pre_tfm_desc, &tfm_desc, &ztensor));
+}
+
+void ggml_zdnn_load_tensor(zdnn_ztensor & ztensor, void * buffer) {
+    ZDNN_CHECK(zdnn_transform_ztensor(&ztensor, buffer));
+}
+
+void ggml_zdnn_init_tensor(ggml_backend_zdnn_buffer * buffer, const ggml_tensor * tensor) {
+    switch (tensor->op) {
+        case GGML_OP_MUL_MAT:
+            {
+                zdnn_init_pre_transformed_desc(
+                    ZDNN_2D,
+                    ggml_zdnn_type_mapping(tensor->type),
+                    &buffer->pre_tfm_desc,
+                    tensor->ne[1], tensor->ne[0]
+                );
+            } break;
+
+        default:
+            {
+                // For 4D tensors, GGML uses NCHW layout. However, because zDNN
+                // automatically transforms everything to NHWC, we will use it
+                // directly to avoid the performance penalty changing the
+                // layout and reshaping the tensor.
+                zdnn_init_pre_transformed_desc(
+                    ZDNN_NHWC,
+                    ggml_zdnn_type_mapping(tensor->type),
+                    &buffer->pre_tfm_desc,
+                    tensor->ne[3], tensor->ne[2], tensor->ne[1], tensor->ne[0]
+                );
+
+                // TODO: Consider adding a ggml check.
+                // TODO: If tensor = 4D, use ZDNN_NCHW by default.
+                // TODO: If tensor = 2D, use ZDNN_NHWC by default.
+            } break;
+    }
+
+    ZDNN_CHECK(zdnn_generate_transformed_desc(&buffer->pre_tfm_desc, &buffer->tfm_desc));
+    ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&buffer->pre_tfm_desc, &buffer->tfm_desc, &buffer->ztensor));
+}
@@ -0,0 +1,19 @@
+#ifndef GGML_ZDNN_UTILITIES_HPP
+#define GGML_ZDNN_UTILITIES_HPP
+
+#include "common.hpp"
+
+zdnn_data_types ggml_zdnn_type_mapping(ggml_type type);
+
+void ggml_zdnn_create_tensor(zdnn_tensor_desc & pre_tfm_desc,
+                             zdnn_tensor_desc & tfm_desc,
+                             zdnn_ztensor     & ztensor,
+                      const ggml_tensor       * src,
+                      const int64_t           * ne,
+                      const zdnn_data_layouts   layout);
+
+void ggml_zdnn_load_tensor(zdnn_ztensor & ztensor, void * buffer);
+
+void ggml_zdnn_init_tensor(ggml_backend_zdnn_buffer * buffer, const ggml_tensor * tensor);
+
+#endif  // GGML_ZDNN_UTILITIES_HPP
--- a/Show More
+++ b/Show More