vulkan : make the vulkan.hpp dynamic dispatcher instance private (#16224 )

* don't use VULKAN_HPP_DEFAULT_DISPATCH_LOADER_DYNAMIC_STORAGE which can cause conflicts if application or other libraries do the same
Show message actions by default (#16289 )
2025-09-27 22:41:03 +02:00 · 2025-09-27 19:56:40 +02:00 · 2025-09-27 18:49:32 +02:00 · 2025-09-27 18:45:07 +02:00 · 2025-09-27 18:26:46 +02:00 · 2025-09-27 19:17:08 +03:00
266 changed files with 12554 additions and 6322 deletions
@@ -1,10 +1,10 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG MUSA_VERSION=rc4.2.0
+ARG MUSA_VERSION=rc4.3.0
 # Target the MUSA build image
-ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64
+ARG BASE_MUSA_DEV_CONTAINER=sh-harbor.mthreads.com/haive/mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64

-ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64
+ARG BASE_MUSA_RUN_CONTAINER=sh-harbor.mthreads.com/haive/mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64

 FROM ${BASE_MUSA_DEV_CONTAINER} AS build

@@ -2,10 +2,10 @@ ARG GCC_VERSION=15.2.0
 ARG UBUNTU_VERSION=24.04

 ### Build Llama.cpp stage
-FROM --platform=linux/s390x gcc:${GCC_VERSION} AS build
+FROM gcc:${GCC_VERSION} AS build

-RUN --mount=type=cache,target=/var/cache/apt \
-    --mount=type=cache,target=/var/lib/apt/lists \
+RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
+    --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
    apt update -y && \
    apt upgrade -y && \
    apt install -y --no-install-recommends \
@@ -40,7 +40,7 @@ COPY requirements     /opt/llama.cpp/gguf-py/requirements


 ### Collect all llama.cpp binaries, libraries and distro libraries
-FROM --platform=linux/s390x scratch AS collector
+FROM scratch AS collector

 # Copy llama.cpp binaries and libraries
 COPY --from=build /opt/llama.cpp/bin     /llama.cpp/bin
@@ -49,13 +49,14 @@ COPY --from=build /opt/llama.cpp/gguf-py /llama.cpp/gguf-py


 ### Base image
-FROM --platform=linux/s390x ubuntu:${UBUNTU_VERSION} AS base
+FROM ubuntu:${UBUNTU_VERSION} AS base

-RUN --mount=type=cache,target=/var/cache/apt \
-    --mount=type=cache,target=/var/lib/apt/lists \
+RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
+    --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
    apt update -y && \
    apt install -y --no-install-recommends \
        # WARNING: Do not use libopenblas-openmp-dev. libopenblas-dev is faster.
+        # See: https://github.com/ggml-org/llama.cpp/pull/15915#issuecomment-3317166506
        curl libgomp1 libopenblas-dev && \
    apt autoremove -y && \
    apt clean -y && \
@@ -68,13 +69,13 @@ COPY --from=collector /llama.cpp/lib /usr/lib/s390x-linux-gnu


 ### Full
-FROM --platform=linux/s390x base AS full
+FROM 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 \
+RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
+    --mount=type=cache,target=/var/lib/apt/lists,sharing=locked \
    apt update -y && \
    apt install -y \
        git cmake libjpeg-dev \
@@ -97,7 +98,7 @@ ENTRYPOINT [ "/app/tools.sh" ]


 ### CLI Only
-FROM --platform=linux/s390x base AS light
+FROM base AS light

 WORKDIR /llama.cpp/bin

@@ -108,7 +109,7 @@ ENTRYPOINT [ "/llama.cpp/bin/llama-cli" ]


 ### Server
-FROM --platform=linux/s390x base AS server
+FROM base AS server

 ENV LLAMA_ARG_HOST=0.0.0.0

@@ -141,97 +141,6 @@ jobs:

  #         cmake --build build --config Release -j $(nproc)

-  ubuntu-24-ppc64el-cpu-cross:
-    runs-on: ubuntu-24.04
-
-    steps:
-      - uses: actions/checkout@v4
-      - name: Setup PowerPC64le
-        run: |
-          sudo dpkg --add-architecture ppc64el
-
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-          EOF
-
-          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
-
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  gcc-14-powerpc64le-linux-gnu \
-                  g++-14-powerpc64le-linux-gnu
-
-      - name: Build
-        run: |
-          cmake -B build -DLLAMA_CURL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
-                         -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
-
-          cmake --build build --config Release -j $(nproc)
-
-  # ubuntu-24-ppc64el-vulkan-cross:
-  #   runs-on: ubuntu-24.04
-
-  #   steps:
-  #     - uses: actions/checkout@v4
-  #     - name: Setup PowerPC64le
-  #       run: |
-  #         sudo dpkg --add-architecture ppc64el
-
-  #         # Add arch-specific repositories for non-amd64 architectures
-  #         cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
-  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-  #         deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-  #         EOF
-
-  #         sudo apt-get update || true    ;# Prevent failure due to missing URLs.
-
-  #         sudo apt-get install -y --no-install-recommends \
-  #                 build-essential \
-  #                 glslc \
-  #                 gcc-14-powerpc64le-linux-gnu \
-  #                 g++-14-powerpc64le-linux-gnu \
-  #                 libvulkan-dev:ppc64el
-
-  #     - name: Build
-  #       run: |
-  #         cmake -B build -DLLAMA_CURL=OFF \
-  #                        -DCMAKE_BUILD_TYPE=Release \
-  #                        -DGGML_VULKAN=ON \
-  #                        -DGGML_OPENMP=OFF \
-  #                        -DLLAMA_BUILD_EXAMPLES=ON \
-  #                        -DLLAMA_BUILD_TOOLS=ON \
-  #                        -DLLAMA_BUILD_TESTS=OFF \
-  #                        -DCMAKE_SYSTEM_NAME=Linux \
-  #                        -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
-  #                        -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
-  #                        -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
-  #                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-  #                        -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
-  #                        -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-  #                        -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-  #                        -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
-
-  #         cmake --build build --config Release -j $(nproc)
-
  debian-13-loongarch64-cpu-cross:
    runs-on: ubuntu-24.04
    container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
@@ -192,6 +192,10 @@ jobs:
            os: ubuntu-22.04
          - build: 'arm64'
            os: ubuntu-22.04-arm
+          - build: 's390x'
+            os: ubuntu-24.04-s390x
+          - build: 'ppc64le'
+            os: ubuntu-24.04-ppc64le

    runs-on: ${{ matrix.os }}

@@ -206,11 +210,28 @@ jobs:
          key: ubuntu-cpu-cmake
          evict-old-files: 1d

-      - name: Dependencies
-        id: depends
+      - name: Build Dependencies
+        id: build_depends
        run: |
          sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install -y --no-install-recommends \
+            python3 python3-pip python3-dev \
+            libjpeg-dev build-essential libcurl4-openssl-dev \
+            git-lfs
+
+      - name: Python Dependencies
+        id: python_depends
+        run: |
+          python3 -m pip install --upgrade pip
+          pip3 install ./gguf-py
+
+      - name: Swap Endianness
+        id: endianness
+        if: ${{ matrix.build == 's390x' }}
+        run: |
+          for f in models/*.gguf; do
+            echo YES | python3 gguf-py/gguf/scripts/gguf_convert_endian.py $f big
+          done

      - name: Build
        id: cmake_build
@@ -228,6 +249,7 @@ jobs:

      - name: Test llama2c conversion
        id: llama2c_test
+        if: ${{ matrix.build != 's390x' }}
        run: |
          cd build
          echo "Fetch tokenizer"
@@ -237,6 +259,15 @@ jobs:
          ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
          ./bin/llama-cli -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256

+      - name: Test llama2c (s390x)
+        id: llama2c_test_s390x
+        if: ${{ matrix.build == 's390x' }}
+        run: |
+          cd build
+          echo "Fetch llama2c big-endian model"
+          wget https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K-be.gguf
+          ./bin/llama-cli -m stories260K-be.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+
  ubuntu-latest-cmake-sanitizer:
    runs-on: ubuntu-latest

@@ -475,7 +506,7 @@ jobs:

  ubuntu-22-cmake-musa:
    runs-on: ubuntu-22.04
-    container: mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
+    container: mthreads/musa:rc4.3.0-devel-ubuntu22.04-amd64

    steps:
      - name: Clone
@@ -1251,56 +1282,129 @@ jobs:
 # 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]
+    runs-on: ubuntu-22.04

    steps:
      - name: Clone
        id: checkout
        uses: actions/checkout@v4

+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ggml-ci-x64-cpu-low-perf
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libcurl4-openssl-dev
+
      - name: Test
        id: ggml-ci
        run: |
-          bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

  ggml-ci-arm64-cpu-low-perf:
-    runs-on: [self-hosted, Linux, ARM64, CPU, low-perf]
+    runs-on: ubuntu-22.04-arm

    steps:
      - name: Clone
        id: checkout
        uses: actions/checkout@v4

+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ggml-ci-arm64-cpu-low-perf
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libcurl4-openssl-dev
+
      - name: Test
        id: ggml-ci
        run: |
-          bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_LOW_PERF=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

  ggml-ci-x64-cpu-high-perf:
-    runs-on: [self-hosted, Linux, X64, CPU, high-perf]
+    runs-on: ubuntu-22.04

    steps:
      - name: Clone
        id: checkout
        uses: actions/checkout@v4

+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ggml-ci-x64-cpu-high-perf
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libcurl4-openssl-dev
+
      - name: Test
        id: ggml-ci
        run: |
-          bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+          LLAMA_ARG_THREADS=$(nproc) bash ./ci/run.sh ./tmp/results ./tmp/mnt

  ggml-ci-arm64-cpu-high-perf:
-    runs-on: [self-hosted, Linux, ARM64, CPU, high-perf]
+    runs-on: ubuntu-22.04-arm

    steps:
      - name: Clone
        id: checkout
        uses: actions/checkout@v4

+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ggml-ci-arm64-cpu-high-perf
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libcurl4-openssl-dev
+
      - 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
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_NO_SVE=1 GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+
+  ggml-ci-arm64-cpu-high-perf-sve:
+    runs-on: ubuntu-22.04-arm
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ggml-ci-arm64-cpu-high-perf-sve
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libcurl4-openssl-dev
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

  ggml-ci-x64-nvidia-cuda:
    runs-on: [self-hosted, Linux, X64, NVIDIA]
@@ -68,22 +68,19 @@ jobs:
          username: ${{ github.repository_owner }}
          password: ${{ secrets.GITHUB_TOKEN }}

-      - name: Determine tag name
+      - name: Determine source tag name
+        id: srctag
+        uses: ./.github/actions/get-tag-name
+        env:
+          BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+
+      - name: Determine image tag name
        id: tag
        shell: bash
        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
          REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
          REPO_NAME="${{ github.event.repository.name }}"

-          # determine tag name postfix (build number, commit hash)
-          if [[ "${{ env.GITHUB_BRANCH_NAME }}" == "master" ]]; then
-            TAG_POSTFIX="-b${BUILD_NUMBER}"
-          else
-            SAFE_NAME=$(echo "${{ env.GITHUB_BRANCH_NAME }}" | tr '/' '-')
-            TAG_POSTFIX="-${SAFE_NAME}-${SHORT_HASH}"
-          fi
          # list all tags possible
          if [[ "${{ matrix.config.tag }}" == "cpu" ]]; then
              TYPE=""
@@ -91,9 +88,9 @@ jobs:
              TYPE="-${{ matrix.config.tag }}"
          fi
          PREFIX="ghcr.io/${REPO_OWNER}/${REPO_NAME}:"
-          FULLTAGS="${PREFIX}full${TYPE},${PREFIX}full${TYPE}${TAG_POSTFIX}"
-          LIGHTTAGS="${PREFIX}light${TYPE},${PREFIX}light${TYPE}${TAG_POSTFIX}"
-          SERVERTAGS="${PREFIX}server${TYPE},${PREFIX}server${TYPE}${TAG_POSTFIX}"
+          FULLTAGS="${PREFIX}full${TYPE},${PREFIX}full${TYPE}-${{ steps.srctag.outputs.name }}"
+          LIGHTTAGS="${PREFIX}light${TYPE},${PREFIX}light${TYPE}-${{ steps.srctag.outputs.name }}"
+          SERVERTAGS="${PREFIX}server${TYPE},${PREFIX}server${TYPE}-${{ steps.srctag.outputs.name }}"
          echo "full_output_tags=$FULLTAGS" >> $GITHUB_OUTPUT
          echo "light_output_tags=$LIGHTTAGS" >> $GITHUB_OUTPUT
          echo "server_output_tags=$SERVERTAGS" >> $GITHUB_OUTPUT
@@ -101,7 +98,6 @@ jobs:
          echo "light_output_tags=$LIGHTTAGS"  # print out for debugging
          echo "server_output_tags=$SERVERTAGS"  # print out for debugging
        env:
-          GITHUB_BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'

      - name: Free Disk Space (Ubuntu)
@@ -177,3 +173,29 @@ jobs:
          # return to this if the experimental github cache is having issues
          #cache-to: type=local,dest=/tmp/.buildx-cache
          #cache-from: type=local,src=/tmp/.buildx-cache
+
+  create_tag:
+    name: Create and push git tag
+    runs-on: ubuntu-22.04
+    permissions:
+      contents: write
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Determine source tag name
+        id: srctag
+        uses: ./.github/actions/get-tag-name
+        env:
+          BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+
+      - name: Create and push git tag
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          git tag ${{ steps.srctag.outputs.name }} || exit 0
+          git push origin ${{ steps.srctag.outputs.name }} || exit 0
@@ -149,6 +149,6 @@ poetry.toml
 /run-chat.sh
 .ccache/

-# Code Workspace
+# IDE
 *.code-workspace
-
+.windsurf/
@@ -1,7 +0,0 @@
---
-trigger: manual
---
-
-#### Tailwind & CSS
-
-   We are using Tailwind v4 which uses oklch colors so we now want to refer to the CSS vars directly, without wrapping it with any color function like `hsla/hsl`, `rgba` etc.
@@ -1,48 +0,0 @@
---
-trigger: manual
---
-
-# Coding rules
-
-## Svelte & SvelteKit
-
-### Services vs Stores Separation Pattern
-
-#### `lib/services/` - Pure Business Logic
-
-   **Purpose**: Stateless business logic and external communication
-   **Contains**:
-    -   API calls to external services (ApiService)
-    -   Pure business logic functions (ChatService, etc.)
-   **Rules**:
-    -   NO Svelte runes ($state, $derived, $effect)
-    -   NO reactive state management
-    -   Pure functions and classes only
-    -   Can import types but not stores
-    -   Focus on "how" - implementation details
-
-#### `lib/stores/` - Reactive State Management
-
-   **Purpose**: Svelte-specific reactive state with runes
-   **Contains**:
-    -   Reactive state classes with $state, $derived, $effect
-    -   Database operations (DatabaseStore)
-    -   UI-focused state management
-    -   Store orchestration logic
-   **Rules**:
-    -   USE Svelte runes for reactivity
-    -   Import and use services for business logic
-    -   NO direct database operations
-    -   NO direct API calls (use services)
-    -   Focus on "what" - reactive state for UI
-
-#### Enforcement
-
-   Services should be testable without Svelte
-   Stores should leverage Svelte's reactivity system
-   Clear separation: services handle data, stores handle state
-   Services can be reused across multiple stores
-
-#### Misc
-
-   Always use `let` for $derived state variables
@@ -1,9 +0,0 @@
---
-trigger: manual
---
-
-# Automated Tests
-
-## General rules
-
-   NEVER include any test code in the production code - we should always have it in a separate dedicated files
@@ -1,7 +0,0 @@
---
-trigger: manual
---
-
-## TypeScript
-
-   Add JSDocs for functions
@@ -92,6 +92,7 @@ option(LLAMA_TOOLS_INSTALL  "llama: install tools"        ${LLAMA_TOOLS_INSTALL_

 # 3rd party libs
 option(LLAMA_CURL       "llama: use libcurl to download model from an URL" ON)
+option(LLAMA_OPENSSL    "llama: use openssl to support HTTPS" OFF)
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)

 # Required for relocatable CMake package
@@ -61,9 +61,10 @@
 /ggml/src/ggml-metal/                   @ggerganov
 /ggml/src/ggml-opt.cpp                  @JohannesGaessler
 /ggml/src/ggml-quants.*                 @ggerganov
+/ggml/src/ggml-rpc/                     @rgerganov
 /ggml/src/ggml-threading.*              @ggerganov @slaren
 /ggml/src/ggml-vulkan/                  @0cc4m
-/ggml/src/ggml-zdnn/                    @taronaeo
+/ggml/src/ggml-zdnn/                    @taronaeo @Andreas-Krebbel @AlekseiNikiforovIBM
 /ggml/src/ggml.c                        @ggerganov @slaren
 /ggml/src/ggml.cpp                      @ggerganov @slaren
 /ggml/src/gguf.cpp                      @JohannesGaessler @Green-Sky
@@ -89,6 +90,7 @@
 /tools/mtmd/                            @ngxson
 /tools/perplexity/                      @ggerganov
 /tools/quantize/                        @ggerganov
+/tools/rpc/                             @rgerganov
 /tools/run/                             @ericcurtin
 /tools/server/*                         @ngxson @ggerganov @ericcurtin # no subdir
 /tools/server/webui/                    @allozaur
@@ -103,4 +105,5 @@
 /LICENSE                                @ggerganov
 /README.md                              @ggerganov
 /SECURITY.md                            @ggerganov
+/build-xcframework.sh                   @danbev
 requirements*.txt                       @CISC
@@ -25,7 +25,7 @@ The project differentiates between 3 levels of contributors:
 - 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
- Let other maintainers, merge their own PRs
+- 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)

@@ -178,6 +178,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
 - React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
 - Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
+- Java: [QuasarByte/llama-cpp-jna](https://github.com/QuasarByte/llama-cpp-jna)
 - Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
 - Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
 - Flutter: [xuegao-tzx/Fllama](https://github.com/xuegao-tzx/Fllama)
@@ -422,6 +422,7 @@ echo "Building for iOS devices..."
 cmake -B build-ios-device -G Xcode \
    "${COMMON_CMAKE_ARGS[@]}" \
    -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
+    -DCMAKE_SYSTEM_NAME=iOS \
    -DCMAKE_OSX_SYSROOT=iphoneos \
    -DCMAKE_OSX_ARCHITECTURES="arm64" \
    -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
@@ -21,7 +21,7 @@ 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
+    mthreads/musa:rc4.3.0-devel-ubuntu22.04-amd64
 ```

 Inside the container, execute the following commands:
@@ -109,6 +109,11 @@ if [ ! -z ${GG_BUILD_MUSA} ]; then
    MUSA_ARCH=${MUSA_ARCH:-21}
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_MUSA=ON -DMUSA_ARCHITECTURES=${MUSA_ARCH}"
 fi
+
+if [ ! -z ${GG_BUILD_NO_SVE} ]; then
+    # arm 9 and newer enables sve by default, adjust these flags depending on the cpu used
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm"
+fi
 ## helpers

 # download a file if it does not exist or if it is outdated
@@ -345,16 +350,16 @@ function gg_run_qwen3_0_6b {

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

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

    (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
@@ -427,7 +432,7 @@ function gg_run_qwen3_0_6b {
 function gg_sum_qwen3_0_6b {
    gg_printf '### %s\n\n' "${ci}"

-    gg_printf 'Pythia 2.8B:\n'
+    gg_printf 'Qwen3 0.6B:\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)"
@@ -87,7 +87,43 @@ if (LLAMA_CURL)
    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL)
    include_directories(${CURL_INCLUDE_DIRS})
    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARIES})
-endif ()
+endif()
+
+if (LLAMA_OPENSSL)
+    find_package(OpenSSL)
+    if (OpenSSL_FOUND)
+        include(CheckCSourceCompiles)
+        set(SAVED_CMAKE_REQUIRED_INCLUDES ${CMAKE_REQUIRED_INCLUDES})
+        set(CMAKE_REQUIRED_INCLUDES ${OPENSSL_INCLUDE_DIR})
+        check_c_source_compiles("
+        #include <openssl/opensslv.h>
+        #if defined(OPENSSL_IS_BORINGSSL) || defined(LIBRESSL_VERSION_NUMBER)
+        #    if OPENSSL_VERSION_NUMBER < 0x1010107f
+        #        error bad version
+        #    endif
+        #else
+        #    if OPENSSL_VERSION_NUMBER < 0x30000000L
+        #        error bad version
+        #    endif
+        #endif
+        int main() { return 0; }
+        " OPENSSL_VERSION_SUPPORTED)
+        set(CMAKE_REQUIRED_INCLUDES ${SAVED_CMAKE_REQUIRED_INCLUDES})
+        if (OPENSSL_VERSION_SUPPORTED)
+            message(STATUS "OpenSSL found: ${OPENSSL_VERSION}")
+            target_compile_definitions(${TARGET} PUBLIC CPPHTTPLIB_OPENSSL_SUPPORT)
+            target_link_libraries(${TARGET} PUBLIC OpenSSL::SSL OpenSSL::Crypto)
+            if (APPLE AND CMAKE_SYSTEM_NAME STREQUAL "Darwin")
+                target_compile_definitions(${TARGET} PUBLIC CPPHTTPLIB_USE_CERTS_FROM_MACOSX_KEYCHAIN)
+                find_library(CORE_FOUNDATION_FRAMEWORK CoreFoundation REQUIRED)
+                find_library(SECURITY_FRAMEWORK Security REQUIRED)
+                target_link_libraries(${TARGET} PUBLIC ${CORE_FOUNDATION_FRAMEWORK} ${SECURITY_FRAMEWORK})
+            endif()
+        endif()
+    else()
+        message(STATUS "OpenSSL not found, SSL support disabled")
+    endif()
+endif()

 if (LLAMA_LLGUIDANCE)
    include(ExternalProject)
@@ -37,6 +37,8 @@
 #if defined(LLAMA_USE_CURL)
 #include <curl/curl.h>
 #include <curl/easy.h>
+#else
+#include <cpp-httplib/httplib.h>
 #endif

 #ifdef __linux__
@@ -572,17 +574,364 @@ 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;
-}
+struct common_url {
+    std::string scheme;
+    std::string user;
+    std::string password;
+    std::string host;
+    std::string path;
+};

-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");
+static common_url parse_url(const std::string & url) {
+    common_url parts;
+    auto scheme_end = url.find("://");
+
+    if (scheme_end == std::string::npos) {
+        throw std::runtime_error("invalid URL: no scheme");
+    }
+    parts.scheme = url.substr(0, scheme_end);
+
+    if (parts.scheme != "http" && parts.scheme != "https") {
+        throw std::runtime_error("unsupported URL scheme: " + parts.scheme);
    }

-    return {};
+    auto rest = url.substr(scheme_end + 3);
+    auto at_pos = rest.find('@');
+
+    if (at_pos != std::string::npos) {
+        auto auth = rest.substr(0, at_pos);
+        auto colon_pos = auth.find(':');
+        if (colon_pos != std::string::npos) {
+            parts.user = auth.substr(0, colon_pos);
+            parts.password = auth.substr(colon_pos + 1);
+        } else {
+            parts.user = auth;
+        }
+        rest = rest.substr(at_pos + 1);
+    }
+
+    auto slash_pos = rest.find('/');
+
+    if (slash_pos != std::string::npos) {
+        parts.host = rest.substr(0, slash_pos);
+        parts.path = rest.substr(slash_pos);
+    } else {
+        parts.host = rest;
+        parts.path = "/";
+    }
+    return parts;
+}
+
+static std::pair<httplib::Client, common_url> http_client(const std::string & url) {
+    common_url parts = parse_url(url);
+
+    if (parts.host.empty()) {
+        throw std::runtime_error("error: invalid URL format");
+    }
+
+    if (!parts.user.empty()) {
+        throw std::runtime_error("error: user:password@ not supported yet"); // TODO
+    }
+
+    httplib::Client cli(parts.scheme + "://" + parts.host);
+    cli.set_follow_location(true);
+
+    // TODO cert
+
+    return { std::move(cli), std::move(parts) };
+}
+
+static std::string show_masked_url(const common_url & parts) {
+    return parts.scheme + "://" + (parts.user.empty() ? "" : "****:****@") + parts.host + parts.path;
+}
+
+static void print_progress(size_t current, size_t total) { // TODO isatty
+    if (!total) {
+        return;
+    }
+
+    size_t width = 50;
+    size_t pct = (100 * current) / total;
+    size_t pos = (width * current) / total;
+
+    std::cout << "["
+              << std::string(pos, '=')
+              << (pos < width ? ">" : "")
+              << std::string(width - pos, ' ')
+              << "] " << std::setw(3) << pct << "%  ("
+              << current / (1024 * 1024) << " MB / "
+              << total / (1024 * 1024) << " MB)\r";
+    std::cout.flush();
+}
+
+struct common_file_metadata {
+    std::string etag;
+    std::string last_modified;
+};
+
+static std::optional<common_file_metadata> read_metadata(const std::string & path) {
+    if (!std::filesystem::exists(path)) {
+        return std::nullopt;
+    }
+
+    nlohmann::json metadata_json;
+    common_file_metadata metadata;
+
+    std::ifstream metadata_in(path);
+    try {
+        metadata_in >> metadata_json;
+        LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, path.c_str(),
+                metadata_json.dump().c_str());
+        if (metadata_json.contains("etag") && metadata_json.at("etag").is_string()) {
+            metadata.etag = metadata_json.at("etag");
+        }
+        if (metadata_json.contains("lastModified") && metadata_json.at("lastModified").is_string()) {
+            metadata.last_modified = metadata_json.at("lastModified");
+        }
+    } catch (const nlohmann::json::exception & e) {
+        LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, path.c_str(), e.what());
+        return std::nullopt;
+    }
+
+    return metadata;
+}
+
+static void write_metadata(const std::string & path,
+                           const std::string & url,
+                           const common_file_metadata & metadata) {
+    nlohmann::json metadata_json = {
+        { "url",          url                    },
+        { "etag",         metadata.etag          },
+        { "lastModified", metadata.last_modified }
+    };
+
+    write_file(path, metadata_json.dump(4));
+    LOG_DBG("%s: file metadata saved: %s\n", __func__, path.c_str());
+}
+
+static bool common_pull_file(httplib::Client & cli,
+                             const std::string & resolve_path,
+                             const std::string & path_tmp,
+                             bool supports_ranges,
+                             size_t existing_size,
+                             size_t & total_size) {
+    std::ofstream ofs(path_tmp, std::ios::binary | std::ios::app);
+    if (!ofs.is_open()) {
+        LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path_tmp.c_str());
+        return false;
+    }
+
+    httplib::Headers headers;
+    if (supports_ranges && existing_size > 0) {
+        headers.emplace("Range", "bytes=" + std::to_string(existing_size) + "-");
+    }
+
+    std::atomic<size_t> downloaded{existing_size};
+
+    auto res = cli.Get(resolve_path, headers,
+        [&](const httplib::Response &response) {
+            if (existing_size > 0 && response.status != 206) {
+                LOG_WRN("%s: server did not respond with 206 Partial Content for a resume request. Status: %d\n", __func__, response.status);
+                return false;
+            }
+            if (existing_size == 0 && response.status != 200) {
+                LOG_WRN("%s: download received non-successful status code: %d\n", __func__, response.status);
+                return false;
+            }
+            if (total_size == 0 && response.has_header("Content-Length")) {
+                try {
+                    size_t content_length = std::stoull(response.get_header_value("Content-Length"));
+                    total_size = existing_size + content_length;
+                } catch (const std::exception &e) {
+                    LOG_WRN("%s: invalid Content-Length header: %s\n", __func__, e.what());
+                }
+            }
+            return true;
+        },
+        [&](const char *data, size_t len) {
+            ofs.write(data, len);
+            if (!ofs) {
+                LOG_ERR("%s: error writing to file: %s\n", __func__, path_tmp.c_str());
+                return false;
+            }
+            downloaded += len;
+            print_progress(downloaded, total_size);
+            return true;
+        },
+        nullptr
+    );
+
+    std::cout << "\n";
+
+    if (!res) {
+        LOG_ERR("%s: error during download. Status: %d\n", __func__, res ? res->status : -1);
+        return false;
+    }
+
+    return 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;
+
+    auto [cli, parts] = http_client(url);
+
+    httplib::Headers default_headers = {{"User-Agent", "llama-cpp"}};
+    if (!bearer_token.empty()) {
+        default_headers.insert({"Authorization", "Bearer " + bearer_token});
+    }
+    cli.set_default_headers(default_headers);
+
+    common_file_metadata last;
+    const bool file_exists = std::filesystem::exists(path);
+    if (file_exists) {
+        if (auto opt = read_metadata(metadata_path)) {
+            last = *opt;
+        }
+    } else {
+        LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
+    }
+
+    for (int i = 0; i < max_attempts; ++i) {
+        auto head = cli.Head(parts.path);
+        bool head_ok = head && head->status >= 200 && head->status < 300;
+        if (!head_ok) {
+            LOG_WRN("%s: HEAD invalid http status code received: %d\n", __func__, head ? head->status : -1);
+            if (file_exists) {
+                LOG_INF("%s: Using cached file (HEAD failed): %s\n", __func__, path.c_str());
+                return true;
+            }
+        }
+
+        common_file_metadata current;
+        if (head_ok) {
+            if (head->has_header("ETag")) {
+                current.etag = head->get_header_value("ETag");
+            }
+            if (head->has_header("Last-Modified")) {
+                current.last_modified = head->get_header_value("Last-Modified");
+            }
+        }
+
+        size_t total_size = 0;
+        if (head_ok && head->has_header("Content-Length")) {
+            try {
+                total_size = std::stoull(head->get_header_value("Content-Length"));
+            } catch (const std::exception& e) {
+                LOG_WRN("%s: Invalid Content-Length in HEAD response: %s\n", __func__, e.what());
+            }
+        }
+
+        bool supports_ranges = false;
+        if (head_ok && head->has_header("Accept-Ranges")) {
+            supports_ranges = head->get_header_value("Accept-Ranges") != "none";
+        }
+
+        bool should_download_from_scratch = false;
+        if (head_ok) {
+            if (!last.etag.empty() && last.etag != current.etag) {
+                LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__,
+                        last.etag.c_str(), current.etag.c_str());
+                should_download_from_scratch = true;
+            } else if (!last.last_modified.empty() && last.last_modified != current.last_modified) {
+                LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__,
+                        last.last_modified.c_str(), current.last_modified.c_str());
+                should_download_from_scratch = true;
+            }
+        }
+
+        if (file_exists) {
+            if (!should_download_from_scratch) {
+                LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
+                return true;
+            }
+            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";
+        size_t existing_size = 0;
+
+        if (std::filesystem::exists(path_temporary)) {
+            if (supports_ranges && !should_download_from_scratch) {
+                existing_size = std::filesystem::file_size(path_temporary);
+            } else if (remove(path_temporary.c_str()) != 0) {
+                LOG_ERR("%s: unable to delete file: %s\n", __func__, path_temporary.c_str());
+                return false;
+            }
+        }
+
+        // start the download
+        LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n",
+                __func__, show_masked_url(parts).c_str(), path_temporary.c_str(),
+                current.etag.c_str(), current.last_modified.c_str());
+        const bool was_pull_successful = common_pull_file(cli, parts.path, path_temporary, supports_ranges, existing_size, total_size);
+        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: download failed after %d attempts\n", __func__, max_attempts);
+            }
+
+            continue;
+        }
+
+        if (std::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;
+        }
+        write_metadata(metadata_path, url, current);
+        break;
+    }
+
+    return true;
+}
+
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string          & url,
+                                                             const common_remote_params & params) {
+    auto [cli, parts] = http_client(url);
+
+    httplib::Headers headers = {{"User-Agent", "llama-cpp"}};
+    for (const auto & header : params.headers) {
+        size_t pos = header.find(':');
+        if (pos != std::string::npos) {
+            headers.emplace(header.substr(0, pos), header.substr(pos + 1));
+        } else {
+            headers.emplace(header, "");
+        }
+    }
+
+    if (params.timeout > 0) {
+        cli.set_read_timeout(params.timeout, 0);
+        cli.set_write_timeout(params.timeout, 0);
+    }
+
+    std::vector<char> buf;
+    auto res = cli.Get(parts.path, headers,
+        [&](const char *data, size_t len) {
+            buf.insert(buf.end(), data, data + len);
+            return params.max_size == 0 ||
+                   buf.size() <= static_cast<size_t>(params.max_size);
+        },
+        nullptr
+    );
+
+    if (!res) {
+        throw std::runtime_error("error: cannot make GET request");
+    }
+
+    return { res->status, std::move(buf) };
 }

 #endif // LLAMA_USE_CURL
@@ -51,6 +51,11 @@
 #include <unistd.h>
 #endif

+#if defined(__linux__)
+#include <sys/types.h>
+#include <pwd.h>
+#endif
+
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
@@ -865,8 +870,20 @@ std::string fs_get_cache_directory() {
 #if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
        if (std::getenv("XDG_CACHE_HOME")) {
            cache_directory = std::getenv("XDG_CACHE_HOME");
-        } else {
+        } else if (std::getenv("HOME")) {
            cache_directory = std::getenv("HOME") + std::string("/.cache/");
+        } else {
+#if defined(__linux__)
+            /* no $HOME is defined, fallback to getpwuid */
+            struct passwd *pw = getpwuid(getuid());
+            if ((!pw) || (!pw->pw_dir)) {
+                throw std::runtime_error("Failed to find $HOME directory");
+            }
+
+            cache_directory = std::string(pw->pw_dir) + std::string("/.cache/");
+#else /* defined(__linux__) */
+            throw std::runtime_error("Failed to find $HOME directory");
+#endif /* defined(__linux__) */
        }
 #elif defined(__APPLE__)
        cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
@@ -961,15 +978,13 @@ struct common_init_result common_init_from_params(common_params & params) {

        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
+        bool has_rerank_prompt = llama_model_chat_template(model, "rerank") != NULL;

-        if (!has_eos && !has_sep) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
+        if (!has_eos && !has_sep && !has_rerank_prompt) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, SEP token, or rerank prompt. Reranking will not work\n", __func__);
            ok = false;
        } else if (!has_eos) {
            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
-        } else if (!has_sep) {
-            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
-            ok = false;
        }

        if (!ok) {
@@ -738,7 +738,7 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 // MoE utils
 //

-const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_exps";
+const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_(ch|)exps";

 static std::string llm_ffn_exps_block_regex(int idx) {
    return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
@@ -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);
    }
 }

@@ -3717,11 +3717,29 @@ class Qwen2MoeModel(TextModel):
 class Qwen3Model(Qwen2Model):
    model_arch = gguf.MODEL_ARCH.QWEN3

+    # extra logic for rerank models
+    is_rerank: bool = False
+    is_tied_embeddings: bool = False
+    token_false_id: int | None = None
+    token_true_id: int | None = None
+
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
+
+        # track for intern-s1-mini
        hparams = ModelBase.load_hparams(self.dir_model, is_mistral_format=False)
        self.origin_hf_arch = hparams.get('architectures', [None])[0]

+        # a bit hacky, but currently the only way to detect if this is a rerank model
+        # ref: https://huggingface.co/Qwen/Qwen3-Reranker-0.6B
+        readme_path = self.dir_model / "README.md"
+        readme_text = ""
+        if readme_path.exists():
+            with readme_path.open("r", encoding="utf-8") as f:
+                readme_text = f.read()
+        if "# Qwen3-Reranker" in readme_text:
+            self._find_rerank_config()
+
    def set_vocab(self):
        # deal with intern-s1-mini
        if self.origin_hf_arch == 'InternS1ForConditionalGeneration':
@@ -3730,6 +3748,53 @@ class Qwen3Model(Qwen2Model):

        super().set_vocab()

+    def _find_rerank_config(self):
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+
+        self.is_rerank = True
+        self.is_tied_embeddings = self.hparams.get("tie_word_embeddings", False)
+        self.token_false_id = tokenizer.convert_tokens_to_ids("no")
+        self.token_true_id = tokenizer.convert_tokens_to_ids("yes")
+        self.sep_token_id = tokenizer.convert_tokens_to_ids("|")
+
+        assert self.token_false_id is not None and self.token_true_id is not None
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if self.is_rerank:
+            self.gguf_writer.add_pooling_type(gguf.PoolingType.RANK)
+            self.gguf_writer.add_classifier_output_labels(["yes", "no"])
+            self.gguf_writer.add_chat_template([{
+                "name": "rerank",
+                "template": "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".<|im_end|>\n"
+                            "<|im_start|>user\n<Instruct>: Given a web search query, retrieve relevant passages that answer the query\n<Query>: {query}\n<Document>: {document}<|im_end|>\n"
+                            "<|im_start|>assistant\n<think>\n\n</think>\n\n"
+            }])
+
+    def _get_cls_out_tensor(self, data_torch: Tensor) -> Tensor:
+        # extract "yes" and "no" tokens from the output lm_head tensor
+        false_row = data_torch[self.token_false_id]
+        true_row = data_torch[self.token_true_id]
+        return torch.stack([true_row, false_row], dim=0)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if self.is_rerank:
+            is_tied_head = self.is_tied_embeddings and "embed_tokens" in name
+            is_real_head = not self.is_tied_embeddings and "lm_head" in name
+            if is_tied_head or is_real_head:
+                cls_out_head = (
+                    gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.CLS_OUT] + ".weight",
+                    self._get_cls_out_tensor(data_torch),
+                )
+                if is_tied_head:
+                    embed = (self.map_tensor_name(name), data_torch)
+                    return [cls_out_head, embed]
+                if is_real_head:
+                    return [cls_out_head]
+
+        return super().modify_tensors(data_torch, name, bid)
+

@ModelBase.register("Qwen3MoeForCausalLM")
 class Qwen3MoeModel(Qwen2MoeModel):
@@ -7930,6 +7995,121 @@ class BailingMoeModel(TextModel):
                raise ValueError(f"Unprocessed experts: {experts}")


+@ModelBase.register("GroveMoeForCausalLM", "modeling_grove_moe.GroveMoeForCausalLM")
+class GroveMoeModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.GROVEMOE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if (n_experts := self.hparams.get("num_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+            logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
+        # FIXME?: Hardcoded https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L299
+        self.gguf_writer.add_expert_chunk_feed_forward_length(self.hparams.get("head_dim") or 128)
+        # FIXME?: Hardcoded https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L298
+        self.gguf_writer.add_experts_per_group(2)
+        # FIXME?: Hardcoded https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L376
+        self.gguf_writer.add_expert_group_scale(0.05)
+        # YaRN is not enabled by default
+        # To enable it, please refer to this guide: https://huggingface.co/Qwen/Qwen3-30B-A3B#processing-long-texts
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+
+    _experts: list[dict[str, Tensor]] | None = None
+    _chunk_experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.endswith(".expert_bias"):
+            # FIXME?: Unused https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L303
+            return []
+
+        # process the experts separately
+        if name.find("chunk_experts") != -1:
+            n_experts = self.hparams["num_experts"] // 2 # see add_experts_per_group
+            assert bid is not None
+
+            if self._chunk_experts is None:
+                self._chunk_experts = [{} for _ in range(self.block_count)]
+
+            self._chunk_experts[bid][name] = data_torch
+
+            if len(self._chunk_experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.chunk_experts.{xid}.{w_name}.weight"
+                        datas.append(self._chunk_experts[bid][ename])
+                        del self._chunk_experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.chunk_experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+        elif name.find("experts") != -1:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._chunk_experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            chunk_experts = [k for d in self._chunk_experts for k in d.keys()]
+            if len(chunk_experts) > 0:
+                raise ValueError(f"Unprocessed adjugate experts: {chunk_experts}")
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
@ModelBase.register("ChameleonForConditionalGeneration")
@ModelBase.register("ChameleonForCausalLM")  # obsolete
 class ChameleonModel(TextModel):
@@ -110,7 +110,7 @@ You may want to pass in some different `ARGS`, depending on the MUSA environment

 The defaults are:

- `MUSA_VERSION` set to `rc4.2.0`
+- `MUSA_VERSION` set to `rc4.3.0`

 The resulting images, are essentially the same as the non-MUSA images:

@@ -95,8 +95,13 @@ int main(int argc, char ** argv) {
        params.n_batch = params.n_ctx;
    }

-    // For non-causal models, batch size must be equal to ubatch size
-    params.n_ubatch = params.n_batch;
+    // for non-causal models, batch size must be equal to ubatch size
+    if (params.attention_type != LLAMA_ATTENTION_TYPE_CAUSAL) {
+        params.n_ubatch = params.n_batch;
+    }
+
+    // get max number of sequences per batch
+    const int n_seq_max = llama_max_parallel_sequences();

    llama_backend_init();
    llama_numa_init(params.numa);
@@ -144,6 +149,7 @@ int main(int argc, char ** argv) {
    // get added sep and eos token, if any
    const std::string added_sep_token = llama_vocab_get_add_sep(vocab) ? llama_vocab_get_text(vocab, llama_vocab_sep(vocab)) : "";
    const std::string added_eos_token = llama_vocab_get_add_eos(vocab) ? llama_vocab_get_text(vocab, llama_vocab_eos(vocab)) : "";
+    const char * rerank_prompt = llama_model_chat_template(model, "rerank");

    // tokenize the prompts and trim
    std::vector<std::vector<int32_t>> inputs;
@@ -153,21 +159,28 @@ int main(int argc, char ** argv) {
        // split classification pairs and insert expected separator tokens
        if (pooling_type == LLAMA_POOLING_TYPE_RANK && prompt.find(params.cls_sep) != std::string::npos) {
            std::vector<std::string> pairs = split_lines(prompt, params.cls_sep);
-            std::string final_prompt;
-
-            for (size_t i = 0; i < pairs.size(); i++) {
-                final_prompt += pairs[i];
-                if (i != pairs.size() - 1) {
-                    if (!added_eos_token.empty()) {
-                        final_prompt += added_eos_token;
-                    }
-                    if (!added_sep_token.empty()) {
-                        final_prompt += added_sep_token;
+            if (rerank_prompt != nullptr) {
+                const std::string query = pairs[0];
+                const std::string doc = pairs[1];
+                std::string final_prompt = rerank_prompt;
+                string_replace_all(final_prompt, "{query}"   , query);
+                string_replace_all(final_prompt, "{document}", doc  );
+                inp = common_tokenize(vocab, final_prompt, true, true);
+            } else {
+                std::string final_prompt;
+                for (size_t i = 0; i < pairs.size(); i++) {
+                    final_prompt += pairs[i];
+                    if (i != pairs.size() - 1) {
+                        if (!added_eos_token.empty()) {
+                            final_prompt += added_eos_token;
+                        }
+                        if (!added_sep_token.empty()) {
+                            final_prompt += added_sep_token;
+                        }
                    }
                }
+                inp = common_tokenize(ctx, final_prompt, true, true);
            }
-
-            inp = common_tokenize(ctx, final_prompt, true, true);
        } else {
            inp = common_tokenize(ctx, prompt, true, true);
        }
@@ -229,7 +242,7 @@ int main(int argc, char ** argv) {
        const uint64_t n_toks = inp.size();

        // encode if at capacity
-        if (batch.n_tokens + n_toks > n_batch) {
+        if (batch.n_tokens + n_toks > n_batch || s >= n_seq_max) {
            float * out = emb + e * n_embd;
            batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
            e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
@@ -5,6 +5,11 @@ target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

 set(TEST_TARGET test-eval-callback)
-add_test(NAME ${TEST_TARGET}
-        COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
+if(NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
+        add_test(NAME ${TEST_TARGET}
+                        COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
+else()
+        add_test(NAME ${TEST_TARGET}
+                        COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K-be.gguf --model stories260K-be.gguf --prompt hello --seed 42 -ngl 0)
+endif()
 set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)
@@ -118,13 +118,17 @@ embedding-convert-model:

 embedding-run-original-model:
 	$(call validate_embedding_model_path,embedding-run-original-model)
-	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/embedding/run-original-model.py
+	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" \
+	./scripts/embedding/run-original-model.py \
+	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")

 embedding-run-converted-model:
-	@CONVERTED_EMBEDDING_MODEL="$(CONVERTED_EMBEDDING_MODEL)" ./scripts/embedding/run-converted-model.sh ${CONVERTED_EMBEDDING_MODEL}
+	@./scripts/embedding/run-converted-model.sh $(CONVERTED_EMBEDDING_MODEL) \
+	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")

 embedding-verify-logits: embedding-run-original-model embedding-run-converted-model
-	@./scripts/embedding/compare-embeddings-logits.sh
+	@./scripts/embedding/compare-embeddings-logits.sh \
+	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")

 embedding-inspect-original-model:
 	$(call validate_embedding_model_path,embedding-inspect-original-model)
@@ -156,7 +160,8 @@ embedding-quantize-model:
 	$(call quantize_model,$(CONVERTED_EMBEDDING_MODEL),QUANTIZED_EMBEDDING_MODEL)

 embedding-run-quantized-model:
-	@./scripts/embedding/run-converted-model.sh ${QUANTIZED_EMBEDDING_MODEL}
+	@./scripts/embedding/run-converted-model.sh $(QUANTIZED_EMBEDDING_MODEL) \
+	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")

 ###
 ### Perplexity targets/recipes
@@ -151,6 +151,35 @@ int main(int argc, char ** argv) {
        logits = llama_get_embeddings(ctx);
        n_logits = llama_model_n_embd(model) * batch.n_tokens;
        type = "-embeddings";
+
+        const int n_embd = llama_model_n_embd(model);
+        const int n_embd_count = batch.n_tokens;
+
+        printf("Embedding dimension: %d\n", n_embd);
+        printf("\n");
+
+        // Print embeddings in the specified format
+        for (int j = 0; j < n_embd_count; j++) {
+            printf("embedding %d: ", j);
+
+            // Print first 3 values
+            for (int i = 0; i < 3 && i < n_embd; i++) {
+                printf("%9.6f ", logits[j * n_embd + i]);
+            }
+
+            printf(" ... ");
+
+            // Print last 3 values
+            for (int i = n_embd - 3; i < n_embd; i++) {
+                if (i >= 0) {
+                    printf("%9.6f ", logits[j * n_embd + i]);
+                }
+            }
+
+            printf("\n");
+        }
+        printf("\n");
+
        printf("Embeddings size: %d\n", n_logits);
    } else {
        logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
@@ -183,22 +212,23 @@ int main(int argc, char ** argv) {
        return 1;
    }
    for (int i = 0; i < n_logits; i++) {
-        fprintf(f, "%d: %.6f\n", i, logits[i]);  // Added index and changed format
+        fprintf(f, "%d: %.6f\n", i, logits[i]);
    }
    fclose(f);

-    // Print first and last 10 logits for quick verification
-    printf("First 10 logits: ");
-    for (int i = 0; i < 10 && i < n_logits; i++) {
-        printf("%.6f ", logits[i]);
-    }
-    printf("\n");
+    if (!embedding_mode) {
+        printf("First 10 logits: ");
+        for (int i = 0; i < 10 && i < n_logits; i++) {
+            printf("%.6f ", logits[i]);
+        }
+        printf("\n");

-    printf("Last 10 logits: ");
-    for (int i = n_logits - 10; i < n_logits; i++) {
-        if (i >= 0) printf("%.6f ", logits[i]);
+        printf("Last 10 logits: ");
+        for (int i = n_logits - 10; i < n_logits; i++) {
+            if (i >= 0) printf("%.6f ", logits[i]);
+        }
+        printf("\n\n");
    }
-    printf("\n\n");

    printf("Logits saved to %s\n", bin_filename);
    printf("Logits saved to %s\n", txt_filename);
@@ -2,8 +2,37 @@

 set -e

-MODEL_PATH="${1:-"$EMBEDDING_MODEL_PATH"}"
-MODEL_NAME="${2:-$(basename "$MODEL_PATH")}"
+# Parse command line arguments
+MODEL_PATH=""
+MODEL_NAME=""
+PROMPTS_FILE=""
+
+# First argument is always model path
+if [ $# -gt 0 ] && [[ "$1" != --* ]]; then
+    MODEL_PATH="$1"
+    shift
+fi
+
+# Parse remaining arguments
+while [[ $# -gt 0 ]]; do
+    case $1 in
+        --prompts-file|-pf)
+            PROMPTS_FILE="$2"
+            shift 2
+            ;;
+        *)
+            # If MODEL_NAME not set and this isn't a flag, use as model name
+            if [ -z "$MODEL_NAME" ] && [[ "$1" != --* ]]; then
+                MODEL_NAME="$1"
+            fi
+            shift
+            ;;
+    esac
+done
+
+# Set defaults
+MODEL_PATH="${MODEL_PATH:-"$EMBEDDING_MODEL_PATH"}"
+MODEL_NAME="${MODEL_NAME:-$(basename "$MODEL_PATH")}"

 if [ -t 0 ]; then
    CPP_EMBEDDINGS="data/llamacpp-${MODEL_NAME}-embeddings.bin"
@@ -35,8 +64,18 @@ with open('$TEMP_FILE', 'wb') as f:
    trap "rm -f $TEMP_FILE" EXIT
 fi

-python scripts/utils/semantic_check.py --model-path $MODEL_PATH \
+# Build the semantic_check.py command
+SEMANTIC_CMD="python scripts/utils/semantic_check.py --model-path $MODEL_PATH \
    --python-embeddings data/pytorch-${MODEL_NAME}-embeddings.bin \
-    --cpp-embeddings $CPP_EMBEDDINGS \
-    --prompt "Hello world today"
+    --cpp-embeddings $CPP_EMBEDDINGS"
+
+# Add prompts file if specified, otherwise use default prompt
+if [ -n "$PROMPTS_FILE" ]; then
+    SEMANTIC_CMD="$SEMANTIC_CMD --prompts-file \"$PROMPTS_FILE\""
+else
+    SEMANTIC_CMD="$SEMANTIC_CMD --prompt \"Hello world today\""
+fi
+
+# Execute the command
+eval $SEMANTIC_CMD

@@ -2,8 +2,27 @@

 set -e

-# First try command line argument, then environment variable, then file
-CONVERTED_MODEL="${1:-"$CONVERTED_EMBEDDING_MODEL"}"
+# Parse command line arguments
+CONVERTED_MODEL=""
+PROMPTS_FILE=""
+
+while [[ $# -gt 0 ]]; do
+    case $1 in
+        -p|--prompts-file)
+            PROMPTS_FILE="$2"
+            shift 2
+            ;;
+        *)
+            if [ -z "$CONVERTED_MODEL" ]; then
+                CONVERTED_MODEL="$1"
+            fi
+            shift
+            ;;
+    esac
+done
+
+# First try command line argument, then environment variable
+CONVERTED_MODEL="${CONVERTED_MODEL:-"$CONVERTED_EMBEDDING_MODEL"}"

 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
@@ -13,8 +32,19 @@ if [ -z "$CONVERTED_MODEL" ]; then
    exit 1
 fi

+# Read prompt from file or use default
+if [ -n "$PROMPTS_FILE" ]; then
+    if [ ! -f "$PROMPTS_FILE" ]; then
+        echo "Error: Prompts file '$PROMPTS_FILE' not found" >&2
+        exit 1
+    fi
+    PROMPT=$(cat "$PROMPTS_FILE")
+else
+    PROMPT="Hello world today"
+fi
+
 echo $CONVERTED_MODEL

 cmake --build ../../build --target llama-logits -j8
-
-../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "Hello world today"
+# TODO: update logits.cpp to accept a --file/-f option for the prompt
+../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "$PROMPT"
@@ -13,14 +13,37 @@ unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')

 parser = argparse.ArgumentParser(description='Process model with specified path')
 parser.add_argument('--model-path', '-m', help='Path to the model')
+parser.add_argument('--prompts-file', '-p', help='Path to file containing prompts (one per line)')
 args = parser.parse_args()

+def read_prompt_from_file(file_path):
+    try:
+        with open(file_path, 'r', encoding='utf-8') as f:
+            return f.read().strip()
+    except FileNotFoundError:
+        print(f"Error: Prompts file '{file_path}' not found")
+        exit(1)
+    except Exception as e:
+        print(f"Error reading prompts file: {e}")
+        exit(1)
+
 model_path = os.environ.get('EMBEDDING_MODEL_PATH', args.model_path)
 if model_path is None:
    parser.error("Model path must be specified either via --model-path argument or EMBEDDING_MODEL_PATH environment variable")

 tokenizer = AutoTokenizer.from_pretrained(model_path)

+config = AutoConfig.from_pretrained(model_path)
+
+# This can be used to override the sliding window size for manual testing. This
+# can be useful to verify the sliding window attention mask in the original model
+# and compare it with the converted .gguf model.
+if hasattr(config, 'sliding_window'):
+    original_sliding_window = config.sliding_window
+    #original_sliding_window = 6
+    print(f"Modified sliding window: {original_sliding_window} -> {config.sliding_window}")
+
+print(f"Using unreleased model: {unreleased_model_name}")
 if unreleased_model_name:
    model_name_lower = unreleased_model_name.lower()
    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
@@ -29,19 +52,28 @@ if unreleased_model_name:

    try:
        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
-        model = model_class.from_pretrained(model_path)  # Note: from_pretrained, not fromPretrained
+        model = model_class.from_pretrained(model_path, config=config)
    except (ImportError, AttributeError) as e:
        print(f"Failed to import or load model: {e}")
        exit(1)
 else:
-    model = AutoModel.from_pretrained(model_path)
+    model = AutoModel.from_pretrained(model_path, config=config)
 print(f"Model class: {type(model)}")
-#print(f"Model file: {type(model).__module__}")
-config = AutoConfig.from_pretrained(model_path)
+print(f"Model file: {type(model).__module__}")
+
+# Verify the model is using the correct sliding window
+if hasattr(model.config, 'sliding_window'):
+    print(f"Model's sliding_window: {model.config.sliding_window}")
+else:
+    print("Model config does not have sliding_window attribute")

 model_name = os.path.basename(model_path)

-texts = [ "Hello world today" ]
+if args.prompts_file:
+    prompt_text = read_prompt_from_file(args.prompts_file)
+    texts = [prompt_text]
+else:
+    texts = ["Hello world today"]

 encoded = tokenizer(
    texts,
@@ -40,7 +40,7 @@ if os.path.exists(index_path):
        file_path = os.path.join(model_path, file_name)
        print(f"\n--- From {file_name} ---")

-        with safe_open(file_path, framework="pt") as f:  # type: ignore
+        with safe_open(file_path, framework="pt") as f:
            for tensor_name in sorted(tensor_names):
                tensor = f.get_tensor(tensor_name)
                print(f"- {tensor_name} : shape = {tensor.shape}, dtype = {tensor.dtype}")
@@ -49,7 +49,7 @@ elif os.path.exists(single_file_path):
    # Single file model (original behavior)
    print("Single-file model detected")

-    with safe_open(single_file_path, framework="pt") as f:  # type: ignore
+    with safe_open(single_file_path, framework="pt") as f:
        keys = f.keys()
        print("Tensors in model:")
        for key in sorted(keys):
@@ -101,6 +101,17 @@ def test_single_prompt_similarity(python_emb, cpp_emb, tokens, prompt):
        'rms_diff': np.sqrt(np.mean(diff_matrix**2))
    }

+def read_prompt_from_file(file_path):
+    try:
+        with open(file_path, 'r', encoding='utf-8') as f:
+            return f.read().strip()
+    except FileNotFoundError:
+        print(f"Error: Prompts file '{file_path}' not found")
+        exit(1)
+    except Exception as e:
+        print(f"Error reading prompts file: {e}")
+        exit(1)
+
 def main():
    parser = argparse.ArgumentParser(description='Test semantic similarity between Python and llama.cpp embeddings')
    parser.add_argument('--model-path', '-m', required=True, help='Path to the original Python model')
@@ -108,14 +119,20 @@ def main():
    parser.add_argument('--cpp-embeddings', '-ce', help='Path to llama.cpp embeddings "logits" binary file')
    parser.add_argument('--causal', '-c', default=False, help='if the model is causal (default: false)', action='store_true')
    parser.add_argument('--prompt', '-p', default='Hello world today', help='Test prompt')
+    parser.add_argument('--prompts-file', '-pf', help='Path to file containing prompts')

    args = parser.parse_args()

+    if args.prompts_file:
+        prompt = read_prompt_from_file(args.prompts_file)
+    else:
+        prompt = args.prompt
+
    print("Semantic Similarity Test Between Python and llama.cpp Embedding Models")
    print("=" * 70)

    # Single prompt detailed comparison
-    print(f"\nTesting with prompt: '{args.prompt}'")
+    print(f"\nTesting with prompt: '{prompt}'")

    # Load the python model to get configuration information and also to load the tokenizer.
    print("Loading model and tokenizer using AutoTokenizer:", args.model_path)
@@ -144,7 +161,7 @@ def main():
        else:
            model = AutoModel.from_pretrained(args.model_path)

-    encoded = tokenizer(args.prompt, return_tensors="pt")
+    encoded = tokenizer(prompt, return_tensors="pt")
    tokens = tokenizer.convert_ids_to_tokens(encoded['input_ids'][0])
    n_tokens = len(tokens)
    print(f"n_tokens: {n_tokens}");
@@ -155,7 +172,7 @@ def main():
    python_embeddings = load_embeddings_from_file(args.python_embeddings, n_tokens, model.config.hidden_size)

    # Run comparison
-    results = test_single_prompt_similarity(python_embeddings, llamacpp_embeddings, tokens, args.prompt)
+    results = test_single_prompt_similarity(python_embeddings, llamacpp_embeddings, tokens, prompt)

    # Summary
    print(f"\n=== SUMMARY ===")
@@ -177,7 +177,7 @@ set(GGML_CPU_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC")


 if (MINGW)
-    set(GGML_WIN_VER "0x602" CACHE STRING   "ggml: Windows version")
+    set(GGML_WIN_VER "0xA00" CACHE STRING   "ggml: Windows version")
 endif()

 # ggml core
@@ -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);
@@ -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
@@ -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);
@@ -160,7 +160,6 @@
 #define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
 #define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
-#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -105,6 +105,18 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128

    return ((v4f32)res)[0];
 }
+
+// multiply int8_t, add results pairwise twice
+static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
+    // Get absolute values of x vectors
+    const __m128i ax = __lsx_vsigncov_b(x, x);
+    // Sign the values of the y vectors
+    const __m128i sy = __lsx_vsigncov_b(x, y);
+    // Perform multiplication and create 16-bit values
+    const __m128i dot = lsx_maddubs_h(ax, sy);
+    const __m128i ones = __lsx_vreplgr2vr_h(1);
+    return lsx_madd_h(ones, dot);
+}
 #endif

 #if defined(__loongarch_asx)
@@ -323,18 +335,6 @@ static inline __m256i lasx_xvandi_b_bit(__m256i a, const unsigned int b) {
    }
 }

-// multiply int8_t, add results pairwise twice
-static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
-    // Get absolute values of x vectors
-    const __m128i ax = __lsx_vsigncov_b(x, x);
-    // Sign the values of the y vectors
-    const __m128i sy = __lsx_vsigncov_b(x, y);
-    // Perform multiplication and create 16-bit values
-    const __m128i dot = lsx_maddubs_h(ax, sy);
-    const __m128i ones = __lsx_vreplgr2vr_h(1);
-    return lsx_madd_h(ones, dot);
-}
-
 // horizontally add 8 floats
 static inline float hsum_float_8(const __m256 x) {
    __m128 res = lasx_extractf128(x, 1);
@@ -75,7 +75,8 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i

        for (int j = 0; j < 8; j++) {
            const float32x4_t v = vec_mul(srcv[j], vec_splats(id));
-            const int32x4_t vi = vec_signed(v);
+            /* Uses non-default rounding for vec_signed or vec_round */
+            const int32x4_t vi = vec_signed(__builtin_s390_vfisb(v, 4, 1));

            y[i].qs[4*j + 0] = vec_extract(vi, 0);
            y[i].qs[4*j + 1] = vec_extract(vi, 1);
@@ -122,7 +123,8 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i

        for (int j = 0; j < 8; j++) {
            const float32x4_t v = vec_mul(srcv[j], vec_splats(id));
-            const int32x4_t vi = vec_signed(v);
+            /* Uses non-default rounding for vec_signed or vec_round */
+            const int32x4_t vi = vec_signed(__builtin_s390_vfisb(v, 4, 1));

            y[i].qs[4*j + 0] = vec_extract(vi, 0);
            y[i].qs[4*j + 1] = vec_extract(vi, 1);
@@ -260,6 +262,101 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 }

+void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_MXFP4 == 0);
+    static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
+
+    const int qk = QK_MXFP4;
+    const int nb = n / qk;
+
+    const block_mxfp4 * GGML_RESTRICT x = vx;
+    const block_q8_0  * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0.0f;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    const int8x16_t  v_k = vec_xl(0, kvalues_mxfp4);
+    const uint8x16_t v_m = vec_splats((const uint8_t)0x0F);
+
+    float32x4_t v_acc = vec_splats(0.0f);
+
+    #pragma GCC unroll 8
+    for (; ib + 1 < nb; ib += 2) {
+        const block_mxfp4 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_mxfp4 * GGML_RESTRICT x1 = &x[ib + 1];
+        const block_q8_0  * GGML_RESTRICT y0 = &y[ib + 0];
+        const block_q8_0  * GGML_RESTRICT y1 = &y[ib + 1];
+
+        const uint8x16_t v_x0 = vec_xl(0, x0->qs);
+        const uint8x16_t v_x1 = vec_xl(0, x1->qs);
+
+        int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m);
+        int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4);
+        int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m);
+        int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4);
+
+        v_x0l = vec_perm(v_k, v_k, (uchar8x16_t)v_x0l);
+        v_x0h = vec_perm(v_k, v_k, (uchar8x16_t)v_x0h);
+        v_x1l = vec_perm(v_k, v_k, (uchar8x16_t)v_x1l);
+        v_x1h = vec_perm(v_k, v_k, (uchar8x16_t)v_x1h);
+
+        const int8x16_t v_y0l = vec_xl(0,       y0->qs);
+        const int8x16_t v_y0h = vec_xl(QK8_0/2, y0->qs);
+        const int8x16_t v_y1l = vec_xl(0,       y1->qs);
+        const int8x16_t v_y1h = vec_xl(QK8_0/2, y1->qs);
+
+        const int32x4_t v_xy0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0l, v_y0l), v_x0h, v_y0h);
+        const int32x4_t v_xy1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1l, v_y1l), v_x1h, v_y1h);
+
+        const float32x4_t v_xy0f = vec_float(v_xy0);
+        const float32x4_t v_xy1f = vec_float(v_xy1);
+
+        const float32x4_t v_d0 = vec_splats(GGML_E8M0_TO_FP32_HALF(x0->e) * GGML_CPU_FP16_TO_FP32(y0->d));
+        const float32x4_t v_d1 = vec_splats(GGML_E8M0_TO_FP32_HALF(x1->e) * GGML_CPU_FP16_TO_FP32(y1->d));
+
+        v_acc = vec_madd(v_xy0f, v_d0, v_acc);
+        v_acc = vec_madd(v_xy1f, v_d1, v_acc);
+    }
+
+    for (; ib < nb; ++ib) {
+        const block_mxfp4 * GGML_RESTRICT x0 = &x[ib + 0];
+        const block_q8_0  * GGML_RESTRICT y0 = &y[ib + 0];
+
+        const uint8x16_t v_x = vec_xl(0, x0->qs);
+
+        int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m);
+        int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4);
+
+        v_xl = vec_perm(v_k, v_k, (uchar8x16_t)v_xl);
+        v_xh = vec_perm(v_k, v_k, (uchar8x16_t)v_xh);
+
+        const int8x16_t v_yl = vec_xl(0,       y0->qs);
+        const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs);
+
+        const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
+        const float32x4_t v_xyf = vec_float(v_xy);
+
+        const float32x4_t v_d = vec_splats(GGML_E8M0_TO_FP32_HALF(x0->e) * GGML_CPU_FP16_TO_FP32(y0->d));
+        v_acc = vec_madd(v_xyf, v_d, v_acc);
+    }
+
+    sumf = vec_hsum_f32x4(v_acc);
+    *s = sumf;
+#else
+    UNUSED(x);
+    UNUSED(y);
+    UNUSED(ib);
+    UNUSED(sumf);
+    ggml_vec_dot_mxfp4_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
    const int qk = QK8_0;
    const int nb = n / qk;
@@ -636,7 +733,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
    uint8x16_t q3h[4];
    uint8x16_t q3b[2];
    int8x16_t q3bytes[4];
-    int8x16_t q8bytes[4];
+    int8x16_t q8bytes[8];
    uint8x16_t qhbits[2];

    float sum = 0;
@@ -998,9 +998,9 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 #define GGML_F32_EPR  4

 #define GGML_F32x4         __m128
-#define GGML_F32x4_ZERO    __lsx_vldi(0)
-#define GGML_F32x4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
-#define GGML_F32x4_LOAD(x) __lsx_vld((x), 0)
+#define GGML_F32x4_ZERO    (__m128)__lsx_vldi(0)
+#define GGML_F32x4_SET1(x) (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
+#define GGML_F32x4_LOAD(x) (__m128)__lsx_vld((x), 0)
 #define GGML_F32x4_STORE(x, y)   __lsx_vst(y, x, 0)
 #define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
 #define GGML_F32x4_ADD     __lsx_vfadd_s
@@ -1022,7 +1022,7 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
    __m128i tmp     = __lsx_vsrli_d((__m128i) x[0], 32);                              \
    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, x[0]);                    \
    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
-    const __m128 t0 = __lsx_vshuf4i_w(tmp, 0x88);                                     \
+    const __m128 t0 = (__m128)__lsx_vshuf4i_w(tmp, 0x88);                                     \
    tmp             = __lsx_vsrli_d((__m128i) t0, 32);                                \
    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, t0);                      \
    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
@@ -1052,7 +1052,7 @@ static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) {
    tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]);
    tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]);

-    return __lsx_vld(tmp, 0);
+    return (__m128)__lsx_vld(tmp, 0);
 }

 static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
@@ -1067,9 +1067,9 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 }

 #define GGML_F32Cx4             __m128
-#define GGML_F32Cx4_ZERO        __lsx_vldi(0)
-#define GGML_F32Cx4_SET1(x)     __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
-#define GGML_F32Cx4_LOAD(x)     __lsx_f16x4_load(x)
+#define GGML_F32Cx4_ZERO        (__m128)__lsx_vldi(0)
+#define GGML_F32Cx4_SET1(x)     (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
+#define GGML_F32Cx4_LOAD(x)     (__m128)__lsx_f16x4_load(x)
 #define GGML_F32Cx4_STORE(x, y) __lsx_f16x4_store(x, y)
 #define GGML_F32Cx4_FMA         GGML_F32x4_FMA
 #define GGML_F32Cx4_ADD         __lsx_vfadd_s
@@ -54,7 +54,7 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
    const uint32_t i2  = fastdiv((blockDim.z * blockIdx.z + threadIdx.z), ne3);
    const uint32_t i3  = (blockDim.z * blockIdx.z + threadIdx.z) - (i2 * ne3.z);

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

@@ -586,17 +586,42 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v,
 #endif // defined(GGML_USE_HIP) && (defined(RDNA2)  || defined(RDNA3) || defined(RDNA4) || defined(GCN5) || defined(CDNA))
 }

+static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v, const half2 u) {
+#ifdef FAST_FP16_AVAILABLE
+    acc += v*u;
+#else
+    const float2 tmpv = __half22float2(v);
+    const float2 tmpu = __half22float2(u);
+    float2 tmpacc = __half22float2(acc);
+    tmpacc.x += tmpv.x * tmpu.x;
+    tmpacc.y += tmpv.y * tmpu.y;
+    acc = make_half2(tmpacc.x, tmpacc.y);
+#endif // FAST_FP16_AVAILABLE
+}
+
 // Aligned memory transfers of 8/16 bytes can be faster than 2 transfers with 4 bytes, especially on AMD.
-template <int nbytes>
+template <int nbytes, int alignment = 0>
 static __device__ __forceinline__ void ggml_cuda_memcpy_1(void * __restrict__ dst, const void * __restrict__ src) {
-    if constexpr (nbytes == 4) {
-        *(int *) dst = *(const int *) src;
-    } else if constexpr (nbytes == 8) {
-        *(int2 *) dst = *(const int2 *) src;
-    } else if constexpr (nbytes == 16) {
-        *(int4 *) dst = *(const int4 *) src;
-    } else {
-        static_assert(nbytes == 0 && nbytes == -1, "bad nbytes");
+    if constexpr (alignment != 0) {
+        static_assert(nbytes % alignment == 0, "bad alignment");
+    }
+    constexpr int nb_per_cpy = alignment == 0 ? nbytes : alignment;
+
+#pragma unroll
+    for (int i = 0; i < nbytes/nb_per_cpy; ++i) {
+        if constexpr (nb_per_cpy == 1) {
+            ((char *) dst)[i] = ((const char *) src)[i];
+        } else if constexpr (nb_per_cpy == 2) {
+            ((short *) dst)[i] = ((const short *) src)[i];
+        } else if constexpr (nb_per_cpy == 4) {
+            ((int *) dst)[i] = ((const int *) src)[i];
+        } else if constexpr (nb_per_cpy == 8) {
+            ((int2 *) dst)[i] = ((const int2 *) src)[i];
+        } else if constexpr (nb_per_cpy == 16) {
+            ((int4 *) dst)[i] = ((const int4 *) src)[i];
+        } else {
+            static_assert(nbytes == 0 && nbytes == -1, "bad nbytes");
+        }
    }
 }

@@ -33,276 +33,230 @@ typedef void (* fattn_kernel_t)(
                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
                            const int32_t nb31, const int32_t nb32, const int64_t nb33);

-typedef half (*vec_dot_KQ_f16_t)(
-    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds);
-typedef float (*vec_dot_KQ_f32_t)(
+typedef float (*vec_dot_KQ_t)(
    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds);

-template<typename T, int D, int warp_size>
-static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0(
-    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
-
-    const block_q4_0 * K_q4_0 = (const block_q4_0 *) K_c;
-    GGML_UNUSED(Q_v);
-
-    T sum = 0.0f;
-
-#pragma unroll
-    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
-        const int k_KQ = k_KQ_0 + threadIdx.x;
-
-        const int ib    = k_KQ /  QI8_1;
-        const int iqs4  = k_KQ %  QI4_0;
-        const int shift = k_KQ & (QI8_1/2);
-
-        const int v = (get_int_b2(K_q4_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F;
-        const int u = Q_q8[k_KQ_0/warp_size];
-
-        const int sumi = ggml_cuda_dp4a(v, u, 0);
-
-#ifdef FP16_AVAILABLE
-        if (std::is_same<T, half>::value) {
-            const half2  * Q_ds = (const half2  *) Q_ds_v;
-
-            const half2 sum2 = __half2half2(K_q4_0[ib].d) * Q_ds[k_KQ_0/warp_size];
-            sum += (T) (((half) sumi)*__low2half(sum2) - __high2half(sum2) /* *8/QI8_1 == 1 */);
-        } else
-#endif // FP16_AVAILABLE
-        {
-            const float2 * Q_ds = (const float2 *) Q_ds_v;
-
-            sum += (T) (__half2float(K_q4_0[ib].d) * (sumi*Q_ds[k_KQ_0/warp_size].x - (8/QI8_1)*Q_ds[k_KQ_0/warp_size].y));
-        }
-    }
-
-    return sum;
-}
-
-template<typename T, int D, int warp_size>
-static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1(
-    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
-
-    const block_q4_1 * K_q4_1 = (const block_q4_1 *) K_c;
-    GGML_UNUSED(Q_v);
-
-    T sum = 0.0f;
-
-#pragma unroll
-    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
-        const int k_KQ = k_KQ_0 + threadIdx.x;
-
-        const int ib    = k_KQ /  QI8_1;
-        const int iqs4  = k_KQ %  QI4_1;
-        const int shift = k_KQ & (QI8_1/2);
-
-        const int v = (get_int_b4(K_q4_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F;
-        const int u = Q_q8[k_KQ_0/warp_size];
-
-        const int sumi = ggml_cuda_dp4a(v, u, 0);
-
-#ifdef FP16_AVAILABLE
-        if (std::is_same<T, half>::value) {
-            const half2  * Q_ds = (const half2  *) Q_ds_v;
-
-            const half2 d4d8_m4s8 = K_q4_1[ib].dm * Q_ds[k_KQ_0/warp_size];
-            const half2 sumid4d8_m4s8scaled = d4d8_m4s8 * make_half2(sumi, 1.0f/QI8_1);
-            sum += (T) (__low2half(sumid4d8_m4s8scaled) + __high2half(sumid4d8_m4s8scaled));
-        } else
-#endif // FP16_AVAILABLE
-        {
-            const float2 * Q_ds = (const float2 *) Q_ds_v;
-
-            const float sumid4d8   =  __low2float(K_q4_1[ib].dm)*Q_ds[k_KQ_0/warp_size].x * sumi;
-            const float m4s8scaled = __high2float(K_q4_1[ib].dm)*Q_ds[k_KQ_0/warp_size].y / QI8_1;
-
-            sum += (T) (sumid4d8 + m4s8scaled);
-        }
-    }
-
-    return sum;
-}
-
-template<typename T, int D, int warp_size>
-static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0(
-    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
-
-    const block_q5_0 * K_q5_0 = (const block_q5_0 *) K_c;
-    GGML_UNUSED(Q_v);
-
-    T sum = 0.0f;
-
-#pragma unroll
-    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
-        const int k_KQ = k_KQ_0 + threadIdx.x;
-
-        const int ib    = k_KQ /  QI8_1;
-        const int iqs4  = k_KQ %  QI5_0;
-        const int iqs8  = k_KQ %  QI8_1;
-        const int shift = k_KQ & (QI8_1/2);
-
-        int v = (get_int_b2(K_q5_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F;
-        const int vh = get_int_b2(K_q5_0[ib].qh, 0) >> (iqs8 * QI5_0);
-        v |= (vh <<  4) & 0x00000010; // 0 ->  4
-        v |= (vh << 11) & 0x00001000; // 1 -> 12
-        v |= (vh << 18) & 0x00100000; // 2 -> 20
-        v |= (vh << 25) & 0x10000000; // 3 -> 28
-
-        const int u = Q_q8[k_KQ_0/warp_size];
-
-        const int sumi = ggml_cuda_dp4a(v, u, 0);
-
-#ifdef FP16_AVAILABLE
-        if (std::is_same<T, half>::value) {
-            const half2  * Q_ds = (const half2  *) Q_ds_v;
-
-            const half2 sum2 = __half2half2(K_q5_0[ib].d) * Q_ds[k_KQ_0/warp_size];
-            sum += (T) (((half) sumi)*__low2half(sum2) - __high2half(sum2)*__float2half(2.0f)) /* *16/QI8_1 == 2 */;
-        } else
-#endif // FP16_AVAILABLE
-        {
-            const float2 * Q_ds = (const float2 *) Q_ds_v;
-
-            sum += (T) (__half2float(K_q5_0[ib].d) * (sumi*Q_ds[k_KQ_0/warp_size].x - (16/QI8_1)*Q_ds[k_KQ_0/warp_size].y));
-        }
-    }
-
-    return sum;
-}
-
-template<typename T, int D, int warp_size>
-static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1(
-    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
-
-    const block_q5_1 * K_q5_1 = (const block_q5_1 *) K_c;
-    GGML_UNUSED(Q_v);
-
-    T sum = 0.0f;
-
-#pragma unroll
-    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
-        const int k_KQ = k_KQ_0 + threadIdx.x;
-
-        const int ib    = k_KQ /  QI8_1;
-        const int iqs4  = k_KQ %  QI5_1;
-        const int iqs8  = k_KQ %  QI8_1;
-        const int shift = k_KQ & (QI8_1/2);
-
-        int v = (get_int_b2(K_q5_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F;
-        const int vh = get_int_b2(K_q5_1[ib].qh, 0) >> (iqs8 * QI5_1);
-        v |= (vh <<  4) & 0x00000010; // 0 ->  4
-        v |= (vh << 11) & 0x00001000; // 1 -> 12
-        v |= (vh << 18) & 0x00100000; // 2 -> 20
-        v |= (vh << 25) & 0x10000000; // 3 -> 28
-
-        const int u = Q_q8[k_KQ_0/warp_size];
-
-        const int sumi = ggml_cuda_dp4a(v, u, 0);
-
-#ifdef FP16_AVAILABLE
-        if (std::is_same<T, half>::value) {
-            const half2  * Q_ds = (const half2  *) Q_ds_v;
-
-            const half2 d5d8_m5s8 = K_q5_1[ib].dm * Q_ds[k_KQ_0/warp_size];
-            const half2 sumid5d8_m5s8scaled = d5d8_m5s8 * make_half2(sumi, 1.0f/QI8_1);
-            sum += (T) (__low2half(sumid5d8_m5s8scaled) + __high2half(sumid5d8_m5s8scaled));
-        } else
-#endif // FP16_AVAILABLE
-        {
-            const float2 * Q_ds = (const float2 *) Q_ds_v;
-
-            const float sumid5d8   =  __low2float(K_q5_1[ib].dm)*Q_ds[k_KQ_0/warp_size].x * sumi;
-            const float m5s8scaled = __high2float(K_q5_1[ib].dm)*Q_ds[k_KQ_0/warp_size].y / QI8_1;
-
-            sum += (T) (sumid5d8 + m5s8scaled);
-        }
-    }
-
-    return sum;
-}
-
-template <typename T, int D, int warp_size>
-static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0(
-    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
-
-    const block_q8_0 * K_q8_0 = (const block_q8_0 *) K_c;
-    GGML_UNUSED(Q_v);
-
-    T sum = 0.0f;
-
-#pragma unroll
-    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) {
-        const int k_KQ = k_KQ_0 + threadIdx.x;
-
-        const int ib  = k_KQ / QI8_0;
-        const int iqs = k_KQ % QI8_0;
-
-        const int v = get_int_b2(K_q8_0[ib].qs, iqs);
-
-        T Q_d;
-        if (std::is_same<T, half>::value) {
-            const half2  * Q_ds = (const half2  *) Q_ds_v;
-            Q_d = __low2half(Q_ds[k_KQ_0/warp_size]);
-        } else {
-            const float2 * Q_ds = (const float2 *) Q_ds_v;
-            Q_d = Q_ds[k_KQ_0/warp_size].x;
-        }
-
-        sum += vec_dot_q8_0_q8_1_impl<T, 1>(&v, &Q_q8[k_KQ_0/warp_size], K_q8_0[ib].d, Q_d);
-    }
-
-    return sum;
-}
-
-template <typename T, int D, int warp_size>
-static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_f16(
+template <int D, int nthreads>
+static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_f16(
    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds_v) {

    const half2 * K_h2 = (const half2 *) K_c;
    GGML_UNUSED(Q_q8);
    GGML_UNUSED(Q_ds_v);

-#ifdef FP16_AVAILABLE
-    if (std::is_same<T, half>::value) {
-        const half2 * Q_h2 = (const half2 *) Q_v;
-
-        half2 sum2 = make_half2(0.0f, 0.0f);
-
-#pragma unroll
-        for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += warp_size) {
-            const int k_KQ = k_KQ_0 + threadIdx.x;
-
-            const half2 K_ik = K_h2[k_KQ];
-            sum2 += K_ik * Q_h2[k_KQ_0/warp_size];
-        }
-
-        return __low2half(sum2) + __high2half(sum2);
-    }
-#endif // FP16_AVAILABLE
-
-    const float2 * Q_f2 = (const float2 *) Q_v;
+    constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
+    constexpr int cpy_ne = cpy_nb / 4;

    float sum = 0.0f;

 #pragma unroll
-    for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += warp_size) {
-        const int k_KQ = k_KQ_0 + threadIdx.x;
-
-        const half2 K_ik = K_h2[k_KQ];
-        sum +=  __low2float(K_ik) * Q_f2[k_KQ_0/warp_size].x;
-        sum += __high2float(K_ik) * Q_f2[k_KQ_0/warp_size].y;
+    for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += nthreads*cpy_ne) {
+        half2 tmp[cpy_ne];
+        ggml_cuda_memcpy_1<sizeof(tmp)>(tmp, K_h2 + k_KQ_0 + (threadIdx.x % nthreads)*cpy_ne);
+#pragma unroll
+        for (int k_KQ_1 = 0; k_KQ_1 < cpy_ne; ++k_KQ_1) {
+#ifdef FAST_FP16_AVAILABLE
+            ggml_cuda_mad(sum,                tmp[k_KQ_1] , ((const half2  *) Q_v)[k_KQ_0/nthreads + k_KQ_1]);
+#else
+            ggml_cuda_mad(sum, __half22float2(tmp[k_KQ_1]), ((const float2 *) Q_v)[k_KQ_0/nthreads + k_KQ_1]);
+#endif // FP16_AVAILABLE
+        }
    }

    return sum;
 }

-template <typename Tds>
+template<int D, int nthreads>
+static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_q4_0(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+
+    const block_q4_0 * K_q4_0 = (const block_q4_0 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    float sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += nthreads) {
+        const int k_KQ = k_KQ_0 + (nthreads == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads);
+
+        const int ib    = k_KQ /  QI8_1;
+        const int iqs4  = k_KQ %  QI4_0;
+        const int shift = k_KQ & (QI8_1/2);
+
+        int v;
+        ggml_cuda_memcpy_1<sizeof(int), 2>(&v, K_q4_0[ib].qs + sizeof(int)*iqs4);
+        v = (v >> shift) & 0x0F0F0F0F;
+        const int u = Q_q8[k_KQ_0/nthreads];
+
+        const int sumi = ggml_cuda_dp4a(v, u, 0);
+
+        const float2 Q_ds = ((const float2 *) Q_ds_v)[k_KQ_0/nthreads];
+        sum += __half2float(K_q4_0[ib].d) * (sumi*Q_ds.x - (8/QI8_1)*Q_ds.y);
+    }
+
+    return sum;
+}
+
+template<int D, int nthreads>
+static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_q4_1(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+
+    const block_q4_1 * K_q4_1 = (const block_q4_1 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    float sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += nthreads) {
+        const int k_KQ = k_KQ_0 + (nthreads == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads);
+
+        const int ib    = k_KQ /  QI8_1;
+        const int iqs4  = k_KQ %  QI4_1;
+        const int shift = k_KQ & (QI8_1/2);
+
+        int v;
+        ggml_cuda_memcpy_1<sizeof(int)>(&v, K_q4_1[ib].qs + sizeof(int)*iqs4);
+        v = (v >> shift) & 0x0F0F0F0F;
+        const int u = Q_q8[k_KQ_0/nthreads];
+
+        const int sumi = ggml_cuda_dp4a(v, u, 0);
+
+        const float2 K_dm = __half22float2(K_q4_1[ib].dm);
+        const float2 Q_ds = ((const float2 *) Q_ds_v)[k_KQ_0/nthreads];
+
+        sum += K_dm.x*Q_ds.x*sumi + K_dm.y*Q_ds.y/QI8_1;
+    }
+
+    return sum;
+}
+
+template<int D, int nthreads>
+static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_q5_0(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+
+    const block_q5_0 * K_q5_0 = (const block_q5_0 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    float sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += nthreads) {
+        const int k_KQ = k_KQ_0 + (nthreads == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads);
+
+        const int ib    = k_KQ /  QI8_1;
+        const int iqs4  = k_KQ %  QI5_0;
+        const int iqs8  = k_KQ %  QI8_1;
+        const int shift = k_KQ & (QI8_1/2);
+
+        int v;
+        ggml_cuda_memcpy_1<sizeof(int), 2>(&v, K_q5_0[ib].qs + sizeof(int)*iqs4);
+        v = (v >> shift) & 0x0F0F0F0F;
+
+        {
+            int vh;
+            ggml_cuda_memcpy_1<sizeof(int), 2>(&vh, K_q5_0[ib].qh);
+            vh >>= iqs8 * QI5_0;
+
+            v |= (vh <<  4) & 0x00000010; // 0 ->  4
+            v |= (vh << 11) & 0x00001000; // 1 -> 12
+            v |= (vh << 18) & 0x00100000; // 2 -> 20
+            v |= (vh << 25) & 0x10000000; // 3 -> 28
+        }
+
+        const int u = Q_q8[k_KQ_0/nthreads];
+
+        const int sumi = ggml_cuda_dp4a(v, u, 0);
+
+        const float2 Q_ds = ((const float2 *) Q_ds_v)[k_KQ_0/nthreads];
+
+        sum += __half2float(K_q5_0[ib].d) * (sumi*Q_ds.x - (16/QI8_1)*Q_ds.y);
+    }
+
+    return sum;
+}
+
+template<int D, int nthreads>
+static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_q5_1(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+
+    const block_q5_1 * K_q5_1 = (const block_q5_1 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    float sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += nthreads) {
+        const int k_KQ = k_KQ_0 + (nthreads == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads);
+
+        const int ib    = k_KQ /  QI8_1;
+        const int iqs4  = k_KQ %  QI5_1;
+        const int iqs8  = k_KQ %  QI8_1;
+        const int shift = k_KQ & (QI8_1/2);
+
+        int v;
+        ggml_cuda_memcpy_1<sizeof(int)>(&v, K_q5_1[ib].qs + sizeof(int)*iqs4);
+        v = (v >> shift) & 0x0F0F0F0F;
+
+        {
+            int vh;
+            ggml_cuda_memcpy_1<sizeof(int)>(&vh, K_q5_1[ib].qh);
+            vh >>= iqs8 * QI5_0;
+
+            v |= (vh <<  4) & 0x00000010; // 0 ->  4
+            v |= (vh << 11) & 0x00001000; // 1 -> 12
+            v |= (vh << 18) & 0x00100000; // 2 -> 20
+            v |= (vh << 25) & 0x10000000; // 3 -> 28
+        }
+
+        const int u = Q_q8[k_KQ_0/nthreads];
+
+        const int sumi = ggml_cuda_dp4a(v, u, 0);
+
+        const float2 K_dm = __half22float2(K_q5_1[ib].dm);
+        const float2 Q_ds = ((const float2 *) Q_ds_v)[k_KQ_0/nthreads];
+
+        sum += K_dm.x*Q_ds.x*sumi + K_dm.y*Q_ds.y/QI8_1;
+    }
+
+    return sum;
+}
+
+template <int D, int nthreads>
+static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_q8_0(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+
+    const block_q8_0 * K_q8_0 = (const block_q8_0 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    float sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += nthreads) {
+        const int k_KQ = k_KQ_0 + (nthreads == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads);
+
+        const int ib  = k_KQ / QI8_0;
+        const int iqs = k_KQ % QI8_0;
+
+        int v;
+        ggml_cuda_memcpy_1<sizeof(v), 2>(&v, K_q8_0[ib].qs + 4*iqs);
+
+        const float2 * Q_ds = (const float2 *) Q_ds_v;
+        const float Q_d = Q_ds[k_KQ_0/nthreads].x;
+
+        sum += vec_dot_q8_0_q8_1_impl<float, 1>(&v, &Q_q8[k_KQ_0/nthreads], K_q8_0[ib].d, Q_d);
+    }
+
+    return sum;
+}
+
+template <typename Tds, int ni>
 static __device__ __forceinline__ void quantize_q8_1_to_shared(
    const float * __restrict__ x, const float scale, int * __restrict__ yq32, void * __restrict__ yds) {

    float vals[sizeof(int)] = {0.0f};
 #pragma unroll
    for (int l = 0; l < int(sizeof(int)); ++l) {
-        vals[l] = scale * x[4*threadIdx.x + l];
+        vals[l] = (ni == WARP_SIZE || threadIdx.x < ni) ? scale * x[4*threadIdx.x + l] : 0.0f;
    }

    float amax = fabsf(vals[0]);
@@ -330,7 +284,7 @@ static __device__ __forceinline__ void quantize_q8_1_to_shared(
    }

    yq32[threadIdx.x] = q32;
-    if (threadIdx.x % QI8_1 == 0) {
+    if (threadIdx.x % QI8_1 == 0 && (ni == WARP_SIZE || threadIdx.x < ni)) {
        if (std::is_same<Tds, half2>::value) {
            ((half2  *) yds)[threadIdx.x/QI8_1] =  make_half2(d, sum);
        } else {
@@ -339,167 +293,276 @@ static __device__ __forceinline__ void quantize_q8_1_to_shared(
    }
 }

-typedef half  (*dequantize_1_f16_t)(const void *, const int64_t);
-typedef float (*dequantize_1_f32_t)(const void *, const int64_t);
+typedef void (*dequantize_V_t)(const void *, void *, const int64_t);

-template <typename T>
-static __device__ __forceinline__ T dequantize_1_q4_0(const void * __restrict__ vx, const int64_t i) {
+template <typename T, int ne>
+static __device__ __forceinline__ void dequantize_V_f16(const void * __restrict__ vx, void * __restrict__ dst, const int64_t i0) {
+    if constexpr (std::is_same_v<T, half>) {
+        ggml_cuda_memcpy_1<ne*sizeof(half)>(dst, (const half *) vx + i0);
+    } else if constexpr (std::is_same_v<T, float>) {
+        static_assert(ne % 2 == 0, "bad ne");
+        half2 tmp[ne/2];
+        ggml_cuda_memcpy_1<ne*sizeof(half)>(tmp, (const half *) vx + i0);
+        float2 * dst_f2 = (float2 *) dst;
+#pragma unroll
+        for (int l = 0; l < ne/2; ++l) {
+            dst_f2[l] = __half22float2(tmp[l]);
+        }
+    } else {
+        static_assert(std::is_same_v<T, void>, "unsupported type");
+    }
+}
+
+template <typename T, int ne>
+static __device__ __forceinline__ void dequantize_V_q4_0(const void * __restrict__ vx, void * __restrict__ dst, const int64_t i0) {
    const block_q4_0 * x = (const block_q4_0 *) vx;

-    const int64_t ib    =  i          /  QK4_0;
-    const int     iqs   =  i          % (QK4_0/2);
-    const int     shift = (i % QK4_0) / (QK4_0/2);
+    const int64_t ib    =  i0          /  QK4_0;
+    const int     iqs   =  i0          % (QK4_0/2);
+    const int     shift = (i0 % QK4_0) / (QK4_0/2);

-    const T   d  = x[ib].d;
-    const int q0 = x[ib].qs[iqs];
-    const int q  = ((q0 >> (4*shift)) & 0x0F) - 8;
+    int q;
+    static_assert(ne == 2 || ne == 4, "bad ne");
+    ggml_cuda_memcpy_1<ne, 2>(&q, x[ib].qs + iqs);
+    q >>= 4*shift;
+    q &= 0x0F0F0F0F;
+    q = __vsubss4(q, 0x08080808);
+
+    const int8_t * q8 = (const int8_t *) &q;

 #ifdef FP16_AVAILABLE
-    if (std::is_same<T, half>::value) {
-        return ((half) d)*((half) q);
-    }
-#endif // FP16_AVAILABLE
+    if constexpr (std::is_same_v<T, half>) {
+        const half2 d = __half2half2(x[ib].d);

-    return ((float) d)*((float) q);
+#pragma unroll
+        for (int l0 = 0; l0 < ne; l0 += 2) {
+            ((half2 *) dst)[l0/2] = d * make_half2(q8[l0 + 0], q8[l0 + 1]);
+        }
+    } else
+#endif // FP16_AVAILABLE
+    if constexpr (std::is_same_v<T, float>) {
+        const float d = x[ib].d;
+
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            ((float *) dst)[l] = d * q8[l];
+        }
+    } else {
+        static_assert(std::is_same_v<T, void>, "bad type");
+    }
 }

-template <typename T>
-static __device__ __forceinline__ T dequantize_1_q4_1(const void * __restrict__ vx, const int64_t i) {
+template <typename T, int ne>
+static __device__ __forceinline__ void dequantize_V_q4_1(const void * __restrict__ vx, void * __restrict__ dst, const int64_t i0) {
    const block_q4_1 * x = (const block_q4_1 *) vx;

-    const int64_t ib    =  i          /  QK4_1;
-    const int     iqs   =  i          % (QK4_1/2);
-    const int     shift = (i % QK4_1) / (QK4_1/2);
+    const int64_t ib    =  i0          /  QK4_1;
+    const int     iqs   =  i0          % (QK4_1/2);
+    const int     shift = (i0 % QK4_1) / (QK4_1/2);

-    const half2 dm = x[ib].dm;
-    const int   q0 = x[ib].qs[iqs];
-    const int   q  = ((q0 >> (4*shift)) & 0x0F);
+    int q;
+    static_assert(ne == 2 || ne == 4, "bad ne");
+    ggml_cuda_memcpy_1<ne>(&q, x[ib].qs + iqs);
+    q >>= 4*shift;
+    q &= 0x0F0F0F0F;
+
+    const int8_t * q8 = (const int8_t *) &q;

 #ifdef FP16_AVAILABLE
-    if (std::is_same<T, half>::value) {
-        return __low2half(dm)*((half) q) + __high2half(dm);
-    }
-#endif // FP16_AVAILABLE
+    if constexpr (std::is_same_v<T, half>) {
+        const half2 dm = x[ib].dm;
+        const half2 d  = __half2half2( __low2half(dm));
+        const half2 m  = __half2half2(__high2half(dm));

-    return __low2float(dm)*((float) q) + __high2float(dm);
+#pragma unroll
+        for (int l0 = 0; l0 < ne; l0 += 2) {
+            ((half2 *) dst)[l0/2] = d * make_half2(q8[l0 + 0], q8[l0 + 1]) + m;
+        }
+    } else
+#endif // FP16_AVAILABLE
+    if constexpr (std::is_same_v<T, float>) {
+        const float2 dm = __half22float2(x[ib].dm);
+
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            ((float *) dst)[l] = dm.x * q8[l] + dm.y;
+        }
+    } else {
+        static_assert(std::is_same_v<T, void>, "bad type");
+    }
 }

-template <typename T>
-static __device__ __forceinline__ T dequantize_1_q5_0(const void * __restrict__ vx, const int64_t i) {
+template <typename T, int ne>
+static __device__ __forceinline__ void dequantize_V_q5_0(const void * __restrict__ vx, void * __restrict__ dst, const int64_t i0) {
    const block_q5_0 * x = (const block_q5_0 *) vx;

-    const int64_t ib    =  i          /  QK5_0;
-    const int     idq   =  i          %  QK5_0;
-    const int     iqs   =  i          % (QK5_0/2);
-    const int     shift = (i % QK5_0) / (QK5_0/2);
+    const int64_t ib    =  i0          /  QK5_0;
+    const int     idq   =  i0          %  QK5_0;
+    const int     iqs   =  i0          % (QK5_0/2);
+    const int     shift = (i0 % QK5_0) / (QK5_0/2);

-    const T   d   = x[ib].d;
-    const int ql0 = x[ib].qs[iqs];
-    const int qh0 = get_int_b2(x[ib].qh, 0);
-    const int ql  = ((ql0 >> (4*shift)) & 0x0F);
-    const int qh  = ((qh0 >> idq) << 4) & 0x10;
-    const int q   = (ql | qh) - 16;
+    int q;
+    static_assert(ne == 2 || ne == 4, "bad ne");
+    ggml_cuda_memcpy_1<ne, 2>(&q, x[ib].qs + iqs);
+    q >>= 4*shift;
+    q &= 0x0F0F0F0F;
+
+    {
+        int qh;
+        ggml_cuda_memcpy_1<ne, 2>(&qh, x[ib].qh);
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            q |= ((qh >> (idq + l)) & 0x00000001) << (8*l + 4);
+        }
+    }
+
+    q = __vsubss4(q, 0x10101010);
+
+    const int8_t * q8 = (const int8_t *) &q;

 #ifdef FP16_AVAILABLE
-    if (std::is_same<T, half>::value) {
-        return ((half) d)*((half) q);
-    }
-#endif // FP16_AVAILABLE
+    if constexpr (std::is_same_v<T, half>) {
+        const half2 d = __half2half2(x[ib].d);

-    return ((float) d)*((float) q);
+#pragma unroll
+        for (int l0 = 0; l0 < ne; l0 += 2) {
+            ((half2 *) dst)[l0/2] = d * make_half2(q8[l0 + 0], q8[l0 + 1]);
+        }
+    } else
+#endif // FP16_AVAILABLE
+    if constexpr (std::is_same_v<T, float>) {
+        const float d = x[ib].d;
+
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            ((float *) dst)[l] = d * q8[l];
+        }
+    } else {
+        static_assert(std::is_same_v<T, void>, "bad type");
+    }
 }

-template <typename T>
-static __device__ __forceinline__ T dequantize_1_q5_1(const void * __restrict__ vx, const int64_t i) {
+template <typename T, int ne>
+static __device__ __forceinline__ void dequantize_V_q5_1(const void * __restrict__ vx, void * __restrict__ dst, const int64_t i0) {
    const block_q5_1 * x = (const block_q5_1 *) vx;

-    const int64_t ib    =  i          /  QK5_1;
-    const int     idq   =  i          %  QK5_1;
-    const int     iqs   =  i          % (QK5_1/2);
-    const int     shift = (i % QK5_1) / (QK5_1/2);
+    const int64_t ib    =  i0          /  QK5_1;
+    const int     idq   =  i0          %  QK5_1;
+    const int     iqs   =  i0          % (QK5_1/2);
+    const int     shift = (i0 % QK5_1) / (QK5_1/2);

-    const half2 dm  = x[ib].dm;
-    const int   ql0 = x[ib].qs[iqs];
-    const int   qh0 = get_int_b4(x[ib].qh, 0);
-    const int   ql  = ((ql0 >> (4*shift)) & 0x0F);
-    const int   qh  = ((qh0 >> idq) << 4) & 0x10;
-    const int   q   = (ql | qh);
+    int q;
+    static_assert(ne == 2 || ne == 4, "bad ne");
+    ggml_cuda_memcpy_1<ne>(&q, x[ib].qs + iqs);
+    q >>= 4*shift;
+    q &= 0x0F0F0F0F;
+
+    {
+        int qh;
+        ggml_cuda_memcpy_1<ne>(&qh, x[ib].qh);
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            q |= ((qh >> (idq + l)) & 0x00000001) << (8*l + 4);
+        }
+    }
+
+    const int8_t * q8 = (const int8_t *) &q;

 #ifdef FP16_AVAILABLE
-    if (std::is_same<T, half>::value) {
-        return __low2half(dm)*((half) q) + __high2half(dm);
-    }
-#endif // FP16_AVAILABLE
+    if constexpr (std::is_same_v<T, half>) {
+        const half2 dm = x[ib].dm;
+        const half2 d  = __half2half2( __low2half(dm));
+        const half2 m  = __half2half2(__high2half(dm));

-    return __low2float(dm)*((float) q) + __high2float(dm);
+#pragma unroll
+        for (int l0 = 0; l0 < ne; l0 += 2) {
+            ((half2 *) dst)[l0/2] = d * make_half2(q8[l0 + 0], q8[l0 + 1]) + m;
+        }
+    } else
+#endif // FP16_AVAILABLE
+    if constexpr (std::is_same_v<T, float>) {
+        const float2 dm = __half22float2(x[ib].dm);
+
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            ((float *) dst)[l] = dm.x * q8[l] + dm.y;
+        }
+    } else {
+        static_assert(std::is_same_v<T, void>, "bad type");
+    }
 }

-template <typename T>
-static __device__ __forceinline__ T dequantize_1_q8_0(const void * __restrict__ vx, const int64_t i) {
+template <typename T, int ne>
+static __device__ __forceinline__ void dequantize_V_q8_0(const void * __restrict__ vx, void * __restrict__ dst, const int64_t i0) {
    const block_q8_0 * x = (const block_q8_0 *) vx;

-    const int64_t ib  = i / QK8_0;
-    const int     iqs = i % QK8_0;
+    const int64_t ib  = i0 / QK8_0;
+    const int     iqs = i0 % QK8_0;

-    const T   d = x[ib].d;
-    const int q = x[ib].qs[iqs];
+    static_assert(ne % 2 == 0, "bad ne");
+    int8_t qs[ne];
+    ggml_cuda_memcpy_1<ne, 2>(qs, x[ib].qs + iqs);

 #ifdef FP16_AVAILABLE
-    if (std::is_same<T, half>::value) {
-        return ((half) d)*((half) q);
-    }
+    if constexpr (std::is_same<T, half>::value) {
+        const half2 d = __half2half2(x[ib].d);
+
+#pragma unroll
+        for (int l0 = 0; l0 < ne; l0 += 2) {
+            ((half2 *) dst)[l0/2] = d * make_half2(qs[l0 + 0], qs[l0 + 1]);
+        }
+    } else
 #endif // FP16_AVAILABLE
+    if constexpr (std::is_same<T, float>::value) {
+        const float d = x[ib].d;

-    return ((float) d)*((float) q);
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            ((float *) dst)[l] = d * qs[l];
+        }
+    } else {
+        static_assert(std::is_same_v<T, void>, "unsupported type");
+    }
 }

-template <typename T>
-static __device__ __forceinline__ T dequantize_1_f16(const void * __restrict__ vx, const int64_t i) {
-    const half * x = (const half *) vx;
-
-    return x[i];
+template <ggml_type type_K, int D, int nthreads>
+constexpr __device__ vec_dot_KQ_t get_vec_dot_KQ() {
+    if constexpr (type_K == GGML_TYPE_F16) {
+        return vec_dot_fattn_vec_KQ_f16<D, nthreads>;
+    } else if constexpr (type_K == GGML_TYPE_Q4_0) {
+        return vec_dot_fattn_vec_KQ_q4_0<D, nthreads>;
+    } else if constexpr (type_K == GGML_TYPE_Q4_1) {
+        return vec_dot_fattn_vec_KQ_q4_1<D, nthreads>;
+    } else if constexpr (type_K == GGML_TYPE_Q5_0) {
+        return vec_dot_fattn_vec_KQ_q5_0<D, nthreads>;
+    } else if constexpr (type_K == GGML_TYPE_Q5_1) {
+        return vec_dot_fattn_vec_KQ_q5_1<D, nthreads>;
+    } else if constexpr (type_K == GGML_TYPE_Q8_0) {
+        return vec_dot_fattn_vec_KQ_q8_0<D, nthreads>;
+    } else {
+        static_assert(type_K == -1, "bad type");
+        return nullptr;
+    }
 }

-template <int D, int warp_size = WARP_SIZE>
-constexpr __device__ vec_dot_KQ_f16_t get_vec_dot_KQ_f16(ggml_type type_K) {
-    return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0<half, D, warp_size> :
-        type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1<half, D, warp_size> :
-        type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0<half, D, warp_size> :
-        type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1<half, D, warp_size> :
-        type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0<half, D, warp_size> :
-        type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16<half, D, warp_size> :
-        nullptr;
-}
-
-template <int D, int warp_size = WARP_SIZE>
-constexpr __device__ vec_dot_KQ_f32_t get_vec_dot_KQ_f32(ggml_type type_K) {
-    return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0<float, D, warp_size> :
-        type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1<float, D, warp_size> :
-        type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0<float, D, warp_size> :
-        type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1<float, D, warp_size> :
-        type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0<float, D, warp_size> :
-        type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16<float, D, warp_size> :
-        nullptr;
-}
-
-constexpr __device__ dequantize_1_f16_t get_dequantize_1_f16(ggml_type type_V) {
-    return type_V == GGML_TYPE_Q4_0 ? dequantize_1_q4_0<half> :
-        type_V == GGML_TYPE_Q4_1 ? dequantize_1_q4_1<half> :
-        type_V == GGML_TYPE_Q5_0 ? dequantize_1_q5_0<half> :
-        type_V == GGML_TYPE_Q5_1 ? dequantize_1_q5_1<half> :
-        type_V == GGML_TYPE_Q8_0 ? dequantize_1_q8_0<half> :
-        type_V == GGML_TYPE_F16 ? dequantize_1_f16<half> :
-        nullptr;
-}
-
-constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
-    return type_V == GGML_TYPE_Q4_0 ? dequantize_1_q4_0<float> :
-        type_V == GGML_TYPE_Q4_1 ? dequantize_1_q4_1<float> :
-        type_V == GGML_TYPE_Q5_0 ? dequantize_1_q5_0<float> :
-        type_V == GGML_TYPE_Q5_1 ? dequantize_1_q5_1<float> :
-        type_V == GGML_TYPE_Q8_0 ? dequantize_1_q8_0<float> :
-        type_V == GGML_TYPE_F16 ? dequantize_1_f16<float> :
-        nullptr;
+template <ggml_type type_V, typename T, int ne>
+constexpr __device__ dequantize_V_t get_dequantize_V() {
+    if constexpr (type_V == GGML_TYPE_F16) {
+        return dequantize_V_f16<T, ne>;
+    } else if constexpr (type_V == GGML_TYPE_Q4_0) {
+        return dequantize_V_q4_0<T, ne>;
+    } else if constexpr (type_V == GGML_TYPE_Q4_1) {
+        return dequantize_V_q4_1<T, ne>;
+    } else if constexpr (type_V == GGML_TYPE_Q5_0) {
+        return dequantize_V_q5_0<T, ne>;
+    } else if constexpr (type_V == GGML_TYPE_Q5_1) {
+        return dequantize_V_q5_1<T, ne>;
+    } else if constexpr (type_V == GGML_TYPE_Q8_0) {
+        return dequantize_V_q8_0<T, ne>;
+    } else {
+        static_assert(type_V == -1, "bad type");
+        return nullptr;
+    }
 }

 template <int ncols1>
@@ -870,7 +933,7 @@ void launch_fattn(
            const int efficiency_percent = 100 * nblocks_total / (nwaves*blocks_per_wave);

            // Stop trying configurations with more waves if we already have good efficiency to avoid excessive overhead.
-            if (efficiency_percent_best >= 90 && nwaves > nwaves_best) {
+            if (efficiency_percent_best >= 95 && nwaves > nwaves_best) {
                break;
            }

@@ -1,495 +0,0 @@
-#include "common.cuh"
-#include "fattn-common.cuh"
-
-// Currenlty llvm with the amdgcn target dose not support unrolling loops
-// that contain a break that can not be resolved at compile time.
-#ifdef __clang__
-#pragma clang diagnostic push
-#pragma clang diagnostic ignored "-Wpass-failed"
-#endif // __clang__
-template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
-#ifndef GGML_USE_HIP
-__launch_bounds__(D, 1)
-#endif // GGML_USE_HIP
-static __global__ void flash_attn_vec_ext_f16(
-        const char * __restrict__ Q,
-        const char * __restrict__ K,
-        const char * __restrict__ V,
-        const char * __restrict__ mask,
-        const char * __restrict__ sinks,
-        const int  * __restrict__ KV_max,
-        float      * __restrict__ dst,
-        float2     * __restrict__ dst_meta,
-        const float scale,
-        const float max_bias,
-        const float m0,
-        const float m1,
-        const uint32_t n_head_log2,
-        const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
-                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
-        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
-                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
-                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
-                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
-                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
-#if defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
-
-    // Skip unused kernel variants for faster compilation:
-    if (use_logit_softcap && !(D == 128 || D == 256)) {
-        NO_DEVICE_CODE;
-        return;
-    }
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    if (ncols > 1) {
-        NO_DEVICE_CODE;
-        return;
-    }
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-
-    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
-
-    constexpr vec_dot_KQ_f16_t vec_dot_KQ = get_vec_dot_KQ_f16<D>(type_K);
-    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
-    constexpr dequantize_1_f16_t dequantize_1_v = get_dequantize_1_f16(type_V);
-
-    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
-
-    const int sequence = blockIdx.z / ne02;
-    const int head = blockIdx.z - sequence*ne02;
-    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    Q += nb03*sequence + nb02* head              + nb01*ic0;
-    K += nb13*sequence + nb12*(head / gqa_ratio);
-    V += nb23*sequence + nb22*(head / gqa_ratio);
-
-    const half  * maskh  = (const half  *) (mask + nb33*(sequence % ne33) + nb31*ic0);
-    const float * sinksf = (const float *) (sinks);
-
-    const float slopef = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
-    const half  slopeh = __float2half(slopef);
-
-    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
-    constexpr int nwarps = D / WARP_SIZE;
-    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
-    __builtin_assume(tid < D);
-
-    __shared__ half KQ[ncols*D];
-    half2 * KQ2 = (half2 *) KQ;
-
-    half kqmax[ncols];
-    half kqsum[ncols];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        kqmax[j] = -HALF_MAX_HALF;
-        kqsum[j] = 0.0f;
-    }
-
-    __shared__ half kqmax_shared[ncols][WARP_SIZE];
-    __shared__ half kqsum_shared[ncols][WARP_SIZE];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        if (threadIdx.y == 0) {
-            kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
-            kqsum_shared[j][threadIdx.x] = 0.0f;
-        }
-    }
-
-    __shared__ half maskh_shared[ncols*D];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        maskh_shared[j*D + tid] = 0.0f;
-    }
-
-    __syncthreads();
-
-    // Convert Q to half2 (f16 K) or q8_1 (quantized K) and store in registers:
-    half2  Q_h2[ncols][D/(2*WARP_SIZE)];
-    int   Q_i32[ncols][D/(sizeof(int)*QK8_1) == 0 ? 1 : D/(sizeof(int)*QK8_1)];
-    half2  Q_ds[ncols][D/QK8_1 == 0 ? 1 : D/QK8_1];
-    if (Q_q8_1) {
-#pragma unroll
-        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-            const int j = j0 + threadIdx.y;
-
-            if (j0 + nwarps > ncols && j >= ncols) {
-                break;
-            }
-
-            // Reuse KQ as temporary storage for converting Q to q8_1:
-            int   * tmp_q_i32 = (int   *) &KQ[j*D];
-            half2 * tmp_q_ds  = (half2 *) (tmp_q_i32 + D/sizeof(int));
-
-            // Set memory to zero if out of bounds:
-            if (ncols > 2 && ic0 + j >= ne01) {
-#pragma unroll
-                for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
-                    const int i = i0 + threadIdx.x;
-
-                    tmp_q_i32[i] = 0;
-                }
-                if (threadIdx.x < D/QK8_1) {
-                    tmp_q_ds[threadIdx.x] = make_half2(0.0f, 0.0f);
-                }
-                continue;
-            }
-
-            const float * Q_f = (const float *) (Q + j*nb01);
-#pragma unroll
-            for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
-                quantize_q8_1_to_shared<half2>(Q_f + 4*i0, scale, tmp_q_i32, tmp_q_ds);
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            int   * tmp_q_i32 = (int   *) &KQ[j*D];
-            half2 * tmp_q_ds  = (half2 *) (tmp_q_i32 + D/sizeof(int));
-
-#pragma unroll
-            for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
-                const int i = i0 + threadIdx.x;
-
-                Q_i32[j][i0/WARP_SIZE] = tmp_q_i32[i];
-                Q_ds[j][i0/WARP_SIZE]  = tmp_q_ds[i/QI8_1];
-            }
-        }
-
-        __syncthreads();
-    } else {
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            const float2 * Q_f2_j = (const float2 *) (Q + j*nb01);
-
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-                const int i = i0 + threadIdx.x;
-
-                const float2 tmp = ncols <= 2 || ic0 + j < ne01 ? Q_f2_j[i] : make_float2(0.0f, 0.0f);
-                Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
-            }
-        }
-    }
-
-
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        KQ[j*D + tid] = -HALF_MAX_HALF;
-    }
-    __syncthreads();
-
-    half2 VKQ[ncols] = {{0.0f, 0.0f}};
-
-    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
-    K     += blockIdx.y*D * nb11;
-    V     += blockIdx.y*D * nb21;
-    maskh += blockIdx.y*D;
-    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*D,
-             // Increment pointers after each loop:
-             K += gridDim.y*D*nb11, V += gridDim.y*D*nb21, maskh += gridDim.y*D) {
-
-        // Calculate KQ tile and keep track of new maximum KQ values:
-
-        if (mask) {
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                maskh_shared[j*D + tid] = slopeh*maskh[j*ne11 + tid];
-            }
-            __syncthreads();
-        }
-
-        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
-        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
-        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
-        half kqmax_new = kqmax[0];
-        half kqmax_new_arr[ncols];
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            kqmax_new_arr[j] = kqmax[j];
-        }
-
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
-            const int i_KQ = i_KQ_0 + threadIdx.y;
-
-            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
-                break;
-            }
-
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                half sum = vec_dot_KQ(K + i_KQ*nb11, Q_h2[j], Q_i32[j], Q_ds[j]);
-                sum = warp_reduce_sum((float)sum);
-
-                if (use_logit_softcap) {
-                    sum = logit_softcap*tanhf(sum);
-                }
-
-                sum += maskh_shared[j*D + i_KQ];
-
-                if (ncols == 1) {
-                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
-                } else {
-                    kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
-                }
-
-                if (threadIdx.x == 0) {
-                    KQ[j*D + i_KQ] = sum;
-                }
-            }
-        }
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
-
-            if (threadIdx.x == 0) {
-                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-
-            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
-            kqmax[j] = kqmax_new_j;
-
-            const half val = hexp(KQ[j*D + tid] - kqmax[j]);
-            kqsum[j] = kqsum[j]*KQ_max_scale + val;
-            KQ[j*D + tid] = val;
-
-            VKQ[j] *= __half2half2(KQ_max_scale);
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int k0 = 0; k0 < D; k0 += 2) {
-            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
-                break;
-            }
-
-            half2 V_k;
-            reinterpret_cast<half&>(V_k.x) = dequantize_1_v(V + (k0 + 0)*nb21, tid);
-            reinterpret_cast<half&>(V_k.y) = dequantize_1_v(V + (k0 + 1)*nb21, tid);
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
-            }
-        }
-
-        __syncthreads();
-    }
-
-    if (sinksf && blockIdx.y == 0) {
-        const half sink = __float2half(sinksf[head]);
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            if (threadIdx.x == 0) {
-                kqmax_shared[j][threadIdx.y] = fmaxf(kqmax[j], sink);
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-
-            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
-            kqmax[j] = kqmax_new_j;
-
-            const half val = hexp(sink - kqmax[j]);
-            kqsum[j] = kqsum[j]*KQ_max_scale;
-
-            if (tid == 0) {
-                kqsum[j] += val;
-            }
-
-            VKQ[j] *= __half2half2(KQ_max_scale);
-        }
-
-        __syncthreads();
-    }
-
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        kqsum[j] = warp_reduce_sum((float)kqsum[j]);
-        if (threadIdx.x == 0) {
-            kqsum_shared[j][threadIdx.y] = kqsum[j];
-        }
-    }
-
-    __syncthreads();
-
-#pragma unroll
-    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
-        if (ncols > 2 && ic0 + j_VKQ >= ne01) {
-            break;
-        }
-
-        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
-        kqsum[j_VKQ] = warp_reduce_sum((float)kqsum[j_VKQ]);
-
-        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
-        if (gridDim.y == 1) {
-            dst_val /= kqsum[j_VKQ];
-        }
-        dst[(((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + tid] = dst_val;
-    }
-
-    if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
-        dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
-    }
-#else
-    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
-        max_bias, m0, m1, n_head_log2, logit_softcap,
-        ne00, ne01, ne02, ne03,
-              nb01, nb02, nb03,
-        ne10, ne11, ne12, ne13,
-              nb11, nb12, nb13,
-              nb21, nb22, nb23,
-              ne31, ne32, ne33,
-              nb31, nb32, nb33);
-    NO_DEVICE_CODE;
-#endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE)
-}
-#ifdef __clang__
-#pragma clang diagnostic pop
-#endif // __clang__
-
-template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
-void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    constexpr int nwarps = D/WARP_SIZE;
-    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, type_K, type_V, use_logit_softcap>;
-    constexpr bool need_f16_K = D != 128;
-    constexpr bool need_f16_V = D != 128 && D != 64;
-    constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
-}
-
-template <int D, ggml_type type_K, ggml_type type_V>
-void ggml_cuda_flash_attn_ext_vec_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * KQV = dst;
-    const ggml_tensor * Q   = dst->src[0];
-    const ggml_tensor * K   = dst->src[1];
-    const ggml_tensor * V   = dst->src[2];
-
-    const int32_t precision = KQV->op_params[3];
-    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
-
-    GGML_ASSERT(K->type == type_K);
-    GGML_ASSERT(V->type == type_V);
-
-    float logit_softcap;
-    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
-
-    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-
-    if (Q->ne[1] == 1 || GGML_CUDA_CC_IS_NVIDIA(cc)) {
-        constexpr int cols_per_block = 1;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    if (Q->ne[1] == 2) {
-        constexpr int cols_per_block = 2;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    if (Q->ne[1] <= 4) {
-        constexpr int cols_per_block = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    constexpr int cols_per_block = 8;
-    if (logit_softcap == 0.0f) {
-        constexpr bool use_logit_softcap = false;
-        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-    } else {
-        constexpr bool use_logit_softcap = true;
-        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-    }
-}
-
-#define DECL_FATTN_VEC_F16_CASE(D, type_K, type_V)                          \
-    template void ggml_cuda_flash_attn_ext_vec_f16_case                     \
-    <D, type_K, type_V>(ggml_backend_cuda_context & ctx, ggml_tensor * dst) \
-
-extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0);
-extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1);
-extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0);
-extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1);
-extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0);
-extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16);
-
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0);
-
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1);
-
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0);
-
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1);
-
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0);
-
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16);
-extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16);
-
-extern DECL_FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16);
@@ -1,486 +0,0 @@
-#include "common.cuh"
-#include "fattn-common.cuh"
-
-// Currenlty llvm with the amdgcn target dose not support unrolling loops
-// that contain a break that can not be resolved at compile time.
-#ifdef __clang__
-#pragma clang diagnostic push
-#pragma clang diagnostic ignored "-Wpass-failed"
-#endif // __clang__
-template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
-#ifndef GGML_USE_HIP
-__launch_bounds__(D, 1)
-#endif // GGML_USE_HIP
-static __global__ void flash_attn_vec_ext_f32(
-        const char * __restrict__ Q,
-        const char * __restrict__ K,
-        const char * __restrict__ V,
-        const char * __restrict__ mask,
-        const char * __restrict__ sinks,
-        const int  * __restrict__ KV_max,
-        float      * __restrict__ dst,
-        float2     * __restrict__ dst_meta,
-        const float scale,
-        const float max_bias,
-        const float m0,
-        const float m1,
-        const uint32_t n_head_log2,
-        const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
-                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
-        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
-                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
-                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
-                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
-                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
-#ifdef FLASH_ATTN_AVAILABLE
-
-    // Skip unused kernel variants for faster compilation:
-    if (use_logit_softcap && !(D == 128 || D == 256)) {
-        GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
-            max_bias, m0, m1, n_head_log2, logit_softcap,
-            ne00, ne01, ne02, ne03,
-                  nb01, nb02, nb03,
-            ne10, ne11, ne12, ne13,
-                  nb11, nb12, nb13,
-                  nb21, nb22, nb23,
-                  ne31, ne32, ne33,
-                  nb31, nb32, nb33);
-        NO_DEVICE_CODE;
-        return;
-    }
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    if (ncols > 1) {
-        NO_DEVICE_CODE;
-        return;
-    }
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-
-    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
-
-    constexpr vec_dot_KQ_f32_t vec_dot_KQ = get_vec_dot_KQ_f32<D>(type_K);
-    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
-    constexpr dequantize_1_f32_t dequantize_1_v = get_dequantize_1_f32(type_V);
-
-    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
-
-    const int sequence = blockIdx.z / ne02;
-    const int head = blockIdx.z - sequence*ne02;
-    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    Q += nb03*sequence + nb02* head              + nb01*ic0;
-    K += nb13*sequence + nb12*(head / gqa_ratio);
-    V += nb23*sequence + nb22*(head / gqa_ratio);
-
-    const half  * maskh  = (const half  *) (mask + nb33*(sequence % ne33) + nb31*ic0);
-    const float * sinksf = (const float *) (sinks);
-
-    const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
-
-    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
-    constexpr int nwarps = D / WARP_SIZE;
-    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
-    __builtin_assume(tid < D);
-
-    __shared__ float KQ[ncols*D];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        KQ[j*D + tid] = -FLT_MAX/2.0f;
-    }
-
-    float kqmax[ncols];
-    float kqsum[ncols];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        kqmax[j] = -FLT_MAX/2.0f;
-        kqsum[j] = 0.0f;
-    }
-
-    __shared__ float kqmax_shared[ncols][WARP_SIZE];
-    __shared__ float kqsum_shared[ncols][WARP_SIZE];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        if (threadIdx.y == 0) {
-            kqmax_shared[j][threadIdx.x] = -FLT_MAX/2.0f;
-            kqsum_shared[j][threadIdx.x] = 0.0f;
-        }
-    }
-
-    __shared__ float maskf_shared[ncols*D];
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        maskf_shared[j*D + tid] = 0.0f;
-    }
-
-    __syncthreads();
-
-    // Convert Q to float2 (f16 K) or q8_1 (quantized K) and store in registers:
-    float2  Q_f2[ncols][D/(2*WARP_SIZE)];
-    int    Q_i32[ncols][D/(sizeof(int)*QK8_1) == 0 ? 1 : D >= D/(sizeof(int)*QK8_1)];
-    float2  Q_ds[ncols][D/QK8_1 == 0 ? 1 : D/QK8_1];
-    if (Q_q8_1) {
-#pragma unroll
-        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-            const int j = j0 + threadIdx.y;
-
-            if (j0 + nwarps > ncols && j >= ncols) {
-                break;
-            }
-
-            // Reuse KQ as temporary storage for converting Q to q8_1:
-            int    * tmp_q_i32 = (int    *) &KQ[j*D];
-            float2 * tmp_q_ds  = (float2 *) (tmp_q_i32 + D/sizeof(int));
-
-            // Set memory to zero if out of bounds:
-            if (ncols > 2 && ic0 + j >= ne01) {
-#pragma unroll
-                for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
-                    const int i = i0 + threadIdx.x;
-
-                    tmp_q_i32[i] = 0;
-                }
-                if (threadIdx.x < D/QK8_1) {
-                    tmp_q_ds[threadIdx.x] = make_float2(0.0f, 0.0f);
-                }
-                continue;
-            }
-
-            const float * Q_f = (const float *) (Q + j*nb01);
-#pragma unroll
-            for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
-                quantize_q8_1_to_shared<float2>(Q_f + 4*i0, scale, tmp_q_i32, tmp_q_ds);
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            int    * tmp_q_i32 = (int    *) &KQ[j*D];
-            float2 * tmp_q_ds  = (float2 *) (tmp_q_i32 + D/sizeof(int));
-
-#pragma unroll
-            for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
-                const int i = i0 + threadIdx.x;
-
-                Q_i32[j][i0/WARP_SIZE] = tmp_q_i32[i];
-                Q_ds[j][i0/WARP_SIZE]  = tmp_q_ds[i/QI8_1];
-            }
-        }
-
-        __syncthreads();
-    } else {
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            const float2 * Q_f2_j = (const float2 *) (Q + j*nb01);
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-                const int i = i0 + threadIdx.x;
-
-                Q_f2[j][i0/WARP_SIZE]    = ncols <= 2 || ic0 + j < ne01 ? Q_f2_j[i] : make_float2(0.0f, 0.0f);
-                Q_f2[j][i0/WARP_SIZE].x *= scale;
-                Q_f2[j][i0/WARP_SIZE].y *= scale;
-            }
-        }
-    }
-
-    float VKQ[ncols] = {0.0f};
-
-    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
-    K     += blockIdx.y*D * nb11;
-    V     += blockIdx.y*D * nb21;
-    maskh += blockIdx.y*D;
-    for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*D,
-             // Increment pointers after each loop:
-             K += gridDim.y*D*nb11, V += gridDim.y*D*nb21, maskh += gridDim.y*D) {
-
-        // Calculate KQ tile and keep track of new maximum KQ values:
-
-        if (mask) {
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                maskf_shared[j*D + tid] = slope*__half2float(maskh[j*ne11 + tid]);
-            }
-            __syncthreads();
-        }
-
-        float kqmax_new_arr[ncols];
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            kqmax_new_arr[j] = kqmax[j];
-        }
-
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
-            const int i_KQ = i_KQ_0 + threadIdx.y;
-
-            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
-                break;
-            }
-
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                float sum = vec_dot_KQ(K + i_KQ*nb11, Q_f2[j], Q_i32[j], Q_ds[j]);
-                sum = warp_reduce_sum(sum);
-
-                if (use_logit_softcap) {
-                    sum = logit_softcap*tanhf(sum);
-                }
-
-                sum += maskf_shared[j*D + i_KQ];
-
-                kqmax_new_arr[j] = fmaxf(kqmax_new_arr[j], sum);
-
-                if (threadIdx.x == 0) {
-                    KQ[j*D + i_KQ] = sum;
-                }
-            }
-        }
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            float kqmax_new_j = kqmax_new_arr[j];
-
-            if (threadIdx.x == 0) {
-                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            float kqmax_new_j = kqmax_shared[j][threadIdx.x];
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-
-            const float KQ_max_scale = expf(kqmax[j] - kqmax_new_j);
-            kqmax[j] = kqmax_new_j;
-
-            const float val = expf(KQ[j*D + tid] - kqmax[j]);
-            kqsum[j] = kqsum[j]*KQ_max_scale + val;
-            KQ[j*D + tid] = val;
-
-            VKQ[j] *= KQ_max_scale;
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int k = 0; k < D; ++k) {
-            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k >= ne11) {
-                break;
-            }
-
-            const float V_ki = dequantize_1_v(V + k*nb21, tid);
-#pragma unroll
-            for (int j = 0; j < ncols; ++j) {
-                VKQ[j] += V_ki*KQ[j*D + k];
-            }
-        }
-
-        __syncthreads();
-    }
-
-    if (sinksf && blockIdx.y == 0) {
-        const float sink = sinksf[head];
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            if (threadIdx.x == 0) {
-                kqmax_shared[j][threadIdx.y] = fmaxf(kqmax[j], sink);
-            }
-        }
-
-        __syncthreads();
-
-#pragma unroll
-        for (int j = 0; j < ncols; ++j) {
-            float kqmax_new_j = kqmax_shared[j][threadIdx.x];
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
-
-            const float KQ_max_scale = expf(kqmax[j] - kqmax_new_j);
-            kqmax[j] = kqmax_new_j;
-
-            const float val = expf(sink - kqmax[j]);
-            kqsum[j] = kqsum[j]*KQ_max_scale;
-
-            if (tid == 0) {
-                kqsum[j] += val;
-            }
-
-            VKQ[j] *= KQ_max_scale;
-        }
-
-        __syncthreads();
-    }
-
-#pragma unroll
-    for (int j = 0; j < ncols; ++j) {
-        kqsum[j] = warp_reduce_sum(kqsum[j]);
-        if (threadIdx.x == 0) {
-            kqsum_shared[j][threadIdx.y] = kqsum[j];
-        }
-    }
-
-    __syncthreads();
-
-#pragma unroll
-    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
-        if (ncols > 2 && ic0 + j_VKQ >= ne01) {
-            break;
-        }
-
-        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
-        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
-
-        float dst_val = VKQ[j_VKQ];
-        if (gridDim.y == 1) {
-            dst_val /= kqsum[j_VKQ];
-        }
-        dst[(((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + tid] = dst_val;
-    }
-
-    if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
-        dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]);
-    }
-#else
-    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
-        max_bias, m0, m1, n_head_log2, logit_softcap,
-        ne00, ne01, ne02, ne03,
-              nb01, nb02, nb03,
-        ne10, ne11, ne12, ne13,
-              nb11, nb12, nb13,
-              nb21, nb22, nb23,
-              ne31, ne32, ne33,
-              nb31, nb32, nb33);
-    NO_DEVICE_CODE;
-#endif // FLASH_ATTN_AVAILABLE
-}
-#ifdef __clang__
-#pragma clang diagnostic pop
-#endif // __clang__
-
-template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
-void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    constexpr int nwarps = D/WARP_SIZE;
-    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, type_K, type_V, use_logit_softcap>;
-    constexpr bool need_f16_K = D != 128;
-    constexpr bool need_f16_V = D != 128 && D != 64;
-    constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
-}
-
-template <int D, ggml_type type_K, ggml_type type_V>
-void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * KQV = dst;
-    const ggml_tensor * Q   = dst->src[0];
-    const ggml_tensor * K   = dst->src[1];
-    const ggml_tensor * V   = dst->src[2];
-
-    GGML_ASSERT(K->type == type_K);
-    GGML_ASSERT(V->type == type_V);
-
-    float logit_softcap;
-    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
-
-    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-
-    if (Q->ne[1] == 1 || GGML_CUDA_CC_IS_NVIDIA(cc)) {
-        constexpr int cols_per_block = 1;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    if (Q->ne[1] == 2) {
-        constexpr int cols_per_block = 2;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    if (Q->ne[1] <= 4) {
-        constexpr int cols_per_block = 4;
-        if (logit_softcap == 0.0f) {
-            constexpr bool use_logit_softcap = false;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-        } else {
-            constexpr bool use_logit_softcap = true;
-            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-        }
-        return;
-    }
-
-    constexpr int cols_per_block = 8;
-    if (logit_softcap == 0.0f) {
-        constexpr bool use_logit_softcap = false;
-        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-    } else {
-        constexpr bool use_logit_softcap = true;
-        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
-    }
-}
-
-#define DECL_FATTN_VEC_F32_CASE(D, type_K, type_V)                          \
-    template void ggml_cuda_flash_attn_ext_vec_f32_case                     \
-    <D, type_K, type_V>(ggml_backend_cuda_context & ctx, ggml_tensor * dst) \
-
-extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0);
-extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1);
-extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0);
-extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1);
-extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0);
-extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16);
-
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0);
-
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1);
-
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0);
-
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1);
-
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0);
-
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16);
-extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16);
-
-extern DECL_FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16);
@@ -0,0 +1,593 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+
+static int ggml_cuda_fattn_vec_get_nthreads_host(const int cc) {
+    return 128;
+    GGML_UNUSED(cc);
+}
+
+static constexpr __device__ int ggml_cuda_fattn_vec_get_nthreads_device() {
+    return 128;
+}
+
+// Currenlty llvm with the amdgcn target dose not support unrolling loops
+// that contain a break that can not be resolved at compile time.
+#ifdef __clang__
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wpass-failed"
+#endif // __clang__
+template<int D, int ncols, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
+__launch_bounds__(ggml_cuda_fattn_vec_get_nthreads_device(), 1)
+static __global__ void flash_attn_ext_vec(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        const char * __restrict__ sinks,
+        const int  * __restrict__ KV_max,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const float logit_softcap,
+        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
+        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
+                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
+                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
+                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
+                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
+#ifdef FLASH_ATTN_AVAILABLE
+
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+            max_bias, m0, m1, n_head_log2, logit_softcap,
+            ne00, ne01, ne02, ne03,
+                  nb01, nb02, nb03,
+            ne10, ne11, ne12, ne13,
+                  nb11, nb12, nb13,
+                  nb21, nb22, nb23,
+                  ne31, ne32, ne33,
+                  nb31, nb32, nb33);
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
+    constexpr int cpy_ne = cpy_nb / 4;
+
+#ifdef GGML_USE_HIP
+#ifdef RDNA
+    constexpr int nthreads_KQ_q = 2;
+#else
+    constexpr int nthreads_KQ_q = 4;
+#endif // RDNA
+    constexpr int nthreads_V_q  = (D/4 < 32 ? D/4 : 32);
+#else
+    constexpr int nthreads_KQ_q = (D/4 < 32 ? D/4 : 32);
+    constexpr int nthreads_V_q  = (D/4 < 32 ? D/4 : 32);
+#endif // GGML_USE_HIP
+
+    constexpr int nthreads    = ggml_cuda_fattn_vec_get_nthreads_device();
+    constexpr int nthreads_KQ = type_K == GGML_TYPE_F16 ? 128 / cpy_nb : nthreads_KQ_q;
+    constexpr int nthreads_V  = type_V == GGML_TYPE_F16 ? 128 / cpy_nb : nthreads_V_q;
+
+    static_assert(WARP_SIZE % nthreads_KQ == 0, "bad nthreads_K");
+    static_assert(WARP_SIZE % nthreads_V  == 0, "bad nthreads_V");
+
+    constexpr int V_rows_per_thread = type_V == GGML_TYPE_F16 ? 2*cpy_ne : 4;
+    constexpr int V_cols_per_iter   = WARP_SIZE / nthreads_V;
+
+    constexpr vec_dot_KQ_t vec_dot_KQ = get_vec_dot_KQ<type_K, D, nthreads_KQ>();
+    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
+#ifdef FAST_FP16_AVAILABLE
+    constexpr dequantize_V_t dequantize_V = get_dequantize_V<type_V, half,  V_rows_per_thread>();
+#else
+    constexpr dequantize_V_t dequantize_V = get_dequantize_V<type_V, float, V_rows_per_thread>();
+#endif // FAST_FP16_AVAILABLE
+
+    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
+
+    const int sequence = blockIdx.z / ne02;
+    const int head = blockIdx.z - sequence*ne02;
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    Q += nb03*sequence + nb02* head              + nb01*ic0;
+    K += nb13*sequence + nb12*(head / gqa_ratio);
+    V += nb23*sequence + nb22*(head / gqa_ratio);
+
+    const half * maskh  = (const half  *) (mask + nb33*(sequence % ne33) + nb31*ic0);
+
+    const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = nthreads / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < nthreads);
+
+    constexpr int ne_KQ      = ncols*D;
+    constexpr int ne_combine = nwarps*V_cols_per_iter*D;
+#ifdef FAST_FP16_AVAILABLE
+    half2            VKQ[ncols][(D/2)/nthreads_V] = {{{0.0f, 0.0f}}};
+    __shared__ half   KQ[ne_KQ > ne_combine ? ne_KQ : ne_combine];
+#else
+    float2           VKQ[ncols][(D/2)/nthreads_V] = {{{0.0f, 0.0f}}};
+    __shared__ float  KQ[ne_KQ > ne_combine ? ne_KQ : ne_combine];
+#endif // FAST_FP16_AVAILABLE
+
+    float KQ_max[ncols];
+    float KQ_sum[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ_max[j] = -FLT_MAX/2.0f;
+        KQ_sum[j] = 0.0f;
+    }
+
+    // Convert Q to float2 (f16 K) or q8_1 (quantized K) and store in registers:
+#ifdef FAST_FP16_AVAILABLE
+    half2  Q_reg[ncols][(D/2)/nthreads_KQ]; // Will be initialized completely.
+#else
+    float2 Q_reg[ncols][(D/2)/nthreads_KQ] = {{{0.0f, 0.0f}}}; // May be only partially initialized.
+#endif // FAST_FP16_AVAILABLE
+    int    Q_i32[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
+    float2  Q_ds[ncols][1 > D/(sizeof(int)*nthreads_KQ) ? 1 : D/(sizeof(int)*nthreads_KQ)];
+    if constexpr (Q_q8_1) {
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (j0 + nwarps > ncols && j >= ncols) {
+                break;
+            }
+
+            // Reuse KQ as temporary storage for converting Q to q8_1:
+            int    * tmp_q_i32 = (int    *) &KQ[j*D];
+            float2 * tmp_q_ds  = (float2 *) (tmp_q_i32 + D/sizeof(int));
+
+            // Set memory to zero if out of bounds:
+            if (ncols > 1 && ic0 + j >= ne01) {
+#pragma unroll
+                for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
+                    const int i = i0 + threadIdx.x;
+
+                    if (i0 + WARP_SIZE <= D/sizeof(int) || i < D/sizeof(int)) {
+                        tmp_q_i32[i] = 0;
+                    }
+                }
+                if (threadIdx.x < D/QK8_1) {
+                    tmp_q_ds[threadIdx.x] = make_float2(0.0f, 0.0f);
+                }
+            } else {
+                const float * Q_f = (const float *) (Q + j*nb01);
+                constexpr int nthreads_quantize = D/sizeof(int) < WARP_SIZE ? D/sizeof(int) : WARP_SIZE;
+#pragma unroll
+                for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += nthreads_quantize) {
+                    quantize_q8_1_to_shared<float2, nthreads_quantize>
+                        (Q_f + i0*sizeof(int), scale, tmp_q_i32 + i0, tmp_q_ds + i0/QI8_1);
+                }
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            int    * tmp_q_i32 = (int    *) &KQ[j*D];
+            float2 * tmp_q_ds  = (float2 *) (tmp_q_i32 + D/sizeof(int));
+
+#pragma unroll
+            for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += nthreads_KQ) {
+                const int i = i0 + (nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ);
+
+                Q_i32[j][i0/nthreads_KQ] = tmp_q_i32[i];
+                Q_ds[j][i0/nthreads_KQ]  = tmp_q_ds[i/QI8_1];
+            }
+        }
+
+        __syncthreads();
+    } else {
+#ifdef FAST_FP16_AVAILABLE
+        const half2 scale_h2 = make_half2(scale, scale);
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            const float2 * Q_j = (const float2 *) (Q + j*nb01);
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += nthreads_KQ*cpy_ne) {
+                const int i = i0 + (nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ)*cpy_ne;
+
+                float2 tmp[cpy_ne] = {{0.0f, 0.0f}};
+                if (ncols == 1 || ic0 + j < ne01) {
+                    ggml_cuda_memcpy_1<cpy_nb>(tmp,            &Q_j[i]);
+                    ggml_cuda_memcpy_1<cpy_nb>(tmp + cpy_ne/2, &Q_j[i + cpy_ne/2]);
+                }
+#pragma unroll
+                for (int i1 = 0; i1 < cpy_ne; ++i1) {
+                    Q_reg[j][i0/nthreads_KQ + i1] = make_half2(tmp[i1].x, tmp[i1].y);
+                }
+            }
+#pragma unroll
+            for (int k = 0; k < (D/2)/nthreads_KQ; ++k) {
+                Q_reg[j][k] *= scale_h2;
+            }
+        }
+#else
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            const float2 * Q_j = (const float2 *) (Q + j*nb01);
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += nthreads_KQ*cpy_ne) {
+                const int i = i0 + (nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ)*cpy_ne;
+                if (ncols == 1 || ic0 + j < ne01) {
+                    ggml_cuda_memcpy_1<cpy_nb>(&Q_reg[j][i0/nthreads_KQ],            &Q_j[i]);
+                    ggml_cuda_memcpy_1<cpy_nb>(&Q_reg[j][i0/nthreads_KQ + cpy_ne/2], &Q_j[i + cpy_ne/2]);
+                }
+            }
+#pragma unroll
+            for (int k = 0; k < (D/2)/nthreads_KQ; ++k) {
+                Q_reg[j][k].x *= scale;
+                Q_reg[j][k].y *= scale;
+            }
+        }
+#endif // FAST_FP16_AVAILABLE
+    }
+
+    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
+    K     += blockIdx.y*nthreads * nb11;
+    V     += blockIdx.y*nthreads * nb21;
+    maskh += blockIdx.y*nthreads;
+    for (int k_VKQ_0 = blockIdx.y*nthreads; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*nthreads,
+             // Increment pointers after each loop:
+             K += gridDim.y*nthreads*nb11, V += gridDim.y*nthreads*nb21, maskh += gridDim.y*nthreads) {
+
+        // Calculate KQ tile and keep track of new maximum KQ values:
+        float KQ_reg[ncols]; // KQ in registers.
+
+        float KQ_max_new[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            KQ_max_new[j] = KQ_max[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < nthreads_KQ; ++i_KQ_0) {
+            const int i_KQ = threadIdx.y*WARP_SIZE + (nthreads_KQ == WARP_SIZE ? 0 : (threadIdx.x & ~(nthreads_KQ-1))) + i_KQ_0;
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                float sum = vec_dot_KQ(K + i_KQ*nb11, Q_reg[j], Q_i32[j], Q_ds[j]);
+                sum = warp_reduce_sum<nthreads_KQ>(sum);
+
+                if (use_logit_softcap) {
+                    sum = logit_softcap*tanhf(sum);
+                }
+
+                if (mask) {
+                    sum += slope*__half2float(maskh[j*ne11 + i_KQ]);
+                }
+
+                KQ_max_new[j] = fmaxf(KQ_max_new[j], sum);
+
+                if ((nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ) == i_KQ_0) {
+                    KQ_reg[j] = sum;
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+            for (int offset = nthreads_KQ; offset < WARP_SIZE; offset <<= 1) {
+                KQ_max_new[j] = fmaxf(KQ_max_new[j], __shfl_xor_sync(0xFFFFFFFF, KQ_max_new[j], offset, WARP_SIZE));
+            }
+            const float KQ_max_scale = expf(KQ_max[j] - KQ_max_new[j]);
+            KQ_max[j] = KQ_max_new[j];
+
+            KQ_reg[j] = expf(KQ_reg[j] - KQ_max[j]);
+            KQ_sum[j] = KQ_sum[j]*KQ_max_scale + KQ_reg[j];
+            KQ[j*nthreads + tid] = KQ_reg[j];
+
+#ifdef FAST_FP16_AVAILABLE
+            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
+#pragma unroll
+            for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V) {
+                VKQ[j][i_VKQ_0/nthreads_V] *= KQ_max_scale_h2;
+            }
+#else
+#pragma unroll
+            for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V) {
+                VKQ[j][i_VKQ_0/nthreads_V].x *= KQ_max_scale;
+                VKQ[j][i_VKQ_0/nthreads_V].y *= KQ_max_scale;
+            }
+#endif // FAST_FP16_AVAILABLE
+        }
+
+#ifndef GGML_USE_HIP
+        __syncwarp();
+#endif // GGML_USE_HIP
+
+#pragma unroll
+        for (int k0 = 0; k0 < WARP_SIZE; k0 += V_cols_per_iter) {
+            const int k = threadIdx.y*WARP_SIZE + k0 + (nthreads_V == WARP_SIZE ? 0 : threadIdx.x / nthreads_V);
+
+#ifdef FAST_FP16_AVAILABLE
+            half2 KQ_k[ncols];
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                KQ_k[j] = __half2half2(KQ[j*nthreads + k]);
+            }
+#pragma unroll
+            for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V*V_rows_per_thread/2) {
+                half2 tmp[V_rows_per_thread/2];
+                dequantize_V(V + k*nb21, tmp,
+                    2*i_VKQ_0 + (nthreads_V == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_V)*V_rows_per_thread);
+#pragma unroll
+                for (int i_VKQ_1 = 0; i_VKQ_1 < V_rows_per_thread/2; ++i_VKQ_1) {
+#pragma unroll
+                    for (int j = 0; j < ncols; ++j) {
+                        VKQ[j][i_VKQ_0/nthreads_V + i_VKQ_1] += tmp[i_VKQ_1]*KQ_k[j];
+                    }
+                }
+            }
+#else
+            float KQ_k[ncols];
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                KQ_k[j] = KQ[j*nthreads + k];
+            }
+#pragma unroll
+            for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V*V_rows_per_thread/2) {
+                float2 tmp[V_rows_per_thread/2];
+                dequantize_V(V + k*nb21, tmp,
+                    2*i_VKQ_0 + (nthreads_V == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_V)*V_rows_per_thread);
+#pragma unroll
+                for (int i_VKQ_1 = 0; i_VKQ_1 < V_rows_per_thread/2; ++i_VKQ_1) {
+#pragma unroll
+                    for (int j = 0; j < ncols; ++j) {
+                        VKQ[j][i_VKQ_0/nthreads_V + i_VKQ_1].x += tmp[i_VKQ_1].x*KQ_k[j];
+                        VKQ[j][i_VKQ_0/nthreads_V + i_VKQ_1].y += tmp[i_VKQ_1].y*KQ_k[j];
+                    }
+                }
+            }
+#endif // FAST_FP16_AVAILABLE
+        }
+    }
+
+    if (sinks && blockIdx.y == 0) {
+        const float sink = ((const float *) sinks)[head];
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (j0 + nwarps > ncols && j >= ncols) {
+                break;
+            }
+
+            const float kqmax_new_j = fmaxf(sink, KQ_max[j]);
+            const float KQ_max_scale = expf(KQ_max[j] - kqmax_new_j);
+            KQ_max[j] = kqmax_new_j;
+
+            KQ_sum[j] = KQ_sum[j]*KQ_max_scale + (threadIdx.x == 0 ? expf(sink - KQ_max[j]) : 0.0f);
+
+#ifdef FAST_FP16_AVAILABLE
+            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
+#pragma unroll
+            for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V) {
+                VKQ[j][i_VKQ_0/nthreads_V] *= KQ_max_scale_h2;
+            }
+#else
+#pragma unroll
+            for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V) {
+                VKQ[j][i_VKQ_0/nthreads_V].x *= KQ_max_scale;
+                VKQ[j][i_VKQ_0/nthreads_V].y *= KQ_max_scale;
+            }
+#endif // FAST_FP16_AVAILABLE
+        }
+    }
+
+    __shared__ float KQ_max_shared[ncols][WARP_SIZE];
+    __shared__ float KQ_sum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            KQ_max_shared[j][threadIdx.x] = -FLT_MAX/2.0f;
+            KQ_sum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.x == 0) {
+            KQ_max_shared[j][threadIdx.y] = KQ_max[j];
+        }
+    }
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        if (ncols > 1 && ic0 + j_VKQ >= ne01) {
+            break;
+        }
+
+        float kqmax_new = KQ_max_shared[j_VKQ][threadIdx.x];
+        kqmax_new = warp_reduce_max(kqmax_new);
+        const float kqmax_scale = expf(KQ_max[j_VKQ] - kqmax_new);
+        KQ_max[j_VKQ] = kqmax_new;
+
+#ifdef FAST_FP16_AVAILABLE
+        half2 * VKQ_tmp = (half2 *) KQ + threadIdx.y*(V_cols_per_iter*D/2)
+            + (nthreads_V == WARP_SIZE ? 0 : threadIdx.x / nthreads_V)*(D/2);
+
+        const half2 kqmax_scale_h2 = make_half2(kqmax_scale, kqmax_scale);
+#pragma unroll
+        for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V) {
+            VKQ[j_VKQ][i_VKQ_0/nthreads_V] *= kqmax_scale_h2;
+        }
+#pragma unroll
+        for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V*V_rows_per_thread/2) {
+            const int i_VKQ = i_VKQ_0 + (nthreads_V == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_V)*(V_rows_per_thread/2);
+
+            ggml_cuda_memcpy_1<V_rows_per_thread*sizeof(half)>(VKQ_tmp + i_VKQ, &VKQ[j_VKQ][i_VKQ_0/nthreads_V]);
+        }
+#else
+        float2 * VKQ_tmp = (float2 *) KQ + threadIdx.y*(V_cols_per_iter*D/2)
+            + (nthreads_V == WARP_SIZE ? 0 : threadIdx.x / nthreads_V)*(D/2);
+
+#pragma unroll
+        for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V) {
+            VKQ[j_VKQ][i_VKQ_0/nthreads_V].x *= kqmax_scale;
+            VKQ[j_VKQ][i_VKQ_0/nthreads_V].y *= kqmax_scale;
+        }
+#pragma unroll
+        for (int i_VKQ_0 = 0; i_VKQ_0 < D/2; i_VKQ_0 += nthreads_V*V_rows_per_thread/2) {
+            const int i_VKQ = i_VKQ_0 + (nthreads_V == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_V)*(V_rows_per_thread/2);
+
+            ggml_cuda_memcpy_1<V_rows_per_thread/2*sizeof(float)>(VKQ_tmp + i_VKQ,                       &VKQ[j_VKQ][i_VKQ_0/nthreads_V]);
+            ggml_cuda_memcpy_1<V_rows_per_thread/2*sizeof(float)>(VKQ_tmp + i_VKQ + V_rows_per_thread/4, &VKQ[j_VKQ][i_VKQ_0/nthreads_V + V_rows_per_thread/4]);
+        }
+#endif // FAST_FP16_AVAILABLE
+
+        KQ_sum[j_VKQ] *= kqmax_scale;
+        KQ_sum[j_VKQ] = warp_reduce_sum(KQ_sum[j_VKQ]);
+        if (threadIdx.x == 0) {
+            KQ_sum_shared[j_VKQ][threadIdx.y] = KQ_sum[j_VKQ];
+        }
+
+        __syncthreads();
+
+        if (nthreads <= D || tid < D) {
+            KQ_sum[j_VKQ] = KQ_sum_shared[j_VKQ][threadIdx.x];
+            KQ_sum[j_VKQ] = warp_reduce_sum(KQ_sum[j_VKQ]);
+
+#pragma unroll
+            for (int i0 = 0; i0 < D; i0 += nthreads) {
+                float dst_val = 0;
+#pragma unroll
+                for (int w = 0; w < nwarps; ++w) {
+#pragma unroll
+                    for (int v = 0; v < V_cols_per_iter; ++v) {
+                        dst_val += float(KQ[w*V_cols_per_iter*D + v*D + i0 + tid]);
+                    }
+                }
+                if (gridDim.y == 1) {
+                    dst_val /= KQ_sum[j_VKQ];
+                }
+                dst[(((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + i0 + tid] = dst_val;
+            }
+        }
+
+        if (j_VKQ < ncols-1) {
+            __syncthreads();
+        }
+
+    }
+
+    if (gridDim.y != 1 && tid < ncols && (ncols == 1 || ic0 + tid < ne01)) {
+        dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(KQ_max[tid], KQ_sum[tid]);
+    }
+#else
+    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
+        max_bias, m0, m1, n_head_log2, logit_softcap,
+        ne00, ne01, ne02, ne03,
+              nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+              nb11, nb12, nb13,
+              nb21, nb22, nb23,
+              ne31, ne32, ne33,
+              nb31, nb32, nb33);
+    NO_DEVICE_CODE;
+#endif // FLASH_ATTN_AVAILABLE
+}
+#ifdef __clang__
+#pragma clang diagnostic pop
+#endif // __clang__
+
+template <int D, int cols_per_block, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
+void ggml_cuda_flash_attn_ext_vec_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    const int nthreads = ggml_cuda_fattn_vec_get_nthreads_host(cc);
+    const int nwarps   = nthreads / WARP_SIZE;
+    fattn_kernel_t fattn_kernel = flash_attn_ext_vec<D, cols_per_block, type_K, type_V, use_logit_softcap>;
+    constexpr bool need_f16_K = false;
+    constexpr bool need_f16_V = false;
+    constexpr size_t nbytes_shared = 0;
+    launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
+}
+
+template <int D, ggml_type type_K, ggml_type type_V>
+void ggml_cuda_flash_attn_ext_vec_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+    const ggml_tensor * K   = dst->src[1];
+    const ggml_tensor * V   = dst->src[2];
+
+    GGML_ASSERT(K->type == type_K);
+    GGML_ASSERT(V->type == type_V);
+
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    if (Q->ne[1] == 1) {
+        constexpr int cols_per_block = 1;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    constexpr int cols_per_block = 2;
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        ggml_cuda_flash_attn_ext_vec_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        ggml_cuda_flash_attn_ext_vec_case_impl<D, cols_per_block, type_K, type_V, use_logit_softcap>(ctx, dst);
+    }
+}
+
+#define DECL_FATTN_VEC_CASE(D, type_K, type_V)                              \
+    template void ggml_cuda_flash_attn_ext_vec_case                         \
+    <D, type_K, type_V>(ggml_backend_cuda_context & ctx, ggml_tensor * dst) \
+
+#define EXTERN_DECL_FATTN_VEC_CASES(D, type_K)             \
+    extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_F16);  \
+    extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q4_0); \
+    extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q4_1); \
+    extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q5_0); \
+    extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q5_1); \
+    extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q8_0); \
+
+EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_F16)
+EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q4_0)
+EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q4_1)
+EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q5_0)
+EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q5_1)
+EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q8_0)
+
+EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_F16)
+EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q4_0)
+EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q4_1)
+EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q5_0)
+EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q5_1)
+EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q8_0)
+
+EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_F16)
+EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q4_0)
+EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q4_1)
+EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q5_0)
+EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q5_1)
+EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q8_0)
@@ -2,8 +2,7 @@
 #include "fattn-common.cuh"
 #include "fattn-mma-f16.cuh"
 #include "fattn-tile.cuh"
-#include "fattn-vec-f16.cuh"
-#include "fattn-vec-f32.cuh"
+#include "fattn-vec.cuh"
 #include "fattn-wmma-f16.cuh"
 #include "fattn.cuh"

@@ -117,151 +116,68 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
    }
 }

-#define FATTN_VEC_F16_CASE(D, type_K, type_V)                               \
-    if (Q->ne[0] == (D) && K->type == (type_K) && V->type == (type_V)) {    \
-        ggml_cuda_flash_attn_ext_vec_f16_case<D, type_K, type_V>(ctx, dst); \
-        return;                                                             \
-    }                                                                       \
+#define FATTN_VEC_CASE(D, type_K, type_V)                                \
+    if (Q->ne[0] == (D) && K->type == (type_K) && V->type == (type_V)) { \
+        ggml_cuda_flash_attn_ext_vec_case<D, type_K, type_V>(ctx, dst);  \
+        return;                                                          \
+    }                                                                    \

-static void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+#define FATTN_VEC_CASES_ALL_D(type_K, type_V) \
+    FATTN_VEC_CASE( 64, type_K, type_V)       \
+    FATTN_VEC_CASE(128, type_K, type_V)       \
+    FATTN_VEC_CASE(256, type_K, type_V)       \
+
+static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    ggml_tensor * Q = dst->src[0];
    ggml_tensor * K = dst->src[1];
    ggml_tensor * V = dst->src[2];

 #ifdef GGML_CUDA_FA_ALL_QUANTS
-    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0)
-    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1)
-    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0)
-    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1)
-    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0)
-    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16 )
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_F16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_F16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_F16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_F16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_F16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_F16)

-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q4_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q4_0)

-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q4_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q4_1)

-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q5_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q5_0)

-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q5_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q5_1)

-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0)
-
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16)
-
-    FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_Q8_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
 #else
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
-
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
-
-    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16)
-    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16)
-    FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
-#endif // GGML_CUDA_FA_ALL_QUANTS
-
-    GGML_ABORT("fatal error");
-}
-
-#define FATTN_VEC_F32_CASE(D, type_K, type_V)                               \
-    if (Q->ne[0] == (D) && K->type == (type_K) && V->type == (type_V)) {    \
-        ggml_cuda_flash_attn_ext_vec_f32_case<D, type_K, type_V>(ctx, dst); \
-        return;                                                             \
-    }                                                                       \
-
-static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * Q = dst->src[0];
-    ggml_tensor * K = dst->src[1];
-    ggml_tensor * V = dst->src[2];
-
-#ifdef GGML_CUDA_FA_ALL_QUANTS
-    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0)
-    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1)
-    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0)
-    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1)
-    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0)
-    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16)
-
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0)
-
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1)
-
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0)
-
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1)
-
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0)
-
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16)
-
-    FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
-#else
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
-
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
-
-    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16)
-    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16)
-    FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16,  GGML_TYPE_F16)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
 #endif // GGML_CUDA_FA_ALL_QUANTS

    GGML_ABORT("fatal error");
@@ -271,8 +187,7 @@ static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, gg
 enum best_fattn_kernel {
    BEST_FATTN_KERNEL_NONE     =   0,
    BEST_FATTN_KERNEL_TILE     = 200,
-    BEST_FATTN_KERNEL_VEC_F32  = 100,
-    BEST_FATTN_KERNEL_VEC_F16  = 110,
+    BEST_FATTN_KERNEL_VEC      = 100,
    BEST_FATTN_KERNEL_WMMA_F16 = 300,
    BEST_FATTN_KERNEL_MMA_F16  = 400,
 };
@@ -283,7 +198,6 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
    return BEST_FATTN_KERNEL_NONE;
 #endif// FLASH_ATTN_AVAILABLE

-    const ggml_tensor * KQV   = dst;
    const ggml_tensor * Q     = dst->src[0];
    const ggml_tensor * K     = dst->src[1];
    const ggml_tensor * V     = dst->src[2];
@@ -293,8 +207,6 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
    GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);

    const int cc = ggml_cuda_info().devices[device].cc;
-    const int warp_size = ggml_cuda_info().devices[device].warp_size;
-    const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);

    switch (K->ne[0]) {
        case  64:
@@ -343,31 +255,6 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
 #endif // GGML_CUDA_FA_ALL_QUANTS
        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q8_0:
-#ifdef GGML_CUDA_FA_ALL_QUANTS
-            if (K->ne[0] != 128 && K->ne[0] != 64) {
-                return BEST_FATTN_KERNEL_NONE;
-            }
-#else
-            if (K->ne[0] != 128) {
-                return BEST_FATTN_KERNEL_NONE;
-            }
-#endif // GGML_CUDA_FA_ALL_QUANTS
-            break;
-        default:
-            return BEST_FATTN_KERNEL_NONE;
-    }
-
-    switch (V->type) {
-        case GGML_TYPE_F16:
-            break;
-        case GGML_TYPE_Q4_1:
-        case GGML_TYPE_Q5_0:
-        case GGML_TYPE_Q5_1:
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q8_0:
-            if (K->ne[0] != 128) {
-                return BEST_FATTN_KERNEL_NONE;
-            }
            break;
        default:
            return BEST_FATTN_KERNEL_NONE;
@@ -377,30 +264,39 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
        return BEST_FATTN_KERNEL_NONE;
    }

-    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0;
+    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0;

    // If Turing tensor cores available, use them except for some cases with batch size 1:
    if (turing_mma_available(cc)) {
-        const bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask; // The mma-based kernels have GQA-specific optimizations
-        const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
-        const bool mma_faster_for_rtx4000 = Q->ne[3] > 1 || (gqa_ratio > 4 && K->ne[1] >= 8192);
-        const bool mma_faster_for_bs1 = gqa_opt_applies && !mma_needs_data_conversion &&
-            (cc < GGML_CUDA_CC_ADA_LOVELACE || mma_faster_for_rtx4000);
-        if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
-            if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
-                return BEST_FATTN_KERNEL_VEC_F16;
+        best_fattn_kernel best = BEST_FATTN_KERNEL_MMA_F16;
+
+        if (can_use_vector_kernel) {
+            if (K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16) {
+                if (cc >= GGML_CUDA_CC_ADA_LOVELACE && Q->ne[1] == 1 && Q->ne[3] == 1 && !(gqa_ratio > 4 && K->ne[1] >= 8192)) {
+                    best = BEST_FATTN_KERNEL_VEC;
+                }
+            } else {
+                if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+                    if (Q->ne[1] <= 2) {
+                        best = BEST_FATTN_KERNEL_VEC;
+                    }
+                } else {
+                    if (Q->ne[1] == 1) {
+                        best = BEST_FATTN_KERNEL_VEC;
+                    }
+                }
+            }
+            if ((gqa_ratio % 2 != 0 || !mask) && Q->ne[1] == 1) {
+                best = BEST_FATTN_KERNEL_VEC; // GQA-specific optimizations in the mma kernel do not apply.
            }
-            return BEST_FATTN_KERNEL_VEC_F32;
        }
-        return BEST_FATTN_KERNEL_MMA_F16;
+
+        return best;
    }

-    // Use kernels specializes for small batch sizes if possible:
+    // Use kernels specialized for small batch sizes if possible:
    if (Q->ne[1] <= 8 && can_use_vector_kernel) {
-        if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
-            return BEST_FATTN_KERNEL_VEC_F16;
-        }
-        return BEST_FATTN_KERNEL_VEC_F32;
+        return BEST_FATTN_KERNEL_VEC;
    }

    // For large batch sizes, use the WMMA kernel if possible:
@@ -420,11 +316,8 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
        case BEST_FATTN_KERNEL_TILE:
            ggml_cuda_flash_attn_ext_tile(ctx, dst);
            break;
-        case BEST_FATTN_KERNEL_VEC_F32:
-            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
-            break;
-        case BEST_FATTN_KERNEL_VEC_F16:
-            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+        case BEST_FATTN_KERNEL_VEC:
+            ggml_cuda_flash_attn_ext_vec(ctx, dst);
            break;
        case BEST_FATTN_KERNEL_WMMA_F16:
            ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
@@ -45,6 +45,7 @@
 #include "ggml-cuda/sumrows.cuh"
 #include "ggml-cuda/mean.cuh"
 #include "ggml-cuda/tsembd.cuh"
+#include "ggml-cuda/topk-moe.cuh"
 #include "ggml-cuda/unary.cuh"
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv.cuh"
@@ -2030,7 +2031,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
            const int cc            = ggml_cuda_info().devices[id].cc;
            const int warp_size     = ggml_cuda_info().devices[id].warp_size;
            use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1]);
+            use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1], /*mul_mat_id=*/false);
            use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
            any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
        }
@@ -2038,7 +2039,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
        const int cc            = ggml_cuda_info().devices[ctx.device].cc;
        const int warp_size     = ggml_cuda_info().devices[ctx.device].warp_size;
        use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1]);
+        use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1], /*mul_mat_id=*/false);
        use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
        any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
    }
@@ -2110,7 +2111,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
            return;
        }

-        if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src1->ne[2])) {
+        if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src1->ne[2], /*mul_mat_id=*/true)) {
            ggml_cuda_mul_mat_f(ctx, src0, src1, ids, dst);
            return;
        }
@@ -2825,6 +2826,44 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
    GGML_ASSERT(unary_ops.size() == num_unary);
 #endif

+    //TODO: remove special case once ggml_can_fuse can handle empty nodes
+    std::initializer_list<enum ggml_op> topk_moe_ops           = ggml_cuda_topk_moe_ops(false);
+    std::initializer_list<enum ggml_op> topk_moe_ops_with_norm = ggml_cuda_topk_moe_ops(true);
+
+    if (ops.size() == topk_moe_ops_with_norm.size() && std::equal(ops.begin(), ops.end(), topk_moe_ops_with_norm.begin())) {
+
+        if (node_idx + topk_moe_ops_with_norm.size() > (size_t)cgraph->n_nodes) {
+            return false;
+        }
+
+        for (size_t i = 0; i < topk_moe_ops_with_norm.size(); i++) {
+            if (cgraph->nodes[node_idx + i]->op != topk_moe_ops_with_norm.begin()[i]) return false;
+        }
+        ggml_tensor * softmax = cgraph->nodes[node_idx];
+        ggml_tensor * weights = cgraph->nodes[node_idx+8];
+
+        if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
+            return true;
+        }
+    }
+
+    if (ops.size() == topk_moe_ops.size() && std::equal(ops.begin(), ops.end(), topk_moe_ops.begin())) {
+
+        if (node_idx + topk_moe_ops.size() > (size_t)cgraph->n_nodes) {
+            return false;
+        }
+
+        for (size_t i = 0; i < topk_moe_ops.size(); i++) {
+            if (cgraph->nodes[node_idx + i]->op != topk_moe_ops.begin()[i]) return false;
+        }
+
+        ggml_tensor * softmax = cgraph->nodes[node_idx];
+        ggml_tensor * weights = cgraph->nodes[node_idx+4];
+        if (ggml_cuda_should_use_topk_moe(softmax, weights)) {
+            return true;
+        }
+    }
+
    if (!ggml_can_fuse(cgraph, node_idx, ops)) {
        return false;
    }
@@ -2915,6 +2954,22 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
                if (!disable_fusion) {

+                    if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ true), {})) {
+                        ggml_tensor * weights = cgraph->nodes[i+8];
+                        ggml_tensor * selected_experts = cgraph->nodes[i+3];
+                        ggml_cuda_op_topk_moe(*cuda_ctx, node, weights, selected_experts, /*with norm*/ true);
+                        i += 8;
+                        continue;
+                    }
+
+                    if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ false), {})) {
+                        ggml_tensor * weights = cgraph->nodes[i+4];
+                        ggml_tensor * selected_experts = cgraph->nodes[i+3];
+                        ggml_cuda_op_topk_moe(*cuda_ctx, node, weights, selected_experts, /*with norm*/ false);
+                        i += 4;
+                        continue;
+                    }
+
                    if (node->op == GGML_OP_ADD) {
                        int n_fuse = 0;
                        ggml_op ops[8];
@@ -84,7 +84,7 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
    }
 }

-bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne, const int src1_ncols) {
+bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne, const int src1_ncols, bool mul_mat_id) {

    if (ggml_is_quantized(type)) {
        return false;
@@ -96,8 +96,18 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
    if (src0_ne[1] % MMF_ROWS_PER_BLOCK != 0) {
        return false;
    }
-    if (src1_ncols > 16) {
-        return false;
+
+    if (mul_mat_id) {
+        if (type == GGML_TYPE_F32 && src1_ncols > 32) {
+            return false;
+        }
+        if ((type == GGML_TYPE_F16 || type == GGML_TYPE_BF16) && src1_ncols > 64) {
+            return false;
+        }
+    } else {
+        if (src1_ncols > 16) {
+            return false;
+        }
    }

    switch (type) {
@@ -9,13 +9,13 @@ using namespace ggml_cuda_mma;

 void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);

-bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, const int src1_ncols);
+bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, const int src1_ncols, bool mul_mat_id);

 template <typename T, int rows_per_block, int cols_per_block, int nwarps, bool has_ids>
 __launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
 static __global__ void mul_mat_f(
        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
-        const int ncols, const int nchannels_dst, const int stride_row, const int stride_col_y, const int stride_col_dst,
+        const int ncols, const int ncols_dst_total, const int nchannels_dst, const int stride_row, const int stride_col_y, const int stride_col_dst,
        const int stride_col_id, const int stride_row_id,
        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
@@ -31,9 +31,20 @@ static __global__ void mul_mat_f(

    const int row0        = blockIdx.x * rows_per_block;

-    const int expert_idx  = has_ids ? blockIdx.y : 0;
+    int expert_idx = 0;
+    int col_base = 0;
+
    const int channel_dst = has_ids ? 0 : blockIdx.y;

+    if constexpr (has_ids) {
+        // experts + tiles of ncols_dst are packed in the y dimension
+        int col_tiles = (ncols_dst_total + cols_per_block - 1) / cols_per_block;
+        const int nchannels_x = gridDim.y / col_tiles;
+        const int tile_idx = blockIdx.y / nchannels_x;
+        expert_idx = blockIdx.y - tile_idx * nchannels_x;
+        col_base = tile_idx * cols_per_block;
+    }
+
    const int channel_x   = has_ids ? expert_idx : (channel_dst / channel_ratio);
    const int channel_y   = channel_dst;
    const int sample_dst  = blockIdx.z;
@@ -44,6 +55,14 @@ static __global__ void mul_mat_f(
    y   += int64_t(sample_y)  *stride_sample_y   + (has_ids ? 0 : channel_y  *stride_channel_y);
    dst += int64_t(sample_dst)*stride_sample_dst + (has_ids ? 0 : channel_dst*stride_channel_dst);

+    if constexpr (has_ids) {
+        constexpr int y_stride_scale = std::is_same_v<T, float> ? 1 : 2;
+        const int64_t col_offset = col_base;
+        y   += col_offset * stride_col_y * y_stride_scale;
+        dst += col_offset * stride_col_dst;
+        ids += col_offset * stride_row_id;
+    }
+
    const float2 * y2 = (const float2 *) y;

    extern __shared__ char data_mmv[];
@@ -61,12 +80,17 @@ static __global__ void mul_mat_f(

        for (int j0 = 0; j0 < cols_per_block; j0 += nwarps) {
            const int j = j0 + threadIdx.y;
-            const int32_t * __restrict__ id_row = ids + j*stride_row_id;

            if (threadIdx.x == 0) {
                slot_map[j] = -1;
            }

+            if (col_base + j >= ncols_dst_total) {
+                continue;
+            }
+
+            const int32_t * __restrict__ id_row = ids + j*stride_row_id;
+
            for (int k = threadIdx.x; k < nchannels_dst; k += warp_size) {
                int match = id_row[k*stride_col_id] == expert_idx;

@@ -108,7 +132,8 @@ static __global__ void mul_mat_f(
                    if constexpr (!has_ids) {
                        tile_xy[j0*tile_k_padded + threadIdx.x] = j < cols_per_block ? y[j*stride_col_y + col] : 0.0f;
                    } else {
-                        tile_xy[j0*tile_k_padded + threadIdx.x] = j < cols_per_block ? y[slot_map[j]*stride_channel_y + j*stride_col_y + col] : 0.0f;
+                        const bool valid = j < cols_per_block && (col_base + j) < ncols_dst_total && slot_map[j] >= 0;
+                        tile_xy[j0*tile_k_padded + threadIdx.x] = valid ? y[slot_map[j]*stride_channel_y + j*stride_col_y + col] : 0.0f;
                    }
                }
            } else if constexpr (std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) {
@@ -120,7 +145,8 @@ static __global__ void mul_mat_f(
                        const float2 tmp = j < cols_per_block ? y2[j*stride_col_y + col] : make_float2(0.0f, 0.0f);
                        tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
                    } else {
-                        float2 tmp = j < cols_per_block && slot_map[j] >= 0 ? *(const float2*) &y[slot_map[j]*stride_channel_y + 2*(j*stride_col_y + col)] : make_float2(0.0f, 0.0f);
+                        const bool valid = j < cols_per_block && (col_base + j) < ncols_dst_total && slot_map[j] >= 0;
+                        float2 tmp = valid ? *(const float2*) &y[slot_map[j]*stride_channel_y + 2*(j*stride_col_y + col)] : make_float2(0.0f, 0.0f);
                        tile_xy[j0*tile_k_padded + threadIdx.x] = {tmp.x, tmp.y};
                    }
                }
@@ -183,14 +209,14 @@ static __global__ void mul_mat_f(
            dst[j*stride_col_dst + row0 + threadIdx.x] = sum;
        } else {
            const int slot = (j < cols_per_block) ? slot_map[j] : -1;
-            if (slot >= 0) {
+            if (slot >= 0 && (col_base + j) < ncols_dst_total) {
                dst[slot*stride_channel_dst + j*stride_col_dst + row0 + threadIdx.x] = sum;
            }
        }
    }
 #else
    GGML_UNUSED_VARS(x, y, ids, dst,
-        ncols, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+        ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
        stride_col_id, stride_row_id,
        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
@@ -201,20 +227,23 @@ static __global__ void mul_mat_f(
 template<typename T, int cols_per_block, int nwarps>
 static inline void mul_mat_f_switch_ids(
        const T * x, const float * y, const int32_t * ids, float * dst,
-        const int64_t ncols_x, const int64_t nchannels_dst,
+        const int64_t ncols_x, const int64_t ncols_dst, const int64_t nchannels_dst,
        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
        const int64_t stride_col_id, const int64_t stride_row_id,
        const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
        const int64_t sample_ratio, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared_total, cudaStream_t stream) {
    if (ids) {
-        mul_mat_f<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps, true><<<block_nums, block_dims, nbytes_shared_total, stream>>>
-             (x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+        const int64_t col_tiles = (ncols_dst + cols_per_block - 1) / cols_per_block;
+        dim3 block_nums_ids = block_nums;
+        block_nums_ids.y *= col_tiles;
+        mul_mat_f<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps, true><<<block_nums_ids, block_dims, nbytes_shared_total, stream>>>
+             (x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
             stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
             sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
    } else {
        mul_mat_f<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps, false><<<block_nums, block_dims, nbytes_shared_total, stream>>>
-            (x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+            (x, y, ids, dst, ncols_x, cols_per_block, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
             stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
             sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
    }
@@ -223,7 +252,8 @@ static inline void mul_mat_f_switch_ids(
 template <typename T, int cols_per_block>
 void mul_mat_f_cuda(
        const T * x, const float * y, const int32_t * ids, float * dst,
-        const int64_t ncols_x, const int64_t nrows_x, const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const int64_t ncols_x, const int64_t nrows_x, const int64_t ncols_dst,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
        const int64_t stride_col_id, const int64_t stride_row_id,
        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
@@ -268,49 +298,49 @@ void mul_mat_f_cuda(
    switch (nwarps_best) {
        case 1: {
            mul_mat_f_switch_ids<T, cols_per_block, 1>(
-                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
        } break;
        case 2: {
            mul_mat_f_switch_ids<T, cols_per_block, 2>(
-                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
        } break;
        case 3: {
            mul_mat_f_switch_ids<T, cols_per_block, 3>(
-                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
        } break;
        case 4: {
            mul_mat_f_switch_ids<T, cols_per_block, 4>(
-                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
        } break;
        case 5: {
            mul_mat_f_switch_ids<T, cols_per_block, 5>(
-                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
        } break;
        case 6: {
            mul_mat_f_switch_ids<T, cols_per_block, 6>(
-                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
        } break;
        case 7: {
            mul_mat_f_switch_ids<T, cols_per_block, 7>(
-                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
        } break;
        case 8: {
            mul_mat_f_switch_ids<T, cols_per_block, 8>(
-                x, y, ids, dst, ncols_x, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
+                x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream);
        } break;
@@ -332,84 +362,89 @@ static void mul_mat_f_switch_cols_per_block(
        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
        cudaStream_t stream) {
-    switch (ncols_dst) {
+
+    const int ncols_case = (ids && ncols_dst > 16) ? 16 : ncols_dst;
+
+    GGML_ASSERT(ids || ncols_dst <= 16);
+
+    switch (ncols_case) {
        case  1: {
-            mul_mat_f_cuda<T,  1>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  1>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case  2: {
-            mul_mat_f_cuda<T,  2>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  2>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case  3: {
-            mul_mat_f_cuda<T,  3>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  3>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case  4: {
-            mul_mat_f_cuda<T,  4>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  4>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case  5: {
-            mul_mat_f_cuda<T,  5>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  5>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y,  stride_sample_dst, stream);
        } break;
        case  6: {
-            mul_mat_f_cuda<T,  6>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  6>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case  7: {
-            mul_mat_f_cuda<T,  7>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  7>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case  8: {
-            mul_mat_f_cuda<T,  8>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  8>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case  9: {
-            mul_mat_f_cuda<T,  9>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  9>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case 10: {
-            mul_mat_f_cuda<T, 10>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, 10>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case 11: {
-            mul_mat_f_cuda<T, 11>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, 11>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case 12: {
-            mul_mat_f_cuda<T, 12>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, 12>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case 13: {
-            mul_mat_f_cuda<T, 13>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, 13>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case 14: {
-            mul_mat_f_cuda<T, 14>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, 14>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case 15: {
-            mul_mat_f_cuda<T, 15>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, 15>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
        case 16: {
-            mul_mat_f_cuda<T, 16>(x, y, ids, dst, ncols_x, nrows_x, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, 16>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
        } break;
@@ -422,7 +457,7 @@ static void mul_mat_f_switch_cols_per_block(
 #define DECL_MMF_CASE_HELPER(T, ncols_dst) \
    template void mul_mat_f_cuda<T, ncols_dst>( \
        const T * x, const float * y, const int32_t * ids, float * dst, \
-        const int64_t ncols_x, const int64_t nrows_x, const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst, \
+        const int64_t ncols_x, const int64_t nrows_x, int64_t ncols_dst_total, const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst, \
        const int64_t stride_col_id, const int64_t stride_row_id, \
        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst, \
        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,\
@@ -81,7 +81,7 @@ static __global__ void mmq_ids_helper(
 #pragma unroll
            for (int offset = neu_padded; offset < warp_size; offset += neu_padded) {
                const int tmp = __shfl_up_sync(0xFFFFFFFF, it_compact_add_self, offset, warp_size);
-                if (threadIdx.x >= offset) {
+                if (threadIdx.x >= static_cast<unsigned int>(offset)) {
                    it_compact_add_lower += tmp;
                }
            }
@@ -110,7 +110,7 @@ static __global__ void mmq_ids_helper(

    expert_bounds[expert] = nex_prev;

-    if (expert < gridDim.x - 1) {
+    if (expert < static_cast<int>(gridDim.x) - 1) {
        return;
    }

@@ -220,7 +220,7 @@ static __global__ void mul_mat_vec_q(
            tmp[j][i] = warp_reduce_sum<warp_size>(tmp[j][i]);
        }

-        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + int(threadIdx.x) < stride_col_dst)) {
+        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
            dst[j*stride_col_dst + threadIdx.x] = tmp[j][threadIdx.x];
        }
    }
@@ -51,6 +51,8 @@ static __global__ __launch_bounds__(CUDA_PAD_REFLECT_1D_BLOCK_SIZE, 1) void
    }
    const float value               = *(const float *) (src0_ptr + src_idx * nb00);
    *(float *) (dst_ptr + i0 * nb0) = value;
+
+    GGML_UNUSED(p1);
 }

 void ggml_cuda_op_pad_reflect_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q5_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q5_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q8_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(64, GGML_TYPE_F16, GGML_TYPE_F16);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q4_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q4_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q5_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q5_1);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f16.cuh"
-
-DECL_FATTN_VEC_F16_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q8_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f32.cuh"
-
-DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f32.cuh"
-
-DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_0);
@@ -1,5 +0,0 @@
-// This file has been autogenerated by generate_cu_files.py, do not edit manually.
-
-#include "../fattn-vec-f32.cuh"
-
-DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_1);
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Acly	8656f5de68	vulkan : make the vulkan.hpp dynamic dispatcher instance private (#16224 ) * don't use VULKAN_HPP_DEFAULT_DISPATCH_LOADER_DYNAMIC_STORAGE which can cause conflicts if application or other libraries do the same	2025-09-27 22:41:03 +02:00
Aleksander Grygier	4807e8f96a	Show message actions by default (#16289 )	2025-09-27 19:56:40 +02:00
Aman Gupta	c0bfc57af4	CUDA: mul_mat_id for mmf for bs <= 64 for f16 and bs <= 32 for f32 (#16277 ) * CUDA: mul_mat_id for mmf for bs <= 64 for f16 and bs <= 32 for f32 This commit adds mul_mat_id support for ncols_dst >= 16. It does this by packing ncols_dst tiles into the blockDim.y. My tests on a RTX 3090 show that this is faster than the cuBLAS fallback for f16 till bs=64, and for f32 till bs=32 * Review: refactor if statement	2025-09-27 18:49:32 +02:00
Johannes Gäßler	75a3a6c2cd	CUDA: refactor and deduplicate vector FA kernels (#16208 ) * CUDA: refactor and deduplicate vector FA kernels	2025-09-27 18:45:07 +02:00
Dmytro Minochkin	0499b29c6f	vulkan: throw system error instead of SIGABRT during init on older devices (#16156 ) * Throw system error on old Vulkan driver rather than SIGABRT * Optionally handle any potential error in vulkan init	2025-09-27 18:26:46 +02:00
Adrien Gallouët	234e2ff8ed	server : remove old LLAMA_SERVER_SSL (#16290 ) Signed-off-by: Adrien Gallouët <angt@huggingface.co>	2025-09-27 19:17:08 +03:00
Jeff Bolz	3f81b4e91c	vulkan: support GET_ROWS for k-quants (#16235 ) The dequantize functions are copy/pasted from mul_mm_funcs.comp with very few changes - add a_offset and divide iqs by 2. It's probably possible to call these functions from mul_mm_funcs and avoid the duplication, but I didn't go that far in this change.	2025-09-27 12:36:11 +02:00
Adrien Gallouët	ace6a54565	build : add LLAMA_OPENSSL option (#16287 ) Introduce a new `LLAMA_OPENSSL` option, enabled by default. This preserves the previous default (libcurl first, OpenSSL as fallback), while allowing OpenSSL to be disabled if desired. Signed-off-by: Adrien Gallouët <angt@huggingface.co>	2025-09-27 12:12:46 +03:00
Vinkal	72b24d96c6	model : make minicpm embedding_scale, residual_scale and logit_scale optional with legacy defaults (#16273 ) * minicpm: make GGUF scaling keys optional with legacy defaults Older MiniCPM GGUFs do not include the scaling metadata keys (minicpm.embedding_scale, minicpm.residual_scale, minicpm.logit_scale). The loader currently treats these as required, so quantization fails with: key not found in model: minicpm.embedding_scale This change restores backward compatibility by treating these keys as optional in the loader and using the older MiniCPM scaling values: embedding_scale = 12.0f residual_scale = 1.4f / sqrt(n_layer) logit_scale = 256.0f / n_embd When the GGUF provides the keys, their values override the defaults; otherwise the legacy defaults are used. Newer GGUFs that already include these keys are unaffected. Fixes: #16192 Signed-off-by: Vinkal Chudgar <vinkal.chudgar@gmail.com> * Update src/llama-model.cpp Committed as suggested. Thanks! Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> --------- Signed-off-by: Vinkal Chudgar <vinkal.chudgar@gmail.com> Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>	2025-09-26 23:28:29 +02:00
Aaron Teo	624207e676	devops: add s390x & ppc64le CI (#15925 ) * devops: move s390x and ppc64le ci build we have access to ubuntu-24.04-s390x and ppc64le images now Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: disable ppc64le for now since they have compiler errors Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: stop warnings as errors Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: switch to non-macro flag Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: going the llama macro route Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: add big-endian gguf test models Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: disable ppc64le to test s390x, check test build Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: dup .gguf.inp files for big-endian tests Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: dup .gguf.out files for big-endian too Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: add python setup and endian byteswap Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: pooring thing does not have s390x python3 Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: add missing rust compiler for s390x Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: try rust actions runner Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Revert "devops: try rust actions runner" This reverts commit 3f8db04356033d6c1d7eccc75ca396bc5298250c. Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: try a different path for rust Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: dump home directory and user info Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: install gguf-py only Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: missed relative path Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: remove big-endian files since local swapping is working Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: revert test-tokenizer-0 cmakelists Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Fix unicode flags conversion from and to uint16_t Bitfields are allocated in different order on s390x Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Simplify byteswap command Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Add byteswapping and git-lfs for test-tokenizers-ggml-vocabs Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Fix endianness detection in vocab loader Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Disable test-thread-safety on s390x In this test a model is downloaded, then immediately loaded to check if more downloads are needed, and then used for test. There is no clean way to separate all those steps to add byteswapping between them, so just skip this test. Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Fix q8_0 test in test-quantize-fns vec_signed uses unexpected rounding mode. Explicitly use different rounding function. Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: add big-endian stories260K Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: add s390x test-eval-callback Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: fix test does not exist Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: fix model not found llama-eval-callback Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Fix q3_K dot product error in test-quantize-fns on s390x Array q8bytes had only 4 elements allocated, but 8 elements accessed. This lead to write out of bounds and later read of overwritten values out of bounds and incorrect result. Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: re-enable ppc64le for testing Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: activate test-thread-safety for s390x Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: disable ppc64le tests for some reason it keeps failing test-thread-safety tests and I do not have a machine that is able to replicate the tests. Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * devops: LLAMA_FATAL_WARNINGS=ON Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Correct repository URL for s390x for test-thread-safety model Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Fix fs_get_cache_directory Ensure it works even if both XDG_CACHE_HOME and HOME are unset. This might happen in containers. Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Re-enable CI for ppc64le Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Fortify ggml_rope_impl Only memcpy data from sections argument if it's non-NULL. Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Add TODO in struct unicode_cpt_flags to reimplement it in endian-independent way * Update URL for big-endian model * Update .github/workflows/build.yml Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update remaining mentions of BE models to ggml-org/models repo --------- Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> Co-authored-by: Aleksei Nikiforov <aleksei.nikiforov@linux.ibm.com> Co-authored-by: Aleksei Nikiforov <103434461+AlekseiNikiforovIBM@users.noreply.github.com> Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>	2025-09-27 02:03:33 +08:00
Aleksander Grygier	807e8c6d31	Enhance text file detection logic for file attachments (#16199 ) * feat: Enhances text file detection logic * chore: Build static `webui` output * chore: update webui build output	2025-09-26 19:25:29 +02:00
Aleksander Grygier	1a18927894	Allow viewing conversations even when llama server is down (#16255 ) * webui: allow viewing conversations and sending messages even if llama-server is down - Cached llama.cpp server properties in browser localStorage on startup, persisting successful fetches and reloading them when refresh attempts fail so the chat UI continues to render while the backend is unavailable. - Cleared the stored server properties when resetting the store to prevent stale capability data after cache-backed operation. - Kept the original error-splash behavior when no cached props exist so fresh installs still surface a clear failure state instead of rendering stale data. * feat: Add UI for `props` endpoint unavailable + cleanup logic * webui: extend cached props fallback to offline errors Treat connection failures (refused, DNS, timeout, fetch) the same way as server 5xx so the warning banner shows up when cache is available, instead of falling back to a full error screen. * webui: Left the chat form enabled when a server warning is present so operators can keep sending messages e.g., to restart the backend over llama-swap, even while cached /props data is in use * chore: update webui build output --------- Co-authored-by: Pascal <admin@serveurperso.com>	2025-09-26 18:35:42 +02:00
Isaac McFadyen	e0539eb6ae	webui: switch to hash-based routing (alternative of #16079 ) (#16157 ) * Switched web UI to hash-based routing * Added hash to missed goto function call * Removed outdated SPA handling code * Fixed broken sidebar home link	2025-09-26 18:36:48 +03:00
Aleksander Grygier	5d0a40f390	Always show message actions for mobile UI + improvements for user message sizing (#16076 )	2025-09-26 15:59:07 +02:00
Radoslav Gerganov	d12a983659	codeowners : add rgerganov as owner of RPC [no ci] (#16279 )	2025-09-26 16:09:34 +03:00
Aleksei Nikiforov	cc1cfa277b	mtmd : fix uninitialized variable in bicubic_resize (#16275 ) Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> Co-authored-by: Aaron Teo <aaron.teo1@ibm.com>	2025-09-26 15:00:44 +02:00
Georgi Gerganov	54dbc37053	metal : report OOM errors (#16274 )	2025-09-26 14:14:28 +03:00
Adrien Gallouët	b995a10760	common : use cpp-httplib as a cURL alternative for downloads (#16185 ) * vendor : update httplib Signed-off-by: Adrien Gallouët <angt@huggingface.co> * common : use cpp-httplib as a cURL alternative for downloads The existing cURL implementation is intentionally left untouched to prevent any regressions and to allow for safe, side-by-side testing by toggling the `LLAMA_CURL` CMake option. Signed-off-by: Adrien Gallouët <angt@huggingface.co> * ggml : Bump to Windows 10 Signed-off-by: Adrien Gallouët <angt@huggingface.co> --------- Signed-off-by: Adrien Gallouët <angt@huggingface.co>	2025-09-26 14:12:19 +03:00
Adrien Gallouët	4710dd31bb	build : fix build-ios-device (#16257 ) Signed-off-by: Adrien Gallouët <angt@huggingface.co>	2025-09-26 13:39:35 +03:00
Aaron Teo	9b26511857	ggml-cpu: implement MXFP4 SIMD for s390x (#16193 ) * ggml-cpu: impl mxfp4 s390x Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-cpu: missing s = sumf Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-cpu: fix incorrect kval_mxfp4 type Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-cpu: rework mxfp4 Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-cpu: missing delta calc Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-cpu: fix typo Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-cpu: fix typo for vec_splats Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-cpu: expand to 2 blocks per loop Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-cpu: add unroll to boost perf Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-cpu: back to 1 block per loop to test perf Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * Revert "ggml-cpu: back to 1 block per loop to test perf" This reverts commit 1fe55724e2dc295701101bf838bdd4a512237492. Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * ggml-cpu: rm unroll from single block Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> --------- Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>	2025-09-26 13:27:25 +03:00
Radoslav Gerganov	00217cd413	ci : create git tags for released docker images (#16008 ) * ci : create git tags for released docker images When releasing a docker image for build number X, we should also create the corresponding git tag. This allows users to easily checkout the corresponding source tree for given docker image. * Update .github/workflows/docker.yml Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Update .github/workflows/docker.yml Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> * Apply suggestion from @CISC Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com> --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>	2025-09-26 10:19:23 +00:00
Daniel Bevenius	3b337b01a1	codeowners : add danbev as owner of build-xcframework.sh [no ci] (#16268 )	2025-09-26 08:53:36 +03:00
R0CKSTAR	a86a580a66	musa: upgrade musa sdk to 4.3.0 (#16240 ) Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>	2025-09-26 02:56:38 +02:00
R0CKSTAR	0f7c69689f	musa: fix build warnings (#15611 ) Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>	2025-09-26 02:56:10 +02:00
Sigbjørn Skjæret	835b2b915c	model : add GroveMoE support (#15510 ) * add GroveMoE support * remove constexpr that fails on certain compilers * revert crude scalar div implementation, use cast * build_attn_inp_kv_unified -> build_attn_inp_kv * fix build_attn * re-apply ffn_exps regex changes	2025-09-25 19:50:28 +02:00
Aaron Teo	b05a9d650f	vendors: update miniaudio version (#16212 ) * vendor: update miniaudio.h Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> * vendor: update miniaudio.h Signed-off-by: Aaron Teo <aaron.teo1@ibm.com> --------- Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>	2025-09-25 23:38:10 +08:00
rtaluyev	27052978e4	readme : update bindings (#16144 ) Link to Java JNA bindings to llama.cpp native libraries	2025-09-25 18:20:34 +03:00
Aman Gupta	077c94d0ca	CUDA: add a fused top-K MoE kernel (#16130 ) * CUDA: add a fused top-K MoE kernel This kernel does the following: 1. softmax over the logits per token [n_experts, n_tokens] 2. argmax reduce over the top-k (n_experts_used) logits 3. write weights + ids to global memory It is intended as fusion of softmax->top-k->get_rows pipeline for MoE models * Refactor into ggml_cuda_should_use_topk_moe * Review: Use better coalescing pattern, use WARP_SIZE, store logits into registers before * Review: format + micro-optimizations * Fix bug: fix tie breakers * Add optional norm + clean-up code * Use smem for final write * Add bounds check * Use better memory pattern for writeback	2025-09-25 16:35:05 +02:00
Daniel Bevenius	aa3ee0eb0b	model-conversion : add embedding prompt file support (#15871 ) This commit adds support for passing a prompt file to the model conversion targets/scripts. It also updates the logits.cpp to print out embedding information in the same format as when running the original embedding model. The motivation for this is that it allows us to pass files of different sizes when running the converted models and validating the logits. This can be particularly important when testing the sliding window functionality of models where the sequence length needs to exceed a certain number of tokens to trigger the sliding window logic.	2025-09-25 12:02:36 +02:00
Daniel Bevenius	d0991da39d	server : add support for external server for tests (#16243 ) This commit adds support for using an externally started llama-server instance for the server tests. This can be enabled by setting the DEBUG_EXTERNAL environment variable. The motivation for this is to allow debugging of the server itself when investigating a test failure. Instructions for how to do this are added to the README.md file in the tests directory.	2025-09-25 11:36:47 +02:00
junchao-zhao	aa719c2f88	ggml : fix loongarch lsx compilation error (#15864 )	2025-09-25 12:22:55 +03:00
Johannes Gäßler	4cdd0bb453	docs: fix typo [no ci] (#16244 )	2025-09-25 12:12:27 +03:00
Douglas Hanley	b5bd037832	llama : add support for qwen3 reranker (#15824 )	2025-09-25 11:53:09 +03:00
Georgi Gerganov	dfcd53f7ec	metal : fuse NORM + MUL + ADD, support non-multiples of 4 (#16220 ) * metal : fuse NORM + MUL + ADD * metal : support norms of non-multiple of 4 * cont : fix comment [no ci]	2025-09-25 11:30:16 +03:00
Georgi Gerganov	4ea00794b8	metal : relax reorder conditions (#16216 )	2025-09-25 11:29:42 +03:00
Georgi Gerganov	02a6a82ae7	metal : restore im2col perf (#16219 )	2025-09-25 11:29:08 +03:00
Radoslav Gerganov	c498fc82fe	rpc : use ggml logging facilities Use RPC_DEBUG environment variable to enable debug messages. Add helper macro LOG_DBG() which does an early check of the env var before calling GGML_LOG_DEBUG(). Make sure we log a debug message for every server function.	2025-09-25 07:20:02 +00:00
Aaron Teo	e7a5130a20	codeowners: add ownership of zdnn backend [no ci] (#16232 ) add @Andreas-Krebbel to owners of zDNN backend Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>	2025-09-25 08:06:30 +03:00
Eve	bee378e098	ci: run the x64 and arm ci on the github machines instead (#16183 ) * run the x64 ci on regular machines * set up the same thing for arm fix test-quantize-perf just like #12306 * try to disable sve * add another sve run	2025-09-25 08:06:06 +03:00
Aaron Teo	5fb557653b	devops: fix s390x docker release failure (#16231 )	2025-09-25 11:36:30 +08:00
Aaron Teo	4ae88d07d0	codeowners: add ownership of zdnn backend [no ci] (#16229 ) add @AlekseiNikiforovIBM to owners of zDNN backend Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>	2025-09-25 00:25:04 +08:00
Johannes Gäßler	e789095502	llama: print memory breakdown on exit (#15860 ) * llama: print memory breakdown on exit	2025-09-24 16:53:48 +02:00
Acly	f2a789e334	ggml : split graph allocations according to backend max buffer size (#15815 ) * ggml : make gallocr respect the backend's max buffer size * if the graph requires more memory than can fit into a single allocation, split it into multiple backend buffers * vulkan: report the actual max allocation size in buffer type interface * fix missing newline, apple-clang warning * track size of individual chunks in ggml_dyn_tallocr and raise max chunks. revert to use suballocation_block_size as max chunk size for vulkan. * track (chunk, offset) pairs instead of "global" offsets through gallocr. * simpler, don't need loops to map between local/global offsets * touches more code * fix dyn_tallocr_max_size and initialization * fix memory leak when buffers are reused due to same buffer type appearing multiple times * make vbuffer allocation follow the same logic as backend_buffer did before * continue to use leftover unallocated space of previous chunks after a new one has been created * treat free blocks of each chunk as separate list * they're still allocated together, but start/end of each chunk is tracked, and allocate/free iterate over sub-ranges * exhaust freed blocks of all chunks before considering their last blocks with unallocated space * start with 0 chunks/blocks and create chunks as needed * allow the last chunk to grow beyond max size * refactor: move adding new free block and new chunk into separate functions * allocate chunks individually with a separate free-blocks list for each one * needs a bit more memory/allocations/indirections, but code is simpler * fix warnings (missing static) & debug checks	2025-09-24 16:17:49 +02:00