metal : pad mm results

* llama : move end-user examples to tools directory

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>

.devops/cpu.Dockerfile

@@ -14,9 +14,9 @@ WORKDIR /app
 COPY . .
 
 RUN if [ "$TARGETARCH" = "amd64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
     elif [ "$TARGETARCH" = "arm64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
     else \
         echo "Unsupported architecture"; \
         exit 1; \

.devops/cuda.Dockerfile

@@ -21,7 +21,7 @@ COPY . .
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
     export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/intel.Dockerfile

@@ -17,7 +17,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
         && export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
     fi && \
     echo "Building with dynamic libs" && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${OPT_SYCL_F16} && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${OPT_SYCL_F16} && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/llama-cli-cann.Dockerfile

@@ -22,7 +22,7 @@ ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
 
 RUN echo "Building with static libs" && \
     source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF  && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF -DLLAMA_BUILD_TESTS=OFF  && \
     cmake --build build --config Release --target llama-cli
 
 # TODO: use image with NNRT

.devops/musa.Dockerfile

@@ -35,7 +35,7 @@ COPY . .
 RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/rocm.Dockerfile

@@ -40,7 +40,7 @@ WORKDIR /app
 COPY . .
 
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
+    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
     && cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib \

.devops/vulkan.Dockerfile

@@ -16,7 +16,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1  -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.editorconfig

@@ -21,15 +21,15 @@ indent_style = tab
 [prompts/*.txt]
 insert_final_newline = unset
 
-[examples/server/public/*]
+[tools/server/public/*]
 indent_size = 2
 
-[examples/server/public/deps_*]
+[tools/server/public/deps_*]
 trim_trailing_whitespace = unset
 indent_style = unset
 indent_size = unset
 
-[examples/server/deps_*]
+[tools/server/deps_*]
 trim_trailing_whitespace = unset
 indent_style = unset
 indent_size = unset
@@ -37,7 +37,7 @@ indent_size = unset
 [examples/llama.swiftui/llama.swiftui.xcodeproj/*]
 indent_style = tab
 
-[examples/cvector-generator/*.txt]
+[tools/cvector-generator/*.txt]
 trim_trailing_whitespace = unset
 insert_final_newline = unset
 

.flake8

@@ -2,8 +2,9 @@
 max-line-length = 125
 ignore = E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704,W503
 exclude =
-    # Do not traverse examples
+    # Do not traverse examples and tools
     examples,
+    tools,
     # Do not include package initializers
     __init__.py,
     # No need to traverse our git directory

.github/labeler.yml

@@ -45,7 +45,9 @@ build:
             - CMakePresets.json
 examples:
     - changed-files:
-        - any-glob-to-any-file: examples/**
+        - any-glob-to-any-file:
+            - examples/**
+            - tools/**
 devops:
     - changed-files:
         - any-glob-to-any-file:
@@ -70,7 +72,7 @@ android:
 server:
     - changed-files:
         - any-glob-to-any-file:
-            - examples/server/**
+            - tools/server/**
 ggml:
     - changed-files:
         - any-glob-to-any-file:

.github/workflows/bench.yml.disabled

@@ -27,10 +27,10 @@ on:
   push:
     branches:
       - master
-    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'tools/server/*.h*', 'tools/server/*.cpp']
   pull_request_target:
     types: [opened, synchronize, reopened]
-    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'tools/server/*.h*', 'tools/server/*.cpp']
   schedule:
     -  cron: '04 2 * * *'
 
@@ -69,7 +69,7 @@ jobs:
       - name: Install python env
         id: pipenv
         run: |
-          cd examples/server/bench
+          cd tools/server/bench
           python3 -m venv venv
           source venv/bin/activate
           pip install -r requirements.txt
@@ -79,7 +79,7 @@ jobs:
         run: |
           wget --quiet https://github.com/prometheus/prometheus/releases/download/v2.51.0/prometheus-2.51.0.linux-amd64.tar.gz
           tar xzf prometheus*.tar.gz --strip-components=1
-          ./prometheus --config.file=examples/server/bench/prometheus.yml &
+          ./prometheus --config.file=tools/server/bench/prometheus.yml &
           while ! nc -z localhost 9090; do
             sleep 0.1
           done
@@ -92,7 +92,7 @@ jobs:
       - name: Install k6 and xk6-sse
         id: k6_installation
         run: |
-          cd examples/server/bench
+          cd tools/server/bench
           go install go.k6.io/xk6/cmd/xk6@latest
           xk6 build master \
               --with github.com/phymbert/xk6-sse
@@ -116,7 +116,7 @@ jobs:
       - name: Download the dataset
         id: download_dataset
         run: |
-          cd examples/server/bench
+          cd tools/server/bench
           wget --quiet https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
 
       - name: Server bench
@@ -126,7 +126,7 @@ jobs:
         run: |
           set -eux
 
-          cd examples/server/bench
+          cd tools/server/bench
           source venv/bin/activate
           python bench.py \
               --runner-label ${{ env.RUNNER_LABEL }} \
@@ -157,9 +157,9 @@ jobs:
           name: bench-server-${{ github.job }}-${{ env.RUNNER_LABEL }}-${{ matrix.model }}-${{ matrix.ftype }}
           compression-level: 9
           path: |
-            examples/server/bench/*.jpg
-            examples/server/bench/*.json
-            examples/server/bench/*.log
+            tools/server/bench/*.jpg
+            tools/server/bench/*.json
+            tools/server/bench/*.log
 
       - name: Commit status
         uses: Sibz/github-status-action@v1
@@ -178,17 +178,17 @@ jobs:
         with:
           client_id: ${{secrets.IMGUR_CLIENT_ID}}
           path: |
-            examples/server/bench/prompt_tokens_seconds.jpg
-            examples/server/bench/predicted_tokens_seconds.jpg
-            examples/server/bench/kv_cache_usage_ratio.jpg
-            examples/server/bench/requests_processing.jpg
+            tools/server/bench/prompt_tokens_seconds.jpg
+            tools/server/bench/predicted_tokens_seconds.jpg
+            tools/server/bench/kv_cache_usage_ratio.jpg
+            tools/server/bench/requests_processing.jpg
 
       - name: Extract mermaid
         id: set_mermaid
         run: |
           set -eux
 
-          cd examples/server/bench
+          cd tools/server/bench
           PROMPT_TOKENS_SECONDS=$(cat prompt_tokens_seconds.mermaid)
           echo "PROMPT_TOKENS_SECONDS<<EOF" >> $GITHUB_ENV
           echo "$PROMPT_TOKENS_SECONDS" >> $GITHUB_ENV

.github/workflows/build-linux-cross.yml

@@ -4,18 +4,25 @@ on:
   workflow_call:
 
 jobs:
-  ubuntu-latest-riscv64-cpu-cross:
-    runs-on: ubuntu-latest
+  ubuntu-24-riscv64-cpu-cross:
+    runs-on: ubuntu-24.04
 
     steps:
       - uses: actions/checkout@v4
       - name: Setup Riscv
         run: |
           sudo dpkg --add-architecture riscv64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=riscv64] 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-riscv64-linux-gnu \
@@ -27,6 +34,7 @@ jobs:
           cmake -B build -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=riscv64 \
@@ -40,21 +48,25 @@ jobs:
 
           cmake --build build --config Release -j $(nproc)
 
-  ubuntu-latest-riscv64-vulkan-cross:
-    runs-on: ubuntu-latest
+  ubuntu-24-riscv64-vulkan-cross:
+    runs-on: ubuntu-24.04
 
     steps:
       - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
       - name: Setup Riscv
         run: |
           sudo dpkg --add-architecture riscv64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=riscv64] 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 \
@@ -69,6 +81,7 @@ jobs:
                          -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=riscv64 \
@@ -82,21 +95,25 @@ jobs:
 
           cmake --build build --config Release -j $(nproc)
 
-  ubuntu-latest-arm64-vulkan-cross:
-    runs-on: ubuntu-latest
+  ubuntu-24-arm64-vulkan-cross:
+    runs-on: ubuntu-24.04
 
     steps:
       - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
       - name: Setup Arm64
         run: |
           sudo dpkg --add-architecture arm64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/arm64-ports.list
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=arm64] 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 \
@@ -110,6 +127,7 @@ jobs:
                          -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=aarch64 \

.github/workflows/build.yml

@@ -601,9 +601,8 @@ jobs:
             -DGGML_SYCL_F16=ON
           cmake --build build --config Release -j $(nproc)
 
-# Disabled for now due to sporadic issue syncing.
-#  build-linux-cross:
-#    uses: ./.github/workflows/build-linux-cross.yml
+  build-linux-cross:
+    uses: ./.github/workflows/build-linux-cross.yml
 
   macOS-latest-cmake-ios:
     runs-on: macos-latest
@@ -634,6 +633,7 @@ jobs:
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_COMMON=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
             -DCMAKE_SYSTEM_NAME=iOS \
@@ -670,6 +670,7 @@ jobs:
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_COMMON=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
             -DCMAKE_SYSTEM_NAME=tvOS \
@@ -700,6 +701,7 @@ jobs:
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_COMMON=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
             -DCMAKE_SYSTEM_NAME=visionOS \
@@ -740,6 +742,7 @@ jobs:
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_CURL=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
             -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
@@ -1418,6 +1421,7 @@ jobs:
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_CURL=OFF \
             -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
             -DCMAKE_SYSTEM_NAME=iOS \

.github/workflows/server.yml

@@ -15,10 +15,10 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
 
 env:
   LLAMA_LOG_COLORS: 1
@@ -74,7 +74,7 @@ jobs:
       - name: Tests dependencies
         id: test_dependencies
         run: |
-          pip install -r examples/server/tests/requirements.txt
+          pip install -r tools/server/tests/requirements.txt
 
       # Setup nodejs (to be used for verifying bundled index.html)
       - uses: actions/setup-node@v4
@@ -84,14 +84,14 @@ jobs:
       - name: WebUI - Install dependencies
         id: webui_lint
         run: |
-          cd examples/server/webui
+          cd tools/server/webui
           npm ci
 
       - name: WebUI - Check code format
         id: webui_format
         run: |
           git config --global --add safe.directory $(realpath .)
-          cd examples/server/webui
+          cd tools/server/webui
           git status
 
           npm run format
@@ -108,7 +108,7 @@ jobs:
         id: verify_server_index_html
         run: |
           git config --global --add safe.directory $(realpath .)
-          cd examples/server/webui
+          cd tools/server/webui
           git status
 
           npm run build
@@ -161,21 +161,21 @@ jobs:
         env:
           GITHUB_ACTIONS: "true"
         run: |
-          cd examples/server/tests
+          cd tools/server/tests
           ./tests.sh
 
       - name: Tests (sanitizers)
         id: server_integration_tests_sanitizers
         if: ${{ matrix.sanitizer != '' }}
         run: |
-          cd examples/server/tests
+          cd tools/server/tests
           LLAMA_SANITIZE=1 ./tests.sh
 
       - name: Slow tests
         id: server_integration_tests_slow
         if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
-          cd examples/server/tests
+          cd tools/server/tests
           SLOW_TESTS=1 ./tests.sh
 
 
@@ -211,7 +211,7 @@ jobs:
       - name: Tests dependencies
         id: test_dependencies
         run: |
-          pip install -r examples/server/tests/requirements.txt
+          pip install -r tools/server/tests/requirements.txt
 
       - name: Copy Libcurl
         id: prepare_libcurl
@@ -224,7 +224,7 @@ jobs:
         id: server_integration_tests
         if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
         run: |
-          cd examples/server/tests
+          cd tools/server/tests
           $env:PYTHONIOENCODING = ":replace"
           pytest -v -x -m "not slow"
 
@@ -232,6 +232,6 @@ jobs:
         id: server_integration_tests_slow
         if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
         run: |
-          cd examples/server/tests
+          cd tools/server/tests
           $env:SLOW_TESTS = "1"
           pytest -v -x

.gitignore

@@ -96,11 +96,11 @@ perf-*.txt
 # Examples
 
 examples/jeopardy/results.txt
-examples/server/*.css.hpp
-examples/server/*.html.hpp
-examples/server/*.js.hpp
-examples/server/*.mjs.hpp
-examples/server/*.gz.hpp
+tools/server/*.css.hpp
+tools/server/*.html.hpp
+tools/server/*.js.hpp
+tools/server/*.mjs.hpp
+tools/server/*.gz.hpp
 !build_64.sh
 !examples/*.bat
 !examples/*/*.kts
@@ -110,7 +110,7 @@ examples/server/*.gz.hpp
 
 # Server Web UI temporary files
 node_modules
-examples/server/webui/dist
+tools/server/webui/dist
 
 # Python
 

CMakeLists.txt

@@ -77,6 +77,7 @@ option(LLAMA_BUILD_COMMON "llama: build common utils library" ${LLAMA_STANDALONE
 
 # extra artifacts
 option(LLAMA_BUILD_TESTS    "llama: build tests"          ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_TOOLS    "llama: build tools"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})
 
@@ -187,6 +188,10 @@ if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_EXAMPLES)
     add_subdirectory(pocs)
 endif()
 
+if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS)
+    add_subdirectory(tools)
+endif()
+
 #
 # install
 #

CODEOWNERS

@@ -2,7 +2,7 @@
 
 /ci/ @ggerganov
 /.devops/*.Dockerfile @ngxson
-/examples/server/ @ngxson
+/tools/server/ @ngxson
 /ggml/src/ggml-cuda/fattn* @JohannesGaessler
 /ggml/src/ggml-cuda/mmq.* @JohannesGaessler
 /ggml/src/ggml-cuda/mmv.* @JohannesGaessler

Makefile

@@ -1156,10 +1156,10 @@ $(LIB_COMMON_S): $(OBJ_COMMON)
 
 # Clean generated server assets
 clean-server-assets:
-	find examples/server -type f -name "*.js.hpp"   -delete
-	find examples/server -type f -name "*.mjs.hpp"  -delete
-	find examples/server -type f -name "*.css.hpp"  -delete
-	find examples/server -type f -name "*.html.hpp" -delete
+	find tools/server -type f -name "*.js.hpp"   -delete
+	find tools/server -type f -name "*.mjs.hpp"  -delete
+	find tools/server -type f -name "*.css.hpp"  -delete
+	find tools/server -type f -name "*.html.hpp" -delete
 
 # Clean rule
 clean: clean-server-assets
@@ -1179,7 +1179,7 @@ clean: clean-server-assets
 # Helper function that replaces .c, .cpp, and .cu file endings with .o:
 GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1))))
 
-llama-cli: examples/main/main.cpp \
+llama-cli: tools/main/main.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1192,7 +1192,7 @@ llama-infill: examples/infill/infill.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-run: examples/run/run.cpp \
+llama-run: tools/run/run.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1207,7 +1207,7 @@ llama-simple-chat: examples/simple-chat/simple-chat.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-tokenize: examples/tokenize/tokenize.cpp \
+llama-tokenize: tools/tokenize/tokenize.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1217,27 +1217,27 @@ llama-batched: examples/batched/batched.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-batched-bench: examples/batched-bench/batched-bench.cpp \
+llama-batched-bench: tools/batched-bench/batched-bench.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-quantize: examples/quantize/quantize.cpp \
+llama-quantize: tools/quantize/quantize.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-quantize-stats: examples/quantize-stats/quantize-stats.cpp \
+llama-quantize-stats: tools/quantize-stats/quantize-stats.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-perplexity: examples/perplexity/perplexity.cpp \
+llama-perplexity: tools/perplexity/perplexity.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-imatrix: examples/imatrix/imatrix.cpp \
+llama-imatrix: tools/imatrix/imatrix.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1279,7 +1279,7 @@ llama-gguf-hash: examples/gguf-hash/gguf-hash.cpp examples/gguf-hash/deps/sha1/s
 	$(CXX) $(CXXFLAGS) -Iexamples/gguf-hash/deps -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-gguf-split: examples/gguf-split/gguf-split.cpp \
+llama-gguf-split: tools/gguf-split/gguf-split.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1289,7 +1289,7 @@ llama-eval-callback: examples/eval-callback/eval-callback.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp \
+llama-cvector-generator: tools/cvector-generator/cvector-generator.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1299,12 +1299,12 @@ llama-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-bench: examples/llama-bench/llama-bench.cpp \
+llama-bench: tools/llama-bench/llama-bench.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-export-lora: examples/export-lora/export-lora.cpp \
+llama-export-lora: tools/export-lora/export-lora.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -1360,17 +1360,17 @@ llama-gbnf-validator: examples/gbnf-validator/gbnf-validator.cpp \
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 ifdef GGML_RPC
-rpc-server: examples/rpc/rpc-server.cpp \
+rpc-server: tools/rpc/rpc-server.cpp \
 	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 endif # GGML_RPC
 
 llama-server: \
-	examples/server/server.cpp \
-	examples/server/utils.hpp \
-	examples/server/httplib.h \
-	examples/server/index.html.hpp \
-	examples/server/loading.html.hpp \
+	tools/server/server.cpp \
+	tools/server/utils.hpp \
+	tools/server/httplib.h \
+	tools/server/index.html.hpp \
+	tools/server/loading.html.hpp \
 	common/chat.cpp \
 	common/chat.h \
 	common/chat-template.hpp \
@@ -1378,10 +1378,10 @@ llama-server: \
 	common/minja.hpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Itools/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
-# Portable equivalent of `cd examples/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
-examples/server/%.hpp: examples/server/public/% FORCE Makefile
+# Portable equivalent of `cd tools/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
+tools/server/%.hpp: tools/server/public/% FORCE Makefile
 	@( export NAME=$(subst .,_,$(subst -,_,$(notdir $<))) && \
 		echo "unsigned char $${NAME}[] = {" && \
 		cat $< | od -v -t x1 -An | sed -E 's/([0-9a-fA-F]+)/0x\1, /g' && \
@@ -1394,36 +1394,36 @@ llama-gen-docs: examples/gen-docs/gen-docs.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-libllava.a: examples/llava/llava.cpp \
-	examples/llava/llava.h \
-	examples/llava/clip.cpp \
-	examples/llava/clip.h \
+libllava.a: tools/llava/llava.cpp \
+	tools/llava/llava.h \
+	tools/llava/clip.cpp \
+	tools/llava/clip.h \
 	common/stb_image.h \
 	common/base64.hpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
 
-llama-llava-cli: examples/llava/llava-cli.cpp \
-	examples/llava/llava.cpp \
-	examples/llava/llava.h \
-	examples/llava/clip.cpp \
-	examples/llava/clip.h \
+llama-llava-cli: tools/llava/llava-cli.cpp \
+	tools/llava/llava.cpp \
+	tools/llava/llava.h \
+	tools/llava/clip.cpp \
+	tools/llava/clip.h \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
-llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \
-	examples/llava/llava.cpp \
-	examples/llava/llava.h \
-	examples/llava/clip.cpp \
-	examples/llava/clip.h \
+llama-minicpmv-cli: tools/llava/minicpmv-cli.cpp \
+	tools/llava/llava.cpp \
+	tools/llava/llava.h \
+	tools/llava/clip.cpp \
+	tools/llava/clip.h \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
-llama-qwen2vl-cli: examples/llava/qwen2vl-cli.cpp \
-	examples/llava/llava.cpp \
-	examples/llava/llava.h \
-	examples/llava/clip.cpp \
-	examples/llava/clip.h \
+llama-qwen2vl-cli: tools/llava/qwen2vl-cli.cpp \
+	tools/llava/llava.cpp \
+	tools/llava/llava.h \
+	tools/llava/clip.cpp \
+	tools/llava/clip.h \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
@@ -1480,12 +1480,12 @@ tests/test-double-float: tests/test-double-float.cpp
 
 tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) -Itools/server -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-chat: tests/test-chat.cpp \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) -Itools/server -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 tests/test-opt: tests/test-opt.cpp \

README.md

@@ -17,7 +17,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 ## Hot topics
 
 - **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
-- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli` and `gemma3-cli` https://github.com/ggml-org/llama.cpp/pull/13012, `libllava` will be deprecated
+- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141]((https://github.com/ggml-org/llama.cpp/pull/13141))), `libllava` will be deprecated
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
@@ -242,7 +242,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [Vulkan](docs/build.md#vulkan) | GPU |
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
-| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) | All |
+| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 
 ## Building the project
 
@@ -276,9 +276,9 @@ The Hugging Face platform provides a variety of online tools for converting, qua
 - Use the [GGUF-editor space](https://huggingface.co/spaces/CISCai/gguf-editor) to edit GGUF meta data in the browser (more info: https://github.com/ggml-org/llama.cpp/discussions/9268)
 - Use the [Inference Endpoints](https://ui.endpoints.huggingface.co/) to directly host `llama.cpp` in the cloud (more info: https://github.com/ggml-org/llama.cpp/discussions/9669)
 
-To learn more about model quantization, [read this documentation](examples/quantize/README.md)
+To learn more about model quantization, [read this documentation](tools/quantize/README.md)
 
-## [`llama-cli`](examples/main)
+## [`llama-cli`](tools/main)
 
 #### A CLI tool for accessing and experimenting with most of `llama.cpp`'s functionality.
 
@@ -341,7 +341,7 @@ To learn more about model quantization, [read this documentation](examples/quant
     </details>
 
 
-## [`llama-server`](examples/server)
+## [`llama-server`](tools/server)
 
 #### A lightweight, [OpenAI API](https://github.com/openai/openai-openapi) compatible, HTTP server for serving LLMs.
 
@@ -411,7 +411,7 @@ To learn more about model quantization, [read this documentation](examples/quant
     </details>
 
 
-## [`llama-perplexity`](examples/perplexity)
+## [`llama-perplexity`](tools/perplexity)
 
 #### A tool for measuring the perplexity [^1][^2] (and other quality metrics) of a model over a given text.
 
@@ -436,10 +436,10 @@ To learn more about model quantization, [read this documentation](examples/quant
 
     </details>
 
-[^1]: [examples/perplexity/README.md](./examples/perplexity/README.md)
+[^1]: [tools/perplexity/README.md](./tools/perplexity/README.md)
 [^2]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity)
 
-## [`llama-bench`](examples/llama-bench)
+## [`llama-bench`](tools/llama-bench)
 
 #### Benchmark the performance of the inference for various parameters.
 
@@ -460,7 +460,7 @@ To learn more about model quantization, [read this documentation](examples/quant
 
     </details>
 
-## [`llama-run`](examples/run)
+## [`llama-run`](tools/run)
 
 #### A comprehensive example for running `llama.cpp` models. Useful for inferencing. Used with RamaLama [^3].
 
@@ -504,8 +504,8 @@ To learn more about model quantization, [read this documentation](examples/quant
 
 ## Other documentation
 
-- [main (cli)](examples/main/README.md)
-- [server](examples/server/README.md)
+- [main (cli)](tools/main/README.md)
+- [server](tools/server/README.md)
 - [GBNF grammars](grammars/README.md)
 
 #### Development documentation

SECURITY.md

@@ -40,7 +40,7 @@ To protect sensitive data from potential leaks or unauthorized access, it is cru
 ### Untrusted environments or networks
 
 If you can't run your models in a secure and isolated environment or if it must be exposed to an untrusted network, make sure to take the following security precautions:
-* Do not use the RPC backend, [rpc-server](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) and [llama-server](https://github.com/ggml-org/llama.cpp/tree/master/examples/server) functionality (see https://github.com/ggml-org/llama.cpp/pull/13061).
+* Do not use the RPC backend, [rpc-server](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) and [llama-server](https://github.com/ggml-org/llama.cpp/tree/master/tools/server) functionality (see https://github.com/ggml-org/llama.cpp/pull/13061).
 * Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value.
 * Encrypt your data if sending it over the network.
 

build-xcframework.sh

@@ -8,6 +8,7 @@ TVOS_MIN_OS_VERSION=16.4
 
 BUILD_SHARED_LIBS=OFF
 LLAMA_BUILD_EXAMPLES=OFF
+LLAMA_BUILD_TOOLS=OFF
 LLAMA_BUILD_TESTS=OFF
 LLAMA_BUILD_SERVER=OFF
 GGML_METAL=ON
@@ -31,6 +32,7 @@ COMMON_CMAKE_ARGS=(
     -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
     -DBUILD_SHARED_LIBS=${BUILD_SHARED_LIBS}
     -DLLAMA_BUILD_EXAMPLES=${LLAMA_BUILD_EXAMPLES}
+    -DLLAMA_BUILD_TOOLS=${LLAMA_BUILD_TOOLS}
     -DLLAMA_BUILD_TESTS=${LLAMA_BUILD_TESTS}
     -DLLAMA_BUILD_SERVER=${LLAMA_BUILD_SERVER}
     -DGGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY}

ci/run.sh

@@ -187,8 +187,8 @@ function gg_run_test_scripts_debug {
 
     set -e
 
-    (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-    (cd ./examples/quantize   && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./tools/quantize   && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
 
     set +e
 }
@@ -211,8 +211,8 @@ function gg_run_test_scripts_release {
 
     set -e
 
-    (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
-    (cd ./examples/quantize   && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./tools/quantize   && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
 
     set +e
 }

cmake/build-info.cmake

@@ -41,14 +41,20 @@ endif()
 
 if(MSVC)
     set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
-    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+    if (CMAKE_VS_PLATFORM_NAME)
+        set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+    else()
+        set(BUILD_TARGET "${CMAKE_SYSTEM_NAME} ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
 else()
     execute_process(
-        COMMAND sh -c "\"$@\" --version | head -1" _ ${CMAKE_C_COMPILER}
+        COMMAND ${CMAKE_C_COMPILER} --version
         OUTPUT_VARIABLE OUT
         OUTPUT_STRIP_TRAILING_WHITESPACE
     )
+    string(REGEX REPLACE " *\n.*" "" OUT "${OUT}")
     set(BUILD_COMPILER ${OUT})
+
     execute_process(
         COMMAND ${CMAKE_C_COMPILER} -dumpmachine
         OUTPUT_VARIABLE OUT

common/CMakeLists.txt

@@ -39,7 +39,9 @@ add_custom_command(
     COMMENT "Generating build details from Git"
     COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
             -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+            -DCMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME} -DCMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}
+            -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
     WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
     DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
     VERBATIM

common/arg.cpp

@@ -43,6 +43,25 @@ std::initializer_list<enum llama_example> mmproj_examples = {
     // TODO: add LLAMA_EXAMPLE_SERVER when it's ready
 };
 
+static std::string read_file(const std::string & fname) {
+    std::ifstream file(fname);
+    if (!file) {
+        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
+    }
+    std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+    file.close();
+    return content;
+}
+
+static void write_file(const std::string & fname, const std::string & content) {
+    std::ofstream file(fname);
+    if (!file) {
+        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
+    }
+    file << content;
+    file.close();
+}
+
 common_arg & common_arg::set_examples(std::initializer_list<enum llama_example> examples) {
     this->examples = std::move(examples);
     return *this;
@@ -198,11 +217,11 @@ struct curl_slist_ptr {
 #define CURL_MAX_RETRY 3
 #define CURL_RETRY_DELAY_SECONDS 2
 
-static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
+static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds, const char * method_name) {
     int remaining_attempts = max_attempts;
 
     while (remaining_attempts > 0) {
-        LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
+        LOG_INF("%s: %s %s (attempt %d of %d)...\n", __func__ , method_name, url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
 
         CURLcode res = curl_easy_perform(curl);
         if (res == CURLE_OK) {
@@ -213,6 +232,7 @@ static bool curl_perform_with_retry(const std::string & url, CURL * curl, int ma
         LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
 
         remaining_attempts--;
+        if (remaining_attempts == 0) break;
         std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
     }
 
@@ -231,8 +251,6 @@ static bool common_download_file_single(const std::string & url, const std::stri
         return false;
     }
 
-    bool force_download = false;
-
     // Set the URL, allow to follow http redirection
     curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
     curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
@@ -256,7 +274,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
 
     // If the file exists, check its JSON metadata companion file.
     std::string metadata_path = path + ".json";
-    nlohmann::json metadata;
+    nlohmann::json metadata; // TODO @ngxson : get rid of this json, use regex instead
     std::string etag;
     std::string last_modified;
 
@@ -266,14 +284,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
         if (metadata_in.good()) {
             try {
                 metadata_in >> metadata;
-                LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
-                if (metadata.contains("url") && metadata.at("url").is_string()) {
-                    auto previous_url = metadata.at("url").get<std::string>();
-                    if (previous_url != url) {
-                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
-                        return false;
-                    }
-                }
+                LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
                 if (metadata.contains("etag") && metadata.at("etag").is_string()) {
                     etag = metadata.at("etag");
                 }
@@ -281,10 +292,10 @@ static bool common_download_file_single(const std::string & url, const std::stri
                     last_modified = metadata.at("lastModified");
                 }
             } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-                return false;
+                LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
             }
         }
+        // if we cannot open the metadata file, we assume that the downloaded file is not valid (etag and last-modified are left empty, so we will download it again)
     } else {
         LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
     }
@@ -296,7 +307,10 @@ static bool common_download_file_single(const std::string & url, const std::stri
     };
 
     common_load_model_from_url_headers headers;
+    bool head_request_ok = false;
+    bool should_download = !file_exists; // by default, we should download if the file does not exist
 
+    // get ETag to see if the remote file has changed
     {
         typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
         auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
@@ -325,23 +339,28 @@ static bool common_download_file_single(const std::string & url, const std::stri
         curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
         curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
 
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        // we only allow retrying once for HEAD requests
+        // this is for the use case of using running offline (no internet), retrying can be annoying
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0, "HEAD");
         if (!was_perform_successful) {
-            return false;
+            head_request_ok = false;
         }
 
         long http_code = 0;
         curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code != 200) {
-            // HEAD not supported, we don't know if the file has changed
-            // force trigger downloading
-            force_download = true;
-            LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+        if (http_code == 200) {
+            head_request_ok = true;
+        } else {
+            LOG_WRN("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+            head_request_ok = false;
         }
     }
 
-    bool should_download = !file_exists || force_download;
-    if (!should_download) {
+    // if head_request_ok is false, we don't have the etag or last-modified headers
+    // we leave should_download as-is, which is true if the file does not exist
+    if (head_request_ok) {
+        // check if ETag or Last-Modified headers are different
+        // if it is, we need to download the file again
         if (!etag.empty() && etag != headers.etag) {
             LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
             should_download = true;
@@ -350,6 +369,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
             should_download = true;
         }
     }
+
     if (should_download) {
         std::string path_temporary = path + ".downloadInProgress";
         if (file_exists) {
@@ -403,7 +423,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
         // start the download
         LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
             llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS, "GET");
         if (!was_perform_successful) {
             return false;
         }
@@ -424,13 +444,15 @@ static bool common_download_file_single(const std::string & url, const std::stri
             {"etag", headers.etag},
             {"lastModified", headers.last_modified}
         });
-        std::ofstream(metadata_path) << metadata.dump(4);
-        LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
+        write_file(metadata_path, metadata.dump(4));
+        LOG_DBG("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
 
         if (rename(path_temporary.c_str(), path.c_str()) != 0) {
             LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
             return false;
         }
+    } else {
+        LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
     }
 
     return true;
@@ -605,16 +627,37 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
     // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
     // User-Agent header is already set in common_remote_get_content, no need to set it here
 
+    // we use "=" to avoid clashing with other component, while still being allowed on windows
+    std::string cached_response_fname = "manifest=" + hf_repo + "=" + tag + ".json";
+    string_replace_all(cached_response_fname, "/", "_");
+    std::string cached_response_path = fs_get_cache_file(cached_response_fname);
+
     // make the request
     common_remote_params params;
     params.headers = headers;
-    auto res = common_remote_get_content(url, params);
-    long res_code = res.first;
-    std::string res_str(res.second.data(), res.second.size());
+    long res_code = 0;
+    std::string res_str;
+    bool use_cache = false;
+    try {
+        auto res = common_remote_get_content(url, params);
+        res_code = res.first;
+        res_str = std::string(res.second.data(), res.second.size());
+    } catch (const std::exception & e) {
+        LOG_WRN("error: failed to get manifest: %s\n", e.what());
+        LOG_WRN("try reading from cache\n");
+        // try to read from cache
+        try {
+            res_str = read_file(cached_response_path);
+            res_code = 200;
+            use_cache = true;
+        } catch (const std::exception & e) {
+            throw std::runtime_error("error: failed to get manifest (check your internet connection)");
+        }
+    }
     std::string ggufFile;
     std::string mmprojFile;
 
-    if (res_code == 200) {
+    if (res_code == 200 || res_code == 304) {
         // extract ggufFile.rfilename in json, using regex
         {
             std::regex pattern("\"ggufFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
@@ -631,6 +674,10 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
                 mmprojFile = match[1].str();
             }
         }
+        if (!use_cache) {
+            // if not using cached response, update the cache file
+            write_file(cached_response_path, res_str);
+        }
     } else if (res_code == 401) {
         throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
     } else {
@@ -1142,6 +1189,9 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
         fprintf(stderr, "%s\n", ex.what());
         ctx_arg.params = params_org;
         return false;
+    } catch (std::exception & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        exit(1); // for other exceptions, we exit with status code 1
     }
 
     return true;
@@ -1442,13 +1492,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-f", "--file"}, "FNAME",
         "a file containing the prompt (default: none)",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
+            params.prompt = read_file(value);
             // store the external file name in params
             params.prompt_file = value;
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
             if (!params.prompt.empty() && params.prompt.back() == '\n') {
                 params.prompt.pop_back();
             }
@@ -1458,11 +1504,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-sysf", "--system-prompt-file"}, "FNAME",
         "a file containing the system prompt (default: none)",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.system_prompt));
+            params.system_prompt = read_file(value);
             if (!params.system_prompt.empty() && params.system_prompt.back() == '\n') {
                 params.system_prompt.pop_back();
             }
@@ -1887,15 +1929,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--grammar-file"}, "FNAME",
         "file to read grammar from",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(
-                std::istreambuf_iterator<char>(file),
-                std::istreambuf_iterator<char>(),
-                std::back_inserter(params.sampling.grammar)
-            );
+            params.sampling.grammar = read_file(value);
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1905,6 +1939,23 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.sampling.grammar = json_schema_to_grammar(json::parse(value));
         }
     ).set_sparam());
+    add_opt(common_arg(
+        {"-jf", "--json-schema-file"}, "FILE",
+        "File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\nFor schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead",
+        [](common_params & params, const std::string & value) {
+            std::ifstream file(value);
+            if (!file) {
+                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
+            }
+            std::string schema;
+            std::copy(
+                std::istreambuf_iterator<char>(file),
+                std::istreambuf_iterator<char>(),
+                std::back_inserter(schema)
+            );
+            params.sampling.grammar = json_schema_to_grammar(json::parse(schema));
+        }
+    ).set_sparam());
     add_opt(common_arg(
         {"--pooling"}, "{none,mean,cls,last,rank}",
         "pooling type for embeddings, use model default if unspecified",
@@ -2160,14 +2211,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONT_BATCHING"));
     add_opt(common_arg(
         {"--mmproj"}, "FILE",
-        "path to a multimodal projector file. see examples/llava/README.md",
+        "path to a multimodal projector file. see tools/llava/README.md",
         [](common_params & params, const std::string & value) {
             params.mmproj.path = value;
         }
     ).set_examples(mmproj_examples));
     add_opt(common_arg(
         {"--mmproj-url"}, "URL",
-        "URL to a multimodal projector file. see examples/llava/README.md",
+        "URL to a multimodal projector file. see tools/llava/README.md",
         [](common_params & params, const std::string & value) {
             params.mmproj.url = value;
         }
@@ -2732,7 +2783,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_THREADS_HTTP"));
     add_opt(common_arg(
         {"--cache-reuse"}, "N",
-        string_format("min chunk size to attempt reusing from the cache via KV shifting (default: %d)", params.n_cache_reuse),
+        string_format(
+            "min chunk size to attempt reusing from the cache via KV shifting (default: %d)\n"
+            "[(card)](https://ggml.ai/f0.png)", params.n_cache_reuse
+        ),
         [](common_params & params, int value) {
             params.n_cache_reuse = value;
         }
@@ -2815,14 +2869,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             "list of built-in templates:\n%s", list_builtin_chat_templates().c_str()
         ),
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(
-                std::istreambuf_iterator<char>(file),
-                std::istreambuf_iterator<char>(),
-                std::back_inserter(params.chat_template));
+            params.chat_template = read_file(value);
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE"));
     add_opt(common_arg(

common/common.h

@@ -340,7 +340,7 @@ struct common_params {
 
     common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
 
-    // multimodal models (see examples/llava)
+    // multimodal models (see tools/llava)
     struct common_params_model mmproj;
     bool mmproj_use_gpu = true;     // use GPU for multimodal model
     bool no_mmproj = false;         // explicitly disable multimodal model
@@ -414,8 +414,8 @@ struct common_params {
     int n_pca_batch = 100;
     int n_pca_iterations = 1000;
     dimre_method cvector_dimre_method = DIMRE_METHOD_PCA;
-    std::string cvector_positive_file = "examples/cvector-generator/positive.txt";
-    std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
+    std::string cvector_positive_file = "tools/cvector-generator/positive.txt";
+    std::string cvector_negative_file = "tools/cvector-generator/negative.txt";
 
     bool spm_infill = false; // suffix/prefix/middle pattern for infill
 

convert_hf_to_gguf.py

@@ -16,6 +16,7 @@ from pathlib import Path
 from hashlib import sha256
 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast
 from itertools import chain
+from transformers import AutoConfig
 
 import math
 import numpy as np
@@ -66,8 +67,6 @@ class ModelBase:
     part_names: list[str]
     is_safetensors: bool
     hparams: dict[str, Any]
-    block_count: int
-    tensor_map: gguf.TensorNameMap
     tensor_names: set[str] | None
     gguf_writer: gguf.GGUFWriter
     model_name: str | None
@@ -78,7 +77,11 @@ class ModelBase:
     # subclasses should define this!
     model_arch: gguf.MODEL_ARCH
 
-    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
+    # subclasses should initialize this!
+    block_count: int
+    tensor_map: gguf.TensorNameMap
+
+    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, *, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
                  split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
@@ -113,8 +116,6 @@ class ModelBase:
             if not self.is_safetensors:
                 self.part_names = ModelBase.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
         self.hparams = ModelBase.load_hparams(self.dir_model) if hparams is None else hparams
-        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
-        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
         self.tensor_names = None
         self.metadata_override = metadata_override
         self.model_name = model_name
@@ -417,15 +418,15 @@ class ModelBase:
 
     @staticmethod
     def load_hparams(dir_model: Path):
-        with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-            hparams = json.load(f)
-            architectures = hparams.get("architectures")
-            if "text_config" in hparams:
-                hparams = {**hparams, **hparams["text_config"]}
-            if architectures is not None:
-                # preserve "architectures" from root level config
-                hparams["architectures"] = architectures
-            return hparams
+        try:
+            # for security reason, we don't allow loading remote code by default
+            # if a model need remote code, we will fallback to config.json
+            return AutoConfig.from_pretrained(dir_model, trust_remote_code=False).to_dict()
+        except Exception as e:
+            logger.warning(f"Failed to load model config from {dir_model}: {e}")
+            logger.warning("Trying to load config.json instead")
+            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+                return json.load(f)
 
     @classmethod
     def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
@@ -454,6 +455,20 @@ class ModelBase:
 
 
 class TextModel(ModelBase):
+    model_type = ModelType.TEXT
+    hf_arch: str
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hf_arch = get_model_architecture(self.hparams, self.model_type)
+
+        if "text_config" in self.hparams:
+            # move the text_config to the root level
+            self.hparams = {**self.hparams, **self.hparams["text_config"]}
+
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
     @classmethod
     def __init_subclass__(cls):
         # can't use an abstract property, because overriding it without type errors
@@ -495,7 +510,7 @@ class TextModel(ModelBase):
     def set_gguf_parameters(self):
         self.gguf_writer.add_block_count(self.block_count)
 
-        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
             logger.info(f"gguf: context length = {n_ctx}")
 
@@ -1064,10 +1079,36 @@ class TextModel(ModelBase):
         if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.ADD_EOS)) is not None:
             self.gguf_writer.add_add_eos_token(field.parts[-1].tolist()[0])
 
+    def _try_set_pooling_type(self) -> None:
+        # get pooling path
+        pooling_path = None
+        module_path = self.dir_model / "modules.json"
+        if module_path.is_file():
+            with open(module_path, encoding="utf-8") as f:
+                modules = json.load(f)
+            for mod in modules:
+                if mod["type"] == "sentence_transformers.models.Pooling":
+                    pooling_path = mod["path"]
+                    break
+
+        # get pooling type
+        if pooling_path is not None:
+            with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
+                pooling = json.load(f)
+            if pooling["pooling_mode_mean_tokens"]:
+                pooling_type = gguf.PoolingType.MEAN
+            elif pooling["pooling_mode_cls_token"]:
+                pooling_type = gguf.PoolingType.CLS
+            elif pooling["pooling_mode_lasttoken"]:
+                pooling_type = gguf.PoolingType.LAST
+            else:
+                raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported")
+            self.gguf_writer.add_pooling_type(pooling_type)
+
 
 class VisionModel(ModelBase):
+    model_type = ModelType.VISION
     model_arch = gguf.MODEL_ARCH.CLIP_VISION
-    n_text_embd = 0
     preprocessor_config: dict[str, Any]
     global_config: dict[str, Any]
 
@@ -1077,9 +1118,11 @@ class VisionModel(ModelBase):
         if self.model_arch != gguf.MODEL_ARCH.CLIP_VISION:
             raise TypeError("VisionModel must be subclassed with model_arch = gguf.MODEL_ARCH.CLIP_VISION")
 
-        # small hack to correct the number of layers
-        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.CLIP_VISION, 128)
-        self.n_embd_text = self.find_hparam(["hidden_size", "n_embd"])
+        # get n_embd of the text model
+        if "text_config" not in self.hparams:
+            self.hparams["text_config"] = {}
+        text_config = {**self.hparams, **self.hparams["text_config"]}
+        self.n_embd_text = text_config.get("hidden_size", text_config.get("n_embd", 0))
         assert self.n_embd_text > 0, "n_embd not found in hparams"
 
         if "vision_config" not in self.hparams:
@@ -1088,6 +1131,9 @@ class VisionModel(ModelBase):
         self.global_config = self.hparams
         self.hparams = self.hparams["vision_config"]
 
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth"])
+        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.CLIP_VISION, self.block_count)
+
         # load preprocessor config
         with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
             self.preprocessor_config = json.load(f)
@@ -1105,12 +1151,12 @@ class VisionModel(ModelBase):
         self.gguf_writer.add_vision_patch_size(self.find_hparam(["patch_size"]))
         self.gguf_writer.add_vision_embedding_length(self.find_hparam(["hidden_size"]))
         self.gguf_writer.add_vision_feed_forward_length(self.find_hparam(["intermediate_size"]))
-        self.gguf_writer.add_vision_block_count(self.find_hparam(["num_hidden_layers"]))
+        self.gguf_writer.add_vision_block_count(self.block_count)
         self.gguf_writer.add_vision_head_count(self.find_hparam(["num_attention_heads"]))
 
         # preprocessor config
         self.gguf_writer.add_vision_image_mean(self.preprocessor_config["image_mean"])
-        self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_mean"])
+        self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_std"])
 
     def write_vocab(self):
         raise ValueError("VisionModel does not support vocab writing")
@@ -1726,23 +1772,12 @@ class StableLMModel(TextModel):
     "LlamaForCausalLM",
     "MistralForCausalLM",
     "MixtralForCausalLM",
-    "Idefics3ForConditionalGeneration",
-    "SmolVLMForConditionalGeneration",
+    "VLlama3ForCausalLM",
     "LlavaForConditionalGeneration")
 class LlamaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.LLAMA
     undo_permute = True
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        # fix for SmolVLM2, missing `num_attention_heads` in config.json
-        if self.hparams["architectures"][0] == "SmolVLMForConditionalGeneration":
-            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
-        # fix for Pixtral, missing `num_attention_heads` in config.json
-        if self.hparams["architectures"][0] == "LlavaForConditionalGeneration" \
-                and self.hparams.get("model_type") == "mistral":
-            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
-
     def set_vocab(self):
         try:
             self._set_vocab_sentencepiece()
@@ -1898,31 +1933,50 @@ class LlamaModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("LlavaForConditionalGeneration")
+@ModelBase.register(
+    "LlavaForConditionalGeneration", # pixtral
+    "Mistral3ForConditionalGeneration", # mistral small 3.1
+)
 class LlavaVisionModel(VisionModel):
     img_break_tok_id = -1
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         if self.hparams["model_type"] == "pixtral":
-            # fix missing config.json values
-            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 16)
-            self.hparams["num_hidden_layers"] = self.hparams.get("num_hidden_layers", 24)
-            self.hparams["intermediate_size"] = self.hparams.get("intermediate_size", 4096)
-            self.hparams["hidden_size"] = self.hparams.get("hidden_size", 1024)
+            # layer_norm_eps is not in config.json, it is hard-coded in modeling_pixtral.py
             self.hparams["layer_norm_eps"] = self.hparams.get("layer_norm_eps", 1e-5)
-            self.img_break_tok_id = 12 # see tokenizer_config.json
+            self.img_break_tok_id = self.get_token_id("[IMG_BREAK]")
+            logger.info(f"Image break token id: {self.img_break_tok_id}")
         else:
             raise ValueError(f"Unsupported model type: {self.hparams['model_type']}")
 
+    def get_token_id(self, token: str) -> int:
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+            added_tokens_decoder = json.load(f)['added_tokens_decoder']
+            for id_, token_data in added_tokens_decoder.items():
+                if token_data["content"] == token:
+                    return int(id_)
+        raise ValueError(f"Token '{token}' not found in tokenizer config.")
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
         if hparams["model_type"] == "pixtral":
             self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.PIXTRAL)
-            # default values below are taken from HF tranformers code
             self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
-            self.gguf_writer.add_vision_use_silu(True)
+
+            # hidden_act
+            if hparams["hidden_act"] == "silu":
+                self.gguf_writer.add_vision_use_silu(True)
+            elif hparams["hidden_act"] == "gelu":
+                self.gguf_writer.add_vision_use_gelu(True)
+            else:
+                raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}")
+
+            # spatial_merge_size
+            if "spatial_merge_size" in self.global_config:
+                self.gguf_writer.add_vision_spatial_merge_size(self.global_config["spatial_merge_size"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -1951,13 +2005,12 @@ class LlavaVisionModel(VisionModel):
 class SmolVLMModel(VisionModel):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        # fix for SmolVLM2, missing some keys in config.json
-        # default values are taken from transformers code
         if self.hparams["model_type"] == "smolvlm_vision":
+            # fix for SmolVLM2, missing some keys in config.json
+            # default values are taken from transformers code
             self.hparams["hidden_size"] = self.hparams.get("hidden_size", 1152)
             self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 16)
             self.hparams["intermediate_size"] = self.hparams.get("intermediate_size", 3072)
-            self.hparams["num_hidden_layers"] = self.hparams.get("num_hidden_layers", 12)
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -2070,6 +2123,9 @@ class DeciModel(TextModel):
             # if n_heads_in_group is not None, then
             # _num_kv_heads[il] is num_attention_head // n_heads_in_group and
             # _num_heads[il] is num_attention_head
+            # ***dummy layer*** for nemotron 253B
+            # if n_heads_in_group is None and ffn_mult is None
+            # then _num_kv_heads[il] is 0 and _num_heads[il] is 0 and _ffn_dims is 0
             for il in range(len(_block_configs)):
                 if _block_configs[il]["attention"]["n_heads_in_group"] is None:
                     if _block_configs[il]["attention"]["replace_with_linear"] is True:
@@ -2081,7 +2137,10 @@ class DeciModel(TextModel):
                 else:
                     self._num_kv_heads.append(self.hparams["num_attention_heads"] // _block_configs[il]["attention"]["n_heads_in_group"])
                     self._num_heads.append(self.hparams["num_attention_heads"])
-                _ffn_multipliers.append(_block_configs[il]["ffn"]["ffn_mult"])
+                if _block_configs[il]["ffn"]["ffn_mult"] is None: # dummy layer
+                    _ffn_multipliers.append(0.0)
+                else:
+                    _ffn_multipliers.append(_block_configs[il]["ffn"]["ffn_mult"])
             assert self.block_count == len(self._num_kv_heads)
             assert self.block_count == len(self._num_heads)
             assert self.block_count == len(_ffn_multipliers)
@@ -2519,7 +2578,7 @@ class QwenModel(TextModel):
         self.gguf_writer.add_file_type(self.ftype)
 
 
-@ModelBase.register("Qwen2ForCausalLM")
+@ModelBase.register("Qwen2Model", "Qwen2ForCausalLM")
 class Qwen2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.QWEN2
 
@@ -2531,12 +2590,18 @@ class Qwen2Model(TextModel):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
+        self._try_set_pooling_type()
         if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
             if self.hparams["rope_scaling"].get("type") == "yarn":
                 self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
                 self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
                 self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"])
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if self.hf_arch == "Qwen2Model":
+            name = f"model.{name}"  # map to Qwen2ForCausalLM tensors
+        yield from super().modify_tensors(data_torch, name, bid)
+
 
 @ModelBase.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
 class Qwen2VLModel(TextModel):
@@ -2562,6 +2627,82 @@ class Qwen2VLModel(TextModel):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
+class Qwen2VLVisionModel(VisionModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hparams["image_size"] = self.hparams.get("image_size", 560)
+        # rename config.json values
+        self.hparams["num_attention_heads"] = self.hparams.get("num_heads")
+        self.hparams["num_hidden_layers"] = self.hparams.get("depth")
+        if "embed_dim" in self.hparams: # qwen2vl
+            self.hparams["intermediate_size"] = self.hparams.get("hidden_size")
+            self.hparams["hidden_size"] = self.hparams.get("embed_dim")
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        if self.global_config['model_type'] == 'qwen2_vl':
+            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN2VL)
+        elif self.global_config['model_type'] == 'qwen2_5_vl':
+            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN25VL)
+            self.gguf_writer.add_vision_use_silu(True)
+            # find n_wa_pattern (window attention pattern)
+            fullatt_block_indexes = hparams.get("fullatt_block_indexes")
+            assert fullatt_block_indexes is not None, "fullatt_block_indexes is required for qwen2_5_vl"
+            n_wa_pattern = fullatt_block_indexes[0] + 1
+            # validate n_wa_pattern
+            for i in range(1, len(fullatt_block_indexes)):
+                if fullatt_block_indexes[i] - fullatt_block_indexes[i - 1] != n_wa_pattern:
+                    raise ValueError(f"Invalid fullatt_block_indexes: {fullatt_block_indexes}")
+            self.gguf_writer.add_vision_n_wa_pattern(n_wa_pattern)
+        else:
+            raise ValueError(f"Unknown QwenVL model type: {self.global_config['model_type']}")
+        # default values below are taken from HF tranformers code
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.global_config.get("rms_norm_eps", 1e-6))
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del bid, name, n_dims  # unused
+        if ".patch_embd." in new_name:
+            return gguf.GGMLQuantizationType.F16
+        if ".position_embd." in new_name:
+            return gguf.GGMLQuantizationType.F32
+        return False
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        if name.startswith("visual."):
+            # process visual tensors
+            # split QKV tensors if needed
+            if ".qkv." in name:
+                if data_torch.ndim == 2: # weight
+                    c3, _ = data_torch.shape
+                else: # bias
+                    c3 = data_torch.shape[0]
+                assert c3 % 3 == 0
+                c = c3 // 3
+                wq = data_torch[:c]
+                wk = data_torch[c: c * 2]
+                wv = data_torch[c * 2:]
+                return [
+                    (self.map_tensor_name(name.replace("qkv", "q")), wq),
+                    (self.map_tensor_name(name.replace("qkv", "k")), wk),
+                    (self.map_tensor_name(name.replace("qkv", "v")), wv),
+                ]
+            elif 'patch_embed.proj.weight' in name:
+                # split Conv3D into Conv2Ds
+                c1, c2, kt, kh, kw = data_torch.shape
+                del c1, c2, kh, kw  # unused
+                assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
+                return [
+                    (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight"  , data_torch[:, :, 0, ...]),
+                    (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch[:, :, 1, ...]),
+                ]
+            else:
+                return [(self.map_tensor_name(name), data_torch)]
+        return [] # skip other tensors
+
+
 @ModelBase.register("WavTokenizerDec")
 class WavTokenizerDecModel(TextModel):
     model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC
@@ -3297,29 +3438,7 @@ class BertModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         self.gguf_writer.add_causal_attention(False)
-
-        # get pooling path
-        pooling_path = None
-        module_path = self.dir_model / "modules.json"
-        if module_path.is_file():
-            with open(module_path, encoding="utf-8") as f:
-                modules = json.load(f)
-            for mod in modules:
-                if mod["type"] == "sentence_transformers.models.Pooling":
-                    pooling_path = mod["path"]
-                    break
-
-        # get pooling type
-        if pooling_path is not None:
-            with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
-                pooling = json.load(f)
-            if pooling["pooling_mode_mean_tokens"]:
-                pooling_type = gguf.PoolingType.MEAN
-            elif pooling["pooling_mode_cls_token"]:
-                pooling_type = gguf.PoolingType.CLS
-            else:
-                raise NotImplementedError("Only MEAN and CLS pooling types supported")
-            self.gguf_writer.add_pooling_type(pooling_type)
+        self._try_set_pooling_type()
 
     def set_vocab(self):
         tokens, toktypes, tokpre = self.get_vocab_base()
@@ -3373,14 +3492,7 @@ class BertModel(TextModel):
 
         return [(self.map_tensor_name(name), data_torch)]
 
-
-@ModelBase.register("RobertaModel")
-class RobertaModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.BERT
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
+    def _xlmroberta_tokenizer_init(self) -> None:
         # we need the pad_token_id to know how to chop down position_embd matrix
         if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
             self._position_offset = 1 + pad_token_id
@@ -3389,82 +3501,7 @@ class RobertaModel(BertModel):
         else:
             self._position_offset = None
 
-    def set_vocab(self):
-        """Support BPE tokenizers for roberta models"""
-        bpe_tok_path = self.dir_model / "tokenizer.json"
-        if bpe_tok_path.exists():
-            self._set_vocab_gpt2()
-            self.gguf_writer.add_add_bos_token(True)
-            self.gguf_writer.add_add_eos_token(True)
-
-            # we need this to validate the size of the token_type embeddings
-            # though currently we are passing all zeros to the token_type embeddings
-            # "Sequence A" or "Sequence B"
-            self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
-
-        else:
-            return super().set_vocab()
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # if name starts with "roberta.", remove the prefix
-        # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
-        if name.startswith("roberta."):
-            name = name[8:]
-
-        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
-        if name == "embeddings.position_embeddings.weight":
-            if self._position_offset is not None:
-                data_torch = data_torch[self._position_offset:,:]
-
-        return super().modify_tensors(data_torch, name, bid)
-
-
-@ModelBase.register("NomicBertModel")
-class NomicBertModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.NOMIC_BERT
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # the HF config claims n_ctx=8192, but it uses RoPE scaling
-        self.hparams["n_ctx"] = 2048
-
-        # SwigLU activation
-        assert self.hparams["activation_function"] == "swiglu"
-        # this doesn't do anything in the HF version
-        assert self.hparams["causal"] is False
-        # no bias tensors
-        assert self.hparams["qkv_proj_bias"] is False
-        assert self.hparams["mlp_fc1_bias"] is False
-        assert self.hparams["mlp_fc2_bias"] is False
-        # norm at end of layer
-        assert self.hparams["prenorm"] is False
-        # standard RoPE
-        assert self.hparams["rotary_emb_fraction"] == 1.0
-        assert self.hparams["rotary_emb_interleaved"] is False
-        assert self.hparams["rotary_emb_scale_base"] is None
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
-
-
-@ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
-class XLMRobertaModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.BERT
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # we need the pad_token_id to know how to chop down position_embd matrix
-        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
-            self._position_offset = 1 + pad_token_id
-            if "max_position_embeddings" in self.hparams:
-                self.hparams["max_position_embeddings"] -= self._position_offset
-        else:
-            self._position_offset = None
-
-    def set_vocab(self):
+    def _xlmroberta_set_vocab(self) -> None:
         # to avoid TypeError: Descriptors cannot be created directly
         # exception when importing sentencepiece_model_pb2
         os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
@@ -3546,6 +3583,145 @@ class XLMRobertaModel(BertModel):
         self.gguf_writer.add_add_bos_token(True)
         self.gguf_writer.add_add_eos_token(True)
 
+
+@ModelBase.register("RobertaModel")
+class RobertaModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # we need the pad_token_id to know how to chop down position_embd matrix
+        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
+            self._position_offset = 1 + pad_token_id
+            if "max_position_embeddings" in self.hparams:
+                self.hparams["max_position_embeddings"] -= self._position_offset
+        else:
+            self._position_offset = None
+
+    def set_vocab(self):
+        """Support BPE tokenizers for roberta models"""
+        bpe_tok_path = self.dir_model / "tokenizer.json"
+        if bpe_tok_path.exists():
+            self._set_vocab_gpt2()
+            self.gguf_writer.add_add_bos_token(True)
+            self.gguf_writer.add_add_eos_token(True)
+
+            # we need this to validate the size of the token_type embeddings
+            # though currently we are passing all zeros to the token_type embeddings
+            # "Sequence A" or "Sequence B"
+            self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
+
+        else:
+            return super().set_vocab()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # if name starts with "roberta.", remove the prefix
+        # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
+        if name.startswith("roberta."):
+            name = name[8:]
+
+        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
+        if name == "embeddings.position_embeddings.weight":
+            if self._position_offset is not None:
+                data_torch = data_torch[self._position_offset:,:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("NomicBertModel")
+class NomicBertModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, **kwargs: Any):
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(dir_model)
+
+        self.is_moe = bool(hparams.get("moe_every_n_layers"))
+        self.model_arch = gguf.MODEL_ARCH.NOMIC_BERT_MOE if self.is_moe else gguf.MODEL_ARCH.NOMIC_BERT
+
+        super().__init__(dir_model, ftype, fname_out, hparams=hparams, **kwargs)
+
+        self._tokenizer_is_xlmroberta = self._is_tokenizer_xlmroberta()
+        if self._tokenizer_is_xlmroberta:
+            self._xlmroberta_tokenizer_init()
+
+        npos, mtp = self.hparams["n_positions"], self.hparams.get("max_trained_positions", 2048)
+        if npos == 8192 and mtp == 2048:
+            self.hparams["n_positions"] = 2048  # nomic-embed-text v1 and v1.5 are trained for 2048 tokens.
+        elif npos == 2048 and mtp == 2048:
+            self.hparams["n_positions"] = 512   # nomic-embed-text-v2-moe is trained for 512 tokens.
+        else:
+            raise ValueError(f"unrecognized parameters: n_positions={npos}, max_trained_positions={mtp}")
+
+        assert self.hparams["activation_function"] == "gelu" if self.is_moe else "swiglu"
+
+        # this doesn't do anything in the HF version
+        assert self.hparams["causal"] is False
+        # no bias tensors unless MoE
+        assert self.hparams["qkv_proj_bias"] == self.is_moe
+        assert self.hparams["mlp_fc1_bias"]  == self.is_moe
+        assert self.hparams["mlp_fc2_bias"]  == self.is_moe
+
+        # norm at end of layer
+        assert self.hparams["prenorm"] is False
+        # standard RoPE
+        assert self.hparams["rotary_emb_fraction"] == 1.0
+        assert self.hparams["rotary_emb_interleaved"] is False
+        assert self.hparams["rotary_emb_scale_base"] is None
+
+    def set_vocab(self) -> None:
+        if self._tokenizer_is_xlmroberta:
+            return self._xlmroberta_set_vocab()
+        return super().set_vocab()
+
+    def modify_tensors(self, data_torch: torch.Tensor, name: str, bid: int | None) -> Iterable[tuple[str, torch.Tensor]]:
+        # If the tensor is an experts bias tensor, skip it by returning an empty list.
+        if "mlp.experts.bias" in name:
+            return []  # Explicitly return an empty list.
+
+        if "mlp.experts.mlp.w1" in name:
+            data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"])
+            name += ".weight"
+
+        if "mlp.experts.mlp.w2" in name:
+            data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"])
+            data_torch = data_torch.transpose(1, 2)
+            name += ".weight"
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
+        if self.is_moe:
+            self.gguf_writer.add_moe_every_n_layers(self.hparams["moe_every_n_layers"])
+            self.gguf_writer.add_expert_count(self.hparams["num_experts"])
+            self.gguf_writer.add_expert_used_count(self.hparams["moe_top_k"])
+
+    def _is_tokenizer_xlmroberta(self) -> bool:
+        with open(self.dir_model / "tokenizer.json") as f:
+            tokenizer_json = json.load(f)
+        toktyp = tokenizer_json["model"]["type"]
+        if toktyp == "Unigram":
+            return True
+        if toktyp == "WordPiece":
+            return False
+        raise ValueError(f"unknown tokenizer: {toktyp}")
+
+
+@ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
+class XLMRobertaModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._xlmroberta_tokenizer_init()
+
+    def set_vocab(self):
+        self._xlmroberta_set_vocab()
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # if name starts with "roberta.", remove the prefix
         # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
@@ -5806,6 +5982,18 @@ def split_str_to_n_bytes(split_str: str) -> int:
     return n
 
 
+def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> str:
+    text_config = hparams.get("text_config", {})
+    vision_config = hparams.get("vision_config", {})
+    arch = hparams["architectures"][0]
+    # if "architectures" is found in the sub-config, use that instead
+    if model_type == ModelType.TEXT and text_config.get("architectures") is not None:
+        arch = text_config["architectures"][0]
+    elif model_type == ModelType.VISION and vision_config.get("architectures") is not None:
+        arch = vision_config["architectures"][0]
+    return arch
+
+
 def main() -> None:
     args = parse_args()
 
@@ -5858,16 +6046,16 @@ def main() -> None:
 
     logger.info(f"Loading model: {dir_model.name}")
 
-    hparams = ModelBase.load_hparams(dir_model)
-
     if args.mmproj:
         if "mmproj" not in fname_out.name:
             fname_out = ModelBase.add_prefix_to_filename(fname_out, "mmproj-")
 
     with torch.inference_mode():
         output_type = ftype_map[args.outtype]
-        model_architecture = hparams["architectures"][0]
         model_type = ModelType.VISION if args.mmproj else ModelType.TEXT
+        hparams = ModelBase.load_hparams(dir_model)
+        model_architecture = get_model_architecture(hparams, model_type)
+        logger.info(f"Model architecture: {model_architecture}")
         try:
             model_class = ModelBase.from_model_architecture(model_architecture, model_type=model_type)
         except NotImplementedError:

docs/development/HOWTO-add-model.md

@@ -9,10 +9,10 @@ Adding a model requires few steps:
 After following these steps, you can open PR.
 
 Also, it is important to check that the examples and main ggml backends (CUDA, METAL, CPU) are working with the new architecture, especially:
-- [main](/examples/main/)
-- [imatrix](/examples/imatrix/)
-- [quantize](/examples/quantize/)
-- [server](/examples/server/)
+- [main](/tools/main/)
+- [imatrix](/tools/imatrix/)
+- [quantize](/tools/quantize/)
+- [server](/tools/server/)
 
 ### 1. Convert the model to GGUF
 

docs/multimodal/MobileVLM.md

@@ -33,13 +33,13 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 2. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
 
 ```sh
-python ./examples/llava/llava_surgery.py -m path/to/MobileVLM-1.7B
+python ./tools/llava/llava_surgery.py -m path/to/MobileVLM-1.7B
 ```
 
 3. Use `convert_image_encoder_to_gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:
 
 ```sh
-python ./examples/llava/convert_image_encoder_to_gguf.py \
+python ./tools/llava/convert_image_encoder_to_gguf.py \
     -m path/to/clip-vit-large-patch14-336 \
     --llava-projector path/to/MobileVLM-1.7B/llava.projector \
     --output-dir path/to/MobileVLM-1.7B \
@@ -47,7 +47,7 @@ python ./examples/llava/convert_image_encoder_to_gguf.py \
 ```
 
 ```sh
-python ./examples/llava/convert_image_encoder_to_gguf.py \
+python ./tools/llava/convert_image_encoder_to_gguf.py \
     -m path/to/clip-vit-large-patch14-336 \
     --llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \
     --output-dir path/to/MobileVLM-1.7B_V2 \
@@ -69,10 +69,10 @@ Now both the LLaMA part and the image encoder is in the `MobileVLM-1.7B` directo
 
 ## Android compile and run
 ### compile
-refer to `examples/llava/android/build_64.sh`
+refer to `tools/llava/android/build_64.sh`
 ```sh
-mkdir examples/llava/android/build_64
-cd examples/llava/android/build_64
+mkdir tools/llava/android/build_64
+cd tools/llava/android/build_64
 ../build_64.sh
 ```
 ### run on Android

docs/multimodal/glmedge.md

@@ -25,13 +25,13 @@ git clone https://huggingface.co/THUDM/glm-edge-v-5b or https://huggingface.co/T
 2. Use `glmedge-surgery.py` to split the GLMV-EDGE model to LLM and multimodel projector constituents:
 
 ```sh
-python ./examples/llava/glmedge-surgery.py -m ../model_path
+python ./tools/llava/glmedge-surgery.py -m ../model_path
 ```
 
 4. Use `glmedge-convert-image-encoder-to-gguf.py` to convert the GLMV-EDGE image encoder to GGUF:
 
 ```sh
-python ./examples/llava/glmedge-convert-image-encoder-to-gguf.py -m ../model_path --llava-projector ../model_path/glm.projector --output-dir ../model_path
+python ./tools/llava/glmedge-convert-image-encoder-to-gguf.py -m ../model_path --llava-projector ../model_path/glm.projector --output-dir ../model_path
 ```
 
 5. Use `examples/convert_hf_to_gguf.py` to convert the LLM part of GLMV-EDGE to GGUF:

docs/multimodal/llava.md

@@ -37,19 +37,19 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 2. Install the required Python packages:
 
 ```sh
-pip install -r examples/llava/requirements.txt
+pip install -r tools/llava/requirements.txt
 ```
 
 3. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
 
 ```sh
-python ./examples/llava/llava_surgery.py -m ../llava-v1.5-7b
+python ./tools/llava/llava_surgery.py -m ../llava-v1.5-7b
 ```
 
 4. Use `convert_image_encoder_to_gguf.py` to convert the LLaVA image encoder to GGUF:
 
 ```sh
-python ./examples/llava/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
+python ./tools/llava/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
 ```
 
 5. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
@@ -69,12 +69,12 @@ git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
 2) Install the required Python packages:
 
 ```sh
-pip install -r examples/llava/requirements.txt
+pip install -r tools/llava/requirements.txt
 ```
 
 3) Use `llava_surgery_v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
 ```console
-python examples/llava/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/
+python tools/llava/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/
 ```
 - you will find a llava.projector and a llava.clip file in your model directory
 
@@ -88,7 +88,7 @@ curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.jso
 
 5) Create the visual gguf model:
 ```console
-python ./examples/llava/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
+python ./tools/llava/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
 ```
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
 

docs/multimodal/minicpmo2.6.md

@@ -29,8 +29,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-2_6-gguf) by us)
 
 ```bash
-python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-o-2_6
-python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4
+python ./tools/llava/minicpmv-surgery.py -m ../MiniCPM-o-2_6
+python ./tools/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4
 python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model
 
 # quantize int4 version

docs/multimodal/minicpmv2.5.md

@@ -28,8 +28,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us)
 
 ```bash
-python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
-python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
+python ./tools/llava/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
+python ./tools/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
 python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model
 
 # quantize int4 version

docs/multimodal/minicpmv2.6.md

@@ -28,8 +28,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us)
 
 ```bash
-python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-V-2_6
-python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3
+python ./tools/llava/minicpmv-surgery.py -m ../MiniCPM-V-2_6
+python ./tools/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3
 python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model
 
 # quantize int4 version

examples/CMakeLists.txt

@@ -12,51 +12,30 @@ llama_add_compile_flags()
 
 # examples
 
-include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-
 if (EMSCRIPTEN)
 else()
-    add_subdirectory(batched-bench)
     add_subdirectory(batched)
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
 
     add_subdirectory(gguf-hash)
-    add_subdirectory(gguf-split)
     add_subdirectory(gguf)
     add_subdirectory(gritlm)
-    add_subdirectory(imatrix)
     add_subdirectory(infill)
-    add_subdirectory(llama-bench)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
-    add_subdirectory(main)
     add_subdirectory(parallel)
     add_subdirectory(passkey)
-    add_subdirectory(perplexity)
-    add_subdirectory(quantize)
     add_subdirectory(retrieval)
-    if (LLAMA_BUILD_SERVER)
-        add_subdirectory(server)
-    endif()
     add_subdirectory(save-load-state)
-    add_subdirectory(run)
     add_subdirectory(simple)
     add_subdirectory(simple-chat)
     add_subdirectory(speculative)
     add_subdirectory(speculative-simple)
-    add_subdirectory(tokenize)
-    add_subdirectory(tts)
     add_subdirectory(gen-docs)
     if (NOT GGML_BACKEND_DL)
-        # these examples use the backends directly and cannot be built with dynamic loading
         add_subdirectory(convert-llama2c-to-ggml)
-        add_subdirectory(cvector-generator)
-        add_subdirectory(export-lora)
-        add_subdirectory(llava)
-        if (GGML_RPC)
-            add_subdirectory(rpc)
-        endif()
+        # these examples use the backends directly and cannot be built with dynamic loading
         if (GGML_SYCL)
             add_subdirectory(sycl)
         endif()

examples/llava/qwen2_vl_surgery.py

@@ -1,217 +0,0 @@
-import argparse
-from typing import Dict, List, Optional
-
-import torch
-import numpy as np
-from gguf import *
-from transformers import (
-    AutoProcessor,
-    Qwen2VLConfig,
-    Qwen2VLProcessor,
-    Qwen2VLForConditionalGeneration,
-    Qwen2_5_VLConfig, # type: ignore[reportAttributeAccessIssue]
-    Qwen2_5_VLForConditionalGeneration, # type: ignore[reportAttributeAccessIssue]
-)
-
-
-VISION = "clip.vision"
-
-
-def k(raw_key: str, arch: str) -> str:
-    return raw_key.format(arch=arch)
-
-
-def get_n_wa_pattern(fullatt_block_indexes: Optional[List[int]]):
-    if fullatt_block_indexes is None:
-        return 0
-    n_wa = fullatt_block_indexes[0]
-    for a, b in zip(fullatt_block_indexes, fullatt_block_indexes[1:]):
-        if b - a - 1 != n_wa:
-            raise ValueError(
-                f"window/full attention layer should have fix pattern of "
-                f"for each full-attention layer followed by {n_wa} window-attention layers"
-            )
-    return n_wa + 1
-
-
-class VL2:
-
-    @staticmethod
-    def to_gguf_name(name: str) -> str:
-        og = name
-        name = name.replace("text_model", "t").replace("vision_model", "v")
-        name = name.replace("blocks", "blk").replace("embeddings.", "")
-        name = name.replace("attn.", "attn_")
-        name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.")
-        # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln")
-        name = name.replace("norm1", "ln1").replace("norm2", "ln2")
-        name = name.replace("merger.mlp", 'mm')
-        print(f"[to_gguf_name] {og} --> {name}")
-        return name
-
-    @classmethod
-    def find_vision_tensors(cls, qwen2vl, dtype) -> Dict[str, np.ndarray]:
-        vision_model = qwen2vl.visual
-        tensor_map = {}
-        for name, ten in vision_model.state_dict().items():
-            ten = ten.numpy()
-            if 'qkv' in name:
-                if ten.ndim == 2: # weight
-                    c3, _ = ten.shape
-                else:             # bias
-                    c3 = ten.shape[0]
-                assert c3 % 3 == 0
-                c = c3 // 3
-                wq = ten[:c]
-                wk = ten[c: c * 2]
-                wv = ten[c * 2:]
-                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq
-                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk
-                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv
-            elif 'merger' in name:
-                if name.endswith("ln_q.weight"):
-                    tensor_map['v.post_ln.weight'] = ten
-                elif name.endswith("ln_q.bias"):
-                    tensor_map['v.post_ln.bias'] = ten
-                else:
-                    # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias"
-                    tensor_map[cls.to_gguf_name(name)] = ten
-            elif 'patch_embed.proj.weight' in name:
-                # NOTE: split Conv3D into Conv2Ds
-                c1, c2, kt, kh, kw = ten.shape
-                assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
-                tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...]
-                tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...]
-            else:
-                tensor_map[cls.to_gguf_name(f"vision_model.{name}")] = ten
-
-        for new_name, ten in tensor_map.items():
-            if ten.ndim <= 1 or new_name.endswith("_norm.weight"):
-                tensor_map[new_name] = ten.astype(np.float32)
-            else:
-                tensor_map[new_name] = ten.astype(dtype)
-        tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32)  # dummy tensor, just here as a placeholder
-        return tensor_map
-
-
-class VL25(VL2):
-
-    @staticmethod
-    def to_gguf_name(name: str) -> str:
-        og = name
-        name = name.replace("text_model", "t").replace("vision_model", "v")
-        name = name.replace("blocks", "blk").replace("embeddings.", "")
-        name = name.replace("attn.", "attn_")
-        name = name.replace("mlp.down_proj", "ffn_down").replace("mlp.up_proj", "ffn_up")
-        name = name.replace("mlp.gate_proj", "ffn_gate").replace("proj.", "out.")
-        name = name.replace("norm1", "ln1").replace("norm2", "ln2")
-        name = name.replace("merger.mlp", 'mm')
-        print(f"[vl25][to_gguf_name] {og} --> {name}")
-        return name
-
-
-def main(args):
-    if args.data_type == 'fp32':
-        dtype = torch.float32
-        np_dtype = np.float32
-        ftype = 0
-    elif args.data_type == 'fp16':
-        dtype = torch.float16
-        np_dtype = np.float16
-        ftype = 1
-    else:
-        raise ValueError()
-
-    local_model = False
-    model_path = ""
-    model_name = args.model_name
-    print("model_name: ", model_name)
-    if args.model_type == "qwen2vl":
-        qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
-            model_name, torch_dtype=dtype, device_map="cpu"
-        )
-        cfg: Qwen2VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
-        vcfg = cfg.vision_config
-    else:
-        qwen2vl = Qwen2_5_VLForConditionalGeneration.from_pretrained(
-            model_name, torch_dtype=dtype, device_map="cpu"
-        )
-        cfg: Qwen2_5_VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
-        vcfg = cfg.vision_config
-
-    if os.path.isdir(model_name):
-        local_model = True
-        if model_name.endswith(os.sep):
-            model_name = model_name[:-1]
-        model_path = model_name
-        model_name = os.path.basename(model_name)
-    fname_out = f"{model_name.replace('/', '-').lower()}-vision.gguf"
-
-    fout = GGUFWriter(path=fname_out, arch="clip")
-    fout.add_description("image encoder for Qwen2VL")
-
-    fout.add_file_type(ftype)
-    fout.add_bool("clip.has_text_encoder", False)
-    fout.add_bool("clip.has_vision_encoder", True)
-    fout.add_bool("clip.has_qwen2vl_merger", True)
-
-    print(cfg.vision_config)
-    if 'silu' in cfg.vision_config.hidden_act.lower():
-        fout.add_bool("clip.use_silu", True)
-        fout.add_bool("clip.use_gelu", False)
-    elif 'gelu' in cfg.vision_config.hidden_act.lower():
-        fout.add_bool("clip.use_silu", False)
-        fout.add_bool("clip.use_gelu", 'quick' not in cfg.vision_config.hidden_act.lower())
-    else:
-        raise ValueError()
-
-    if args.model_type == "qwen2.5vl":
-        fout.add_uint32("clip.vision.n_wa_pattern", get_n_wa_pattern(vcfg.fullatt_block_indexes))
-        fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.hidden_size)
-        fout.add_uint32("clip.vision.projection_dim", vcfg.out_hidden_size)
-        fout.add_string("clip.projector_type", "qwen2.5vl_merger")
-    else:
-        fout.add_string("clip.projector_type", "qwen2vl_merger")
-        fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim)
-        fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size)
-
-    if args.model_type == "qwen2.5vl":
-        tensor_map = VL25.find_vision_tensors(qwen2vl, np_dtype)
-    else:
-        tensor_map = VL2.find_vision_tensors(qwen2vl, np_dtype)
-    for name, data in tensor_map.items():
-        fout.add_tensor(name, data)
-
-    fout.add_uint32("clip.vision.patch_size", vcfg.patch_size)
-    fout.add_uint32("clip.vision.image_size", 14 * 40)  # some reasonable size that is divable by (14*2)
-    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), vcfg.num_heads)
-    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
-    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), vcfg.depth)
-    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), 0)  # not sure what this does, put 0 here as a placeholder
-    fout.add_name(model_name)
-    """
-    HACK: Since vision rope related parameter aren't stored in the `Qwen2VLConfig,
-            it will be hardcoded in the `clip_image_build_graph` from `clip.cpp`.
-    """
-
-    if local_model:
-        processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_path)
-    else:
-        processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_name)
-    fout.add_array("clip.vision.image_mean", processor.image_processor.image_mean) # type: ignore[reportAttributeAccessIssue]
-    fout.add_array("clip.vision.image_std", processor.image_processor.image_std) # type: ignore[reportAttributeAccessIssue]
-
-    fout.write_header_to_file()
-    fout.write_kv_data_to_file()
-    fout.write_tensors_to_file()
-    fout.close()
-    print("save model as: ", fname_out)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument("model_name", nargs='?', default="Qwen/Qwen2-VL-2B-Instruct")
-    parser.add_argument("--model_type", nargs='?', choices=['qwen2vl', 'qwen2.5vl'], default="qwen2vl")
-    parser.add_argument("--data_type", nargs='?', choices=['fp32', 'fp16'], default="fp32")
-    args = parser.parse_args()
-    main(args)

examples/pydantic_models_to_grammar_examples.py

@@ -23,7 +23,7 @@ def create_completion(host, prompt, gbnf_grammar):
     """Calls the /completion API on llama-server.
 
     See
-    https://github.com/ggml-org/llama.cpp/tree/HEAD/examples/server#api-endpoints
+    https://github.com/ggml-org/llama.cpp/tree/HEAD/tools/server#api-endpoints
     """
     print(f"  Request:\n    Grammar:\n{textwrap.indent(gbnf_grammar, '      ')}\n    Prompt:\n{textwrap.indent(prompt.rstrip(), '      ')}")
     headers = {"Content-Type": "application/json"}

ggml/CMakeLists.txt

@@ -360,3 +360,27 @@ write_basic_package_version_file(
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
               ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)
+
+if (MSVC)
+    set(MSVC_WARNING_FLAGS
+        /wd4005  # Macro redefinition
+        /wd4244  # Conversion from one type to another type, possible loss of data
+        /wd4267  # Conversion from 'size_t' to a smaller type, possible loss of data
+    )
+    function(disable_msvc_warnings target_name)
+        if(TARGET ${target_name})
+            target_compile_options(${target_name} PRIVATE ${MSVC_WARNING_FLAGS})
+        endif()
+    endfunction()
+
+    disable_msvc_warnings(ggml-base)
+    disable_msvc_warnings(ggml)
+    disable_msvc_warnings(ggml-cpu)
+    disable_msvc_warnings(ggml-cpu-x64)
+    disable_msvc_warnings(ggml-cpu-sse42)
+    disable_msvc_warnings(ggml-cpu-sandybridge)
+    disable_msvc_warnings(ggml-cpu-haswell)
+    disable_msvc_warnings(ggml-cpu-skylakex)
+    disable_msvc_warnings(ggml-cpu-icelake)
+    disable_msvc_warnings(ggml-cpu-alderlake)
+endif()

ggml/include/ggml-cpp.h

@@ -24,7 +24,7 @@ typedef std::unique_ptr<gguf_context, gguf_context_deleter> gguf_context_ptr;
 
 struct ggml_gallocr_deleter { void operator()(ggml_gallocr_t galloc) { ggml_gallocr_free(galloc); } };
 
-typedef std::unique_ptr<ggml_gallocr_t, ggml_gallocr_deleter> ggml_gallocr_ptr;
+typedef std::unique_ptr<ggml_gallocr, ggml_gallocr_deleter> ggml_gallocr_ptr;
 
 // ggml-backend
 

ggml/src/ggml-alloc.c

@@ -816,7 +816,10 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
 static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
     size_t node_size = 0;
     if (!node->data && !node->view_src) {
-        GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
+        // If we previously had data but don't now then reallocate
+        if (talloc->buffer_id < 0) {
+            return false;
+        }
         node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
     }
     return talloc->size_max >= node_size;

ggml/src/ggml-cpu/CMakeLists.txt

@@ -352,10 +352,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         # TODO: Separation to determine activation of VX/VXE/VXE2
         if (${S390X_M} MATCHES "8561|8562")
             message(STATUS "z15 target")
-            list(APPEND ARCH_FLAGS -march=z15 -mtune=z15)
+            list(APPEND ARCH_FLAGS -march=z15)
         elseif (${S390X_M} MATCHES "3931")
             message(STATUS "z16 target")
-            list(APPEND ARCH_FLAGS -march=z16 -mtune=z16)
+            list(APPEND ARCH_FLAGS -march=z16)
+        elseif (${S390X_M} MATCHES "9175|9176")
+            # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
+            message(STATUS "z17 target")
+            list(APPEND ARCH_FLAGS -march=z17)
         else()
             message(STATUS "Unknown target")
             message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -1054,6 +1054,493 @@ class tinyBLAS_Q0_AVX {
       } \
    } \
 
+template <typename TA, typename TB, typename TC>
+class tinyBLAS_BF16_PPC {
+  public:
+    tinyBLAS_BF16_PPC(int64_t k,
+                const TA *A, int64_t lda,
+                const TB *B, int64_t ldb,
+                TC *C, int64_t ldc,
+                int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+        mnpack(0, m, 0, n);
+    }
+
+  private:
+    void vector_permute_store(vec_t *c, int numVec, unsigned char *vecOffset) {
+        vec_t t[8], s[8];
+        vec_t swiz1 = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23};
+        vec_t swiz2 = {8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31};
+        vec_t swiz3 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        vec_t swiz4 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+
+        if (numVec == 2) {
+            t[0] = vec_perm(c[0], c[1], swiz1);
+            t[1] = vec_perm(c[2], c[3], swiz1);
+            s[0] = vec_perm(t[0], t[1], swiz3);
+            s[1] = vec_perm(t[0], t[1], swiz4);
+            vec_xst(s[0], 0, (vec_t*)vecOffset);
+            vec_xst(s[1], 0, (vec_t*)(vecOffset + 16));
+        } else if (numVec == 4) {
+            t[0] = vec_perm(c[0], c[1], swiz1);
+            t[1] = vec_perm(c[0], c[1], swiz2);
+            t[2] = vec_perm(c[2], c[3], swiz1);
+            t[3] = vec_perm(c[2], c[3], swiz2);
+            s[0] = vec_perm(t[0], t[2], swiz3);
+            s[1] = vec_perm(t[0], t[2], swiz4);
+            s[2] = vec_perm(t[1], t[3], swiz3);
+            s[3] = vec_perm(t[1], t[3], swiz4);
+            for (int i = 0; i < 4; ++i)
+                vec_xst(s[i], 0, (vec_t*)(vecOffset + i * 16));
+        } else if (numVec == 8) {
+            for (int i = 0; i < 4; i += 2) {
+                t[i+0] = vec_perm(c[i+0], c[i+1], swiz1);
+                t[i+1] = vec_perm(c[i+0], c[i+1], swiz2);
+            }
+            for (int i = 4; i < 8; i += 2) {
+                t[i+0] = vec_perm(c[i+0], c[i+1], swiz1);
+                t[i+1] = vec_perm(c[i+0], c[i+1], swiz2);
+            }
+            s[0] = vec_perm(t[0], t[2], swiz3);
+            s[1] = vec_perm(t[0], t[2], swiz4);
+            s[2] = vec_perm(t[1], t[3], swiz3);
+            s[3] = vec_perm(t[1], t[3], swiz4);
+            s[4] = vec_perm(t[4], t[6], swiz3);
+            s[5] = vec_perm(t[4], t[6], swiz4);
+            s[6] = vec_perm(t[5], t[7], swiz3);
+            s[7] = vec_perm(t[5], t[7], swiz4);
+            for (int i = 0; i < 8; ++i)
+                vec_xst(s[i], 0, (vec_t*)(vecOffset + i * 16));
+        }
+    }
+
+    void packNormal(const TA* a, int64_t lda, int rows, int cols, unsigned char* vec) {
+        int64_t i, j;
+        TA *aoffset = NULL;
+        unsigned char *vecOffset = NULL;
+        TA * aoffsets[8];
+        vector unsigned char c_arr[8];
+        aoffset = const_cast<TA*>(a);
+        vecOffset = vec;
+        j = (rows >> 3);
+        if (j > 0) {
+            do {
+                if (cols == 4) {
+                    aoffsets[0] = aoffset;
+                    for (int it = 1; it < 4; ++it)
+                        aoffsets[it] = aoffsets[it-1] + lda;
+                    aoffset += 4 * lda;
+                    for (int i = 0; i < 4; ++i)
+                        c_arr[i] = vec_xl(0, (vector unsigned char*)aoffsets[i]);
+                    vector_permute_store(c_arr, 4, vecOffset);
+                    for (int i = 0; i<4; i++)
+                        aoffsets[i] = aoffsets[i]+lda;
+                    vecOffset +=64;
+                }
+                i = (cols >> 3);
+                if (i > 0) {
+                    aoffsets[0] = aoffset;
+                    for (int it = 1; it < 8; ++it) {
+                        aoffsets[it] = aoffsets[it-1] + lda;
+                    }
+                    aoffset += 8 * lda;
+                    do {
+                        for (int it = 0; it < 8; ++it)
+                            c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]);
+                        vector_permute_store(c_arr, 8, vecOffset);
+                        for (int it = 0; it < 8; ++it)
+                            aoffsets[it] = aoffsets[it] + 8*lda;
+                        vecOffset += 128;
+                        i--;
+                    } while(i > 0);
+                }
+                j--;
+            } while(j > 0);
+        }
+        if (rows & 4) {
+            aoffsets[0] = aoffset;
+            for (int it = 1; it < 4; ++it)
+                aoffsets[it] = aoffsets[it-1] + lda;
+            aoffset += 4 * lda;
+            if (cols == 4) {
+                for (int it = 0; it < 4; ++it)
+                    c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]);
+                vector_permute_store(c_arr, 2, vecOffset);
+                for (int it = 0; it< 4; it++)
+                    aoffsets[it] = aoffsets[it] + lda;
+                vecOffset += 32;
+            }
+            i = (cols >> 3);
+            if (i > 0) {
+                do {
+                    for (int it = 0; it < 4; ++it)
+                        c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]);
+                    vector_permute_store(c_arr, 4, vecOffset);
+                    for (int it = 0; it< 4; it++)
+                        aoffsets[it] = aoffsets[it] + 8*lda;
+                    vecOffset += 64;
+                    i--;
+                } while(i > 0);
+            }
+        }
+        if (rows & 3) {
+            aoffsets[0] = aoffset;
+            for (int it = 1; it < 4; ++it)
+                aoffsets[it] = aoffsets[it-1] + lda;
+            if (cols == 4) {
+                switch(rows) {
+                    case 3: c_arr[2] = vec_xl(0, (vector unsigned char*)aoffsets[2]);
+                    case 2: c_arr[1] = vec_xl(0, (vector unsigned char*)aoffsets[1]);
+                    case 1: c_arr[0] = vec_xl(0, (vector unsigned char*)aoffsets[0]);
+                        break;
+                }
+                vector_permute_store(c_arr, 2, vecOffset);
+                for (int it = 0; it< 4; it++)
+                     aoffsets[it] = aoffsets[it] + lda;
+                vecOffset += 32;
+            }
+            i = (cols >> 3);
+            if (i > 0) {
+                do {
+                    switch(rows) {
+                        case 3: c_arr[2] = vec_xl(0, (vector unsigned char*)aoffsets[2]);
+                        case 2: c_arr[1] = vec_xl(0, (vector unsigned char*)aoffsets[1]);
+                        case 1: c_arr[0] = vec_xl(0, (vector unsigned char*)aoffsets[0]);
+                            break;
+                    }
+                    vector_permute_store(c_arr, 4, vecOffset);
+                    for (int it = 0; it <4; it++)
+                         aoffsets[it] = aoffsets[it] + 8* lda;
+                    vecOffset += 64;
+                    i--;
+                } while(i > 0);
+            }
+        }
+    }
+
+    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        int m_rem = MIN(m - m0, 8);
+        int n_rem = MIN(n - n0, 8);
+
+        if (m_rem >= 8 && n_rem >= 8) {
+            mc = 8;
+            nc = 8;
+            gemm<8,8>(m0, m, n0, n);
+        } else if (m_rem >= 4 && n_rem >= 8) {
+            mc = 4;
+            nc = 8;
+            gemm<4,8>(m0, m, n0, n);
+        } else if (m_rem >=8 && n_rem >=4){
+                mc = 8;
+                nc = 4;
+                gemm<8,4>(m0, m, n0, n);
+        } else if ((m_rem < 4) && (n_rem >= 8)) {
+            nc = 8;
+            switch(m_rem) {
+                case 1:
+                    mc = 1;
+                    gemm_Mx8<1>(m0, m, n0, n);
+                    break;
+                case 2:
+                    mc = 2;
+                    gemm_Mx8<2>(m0, m, n0, n);
+                    break;
+                case 3:
+                    mc = 3;
+                    gemm_Mx8<3>(m0, m, n0, n);
+                    break;
+                default:
+                    return;
+            }
+        } else if (m_rem >= 4 && n_rem >= 4) {
+            mc = 4;
+            nc = 4;
+            gemm_small<4, 4>(m0, m, n0, n);
+        } else if ((m_rem > 4) && (n_rem < 4)) {
+            mc = 4;
+            switch(n_rem) {
+                case 1:
+                    nc = 1;
+                    gemm_small<4, 1>(m0, m, n0, n);
+                    break;
+                case 2:
+                    nc = 2;
+                    gemm_small<4, 2>(m0, m, n0, n);
+                    break;
+                case 3:
+                    nc = 3;
+                    gemm_small<4, 3>(m0, m, n0, n);
+                    break;
+
+                default:
+                    return;
+            }
+        } else {
+            switch((m_rem << 4) | n_rem) {
+                case 0x43:
+                    mc = 4;
+                    nc = 3;
+                    gemm_small<4, 3>(m0, m, n0, n);
+                    break;
+                case 0x42:
+                    mc = 4;
+                    nc = 2;
+                    gemm_small<4, 2>(m0, m, n0, n);
+                    break;
+                case 0x41:
+                    mc = 4;
+                    nc = 1;
+                    gemm_small<4, 1>(m0, m, n0, n);
+                    break;
+                case 0x34:
+                    mc = 3;
+                    nc = 4;
+                    gemm_small<3, 4>(m0, m, n0, n);
+                    break;
+                case 0x33:
+                    mc = 3;
+                    nc = 3;
+                    gemm_small<3, 3>(m0, m, n0, n);
+                    break;
+                case 0x32:
+                    mc = 3;
+                    nc = 2;
+                    gemm_small<3, 2>(m0, m, n0, n);
+                    break;
+                case 0x31:
+                    mc = 3;
+                    nc = 1;
+                    gemm_small<3, 1>(m0, m, n0, n);
+                    break;
+                case 0x24:
+                    mc = 2;
+                    nc = 4;
+                    gemm_small<2,4>(m0, m, n0, n);
+                    break;
+                case 0x23:
+                    mc = 2;
+                    nc = 3;
+                    gemm_small<2, 3>(m0, m, n0, n);
+                    break;
+                case 0x22:
+                    mc = 2;
+                    nc = 2;
+                    gemm_small<2, 2>(m0, m, n0, n);
+                    break;
+                case 0x21:
+                    mc = 2;
+                    nc = 1;
+                    gemm_small<2, 1>(m0, m, n0, n);
+                    break;
+                case 0x14:
+                    mc = 1;
+                    nc = 4;
+                    gemm_small<1, 4>(m0, m, n0, n);
+                    break;
+                case 0x13:
+                    mc = 1;
+                    nc = 3;
+                    gemm_small<1, 3>(m0, m, n0, n);
+                    break;
+                case 0x12:
+                    mc = 1;
+                    nc = 2;
+                    gemm_small<1, 2>(m0, m, n0, n);
+                    break;
+                case 0x11:
+                    mc = 1;
+                    nc = 1;
+                    gemm_small<1, 1>(m0, m, n0, n);
+                    break;
+                default:
+                    return;
+            }
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+    void KERNEL_4x8(int64_t ii, int64_t jj) {
+        vec_t vec_A[4], vec_B[8] , vec_C[4];
+        acc_t acc_0, acc_1;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        for (int l = 0; l < k; l+=8) {
+            packNormal((A+(ii*lda)+l), lda, 4, 8, (uint8_t*)vec_A);
+            packNormal((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B);
+            for (int x = 0; x < 4; x++) {
+                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]);
+            }
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii, jj+4);
+    }
+
+    void KERNEL_8x4(int64_t ii, int64_t jj) {
+        vec_t vec_A[8], vec_B[4] , vec_C[4];
+        acc_t acc_0, acc_1;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        for (int l = 0; l < k; l+=8) {
+            packNormal((A+(ii*lda)+l), lda, 8, 8, (uint8_t*)vec_A);
+            packNormal((B+(jj*ldb)+l), ldb, 8, 4, (uint8_t*)vec_B);
+            for (int x = 0; x < 4; x++) {
+                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x+4], vec_B[x]);
+            }
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii+4, jj);
+    }
+
+
+    void KERNEL_8x8(int64_t ii, int64_t jj) {
+        vec_t vec_A[8], vec_B[8], vec_C[4];
+        acc_t acc_0, acc_1, acc_2, acc_3;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        __builtin_mma_xxsetaccz(&acc_2);
+        __builtin_mma_xxsetaccz(&acc_3);
+        for (int l = 0; l < k; l+=8) {
+            packNormal(A+(ii*lda)+l, lda, 8, 8, (uint8_t*)vec_A);
+            packNormal(B+(jj*ldb)+l, ldb, 8, 8, (uint8_t*)vec_B);
+            for (int x = 0; x < 4; x++) {
+                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvbf16ger2pp(&acc_1, (vec_t)vec_A[x], (vec_t)vec_B[x+4]);
+                __builtin_mma_xvbf16ger2pp(&acc_2, (vec_t)vec_A[x+4], (vec_t)vec_B[x]);
+                __builtin_mma_xvbf16ger2pp(&acc_3, (vec_t)vec_A[x+4], (vec_t)vec_B[x+4]);
+            }
+        }
+
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii, jj+4);
+        SAVE_ACC(&acc_2, ii+4, jj);
+        SAVE_ACC(&acc_3, ii+4, jj+4);
+    }
+
+    template<int RM, int RN>
+    void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            vec_t vec_C[4];
+            acc_t acc_0;
+            __builtin_mma_xxsetaccz(&acc_0);
+            vec_t vec_A[2], vec_B[2];
+            for (int l=0; l<k; l+=4) {
+                packNormal(A+(ii*lda)+l, lda, RM, 4, (uint8_t*)vec_A);
+                packNormal(B+(jj*ldb)+l, ldb, RN, 4, (uint8_t*)vec_B);
+                for (int x = 0; x<2; x++) {
+                    __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                }
+            }
+            __builtin_mma_disassemble_acc(vec_C, &acc_0);
+            for (int I = 0; I < RM; I++) {
+                for (int J = 0; J < RN; J++) {
+                    *((TC*)(C+ii+((jj+J)*ldc)+I)) = *((TC*)&vec_C[I]+J);
+                }
+            }
+        }
+    }
+
+    template<int RM>
+    void gemm_Mx8(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int RN = 8;
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            vec_t vec_C[4];
+            acc_t acc_0, acc_1;
+            __builtin_mma_xxsetaccz(&acc_0);
+            __builtin_mma_xxsetaccz(&acc_1);
+            vec_t vec_A[4], vec_B[8];
+            for (int l=0; l<k; l+=8) {
+                packNormal(A+(ii*lda)+l, lda, RM, 8, (uint8_t*)vec_A);
+                packNormal(B+(jj*ldb)+l, ldb, RN, 8, (uint8_t*)vec_B);
+                for (int x = 0; x<4; x++) {
+                    __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                    __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]);
+                }
+            }
+            __builtin_mma_disassemble_acc(vec_C, &acc_0);
+            for (int I = 0; I < RM; I++) {
+                for (int J = 0; J < 4; J++) {
+                    *((TC*)(C+ii+((jj+J)*ldc)+I)) = *((TC*)&vec_C[I]+J);
+                }
+            }
+            __builtin_mma_disassemble_acc(vec_C, &acc_1);
+            for (int I = 0; I < RM; I++) {
+                for (int J = 0; J < 4; J++) {
+                    *((TC*)(C+ii+((jj+4+J)*ldc)+I)) = *((TC*)&vec_C[I]+J);
+                }
+            }
+        }
+    }
+
+    template<int RM, int RN>
+    inline void kernel(int64_t ii, int64_t jj) {
+       if constexpr(RM == 4 && RN == 8) {
+          KERNEL_4x8(ii,jj);
+       } else if constexpr(RM == 8 && RN == 8) {
+          KERNEL_8x8(ii,jj);
+       } else if constexpr(RM == 8 && RN == 4) {
+          KERNEL_8x4(ii,jj);
+       } else {
+          static_assert(false, "RN/RM values not supported");
+       }
+    }
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            kernel<RM, RN>(ii, jj);
+        }
+    }
+
+    const TA *const A;
+    const TB *const B;
+    TC *C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+
 template <typename TA, typename TB, typename TC>
 class tinyBLAS_Q0_PPC {
   public:
@@ -2202,6 +2689,7 @@ class tinyBLAS_PPC {
         boffset = vec;
         j = (rows >> 3);
         if (j > 0) {
+
             do {
                 aoffset1 = aoffset;
                 aoffset2 = aoffset1 + lda;
@@ -2875,9 +3363,22 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
                 (float *)C, ldc};
             return tb.matmul(m, n);
         }
+#elif defined(__MMA__)
+        if ((k % 8))
+                return false;
+        if(Btype == GGML_TYPE_BF16) {
+           tinyBLAS_BF16_PPC<ggml_bf16_t, ggml_bf16_t, float> tb{ k,
+            (const ggml_bf16_t *)A, lda,
+            (const ggml_bf16_t *)B, ldb,
+            (float *)C, ldc,
+            params->ith, params->nth};
+        tb.matmul(m, n);
+        return true;
+        }
 #endif
         return false;
     }
+
     case GGML_TYPE_F16: {
 #if defined(__AVX512F__)
         if (Btype == GGML_TYPE_F16) {

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

@@ -341,7 +341,7 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
 #define GGML_F32_EPR  4
 
 #define GGML_F32x4              vector float
-#define GGML_F32x4_ZERO         0.0f
+#define GGML_F32x4_ZERO         {0.0f}
 #define GGML_F32x4_SET1         vec_splats
 #define GGML_F32x4_LOAD(p)      vec_xl(0, p)
 #define GGML_F32x4_STORE(p, r)  vec_xst(r, 0, p)

ggml/src/ggml-cuda/CMakeLists.txt

@@ -133,6 +133,7 @@ if (CUDAToolkit_FOUND)
                 COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
                 OUTPUT_VARIABLE CUDA_CCVER
                 ERROR_QUIET
+                OUTPUT_STRIP_TRAILING_WHITESPACE
             )
         else()
             if (CUDA_CCFULLVER MATCHES Apple)
@@ -143,7 +144,7 @@ if (CUDAToolkit_FOUND)
             string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
         endif()
 
-        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+        message(STATUS "CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
 
         ggml_get_flags(${CUDA_CCID} ${CUDA_CCVER})
         list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later

ggml/src/ggml-cuda/convert.cu

@@ -1,6 +1,8 @@
 #include "convert.cuh"
 #include "dequantize.cuh"
 
+#include <cstdint>
+
 #define CUDA_Q8_0_NE_ALIGN 2048
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
@@ -570,30 +572,46 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t
 }
 
 template <typename src_t, typename dst_t>
-static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
-    const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
+static __global__ void convert_unary(
+        const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t s01, const int64_t s02, const int64_t s03) {
+    const int64_t i00 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
-    if (i >= k) {
+    if (i00 >= ne00) {
         return;
     }
 
+    const int64_t i01 = blockIdx.y;
+    const int64_t i02 = blockIdx.z % ne02;
+    const int64_t i03 = blockIdx.z / ne02;
+
     const src_t * x = (const src_t *) vx;
 
-    y[i] = float(x[i]);
+    const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
+    const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
+    y[iy] = float(x[ix]);
 }
 
 template <typename src_t, typename dst_t>
-static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
-    convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+static void convert_unary_cuda(const void * vx, dst_t * y,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
+    const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, ne02*ne03);
+    convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
+        (vx, y, ne00, ne01, ne02, s01, s02, s03);
+}
+
+template <typename src_t, typename dst_t>
+static void convert_unary_cont_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    convert_unary_cuda<src_t>(vx, y, k, 1, 1, 1, k, k, k, stream);
 }
 
 to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F32:
-            return convert_unary_cuda<float>;
+            return convert_unary_cont_cuda<float>;
         case GGML_TYPE_F16:
-            return convert_unary_cuda<half>;
+            return convert_unary_cont_cuda<half>;
         default:
             return nullptr;
     }
@@ -643,9 +661,9 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
         case GGML_TYPE_IQ3_S:
             return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_F32:
-            return convert_unary_cuda<float>;
+            return convert_unary_cont_cuda<float>;
         case GGML_TYPE_BF16:
-            return convert_unary_cuda<nv_bfloat16>;
+            return convert_unary_cont_cuda<nv_bfloat16>;
         default:
             return nullptr;
     }
@@ -692,7 +710,18 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
         case GGML_TYPE_IQ3_S:
             return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_F16:
-            return convert_unary_cuda<half>;
+            return convert_unary_cont_cuda<half>;
+        case GGML_TYPE_BF16:
+            return convert_unary_cont_cuda<nv_bfloat16>;
+        default:
+            return nullptr;
+    }
+}
+
+to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return convert_unary_cuda<float>;
         case GGML_TYPE_BF16:
             return convert_unary_cuda<nv_bfloat16>;
         default:

ggml/src/ggml-cuda/convert.cuh

@@ -3,7 +3,7 @@
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
 
 template<typename T>
-using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int64_t k, cudaStream_t stream);
+using to_t_cuda_t = void (*)(const void * x, T * y, int64_t k, cudaStream_t stream);
 
 typedef to_t_cuda_t<float> to_fp32_cuda_t;
 typedef to_t_cuda_t<half> to_fp16_cuda_t;
@@ -14,3 +14,13 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type);
 to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type);
 
 to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type);
+
+// TODO more general support for non-contiguous inputs
+
+template<typename T>
+using to_t_nc_cuda_t = void (*)(const void * x, T * y,
+    int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03,
+    int64_t s01, int64_t s02, int64_t s03, cudaStream_t stream);
+
+typedef to_t_nc_cuda_t<half> to_fp16_nc_cuda_t;
+to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);

ggml/src/ggml-cuda/cpy.cu

@@ -592,6 +592,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
         graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
     }
+#else
+    GGML_UNUSED(disable_indirection_for_this_node);
 #endif
     if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));

ggml/src/ggml-cuda/getrows.cu

@@ -33,8 +33,8 @@ static __global__ void k_get_rows(
     dfloat2 v;
     dequantize_kernel(src0_row, ib, iqs, v);
 
-    dst_row[iybs + iqs + 0]        = v.x;
-    dst_row[iybs + iqs + y_offset] = v.y;
+    dst_row[iybs + iqs + 0]        = float(v.x);
+    dst_row[iybs + iqs + y_offset] = float(v.y);
 }
 
 template<typename src0_t, typename dst_t>
@@ -60,7 +60,7 @@ static __global__ void k_get_rows_float(
     dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
     const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);
 
-    dst_row[i00] = src0_row[i00];
+    dst_row[i00] = float(src0_row[i00]);
 }
 
 template<typename grad_t, typename dst_t>
@@ -86,120 +86,159 @@ static __global__ void k_get_rows_back_float(
     dst[dst_row*ncols + col] = sum;
 }
 
-template<int qk, int qr, dequantize_kernel_t dq>
-static void get_rows_cuda(
-        const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-        const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
+template<int qk, int qr, dequantize_kernel_t dq, typename dst_t>
+static void get_rows_cuda_q(
+        const void * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
     const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
     const dim3 block_nums(block_num_x, ne10, ne11*ne12);
 
     // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
+    // const size_t s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);
 
-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);
 
     GGML_ASSERT(ne00 % 2 == 0);
 
     k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
-        src0_dd, src1_dd, dst_dd,
+        src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
         /*ne10, ne11,*/ ne12, /*ne13,*/
         /* s0,*/ s1, s2, s3,
         /* nb00,*/ nb01, nb02, nb03,
         s10, s11, s12/*, s13*/);
-
-    GGML_UNUSED(dst);
 }
 
-template<typename src0_t>
+template<typename src0_t, typename dst_t>
 static void get_rows_cuda_float(
-        const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-        const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    GGML_ASSERT(ne13 == 1);
-
+        const src0_t * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
     const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
     const dim3 block_nums(block_num_x, ne10, ne11*ne12);
 
     // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
+    // const size_t s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);
 
-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);
 
     k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
-        src0_dd, src1_dd, dst_dd,
+        src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
         /*ne10, ne11,*/ ne12, /*ne13,*/
         /* s0,*/ s1, s2, s3,
         /* nb00,*/ nb01, nb02, nb03,
         s10, s11, s12/*, s13*/);
+}
 
-    GGML_UNUSED(dst);
+template <typename dst_t>
+static void ggml_cuda_get_rows_switch_src0_type(
+        const void * src0_d, const ggml_type src0_type, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
+    switch (src0_type) {
+        case GGML_TYPE_F16:
+            get_rows_cuda_float((const half *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_F32:
+            get_rows_cuda_float((const float *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_BF16:
+            get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            get_rows_cuda_q<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            get_rows_cuda_q<QK4_1, QR4_1, dequantize_q4_1>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            get_rows_cuda_q<QK5_0, QR5_0, dequantize_q5_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            get_rows_cuda_q<QK5_1, QR5_1, dequantize_q5_1>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            get_rows_cuda_q<QK8_0, QR8_0, dequantize_q8_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        default:
+            // TODO: k-quants
+            GGML_ABORT("%s: unsupported src0 type: %s\n", __func__, ggml_type_name(src0_type));
+            break;
+    }
+}
+
+void get_rows_cuda(
+        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
+        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
+        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
+        size_t nb1, size_t nb2, size_t nb3,
+        cudaStream_t stream) {
+    switch (dst_type) {
+        case GGML_TYPE_F32:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (float *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_F16:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (half *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_BF16:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (nv_bfloat16 *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        default:
+            GGML_ABORT("%s: unsupported dst type: %s\n", __func__, ggml_type_name(dst_type));
+            break;
+    }
 }
 
 void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    const void    * src0_d = (const void    *) src0->data;
-    const int32_t * src1_d = (const int32_t *) src1->data;
-    float         * dst_d  = (float         *) dst->data;
-
     cudaStream_t stream = ctx.stream();
 
+    GGML_TENSOR_BINARY_OP_LOCALS
+
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(ne13 == 1);
 
     GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
     GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
     GGML_ASSERT(dst->nb[0]  == ggml_type_size(dst->type));
 
-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            get_rows_cuda_float(src0, src1, dst, (const half *) src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_F32:
-            get_rows_cuda_float(src0, src1, dst, (const float *) src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_0:
-            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        default:
-            // TODO: k-quants
-            GGML_ABORT("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
-            break;
-    }
+    get_rows_cuda(src0->data, src0->type, (const int32_t *) src1->data, dst->data, dst->type,
+        ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
 }
 
 void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/getrows.cuh

@@ -3,6 +3,13 @@
 #define CUDA_GET_ROWS_BLOCK_SIZE 256
 #define CUDA_GET_ROWS_BACK_BLOCK_SIZE 256
 
+void get_rows_cuda(
+        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
+        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
+        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
+        size_t nb1, size_t nb2, size_t nb3,
+        cudaStream_t stream);
+
 void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -1551,7 +1551,7 @@ static void ggml_cuda_op_mul_mat(
 
             if (src1_on_device && src1_is_contiguous) {
                 quantize_src1(
-                    dev[id].src1_ddf, dev[id].src1_ddq, src0->type, ne10,
+                    dev[id].src1_ddf, nullptr, dev[id].src1_ddq, src0->type, ne10,
                     nb11/sizeof(float), nb12/sizeof(float), nb13/sizeof(float),
                     src1_padded_col_size, ne11, ne12, ne13, stream);
                 CUDA_CHECK(cudaGetLastError());
@@ -1649,7 +1649,7 @@ static void ggml_cuda_op_mul_mat(
 
                 if (quantize_src1 && !src1_is_contiguous) {
                     quantize_src1(
-                        src1_ddf_i, src1_ddq_i, src0->type, ne10, ne10, ne11*ne10, ne12*ne11*ne10,
+                        src1_ddf_i, nullptr, src1_ddq_i, src0->type, ne10, ne10, ne11*ne10, ne12*ne11*ne10,
                         src1_padded_col_size, src1_ncols, 1, 1, stream);
                     CUDA_CHECK(cudaGetLastError());
                 }
@@ -1720,15 +1720,15 @@ static __global__ void k_compute_batched_ptrs(
         size_t  nb12, size_t  nb13,
         size_t  nbd2, size_t  nbd3,
         int64_t r2,   int64_t r3) {
-    int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
-    int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
+    const int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
+    const int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
 
     if (i13 >= ne13 || i12 >= ne12) {
         return;
     }
 
-    int64_t i03 = i13 / r3;
-    int64_t i02 = i12 / r2;
+    const int64_t i03 = i13 / r3;
+    const int64_t i02 = i12 / r2;
 
     ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
     ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
@@ -1742,6 +1742,10 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
 
+    // Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.
+    // As long as dst is contiguous this does not matter though.
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t ne_dst = ggml_nelements(dst);
@@ -1750,21 +1754,31 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 
     CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
 
-    void * src0_ddq = src0->data;
-    half * src0_f16 = (half *) src0_ddq;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf  = (float *) dst->data;
+    const half * src0_f16 = (const half *) src0->data;
+    float * dst_ddf = (float *) dst->data;
+
+    const half * src1_f16 = (const half *) src1->data;
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    GGML_ASSERT(nb10 == ts_src1);
+    int64_t s11 = nb11 / ts_src1;
+    int64_t s12 = nb12 / ts_src1;
+    int64_t s13 = nb13 / ts_src1;
+    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
 
     // convert src1 to fp16
-    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
     if (src1->type != GGML_TYPE_F16) {
-        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+        const to_fp16_nc_cuda_t to_fp16_cuda = ggml_get_to_fp16_nc_cuda(src1->type);
         const int64_t ne_src1 = ggml_nelements(src1);
         src1_f16_alloc.alloc(ne_src1);
         GGML_ASSERT(to_fp16_cuda != nullptr);
-        to_fp16_cuda(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
+
+        to_fp16_cuda(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
+
+        src1_f16 = src1_f16_alloc.get();
+        s11 = ne10;
+        s12 = ne11*s11;
+        s13 = ne12*s12;
     }
-    half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get();
 
     ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
     char * dst_t;
@@ -1824,13 +1838,13 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
                 int i02 = i12 / r2;
 
                 CUBLAS_CHECK(
-                        cublasGemmEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
-                            ne01, ne11, ne10,
-                            alpha, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , CUDA_R_16F,   nb01/sizeof(half),
-                                   (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F,   nb11/sizeof(float),
-                            beta,  (      char *)       dst_t + i12*nbd2          + i13*nbd3,          cu_data_type, ne01,
-                            cu_compute_type,
-                            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+                cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+                    ne01, ne11, ne10,
+                    alpha, (const char *) src0_f16 + i03*nb03 + i02*nb02, CUDA_R_16F,   nb01/sizeof(half),
+                                          src1_f16 + i13*s13  + i12*s12,  CUDA_R_16F,   s11,
+                    beta,  (      char *)    dst_t + i13*nbd3 + i12*nbd2, cu_data_type, ne0,
+                    cu_compute_type,
+                    CUBLAS_GEMM_DEFAULT_TENSOR_OP));
             }
         }
     }
@@ -1841,15 +1855,15 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
         CUBLAS_CHECK(
         cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                alpha, (const char *) src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00,  // strideA
-                       (const char *) src1_f16, CUDA_R_16F,   nb11/nb10, nb12/nb10,  // strideB
-                beta,  (      char *)    dst_t, cu_data_type, ne01,       nb2/nb0,   // strideC
+                alpha, src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00, // strideA
+                       src1_f16, CUDA_R_16F,   s11,       s12,       // strideB
+                beta,     dst_t, cu_data_type, ne0,       ne1*ne0,   // strideC
                 ne12*ne13,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
     } else {
         // use cublasGemmBatchedEx
-        const int ne23 = ne12*ne13;
+        const int64_t ne23 = ne12*ne13;
 
         ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
         ggml_cuda_pool_alloc<      void *> ptrs_dst(ctx.pool(), 1*ne23);
@@ -1861,8 +1875,8 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
                 ne12, ne13,
                 ne23,
                 nb02, nb03,
-                src1->type == GGML_TYPE_F16 ? nb12 : nb12/2,
-                src1->type == GGML_TYPE_F16 ? nb13 : nb13/2,
+                src1->type == GGML_TYPE_F16 ? nb12 : s12*sizeof(half),
+                src1->type == GGML_TYPE_F16 ? nb13 : s13*sizeof(half),
                 nbd2, nbd3,
                 r2, r3);
         CUDA_CHECK(cudaGetLastError());
@@ -1871,8 +1885,8 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
         cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
                 alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F,   nb01/nb00,
-                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   nb11/nb10,
-                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01,
+                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   s11,
+                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
                 ne23,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@@ -1935,8 +1949,10 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst);
     } else if (!split && use_mul_mat_vec_q) {
         ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
+    } else if (!split && use_mul_mat_q) {
+        ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst);
     } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
-            dst->op_params[0] == GGML_PREC_DEFAULT && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+            !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // general KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_mul_mat_vec) {
@@ -1950,183 +1966,145 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     }
 }
 
-struct mmid_row_mapping {
-    int32_t i1;
-    int32_t i2;
-};
-
-static __global__ void k_copy_src1_to_contiguous(const char * __restrict__ src1_original, char * __restrict__ src1_contiguous,
-                                                 int * __restrict__ cur_src1_row, mmid_row_mapping * __restrict__ row_mapping,
-                                                 const char * __restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0,
-                                                 int64_t ne11, int64_t ne10,
-                                                 size_t nb11, size_t nb12) {
-    int32_t iid1 = blockIdx.x;
-    int32_t id = blockIdx.y;
-
-    const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0);
-
-    if (row_id_i != i02) {
-        return;
-    }
-
-    const int64_t i11 = id % ne11;
-    const int64_t i12 = iid1;
-
-    __shared__ int src1_row;
-    if (threadIdx.x == 0) {
-        src1_row = atomicAdd(cur_src1_row, 1);
-        row_mapping[src1_row] = {id, iid1};
-    }
-    __syncthreads();
-
-    const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12);
-    float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11);
-
-    for (int i = threadIdx.x; i < ne10; i += blockDim.x) {
-        src1_row_contiguous[i] = src1_row_original[i];
-    }
-}
-
-static __global__ void k_copy_dst_from_contiguous(char * __restrict__ dst_original, const char * __restrict__ dst_contiguous,
-                                                  const mmid_row_mapping * __restrict__ row_mapping,
-                                                  int64_t ne0,
-                                                  size_t nb1, size_t nb2) {
-    int32_t i = blockIdx.x;
-
-    const int32_t i1 = row_mapping[i].i1;
-    const int32_t i2 = row_mapping[i].i2;
-
-    const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1);
-    float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2);
-
-    for (int j = threadIdx.x; j < ne0; j += blockDim.x) {
-        dst_row_original[j] = dst_row_contiguous[j];
-    }
-}
-
 static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
     const ggml_tensor * ids  = dst->src[2];
 
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");
+
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && ne2 == 1) {
-        if (ggml_is_quantized(src0->type)) {
-            ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
-        } else {
-            ggml_cuda_mul_mat_vec(ctx, src0, src1, ids, dst);
-        }
-        return;
-    }
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
 
-    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");
+    if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        if (ne2 == 1) {
+            if (ggml_is_quantized(src0->type)) {
+                ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+            } else {
+                ggml_cuda_mul_mat_vec(ctx, src0, src1, ids, dst);
+            }
+            return;
+        }
+
+        if (ggml_cuda_should_use_mmq(src0->type, cc, ne12)) {
+            ggml_cuda_mul_mat_q(ctx, src0, src1, ids, dst);
+            return;
+        }
+    }
 
     cudaStream_t stream = ctx.stream();
 
-    const int64_t n_as = ne02;
-    const int64_t n_ids = ids->ne[0];
+    GGML_ASSERT(nb12 % nb11 == 0);
+    GGML_ASSERT(nb2  % nb1  == 0);
+
+    const ggml_type type_src1_sorted = (src0->type == GGML_TYPE_F16 && !fast_fp16_hardware_available(cc))
+        || ggml_is_quantized(src0->type) ? GGML_TYPE_F32 : src0->type;
+    const ggml_type type_dst_sorted  = GGML_TYPE_F32;
+    const size_t ts_src1_sorted = ggml_type_size(type_src1_sorted);
+    const size_t ts_dst_sorted  = ggml_type_size(type_dst_sorted);
+
+    const int64_t n_expert_used = ids->ne[0];
+    const int64_t ne_get_rows = ne12 * n_expert_used;
+
+    std::vector<int32_t> ids_to_sorted_host;
+    ids_to_sorted_host.reserve(2*ne_get_rows);
+    std::vector<int32_t> ids_from_sorted_host(ne_get_rows);
+
+    ggml_cuda_pool_alloc<int32_t> ids_buf_dev(ctx.pool(), 2*ne_get_rows);
+
+    std::vector<int32_t> tokens_per_expert(ne02);
+
+    ggml_cuda_pool_alloc<char> src1_sorted(ctx.pool(), ne12*n_expert_used*ne10*ts_src1_sorted);
+    ggml_cuda_pool_alloc<char>  dst_sorted(ctx.pool(), ne2 *n_expert_used* ne0*ts_dst_sorted);
 
     std::vector<char> ids_host(ggml_nbytes(ids));
-    const char * ids_dev = (const char *) ids->data;
-    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
+    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
     CUDA_CHECK(cudaStreamSynchronize(stream));
 
-    ggml_tensor src0_row = *src0;
-    ggml_tensor src1_row = *src1;
-    ggml_tensor dst_row  = *dst;
-
-    char * src0_original = (char *) src0->data;
-    char * src1_original = (char *) src1->data;
-    char * dst_original  = (char *)  dst->data;
-
-    src0_row.ne[2] = 1;
-    src0_row.ne[3] = 1;
-    src0_row.nb[3] = nb02;
-
-    src1_row.ne[1] = 1;
-    src1_row.ne[2] = 1;
-    src1_row.ne[3] = 1;
-    src1_row.nb[2] = nb11;
-    src1_row.nb[3] = nb11;
-
-    dst_row.ne[1] = 1;
-    dst_row.ne[2] = 1;
-    dst_row.ne[3] = 1;
-    dst_row.nb[2] = nb1;
-    dst_row.nb[3] = nb1;
-
-    ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
-    ggml_cuda_pool_alloc<char>  dst_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst));
-
-    src1_row.data = src1_contiguous.get();
-    dst_row.data  =  dst_contiguous.get();
-
-    for (int64_t i02 = 0; i02 < n_as; i02++) {
-        int64_t num_src1_rows = 0;
-
-        for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-            for (int64_t id = 0; id < n_ids; id++) {
-                const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
-
-                GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
-
-                if (row_id_i != i02) {
-                    continue;
+    for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices
+        for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens
+            for (int64_t iex = 0; iex < n_expert_used; ++iex) {
+                const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]);
+                assert(expert_to_use >= 0 && expert_to_use < ne02);
+                if (expert_to_use == i02) {
+                    ids_from_sorted_host[i12*n_expert_used + iex] = ids_to_sorted_host.size();
+                    ids_to_sorted_host.push_back(i12*ne11 + iex % ne11);
+                    tokens_per_expert[i02]++;
+                    break;
                 }
-
-                num_src1_rows++;
             }
         }
+    }
+    GGML_ASSERT(ids_to_sorted_host.size() == size_t(ne_get_rows));
 
-        if (num_src1_rows == 0) {
+    ids_to_sorted_host.insert(ids_to_sorted_host.end(), ids_from_sorted_host.begin(), ids_from_sorted_host.end());
+
+    CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_to_sorted_host.data(), 2*ne_get_rows*sizeof(int32_t), cudaMemcpyHostToDevice, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    const int32_t * ids_to_sorted   = ids_buf_dev.ptr + 0*ne_get_rows;
+    const int32_t * ids_from_sorted = ids_buf_dev.ptr + 1*ne_get_rows;
+
+    get_rows_cuda(src1->data, src1->type, ids_to_sorted, src1_sorted.ptr, type_src1_sorted,
+        ne10, nb11, nb12, nb13,
+        ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t),
+        ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, stream);
+    CUDA_CHECK(cudaGetLastError());
+
+    char * src1_data_cur = (char *) src1_sorted.ptr;
+    char *  dst_data_cur = (char *)  dst_sorted.ptr;
+    for (int64_t i02 = 0; i02 < ne02; ++i02) {
+        if (tokens_per_expert[i02] == 0) {
             continue;
         }
 
-        ggml_cuda_pool_alloc<int> dev_cur_src1_row(ctx.pool(), 1);
-        ggml_cuda_pool_alloc<mmid_row_mapping> dev_row_mapping(ctx.pool(), num_src1_rows);
-        CUDA_CHECK(cudaMemsetAsync(dev_cur_src1_row.get(), 0, sizeof(int), stream));
+        ggml_tensor src0_slice = *src0;
+        src0_slice.ne[2] = 1;
+        src0_slice.nb[3] = src0_slice.nb[2];
+        src0_slice.data  = (char *) src0->data + i02*nb02;
 
-        {
-            dim3 block_dims(std::min((unsigned int)ne10, 768u));
-            dim3 grid_dims(ids->ne[1], n_ids);
-            k_copy_src1_to_contiguous<<<grid_dims, block_dims, 0, stream>>>(
-                    src1_original, src1_contiguous.get(),
-                    dev_cur_src1_row.get(), dev_row_mapping.get(),
-                    ids_dev, i02, ids->nb[1], ids->nb[0],
-                    ne11, ne10,
-                    nb11, nb12);
-            CUDA_CHECK(cudaGetLastError());
-        }
+        ggml_tensor src1_slice;
+        memset(&src1_slice, 0, sizeof(src1_slice));
+        src1_slice.buffer = src1->buffer;
+        src1_slice.type   = type_src1_sorted;
+        src1_slice.ne[0]  = ne10;
+        src1_slice.ne[1]  = tokens_per_expert[i02];
+        src1_slice.ne[2]  = 1;
+        src1_slice.ne[3]  = 1;
+        src1_slice.nb[0]  = ts_src1_sorted;
+        src1_slice.nb[1]  = src1_slice.ne[0] * src1_slice.nb[0];
+        src1_slice.nb[2]  = src1_slice.ne[1] * src1_slice.nb[1];
+        src1_slice.nb[3]  = src1_slice.ne[2] * src1_slice.nb[2];
+        src1_slice.data   = src1_data_cur;
 
-        src0_row.data = src0_original + i02*nb02;
+        ggml_tensor dst_slice;
+        memset(&dst_slice, 0, sizeof(dst_slice));
+        dst_slice.buffer = dst->buffer;
+        dst_slice.type   = type_dst_sorted;
+        dst_slice.ne[0]  = ne0;
+        dst_slice.ne[1]  = tokens_per_expert[i02];
+        dst_slice.ne[2]  = 1;
+        dst_slice.ne[3]  = 1;
+        dst_slice.nb[0]  = ts_dst_sorted;
+        dst_slice.nb[1]  = dst_slice.ne[0] * dst_slice.nb[0];
+        dst_slice.nb[2]  = dst_slice.ne[1] * dst_slice.nb[1];
+        dst_slice.nb[3]  = dst_slice.ne[2] * dst_slice.nb[2];
+        dst_slice.data   = dst_data_cur;
 
-        GGML_ASSERT(nb11 == sizeof(float)*ne10);
-        GGML_ASSERT(nb1 == sizeof(float)*ne0);
+        ggml_cuda_mul_mat(ctx, &src0_slice, &src1_slice, &dst_slice);
+        CUDA_CHECK(cudaGetLastError());
 
-        src1_row.ne[1] = num_src1_rows;
-        src1_row.nb[1] = nb11;
-        src1_row.nb[2] = num_src1_rows*nb11;
-        src1_row.nb[3] = num_src1_rows*nb11;
-
-        dst_row.ne[1] = num_src1_rows;
-        dst_row.nb[1] = nb1;
-        dst_row.nb[2] = num_src1_rows*nb1;
-        dst_row.nb[3] = num_src1_rows*nb1;
-
-        ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
-
-        {
-            dim3 block_dims(std::min((unsigned int)ne0, 768u));
-            dim3 grid_dims(num_src1_rows);
-            k_copy_dst_from_contiguous<<<grid_dims, block_dims, 0, stream>>>(
-                    dst_original, dst_contiguous.get(),
-                    dev_row_mapping.get(),
-                    ne0,
-                    nb1, nb2);
-            CUDA_CHECK(cudaGetLastError());
-        }
+        src1_data_cur += src1_slice.nb[2];
+        dst_data_cur  +=  dst_slice.nb[2];
     }
+
+    get_rows_cuda(dst_sorted.ptr, type_dst_sorted, ids_from_sorted, dst->data, dst->type,
+        ne0, ne0*ts_dst_sorted, ne_get_rows*ne0*ts_dst_sorted, ne_get_rows*ne0*ts_dst_sorted,
+        ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t),
+        nb1, nb2, nb3, stream);
 }
 
 static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) {

ggml/src/ggml-cuda/mmq.cu

@@ -1,37 +1,10 @@
 #include "mmq.cuh"
+#include "quantize.cuh"
 
-void ggml_cuda_op_mul_mat_q(
-    ggml_backend_cuda_context & ctx,
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream) {
+#include <vector>
 
-    const int64_t ne00 = src0->ne[0];
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    GGML_ASSERT(ne10 % QK8_1 == 0);
-
-    const int64_t ne0 = dst->ne[0];
-
-    const int64_t row_diff = row_high - row_low;
-    const int64_t stride00 = ne00 / ggml_blck_size(src0->type);
-
-    int id = ggml_cuda_get_device();
-    const int cc = ggml_cuda_info().devices[id].cc;
-
-    // the main device has a larger memory buffer to hold the results from all GPUs
-    // nrows_dst == nrows of the matrix that the kernel writes into
-    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
-
-    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
-    // Also its fixup needs to allocate a temporary buffer in the memory pool.
-    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
-    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
-        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
-    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
-
-    switch (src0->type) {
+static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
+    switch (args.type_x) {
         case GGML_TYPE_Q4_0:
             mul_mat_q_case<GGML_TYPE_Q4_0>(ctx, args, stream);
             break;
@@ -90,10 +63,195 @@ void ggml_cuda_op_mul_mat_q(
             GGML_ABORT("fatal error");
             break;
     }
+}
+
+void ggml_cuda_mul_mat_q(
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(        dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(!ids || ids->type  == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID.
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    cudaStream_t stream = ctx.stream();
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    const size_t ts_src0 = ggml_type_size(src0->type);
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    const size_t ts_dst  = ggml_type_size(dst->type);
+
+    GGML_ASSERT(        nb00       == ts_src0);
+    GGML_ASSERT(        nb10       == ts_src1);
+    GGML_ASSERT(        nb0        == ts_dst);
+    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
+
+    const char  * src0_d = (const char  *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       *  dst_d = (float       *)  dst->data;
+
+    const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING);
+
+    const int64_t s01 = src0->nb[1] / ts_src0;
+    const int64_t s1  =  dst->nb[1] / ts_dst;
+    const int64_t s02 = src0->nb[2] / ts_src0;
+    const int64_t s2  =  dst->nb[2] / ts_dst;
+    const int64_t s03 = src0->nb[3] / ts_src0;
+    const int64_t s3  =  dst->nb[3] / ts_dst;
+
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;
+
+    if (!ids) {
+        const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 +
+            get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
+        ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), nbytes_src1_q8_1);
+
+        {
+            const int64_t s11 = src1->nb[1] / ts_src1;
+            const int64_t s12 = src1->nb[2] / ts_src1;
+            const int64_t s13 = src1->nb[3] / ts_src1;
+            quantize_mmq_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type,
+                ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+        }
+
+        const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
+        const int64_t s13 = ne12*s12;
+
+        const mmq_args args = {
+            src0_d, src0->type, (const int *) src1_q8_1.ptr, nullptr, nullptr, dst_d,
+            ne00, ne01, ne1, s01, s1,
+            ne02, ne12, s02, s12, s2,
+            ne03, ne13, s03, s13, s3,
+            use_stream_k};
+        ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
+        return;
+    }
+
+    GGML_ASSERT(ne13 == 1);
+    GGML_ASSERT(nb12 % nb11 == 0);
+    GGML_ASSERT(nb2  % nb1  == 0);
+
+    const int64_t n_expert_used = ids->ne[0];
+    const int64_t ne_get_rows = ne12 * n_expert_used;
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+    std::vector<int32_t> ids_src1_host;
+    ids_src1_host.reserve(ne_get_rows);
+    std::vector<int32_t> ids_dst_host;
+    ids_dst_host.reserve(ne_get_rows);
+    std::vector<int32_t> tokens_per_expert_host(ne02);
+    std::vector<int32_t> expert_bounds_host(ne02 + 1);
+    ggml_cuda_pool_alloc<int32_t> ids_buf_dev(ctx.pool());
+
+    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices
+        for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens
+            for (int64_t iex = 0; iex < n_expert_used; ++iex) {
+                const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]);
+                assert(expert_to_use >= 0 && expert_to_use < ne02);
+                if (expert_to_use == i02) {
+                    ids_src1_host.push_back(i12*(nb12/nb11) + iex % ne11);
+                    ids_dst_host.push_back(i12*ne1 + iex);
+                    tokens_per_expert_host[i02]++;
+                    break;
+                }
+            }
+        }
+    }
+
+    int32_t cumsum = 0;
+    for (int64_t i = 0; i < ne02; ++i) {
+        expert_bounds_host[i] = cumsum;
+        cumsum += tokens_per_expert_host[i];
+    }
+    expert_bounds_host[ne02] = cumsum;
+
+    std::vector<int32_t> ids_buf_host;
+    ids_buf_host.reserve(ids_src1_host.size() + ids_dst_host.size() + expert_bounds_host.size());
+    ids_buf_host.insert(ids_buf_host.end(), ids_src1_host.begin(), ids_src1_host.end());
+    ids_buf_host.insert(ids_buf_host.end(), ids_dst_host.begin(), ids_dst_host.end());
+    ids_buf_host.insert(ids_buf_host.end(), expert_bounds_host.begin(), expert_bounds_host.end());
+    ids_buf_dev.alloc(ids_buf_host.size() + get_mmq_x_max_host(cc)); // Expert bounds are padded on device.
+    CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_buf_host.data(), ids_buf_host.size()*sizeof(int32_t), cudaMemcpyHostToDevice, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    const int32_t * ids_src1_dev      = ids_buf_dev.ptr;
+    const int32_t * ids_dst_dev       = ids_src1_dev + ids_src1_host.size();
+    const int32_t * expert_bounds_dev = ids_dst_dev + ids_dst_host.size();
+
+    const size_t nbytes_src1_q8_1 = ne12*n_expert_used*ne10_padded * sizeof(block_q8_1)/QK8_1 +
+        get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
+    ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), nbytes_src1_q8_1);
+
+    const int64_t ne11_flat = ne12*n_expert_used;
+    const int64_t ne12_flat = 1;
+    const int64_t ne13_flat = 1;
+
+    {
+        const int64_t s11 = src1->nb[1] / ts_src1;
+        const int64_t s12 = src1->nb[2] / ts_src1;
+        const int64_t s13 = src1->nb[2] / ts_src1;
+        quantize_mmq_q8_1_cuda(src1_d, ids_src1_dev, src1_q8_1.get(), src0->type,
+            ne10, s11, s12, s13, ne10_padded, ne11_flat, ne12_flat, ne13_flat, stream);
+    }
+
+    const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
+    const int64_t s13 = ne12*s12;
+
+    // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid.
+    const mmq_args args = {
+        src0_d, src0->type, (const int *) src1_q8_1.ptr, ids_dst_dev, expert_bounds_dev, dst_d,
+        ne00, ne01, ne_get_rows, s01, s1,
+        ne02, ne02, s02, s12, s2,
+        ne03, ne13, s03, s13, s3,
+        use_stream_k};
+
+    ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
+}
+
+void ggml_cuda_op_mul_mat_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    const int64_t ne00 = src0->ne[0];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne0 = dst->ne[0];
+
+    const int64_t row_diff = row_high - row_low;
+    const int64_t stride01 = ne00 / ggml_blck_size(src0->type);
+
+    const int id = ggml_cuda_get_device();
+    const int cc = ggml_cuda_info().devices[id].cc;
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the kernel writes into
+    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
+
+    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
+    // Also its fixup needs to allocate a temporary buffer in the memory pool.
+    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
+        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
+    const mmq_args args = {
+        src0_dd_i, src0->type, (const int *) src1_ddq_i, nullptr, nullptr, dst_dd_i,
+        ne00, row_diff, src1_ncols, stride01, nrows_dst,
+        1, 1, 0, 0, 0,
+        1, 1, 0, 0, 0,
+        use_stream_k};
+
+    ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
 
     GGML_UNUSED(src1);
     GGML_UNUSED(dst);
     GGML_UNUSED(src1_ddf_i);
+    GGML_UNUSED(src1_padded_row_size);
 }
 
 bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {

ggml/src/ggml-cuda/mmq.cuh

@@ -13,9 +13,10 @@ using namespace ggml_cuda_mma;
 #define MMQ_ITER_K 256
 #define MMQ_NWARPS 8
 
-typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int & kbx0, const int & i_max, const int & stride);
-typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00);
-typedef void (*mmq_write_back_t)(const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max);
+typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int kbx0, const int i_max, const int stride);
+typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00);
+typedef void (*mmq_write_back_t)(const float * __restrict__ sum, const int32_t * __restrict__ get_rows_to_sorted,
+    float * __restrict__ dst, const int stride, const int i_max, const int j_max);
 
 enum mmq_q8_1_ds_layout {
     MMQ_Q8_1_DS_LAYOUT_D4,
@@ -233,7 +234,7 @@ static constexpr __device__ int mmq_get_granularity_device(const int /* mmq_x */
 // ------------------------------------------------------------
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_0(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -289,7 +290,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -328,7 +329,7 @@ static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_1(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -384,7 +385,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -423,7 +424,7 @@ static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_0(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -495,7 +496,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_1(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -565,7 +566,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q8_0(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -621,7 +622,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -651,7 +652,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps, mmq_q8_1_ds_layout ds_layout>
 static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     typedef tile<16, 8, int> tile_A;
     typedef tile< 8, 8, int> tile_B;
@@ -732,7 +733,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -762,7 +763,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     typedef tile<16, 8, int> tile_A;
     typedef tile< 8, 8, int> tile_B;
@@ -839,7 +840,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16;
     const int   * x_qs = (const int   *) x;
@@ -871,7 +872,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 #ifdef NEW_MMA_AVAILABLE
 
     typedef tile<16, 4, int> tile_A;
@@ -955,7 +956,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q2_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1011,7 +1012,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1074,7 +1075,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 #ifdef NEW_MMA_AVAILABLE
 
     typedef tile<16, 4, int> tile_A;
@@ -1201,7 +1202,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q3_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1298,7 +1299,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q3_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1340,7 +1341,7 @@ static __device__ __forceinline__ int unpack_scales_q45_K(const int * scales, co
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1437,7 +1438,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1469,7 +1470,7 @@ static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1578,7 +1579,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1610,7 +1611,7 @@ static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q6_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1693,7 +1694,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1726,7 +1727,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 #ifdef NEW_MMA_AVAILABLE
 
     typedef tile<16, 4, int> tile_A;
@@ -1835,7 +1836,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_nl(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1893,7 +1894,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_xxs(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1951,7 +1952,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_xs(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2007,7 +2008,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_s(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2070,7 +2071,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq3_xxs(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2126,7 +2127,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq3_s(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2189,7 +2190,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq1_s(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2245,7 +2246,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_xs(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2306,8 +2307,8 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template<int mmq_x, int mmq_y, int nwarps, bool need_check>
 static __device__ __forceinline__ void mmq_write_back_dp4a(
-    const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
-
+        const float * __restrict__ sum, const int32_t * __restrict__ ids_dst, float * __restrict__ dst,
+        const int stride, const int i_max, const int j_max) {
 #pragma unroll
     for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
         const int j = j0 + threadIdx.y;
@@ -2324,15 +2325,15 @@ static __device__ __forceinline__ void mmq_write_back_dp4a(
                 continue;
             }
 
-            dst[j*stride + i] = sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+            dst[ids_dst[j]*stride + i] = sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
         }
     }
 }
 
 template<int mmq_x, int mmq_y, int nwarps, bool need_check>
 static __device__ __forceinline__ void mmq_write_back_mma(
-    const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
-
+        const float * __restrict__ sum, const int * __restrict__ ids_dst, float * __restrict__ dst,
+        const int stride, const int i_max, const int j_max) {
     typedef tile<16, 8, int> tile_C;
 
     constexpr int granularity = mmq_get_granularity_device(mmq_x);
@@ -2362,7 +2363,7 @@ static __device__ __forceinline__ void mmq_write_back_mma(
                     continue;
                 }
 
-                dst[j*stride + i] = sum[(j0/tile_C::J + n)*tile_C::ne + l];
+                dst[ids_dst[j]*stride + i] = sum[(j0/tile_C::J + n)*tile_C::ne + l];
             }
         }
     }
@@ -2518,17 +2519,18 @@ struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ4_XS> {
 };
 
 template <ggml_type type, int mmq_x, int nwarps, bool need_check, bool fixup>
-static __device__ void mul_mat_q_process_tile(
-    const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
-    const int & ne00, const int & ne01, const int & stride01, const int & ne10, const int & ne11, const int & stride11, const int & ne0,
-    const int & it, const int & jt, const int & kb0_start, const int & kb0_stop) {
+static __device__ __forceinline__ void mul_mat_q_process_tile(
+        const char * __restrict__ x, const int offset_x, const int * __restrict__ y,
+        const int * __restrict__ ids_dst, float * __restrict__ dst, float * __restrict__ tmp_fixup,
+        const int nrows_x, const int ncols_y, const int stride_row_x, const int stride_col_dst,
+        const int tile_x_max_i, const int tile_y_max_j, const int kb0_start, const int kb0_stop) {
 
     constexpr int              qk         = ggml_cuda_type_traits<type>::qk;
     constexpr int              mmq_y      = get_mmq_y_device();
     constexpr load_tiles_mmq_t load_tiles = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::load_tiles;
 
-    extern __shared__ char data_mul_mat_q[];
-    int * tile_y = (int *) data_mul_mat_q;
+    extern __shared__ int data_mul_mat_q[];
+    int * tile_y = data_mul_mat_q + mmq_x;
     int * tile_x = tile_y + GGML_PAD(mmq_x*(WARP_SIZE + WARP_SIZE/QI8_1), nwarps*WARP_SIZE);
 
 #ifdef NEW_MMA_AVAILABLE
@@ -2543,16 +2545,11 @@ static __device__ void mul_mat_q_process_tile(
 
     float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
 
-    const int tile_x_max_i = ne01 - it*mmq_y - 1;
-    const int tile_y_max_j = ne11 - jt*mmq_x - 1;
-
-    const int * y = (const int *) yc + jt*(mmq_x*sizeof(block_q8_1_mmq)/sizeof(int));
-
     for (int kb0 = kb0_start; kb0 < kb0_stop; kb0 += blocks_per_iter) {
-        load_tiles(x, tile_x, stride01*it*mmq_y + kb0, tile_x_max_i, stride01);
+        load_tiles(x, tile_x, offset_x + kb0, tile_x_max_i, stride_row_x);
 
         {
-            const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int));
+            const int * by0 = y + ncols_y*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int));
 #pragma unroll
             for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) {
                 int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x;
@@ -2568,7 +2565,7 @@ static __device__ void mul_mat_q_process_tile(
         __syncthreads();
 
         {
-            const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 1*sizeof(block_q8_1_mmq)/sizeof(int));
+            const int * by0 = y + ncols_y*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 1*sizeof(block_q8_1_mmq)/sizeof(int));
 #pragma unroll
             for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) {
                 int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x;
@@ -2585,12 +2582,10 @@ static __device__ void mul_mat_q_process_tile(
     }
 
     if (fixup) {
-        write_back(sum, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x);
+        write_back(sum, ids_dst, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x);
     } else {
-        write_back(sum, dst + jt*mmq_x*ne0 + it*mmq_y, ne0, tile_x_max_i, tile_y_max_j);
+        write_back(sum, ids_dst, dst, stride_col_dst, tile_x_max_i, tile_y_max_j);
     }
-
-    GGML_UNUSED(ne00); GGML_UNUSED(ne10);
 }
 
 
@@ -2609,8 +2604,11 @@ template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 static __global__ void mul_mat_q(
-    const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
-    const int ne00, const int ne01, const int stride01, const int ne10, const int ne11, const int stride11, const int ne0) {
+        const char * __restrict__ x, const int * __restrict__ y, const int32_t * __restrict__ ids_dst,
+        const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_fixup,
+        const int ncols_x, const int nrows_x, const int ncols_y, const int stride_row_x, const int stride_col_dst,
+        const int channel_ratio, const int nchannels_y, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int nsamples_y, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
 
     // Skip unused template specializations for faster compilation:
     if (mmq_x > get_mmq_x_max_device() || mmq_x % mmq_get_granularity_device(mmq_x) != 0) {
@@ -2621,26 +2619,85 @@ static __global__ void mul_mat_q(
     constexpr int qk    = ggml_cuda_type_traits<type>::qk;
     constexpr int mmq_y = get_mmq_y_device();
 
+    const int ntx = (ncols_y + mmq_x - 1) / mmq_x; // Number of tiles x
+    const int nty = (nrows_x + mmq_y - 1) / mmq_y; // Number of tiles y
+
+    // Initialize the ids for writing back data with just the index.
+    // For regular matrix multiplications this is never changed.
+    // For MoE the correct indices are loaded from ids_dst.
+    extern __shared__ int ids_dst_shared[]; // Stored at beginning of shared memory.
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) {
+        const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+        if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) {
+            break;
+        }
+
+        ids_dst_shared[j] = j;
+    }
+
     // On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
 #if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
     {
+        const int wt = blockIdx.z / nchannels_y;
+        const int zt = blockIdx.z - wt*nchannels_y;
+        const int jt = blockIdx.y;
+        const int it = blockIdx.x;
+
+        // Defaults for regular matrix multiplication:
+        int col_low    = 0;
+        int col_high   = ncols_y;
+        int col_diff   = ncols_y;
+        int offset_y   = wt*stride_sample_y   + zt*stride_channel_y;
+        int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst;
+
+        if (ids_dst) {
+            col_low  = expert_bounds[zt + 0];
+            col_high = expert_bounds[zt + 1];
+            col_diff = col_high - col_low;
+
+            offset_y   = 0;
+            offset_dst = 0;
+
+            if (jt*mmq_x >= col_diff) {
+                return;
+            }
+
+#pragma unroll
+            for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) {
+                const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+                if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) {
+                    break;
+                }
+
+                ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j];
+            }
+        }
+
+        offset_y   += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int));
+        offset_dst += it*mmq_y;
+
+        const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
+        const int tile_y_max_j = col_diff - jt*mmq_x - 1;
+
+        const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
+
         constexpr bool fixup = false;
         mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
-            (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
-                blockIdx.x, blockIdx.y, 0, ne00/qk);
+            (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, nrows_x, ncols_y, stride_row_x, stride_col_dst,
+             tile_x_max_i, tile_y_max_j, 0, ncols_x/qk);
         return;
     }
 #endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
 
-    const     int64_t blocks_per_ne00 = ne00 / qk;
+    const     int64_t blocks_per_ne00 = ncols_x / qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
 
-    const int ntx = (ne11 + mmq_x - 1) / mmq_x; // Number of tiles x
-    const int nty = (ne01 + mmq_y - 1) / mmq_y; // Number of tiles y
-
     // kbc == k block continuous, current index in continuous ijk space.
-    int64_t kbc      = (int64_t) blockIdx.x     *blocks_per_ne00*ntx*nty / gridDim.x;
-    int64_t kbc_stop = (int64_t)(blockIdx.x + 1)*blocks_per_ne00*ntx*nty / gridDim.x;
+    int64_t kbc      = (int64_t) blockIdx.x     *nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+    int64_t kbc_stop = (int64_t)(blockIdx.x + 1)*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
 
     kbc      -= (kbc      % blocks_per_ne00) % blocks_per_iter;
     kbc_stop -= (kbc_stop % blocks_per_ne00) % blocks_per_iter;
@@ -2649,13 +2706,64 @@ static __global__ void mul_mat_q(
     int kb0_start = kbc % blocks_per_ne00;
     int kb0_stop  = min(blocks_per_ne00, kb0_start + kbc_stop - kbc);
     while (kbc < kbc_stop && kb0_stop == blocks_per_ne00) {
-        const int jt =  kbc /    (blocks_per_ne00*nty);                    // j index of current tile.
-        const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00; // i index of current tile.
+        int tmp = kbc;
+        const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+        tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+        const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00);
+        tmp -= wt * (nchannels_y*ntx*blocks_per_ne00);
+        const int zt = tmp / (ntx*blocks_per_ne00);
+        tmp -= zt * (ntx*blocks_per_ne00);
+        const int jt = tmp / blocks_per_ne00;
+
+        // Defaults for regular matrix multiplication:
+        int col_low    = 0;
+        int col_high   = ncols_y;
+        int col_diff   = ncols_y;
+        int offset_y   = wt*stride_sample_y   + zt*stride_channel_y;
+        int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst;
+
+        if (ids_dst) {
+            col_low  = expert_bounds[zt + 0];
+            col_high = expert_bounds[zt + 1];
+            col_diff = col_high - col_low;
+
+            offset_y   = 0;
+            offset_dst = 0;
+
+            if (jt*mmq_x >= col_diff) {
+                kbc += blocks_per_ne00;
+                kbc -= kbc % blocks_per_ne00;
+
+                kb0_start = 0;
+                kb0_stop  = min(blocks_per_ne00, kbc_stop - kbc);
+
+                continue;
+            }
+
+#pragma unroll
+            for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) {
+                const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+                if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) {
+                    break;
+                }
+
+                ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j];
+            }
+        }
+
+        offset_y   += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int));
+        offset_dst += it*mmq_y;
+
+        const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
+        const int tile_y_max_j = col_diff - jt*mmq_x - 1;
+
+        const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
 
         constexpr bool fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
         mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
-            (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
-             it, jt, kb0_start, kb0_stop);
+            (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, nrows_x, ncols_y, stride_row_x, stride_col_dst,
+             tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 
         kbc += blocks_per_ne00;
         kbc -= kbc % blocks_per_ne00;
@@ -2668,55 +2776,106 @@ static __global__ void mul_mat_q(
         return;
     }
 
-    const int jt =  kbc /    (blocks_per_ne00*nty);
-    const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
+    int tmp = kbc;
+    const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+    tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+    const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00);
+    tmp -= wt * (nchannels_y*ntx*blocks_per_ne00);
+    const int zt = tmp / (ntx*blocks_per_ne00);
+    tmp -= zt * (ntx*blocks_per_ne00);
+    const int jt = tmp / blocks_per_ne00;
+
+    // Defaults for regular matrix multiplication:
+    int col_low    = 0;
+    int col_high   = ncols_y;
+    int col_diff   = ncols_y;
+    int offset_y   = wt*stride_sample_y   + zt*stride_channel_y;
+    int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst;
+
+    if (ids_dst) {
+        col_low  = expert_bounds[zt + 0];
+        col_high = expert_bounds[zt + 1];
+        col_diff = col_high - col_low;
+
+        offset_y   = 0;
+        offset_dst = 0;
+
+        if (jt*mmq_x >= col_diff) {
+            return;
+        }
+
+        // The memory layout for the fixup buffer is always contiguous, therefore reset ids:
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) {
+            const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+            if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) {
+                break;
+            }
+
+            ids_dst_shared[j] = j;
+        }
+    }
+
+    offset_y   += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int));
+    offset_dst += it*mmq_y;
+
+    const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
+    const int tile_y_max_j = col_diff - jt*mmq_x - 1;
+
+    const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
 
     constexpr bool fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
     mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
-        (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
-            it, jt, kb0_start, kb0_stop);
+        (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, nrows_x, ncols_y, stride_row_x, stride_col_dst,
+         tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 }
 
 
 template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 static __global__ void mul_mat_q_stream_k_fixup(
-    float * __restrict__ dst, const float * __restrict__ tmp_last_tile, const int ne00, const int ne01, const int ne11, const int ne0, const int block_num_mmq) {
-
+        const int32_t * ids_dst, const int32_t * expert_bounds, float * __restrict__ dst, const float * __restrict__ tmp_last_tile,
+        const int ncols_x, const int nrows_x, const int ncols_y, const int stride_col_dst,
+        const int nchannels_y, const int stride_channel_dst, const int nsamples_y, const int stride_sample_dst) {
     constexpr int     mmq_y           = get_mmq_y_device();
     constexpr int     qk              = ggml_cuda_type_traits<type>::qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
-    const     int64_t blocks_per_ne00 = ne00 / qk;
+    const     int64_t blocks_per_ne00 = ncols_x / qk;
 
     float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
 
-    const int ntx = (ne11 + mmq_x - 1) / mmq_x;
-    const int nty = (ne01 + mmq_y - 1) / mmq_y;
+    const int ntx  = (ncols_y + mmq_x - 1) / mmq_x;
+    const int nty  = (nrows_x + mmq_y - 1) / mmq_y;
+
+    const int bidx0 = blockIdx.x;
+
+    // kbc == k block continuous, current index in continuous ijk space.
+    int64_t kbc0      = (int64_t) bidx0     *nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+    int64_t kbc0_stop = (int64_t)(bidx0 + 1)*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+
+    kbc0      -= (kbc0      % blocks_per_ne00) % blocks_per_iter;
+    kbc0_stop -= (kbc0_stop % blocks_per_ne00) % blocks_per_iter;
+
+    const bool did_not_have_any_data   = kbc0 == kbc0_stop;
+    const bool wrote_beginning_of_tile = kbc0 % blocks_per_ne00 == 0;
+    const bool did_not_write_last      = kbc0/blocks_per_ne00 == kbc0_stop/blocks_per_ne00 && kbc0_stop % blocks_per_ne00 != 0;
+    if (did_not_have_any_data || wrote_beginning_of_tile || did_not_write_last) {
+        return;
+    }
 
     bool any_fixup = false;
 
-    const int bidx_start = ((blockIdx.y*nty + blockIdx.x)     * block_num_mmq)                           / (gridDim.y*gridDim.x);
-    const int bidx_stop  = ((blockIdx.y*nty + blockIdx.x + 1) * block_num_mmq + gridDim.y*gridDim.x - 1) / (gridDim.y*gridDim.x);
+    // Iterate over previous blocks and sum up partial sums written to fixup buffer.
+    // All CUDA blocks that get here must have a previous block that needs a fixup.
+    int64_t bidx = bidx0 - 1;
+    int64_t kbc_stop = kbc0;
+    while(true) {
+        int64_t kbc = bidx*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+        kbc -= (kbc % blocks_per_ne00) % blocks_per_iter;
 
-    int64_t kbc_0;
-    int64_t kbc_stop_0 = (int64_t) bidx_start*blocks_per_ne00*ntx*nty / block_num_mmq;
-
-    for (int bidx = bidx_start; bidx < bidx_stop; ++bidx) {
-        kbc_0 = kbc_stop_0;
-        kbc_stop_0 = (int64_t) (bidx + 1)*blocks_per_ne00*ntx*nty / block_num_mmq;
-
-        const int64_t kbc      = kbc_0      - (kbc_0      % blocks_per_ne00) % blocks_per_iter;
-        const int64_t kbc_stop = kbc_stop_0 - (kbc_stop_0 % blocks_per_ne00) % blocks_per_iter;
-
-        // Skip fixup tile if the MMQ CUDA block never wrote anything to it:
-        if (kbc == kbc_stop || kbc_stop % blocks_per_ne00 == 0) {
-            continue;
-        }
-
-        const int jt =  kbc_stop /    (blocks_per_ne00*nty);
-        const int it = (kbc_stop - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
-
-        // Skip fixup tile if it's unrelated to the output tile assigned to this CUDA block:
-        if ((unsigned)it != blockIdx.x || (unsigned)jt != blockIdx.y) {
+        if (kbc == kbc_stop) { // Did not have any data.
+            bidx--;
+            kbc_stop = kbc;
             continue;
         }
 
@@ -2733,16 +2892,71 @@ static __global__ void mul_mat_q_stream_k_fixup(
                 sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE] += tmp_last_tile[bidx*(mmq_x*mmq_y) + j*mmq_y + i];
             }
         }
+
+        // If this block started in a previous tile we are done and don't need to combine additional partial results.
+        if (kbc % blocks_per_ne00 == 0 || kbc/blocks_per_ne00 < kbc0/blocks_per_ne00) {
+            break;
+        }
+        bidx--;
+        kbc_stop = kbc;
     }
 
     if (!any_fixup) {
         return;
     }
 
-    dst += blockIdx.y*mmq_x*ne0 + blockIdx.x*mmq_y;
+    int tmp = kbc0;
+    const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+    tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+    const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00);
+    tmp -= wt * (nchannels_y*ntx*blocks_per_ne00);
+    const int zt = tmp / (ntx*blocks_per_ne00);
+    tmp -= zt * (ntx*blocks_per_ne00);
+    const int jt = tmp / blocks_per_ne00;
 
-    const int i_max = ne01 - blockIdx.x*mmq_y - 1;
-    const int j_max = ne11 - blockIdx.y*mmq_x - 1;
+    if (!ids_dst) {
+        const int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst + it*mmq_y;
+        dst += offset_dst;
+
+        const int i_max = nrows_x - it*mmq_y - 1;
+        const int j_max = ncols_y - jt*mmq_x - 1;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (j > j_max) {
+                return;
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                if (need_check && i > i_max) {
+                    continue;
+                }
+
+                dst[j*stride_col_dst + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+            }
+        }
+        return;
+    }
+
+    __shared__ int ids_dst_shared[mmq_x];
+    const int col_low  = expert_bounds[zt + 0];
+    const int col_high = expert_bounds[zt + 1];
+    const int col_diff = col_high - col_low;
+
+    for (int j = threadIdx.y*WARP_SIZE + threadIdx.x; j < mmq_x; j += nwarps*WARP_SIZE) {
+        ids_dst_shared[j] = ids_dst[col_low + j];
+    }
+
+    const int offset_dst = it*mmq_y;
+    dst += offset_dst;
+
+    const int i_max = nrows_x  - it*mmq_y - 1;
+    const int j_max = col_diff - jt*mmq_x - 1;
 
 #pragma unroll
     for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
@@ -2760,26 +2974,27 @@ static __global__ void mul_mat_q_stream_k_fixup(
                 continue;
             }
 
-            dst[j*ne0 + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+            dst[ids_dst_shared[j]*stride_col_dst + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
         }
     }
 }
 
 struct mmq_args {
-    const char * x; const char * y; float * dst;
-    int64_t ne00; int64_t ne01; int64_t stride01;
-    int64_t ne10; int64_t ne11; int64_t stride11;
-    int64_t ne0;
+    const char * x; ggml_type type_x; const int * y; const int32_t * ids_dst; const int32_t * expert_bounds; float * dst;
+    int64_t ncols_x; int64_t nrows_x; int64_t ncols_y; int64_t stride_row_x; int64_t nrows_dst;
+    int64_t nchannels_x; int64_t nchannels_y; int64_t stride_channel_x; int64_t stride_channel_y; int64_t stride_channel_dst;
+    int64_t nsamples_x; int64_t nsamples_y; int64_t stride_sample_x; int64_t stride_sample_y; int64_t stride_sample_dst;
     bool use_stream_k;
 };
 
 template<ggml_type type>
-static int mmq_get_shmem(const int mmq_x, const int mmq_y, const int cc) {
+static size_t mmq_get_nbytes_shared(const int mmq_x, const int mmq_y, const int cc) {
     const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, mmq_y);
     const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type);
-    const int shmem_x = new_mma_available(cc) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
-    const int shmem_y = mmq_x*sizeof(block_q8_1_mmq);
-    return shmem_x + GGML_PAD(shmem_y, MMQ_NWARPS*WARP_SIZE*sizeof(int));
+    const size_t nbs_ids = mmq_x*sizeof(int);
+    const size_t nbs_x = new_mma_available(cc) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
+    const size_t nbs_y = mmq_x*sizeof(block_q8_1_mmq);
+    return nbs_ids + nbs_x + GGML_PAD(nbs_y, MMQ_NWARPS*WARP_SIZE*sizeof(int));
 }
 
 template <ggml_type type, int mmq_x>
@@ -2791,86 +3006,114 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
     const dim3 block_dims(WARP_SIZE, MMQ_NWARPS, 1);
 
-    const int shmem = mmq_get_shmem<type>(mmq_x, mmq_y, cc);
+    const int nbytes_shared = mmq_get_nbytes_shared<type>(mmq_x, mmq_y, cc);
 
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
-    static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-    if (!shmem_limit_raised[id]) {
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
-        shmem_limit_raised[id] = true;
+    static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
+    if (!shared_memory_limit_raised[id]) {
+        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
+        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
+        shared_memory_limit_raised[id] = true;
     }
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
 
-    const int nty = (args.ne01 + mmq_y - 1) / mmq_y;
-    const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
-    const dim3 block_nums_xy_tiling(nty, ntx, 1);
+    const int nty  = (args.nrows_x + mmq_y - 1) / mmq_y;
+    const int ntx  = (args.ncols_y + mmq_x - 1) / mmq_x;
+    const int ntzw = args.nchannels_y * args.nsamples_y;
+    const dim3 block_nums_xy_tiling(nty, ntx, ntzw);
+
+    GGML_ASSERT(args.nchannels_y % args.nchannels_x == 0);
+    GGML_ASSERT(args.nsamples_y  % args.nsamples_x  == 0);
+    const int channel_ratio = args.nchannels_y / args.nchannels_x;
+    const int sample_ratio  = args.nsamples_y  / args.nsamples_x;
 
     if (!args.use_stream_k) {
-        if (args.ne01 % mmq_y == 0) {
+        if (args.nrows_x % mmq_y == 0) {
             constexpr bool need_check = false;
-            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
-                (args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
+                (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
+                 args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+                 channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
         } else {
             constexpr bool need_check = true;
-            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
-                (args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
+                (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
+                 args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+                 channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
         }
         return;
     }
 
-    const dim3 block_nums_mmq(nsm, 1, 1);
+    const dim3 block_nums_stream_k(nsm, 1, 1);
+    const bool fixup_needed = ntx*nty*ntzw % nsm != 0;
 
     ggml_cuda_pool & pool = ctx.pool(id);
-    ggml_cuda_pool_alloc<float> tmp_fixup(pool, block_nums_mmq.x * mmq_x*mmq_y);
+    ggml_cuda_pool_alloc<float> tmp_fixup(pool);
+    if (fixup_needed) {
+        tmp_fixup.alloc(block_nums_stream_k.x * mmq_x*mmq_y);
+    }
 
-    if (args.ne01 % mmq_y == 0) {
+    if (args.nrows_x % mmq_y == 0) {
         constexpr bool need_check = false;
 
-        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_mmq, block_dims, shmem, stream>>>
-            (args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
+            (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
+             args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+             channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
 
-        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, 0, stream>>>
-            (args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x);
+        if (!fixup_needed) {
+            return;
+        }
+
+        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
+            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_y,
+             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst);
     } else {
         constexpr bool need_check = true;
 
-        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_mmq, block_dims, shmem, stream>>>
-            (args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
+            (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
+             args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+             channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
 
-        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, 0, stream>>>
-            (args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x);
+        if (!fixup_needed) {
+            return;
+        }
+
+        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
+            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_y,
+             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst);
     }
 }
 
 template <ggml_type type>
 void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
-    const int id    = ggml_cuda_get_device();
-    const int cc    = ggml_cuda_info().devices[id].cc;
-    const int smpbo = ggml_cuda_info().devices[id].smpbo;
+    const int    id    = ggml_cuda_get_device();
+    const int    cc    = ggml_cuda_info().devices[id].cc;
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
 
     const int mmq_x_max = get_mmq_x_max_host(cc);
     const int mmq_y = get_mmq_y_host(cc);
-    const int block_num_y = (args.ne01 + mmq_y - 1) / mmq_y;
-    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;
 
     int mmq_x_best  = 0;
-    int nparts_best = INT_MAX;
+    int ntiles_x_best = INT_MAX;
 
-    for (int mmq_x = 8; mmq_x <= mmq_x_max && nparts_best > 1; mmq_x += 8) {
+    for (int mmq_x = 8; mmq_x <= mmq_x_max && ntiles_x_best > 1; mmq_x += 8) {
         const int granularity = mmq_get_granularity_host(mmq_x, cc);
 
-        if (mmq_x % granularity != 0 || mmq_get_shmem<type>(mmq_x, mmq_y, cc) > smpbo) {
+        if (mmq_x % granularity != 0 || mmq_get_nbytes_shared<type>(mmq_x, mmq_y, cc) > smpbo) {
             continue;
         }
 
-        const int ntiles_x = (args.ne11 + mmq_x - 1) / mmq_x;
-        const int nwaves_xy_tiling = ntiles_x*block_num_y;
-        const int nparts = use_stream_k ? ntiles_x : nwaves_xy_tiling;
+        const int ntiles_x = (args.ncols_y + mmq_x - 1) / mmq_x;
 
-        if (nparts < nparts_best) {
-            mmq_x_best  = mmq_x;
-            nparts_best = nparts;
+        if (ntiles_x < ntiles_x_best) {
+            mmq_x_best = mmq_x;
+            ntiles_x_best = ntiles_x;
         }
     }
 
@@ -2954,6 +3197,9 @@ extern DECL_MMQ_CASE(GGML_TYPE_IQ4_XS);
 
 // -------------------------------------------------------------------------------------------------------------------------
 
+void ggml_cuda_mul_mat_q(
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+
 void ggml_cuda_op_mul_mat_q(
     ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,

ggml/src/ggml-cuda/mmvq.cu

@@ -158,7 +158,7 @@ static __global__ void mul_mat_vec_q(
     const     int blocks_per_row_x = ncols_x / qk;
     constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;
 
-    // The MUL_MAT_ID code path with ids != nullptr is only implemetned for ncols_dst == 1.
+    // The MUL_MAT_ID code path with ids != nullptr is only implemented for ncols_dst == 1.
     const int channel_dst = blockIdx.y;
     const int channel_x   = ncols_dst == 1 && ids ? ids[channel_dst]          : channel_dst / channel_ratio;
     const int channel_y   = ncols_dst == 1 && ids ? channel_dst % nchannels_y : channel_dst;
@@ -507,7 +507,7 @@ void ggml_cuda_mul_mat_vec_q(
     GGML_ASSERT(        nb0        == ts_dst);
     GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
 
-    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for  batch size 1.
+    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1.
 
     const float   * src1_d =       (const float   *) src1->data;
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
@@ -519,7 +519,7 @@ void ggml_cuda_mul_mat_vec_q(
         const int64_t s11 = src1->nb[1] / ts_src1;
         const int64_t s12 = src1->nb[2] / ts_src1;
         const int64_t s13 = src1->nb[3] / ts_src1;
-        quantize_row_q8_1_cuda(src1_d, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+        quantize_row_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
     }
 
     const int64_t s01 = src0->nb[1] / ts_src0;

ggml/src/ggml-cuda/quantize.cu

@@ -49,29 +49,38 @@ static __global__ void quantize_q8_1(
 
 template <mmq_q8_1_ds_layout ds_layout>
 static __global__ void quantize_mmq_q8_1(
-    const float * __restrict__ x, void * __restrict__ vy, const int64_t kx0, const int64_t kx1, const int64_t kx0_padded) {
+        const float * __restrict__ x, const int32_t * __restrict__ ids, void * __restrict__ vy,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int ne1, const int ne2) {
 
     constexpr int vals_per_scale = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 64 : 32;
     constexpr int vals_per_sum   = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 16 : 32;
 
-    const int64_t ix0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
+    const int64_t i0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
 
-    if (ix0 >= kx0_padded) {
+    if (i0 >= ne0) {
         return;
     }
 
-    const float4 * x4 = (const float4 *) x;
+    const int64_t i1 = blockIdx.y;
+    const int64_t i2 = blockIdx.z % ne2;
+    const int64_t i3 = blockIdx.z / ne2;
 
-    const int64_t ix1 = kx1*blockIdx.z + blockIdx.y;
+    const int64_t i00 = i0;
+    const int64_t i01 = ids ? ids[i1] : i1;
+    const int64_t i02 = i2;
+    const int64_t i03 = i3;
+
+    const float4 * x4 = (const float4 *) x;
 
     block_q8_1_mmq * y = (block_q8_1_mmq *) vy;
 
     const int64_t ib0 = blockIdx.z*((int64_t)gridDim.y*gridDim.x*blockDim.x/QK8_1); // first block of channel
-    const int64_t ib  = ib0 + (ix0 / (4*QK8_1))*kx1 + blockIdx.y;                   // block index in channel
-    const int64_t iqs = ix0 % (4*QK8_1);                                            // quant index in block
+    const int64_t ib  = ib0 + (i0 / (4*QK8_1))*ne1 + blockIdx.y;                    // block index in channel
+    const int64_t iqs = i0 % (4*QK8_1);                                             // quant index in block
 
     // Load 4 floats per thread and calculate max. abs. value between them:
-    const float4 xi = ix0 < kx0 ? x4[(ix1*kx0 + ix0)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+    const float4 xi = i0 < ne00 ? x4[(i03*s03 + i02*s02 + i01*s01 + i00)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
     float amax = fabsf(xi.x);
     amax = fmaxf(amax, fabsf(xi.y));
     amax = fmaxf(amax, fabsf(xi.z));
@@ -87,7 +96,7 @@ static __global__ void quantize_mmq_q8_1(
     if (ds_layout != MMQ_Q8_1_DS_LAYOUT_D4) {
         sum = xi.x + xi.y + xi.z + xi.w;
 
-        // Exchange calculate sum across vals_per_sum/4 threads.
+        // Calculate sums across vals_per_sum/4 threads.
 #pragma unroll
         for (int offset = vals_per_sum/8; offset > 0; offset >>= 1) {
             sum += __shfl_xor_sync(0xFFFFFFFF, sum, offset, WARP_SIZE);
@@ -137,9 +146,10 @@ static __global__ void quantize_mmq_q8_1(
 }
 
 void quantize_row_q8_1_cuda(
-    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
-
+        const float * x, const int32_t * ids, void * vy, const ggml_type type_src0,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
+    GGML_ASSERT(!ids);
     GGML_ASSERT(ne0 % QK8_1 == 0);
 
     const int64_t block_num_x = (ne0 + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
@@ -150,9 +160,9 @@ void quantize_row_q8_1_cuda(
 }
 
 void quantize_mmq_q8_1_cuda(
-    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
-
+        const float * x, const int32_t * ids, void * vy, const ggml_type type_src0,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
     GGML_ASSERT(ne0 % (4*QK8_1) == 0);
 
     const int64_t block_num_x = (ne0 + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
@@ -161,21 +171,18 @@ void quantize_mmq_q8_1_cuda(
     switch (mmq_get_q8_1_ds_layout(type_src0)) {
         case MMQ_Q8_1_DS_LAYOUT_D4:
             quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D4>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, ne1, ne0);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
             break;
         case MMQ_Q8_1_DS_LAYOUT_DS4:
             quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_DS4>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, ne1, ne0);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
             break;
         case MMQ_Q8_1_DS_LAYOUT_D2S6:
             quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D2S6>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, ne1, ne0);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
             break;
         default:
             GGML_ABORT("fatal error");
             break;
     }
-    GGML_UNUSED(s01);
-    GGML_UNUSED(s02);
-    GGML_UNUSED(s03);
 }

ggml/src/ggml-cuda/quantize.cuh

@@ -12,13 +12,16 @@ static_assert(MATRIX_ROW_PADDING %    CUDA_QUANTIZE_BLOCK_SIZE      == 0, "Risk
 static_assert(MATRIX_ROW_PADDING % (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ) == 0, "Risk of out-of-bounds access.");
 
 typedef void (*quantize_cuda_t)(
-    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
 
 void quantize_row_q8_1_cuda(
-    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
 
 void quantize_mmq_q8_1_cuda(
-    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);

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

@@ -2443,7 +2443,7 @@ static bool ggml_metal_encode_node(
 #if 0
                 // cpy to tmp buffer in MTLHeap
 
-                id<MTLBuffer> h_src0 = h_src0 = ggml_metal_mem_pool_alloc(mem_pool, ggml_nbytes(src0));
+                id<MTLBuffer> h_src0 = ggml_metal_mem_pool_alloc(mem_pool, ggml_nbytes(src0));
                 if (!h_src0) {
                     GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, ggml_nbytes(src0));
                     return false;
@@ -2947,6 +2947,12 @@ static bool ggml_metal_encode_node(
                         default: break;
                     }
 
+                    id<MTLBuffer> h_dst = ggml_metal_mem_pool_alloc(mem_pool, sizeof(float)*GGML_PAD(ne01, 64)*GGML_PAD(ne11, 32)*ne12*ne13);
+                    if (!h_dst) {
+                        GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, ggml_nbytes(src0));
+                        return false;
+                    }
+
                     id<MTLComputePipelineState> pipeline = nil;
 
                     switch (src0->type) {
@@ -2986,8 +2992,8 @@ static bool ggml_metal_encode_node(
                         /*.nb11 =*/ nb11,
                         /*.nb12 =*/ nb12,
                         /*.nb13 =*/ nb13,
-                        /*.ne0  =*/ ne0,
-                        /*.ne1  =*/ ne1,
+                        /*.ne0  =*/ GGML_PAD(ne01, 64),
+                        /*.ne1  =*/ GGML_PAD(ne11, 32),
                         /*.r2   =*/ r2,
                         /*.r3   =*/ r3,
                     };
@@ -2996,10 +3002,40 @@ static bool ggml_metal_encode_node(
                     [encoder setBytes:&args    length:sizeof(args) atIndex:0];
                     [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
                     [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
-                    [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+                  //[encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+                    [encoder setBuffer:h_dst   offset:0            atIndex:3];
 
                     [encoder setThreadgroupMemoryLength:8192 atIndex:0];
                     [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+
+                    ggml_metal_kargs_cpy args_cpy = {
+                        /*.ne00 =*/ ne0,
+                        /*.ne01 =*/ ne1,
+                        /*.ne02 =*/ ne2,
+                        /*.ne03 =*/ ne3,
+                        /*.nb00 =*/ nb0,
+                        /*.nb01 =*/ nb0*GGML_PAD(ne01, 64),
+                        /*.nb02 =*/ nb0*GGML_PAD(ne01, 64)*GGML_PAD(ne11, 32),
+                        /*.nb03 =*/ nb0*GGML_PAD(ne01, 64)*GGML_PAD(ne11, 32)*ne12,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                        /*.ne2  =*/ ne2,
+                        /*.ne3  =*/ ne3,
+                        /*.nb0  =*/ nb0,
+                        /*.nb1  =*/ nb1,
+                        /*.nb2  =*/ nb2,
+                        /*.nb3  =*/ nb3,
+                    };
+
+                    [encoder setComputePipelineState:ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline];
+
+                    [encoder setBytes:&args_cpy length:sizeof(args_cpy) atIndex:0];
+                    [encoder setBuffer:h_dst  offset:0                  atIndex:1];
+                    [encoder setBuffer:id_dst offset:offs_dst           atIndex:2];
+
+                    const int nth_cpy = MIN(1024, ne0);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth_cpy, 1, 1)];
                 } else {
                     id<MTLComputePipelineState> pipeline = nil;
 

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

@@ -6305,34 +6305,34 @@ kernel void kernel_mul_mm(
         }
     } else {
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *) shmem) \
-                                     + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
-        for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
-        }
+        //threadgroup_barrier(mem_flags::mem_threadgroup);
+        //threadgroup float * temp_str = ((threadgroup float *) shmem) \
+        //                             + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
+        //for (short i = 0; i < 8; i++) {
+        //    simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
+        //}
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+        //threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        if (sgitg == 0) {
-            for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
-                device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*args.ne0 + im*args.ne1*args.ne0;
-                device float4 * D4 = (device float4 *) D;
+        //if (sgitg == 0) {
+        //    for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
+        //        device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*args.ne0 + im*args.ne1*args.ne0;
+        //        device float4 * D4 = (device float4 *) D;
 
-                threadgroup float  * C  = temp_str + (j*BLOCK_SIZE_M);
-                threadgroup float4 * C4 = (threadgroup float4 *) C;
+        //        threadgroup float  * C  = temp_str + (j*BLOCK_SIZE_M);
+        //        threadgroup float4 * C4 = (threadgroup float4 *) C;
 
-                int i = 0;
-                for (; i < n_rows/4; i++) {
-                    *(D4 + i) = *(C4 + i);
-                }
+        //        int i = 0;
+        //        for (; i < n_rows/4; i++) {
+        //            *(D4 + i) = *(C4 + i);
+        //        }
 
-                i *= 4;
-                for (; i < n_rows; i++) {
-                    *(D + i) = *(C + i);
-                }
-            }
-        }
+        //        i *= 4;
+        //        for (; i < n_rows; i++) {
+        //            *(D + i) = *(C + i);
+        //        }
+        //    }
+        //}
     }
 }
 

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -518,6 +518,11 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
     result.view_src = reinterpret_cast<uint64_t>(tensor->view_src);
     result.view_offs = tensor->view_offs;
     result.data = reinterpret_cast<uint64_t>(tensor->data);
+
+    // Avoid sending uninitialized data over the wire
+    memset(result.name, 0, sizeof(result.name));
+    memset(result.padding, 0, sizeof(result.padding));
+
     snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
     return result;
 }
@@ -982,8 +987,21 @@ bool rpc_server::buffer_clear(const rpc_msg_buffer_clear_req & request) {
 }
 
 ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor) {
+    // Validate tensor type before using it
+    if (tensor->type >= GGML_TYPE_COUNT) {
+        GGML_LOG_ERROR("[%s] invalid tensor type received: %u\n", __func__, tensor->type);
+        return nullptr;
+    }
+
     ggml_tensor * result = ggml_new_tensor_4d(ctx, (ggml_type) tensor->type,
         tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+
+    // ggml_new_tensor_4d might fail if dimensions are invalid, although less likely to crash than invalid type
+    if (result == nullptr) {
+        GGML_LOG_ERROR("[%s] ggml_new_tensor_4d failed for type %u\\n", __func__, tensor->type);
+        return nullptr;
+    }
+
     for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
         result->nb[i] = tensor->nb[i];
     }
@@ -1043,7 +1061,9 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
         const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
 
         if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
-            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+            GGML_LOG_ERROR("[%s] tensor data region (data=0x%" PRIx64 ", offset=%" PRIu64 ", size=%zu) out of buffer bounds [0x%zx, 0x%zx)\n",
+                           __func__, in_tensor->data, offset, size, p0, p1);
+            return false;
         }
     }
 
@@ -1118,7 +1138,9 @@ bool rpc_server::set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set
         const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
 
         if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
-            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+            GGML_LOG_ERROR("[%s] tensor data region (data=0x%" PRIx64 ", offset=%" PRIu64 ", size=%zu, hash=0x%" PRIx64 ") out of buffer bounds [0x%zx, 0x%zx)\n",
+                           __func__, in_tensor->data, offset, size, *hash, p0, p1);
+            return false;
         }
     }
     ggml_backend_tensor_set(tensor, cached_file.data(), offset, size);
@@ -1183,7 +1205,9 @@ bool rpc_server::get_tensor(const rpc_msg_get_tensor_req & request, std::vector<
         if (request.tensor.data + request.offset < p0 ||
             request.tensor.data + request.offset >= p1 ||
             request.size > (p1 - request.tensor.data - request.offset)) {
-                GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+                GGML_LOG_ERROR("[%s] requested tensor region (data=0x%" PRIx64 ", offset=%" PRIu64 ", size=%" PRIu64 ") out of buffer bounds [0x%zx, 0x%zx)\n",
+                               __func__, request.tensor.data, request.offset, request.size, p0, p1);
+                return false;
         }
     }
 
@@ -1237,22 +1261,50 @@ ggml_tensor * rpc_server::create_node(uint64_t id,
                                       struct ggml_context * ctx,
                                       const std::unordered_map<uint64_t, const rpc_tensor*> & tensor_ptrs,
                                       std::unordered_map<uint64_t, struct ggml_tensor*> & tensor_map) {
-    if (id == 0) {
-        return nullptr;
-    }
     if (tensor_map.find(id) != tensor_map.end()) {
         return tensor_map[id];
     }
-    const rpc_tensor * tensor = tensor_ptrs.at(id);
+    // Safely find the tensor pointer
+    auto it_ptr = tensor_ptrs.find(id);
+    if (it_ptr == tensor_ptrs.end()) {
+        return nullptr;
+    }
+    const rpc_tensor * tensor = it_ptr->second;
+
     struct ggml_tensor * result = deserialize_tensor(ctx, tensor);
     if (result == nullptr) {
         return nullptr;
     }
     tensor_map[id] = result;
     for (int i = 0; i < GGML_MAX_SRC; i++) {
-        result->src[i] = create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map);
+        // Check if the source ID is 0 before calling create_node recursively
+        if (tensor->src[i] == 0) {
+            result->src[i] = nullptr;
+        } else {
+            result->src[i] = create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map);
+            // If the recursive call failed for a non-zero ID, propagate the error
+            if (result->src[i] == nullptr) {
+                GGML_LOG_ERROR("[%s] failed to create source node %d (src_id=%" PRIu64 ") for node id %" PRIu64 "\n",
+                               __func__, i, tensor->src[i], id);
+                // Must return nullptr to signal failure up the call stack
+                return nullptr;
+            }
+        }
+    }
+
+    // Handle view_src similarly
+    if (tensor->view_src == 0) {
+        result->view_src = nullptr;
+    } else {
+        result->view_src = create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map);
+        // If the recursive call failed for a non-zero ID, propagate the error
+        if (result->view_src == nullptr) {
+            GGML_LOG_ERROR("[%s] failed to create view_src node (view_src_id=%" PRIu64 ") for node id %" PRIu64 "\n",
+                           __func__, tensor->view_src, id);
+            // Must return nullptr to signal failure up the call stack
+            return nullptr;
+        }
     }
-    result->view_src = create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map);
     result->view_offs = tensor->view_offs;
     return result;
 }
@@ -1278,6 +1330,7 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph
     GGML_PRINT_DEBUG("[%s] n_nodes: %u, n_tensors: %u\n", __func__, n_nodes, n_tensors);
 
     size_t buf_size = ggml_tensor_overhead()*(n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
+
     struct ggml_init_params params = {
         /*.mem_size   =*/ buf_size,
         /*.mem_buffer =*/ NULL,
@@ -1297,6 +1350,14 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph
         int64_t id;
         memcpy(&id, &nodes[i], sizeof(id));
         graph->nodes[i] = create_node(id, ctx, tensor_ptrs, tensor_map);
+
+        // Check if create_node failed for a *non-zero* ID.
+        // If id was 0, create_node returning nullptr is expected.
+        // If id was non-zero and create_node returned nullptr, it indicates a deserialization error.
+        if (graph->nodes[i] == nullptr && id != 0) {
+            GGML_LOG_ERROR("[%s] failed to create graph node %d (id=%" PRId64 ")\n", __func__, i, id);
+            return false;
+        }
     }
     ggml_status status = ggml_backend_graph_compute(backend, graph);
     response.result = status;
@@ -1361,7 +1422,9 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir,
                     return;
                 }
                 rpc_msg_get_alloc_size_rsp response;
-                server.get_alloc_size(request, response);
+                if (!server.get_alloc_size(request, response)) {
+                    return;
+                }
                 if (!send_msg(sockfd, &response, sizeof(response))) {
                     return;
                 }

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -71,6 +71,22 @@ if (Vulkan_FOUND)
         add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
     endif()
 
+    # Compile a test shader to determine whether GL_EXT_bfloat16 is supported.
+    # If it's not, there will be an error to stderr.
+    # If it's supported, set a define to indicate that we should compile those shaders
+    execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_bfloat16_support.comp"
+                    OUTPUT_VARIABLE glslc_output
+                    ERROR_VARIABLE glslc_error)
+
+    if (${glslc_error} MATCHES ".*extension not supported: GL_EXT_bfloat16.*")
+        message(STATUS "GL_EXT_bfloat16 not supported by glslc")
+        set(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT OFF)
+    else()
+        message(STATUS "GL_EXT_bfloat16 supported by glslc")
+        set(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT ON)
+        add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+    endif()
+
     target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
     target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
 
@@ -142,6 +158,7 @@ if (Vulkan_FOUND)
                     -DGGML_VULKAN_COOPMAT_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT_GLSLC_SUPPORT}
                     -DGGML_VULKAN_COOPMAT2_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT}
                     -DGGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT=${GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT}
+                    -DGGML_VULKAN_BFLOAT16_GLSLC_SUPPORT=${GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT}
             BUILD_COMMAND ${CMAKE_COMMAND} --build .
             INSTALL_COMMAND ${CMAKE_COMMAND} --install .
             INSTALL_DIR ${CMAKE_BINARY_DIR}

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

@@ -51,6 +51,24 @@
 
 #include "ggml-vulkan-shaders.hpp"
 
+// remove this once it's more widely available in the SDK
+#if !defined(VK_KHR_shader_bfloat16)
+
+#define VK_KHR_shader_bfloat16 1
+#define VK_KHR_SHADER_BFLOAT16_SPEC_VERSION                          1
+#define VK_KHR_SHADER_BFLOAT16_EXTENSION_NAME                        "VK_KHR_shader_bfloat16"
+#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR ((VkStructureType)1000141000)
+#define VK_COMPONENT_TYPE_BFLOAT16_KHR                               ((VkComponentTypeKHR)1000141000)
+
+typedef struct VkPhysicalDeviceShaderBfloat16FeaturesKHR {
+    VkStructureType                       sType;
+    void*                                 pNext;
+    VkBool32                              shaderBFloat16Type;
+    VkBool32                              shaderBFloat16DotProduct;
+    VkBool32                              shaderBFloat16CooperativeMatrix;
+} VkPhysicalDeviceShaderBfloat16FeaturesKHR;
+#endif
+
 #define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1))
 #define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
 static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
@@ -266,8 +284,9 @@ struct vk_device_struct {
     bool subgroup_require_full_support;
 
     bool coopmat_support;
-    bool coopmat_acc_f32_support;
-    bool coopmat_acc_f16_support;
+    bool coopmat_acc_f32_support {};
+    bool coopmat_acc_f16_support {};
+    bool coopmat_bf16_support {};
     uint32_t coopmat_m;
     uint32_t coopmat_n;
     uint32_t coopmat_k;
@@ -293,6 +312,7 @@ struct vk_device_struct {
 
     vk_matmul_pipeline pipeline_matmul_f32 {};
     vk_matmul_pipeline pipeline_matmul_f32_f16 {};
+    vk_matmul_pipeline pipeline_matmul_bf16 {};
     vk_matmul_pipeline2 pipeline_matmul_f16;
     vk_matmul_pipeline2 pipeline_matmul_f16_f32;
 
@@ -301,6 +321,7 @@ struct vk_device_struct {
     vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_COUNT];
 
     vk_matmul_pipeline pipeline_matmul_id_f32 {};
+    vk_matmul_pipeline pipeline_matmul_id_bf16 {};
     vk_matmul_pipeline2 pipeline_matmul_id_f16;
     vk_matmul_pipeline2 pipeline_matmul_id_f16_f32;
 
@@ -333,8 +354,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_clamp_f32;
     vk_pipeline pipeline_pad_f32;
     vk_pipeline pipeline_repeat_f32, pipeline_repeat_back_f32;
-    vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16;
-    vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16;
+    vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16, pipeline_cpy_f32_bf16;
+    vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16, pipeline_contig_cpy_f32_bf16;
     vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
     vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
     vk_pipeline pipeline_norm_f32;
@@ -368,6 +389,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_rwkv_wkv6_f32;
     vk_pipeline pipeline_rwkv_wkv7_f32;
     vk_pipeline pipeline_opt_step_adamw_f32;
+    vk_pipeline pipeline_conv2d_dw_whcn_f32;
+    vk_pipeline pipeline_conv2d_dw_cwhn_f32;
 
     // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
     vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2];
@@ -680,6 +703,24 @@ struct vk_op_rwkv_wkv7_push_constants {
     uint32_t H;
 };
 
+struct vk_op_conv2d_dw_push_constants {
+    uint32_t ne;
+    uint32_t batches;
+    uint32_t channels;
+    uint32_t dst_w;
+    uint32_t dst_h;
+    uint32_t src_w;
+    uint32_t src_h;
+    uint32_t knl_w;
+    uint32_t knl_h;
+    int32_t stride_x;
+    int32_t stride_y;
+    int32_t pad_x;
+    int32_t pad_y;
+    int32_t dilation_x;
+    int32_t dilation_y;
+};
+
 struct vk_op_upscale_push_constants {
     uint32_t ne; uint32_t a_offset; uint32_t d_offset;
     uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
@@ -1791,6 +1832,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     if (!device->pipeline_matmul_id_f32) {
         device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
     }
+    if (!device->pipeline_matmul_bf16) {
+        device->pipeline_matmul_bf16 = std::make_shared<vk_matmul_pipeline_struct>();
+    }
+    if (!device->pipeline_matmul_id_bf16) {
+        device->pipeline_matmul_id_bf16 = std::make_shared<vk_matmul_pipeline_struct>();
+    }
 
     std::vector<std::future<void>> compiles;
     auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint,
@@ -1900,6 +1947,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT)   \
 
         CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3)
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (device->coopmat_bf16_support) {
+            CREATE_MM(pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3)
+        }
+#endif
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
@@ -1921,6 +1973,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL].f16acc,  matmul_iq4_nl_f16,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
 
         CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (device->coopmat_bf16_support) {
+            CREATE_MM(pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
+        }
+#endif
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
@@ -1974,6 +2031,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (device->coopmat_bf16_support) {
+            CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, )
+        }
+#endif
 
         if (device->coopmat_acc_f16_support) {
             CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -2022,6 +2084,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (device->coopmat_bf16_support) {
+            CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+        }
+#endif
 
         if (device->coopmat_acc_f16_support) {
             CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
@@ -2104,6 +2171,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
 
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+
         CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -2139,6 +2208,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
 
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id);
+
         CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
@@ -2191,6 +2262,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
 
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+
         CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -2226,6 +2299,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
 
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id);
+
         CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
@@ -2246,8 +2321,26 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc,   matmul_id_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc,  matmul_id_iq4_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc,  matmul_id_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-#undef CREATE_MM
     }
+    // reusing CREATE_MM from the fp32 path
+    if ((device->coopmat2 || device->coopmat_support)
+#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
+        && !device->coopmat_bf16_support
+#endif
+        ) {
+        // use scalar tile sizes
+        l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 };
+        m_warptile = { 128,  64,  64, 16, subgroup_size_8, 32, 2, 4, 2, 1, subgroup_size_8 };
+        s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, 2, 2, 1, subgroup_size_8 };
+
+        l_wg_denoms = {128, 128, 1 };
+        m_wg_denoms = { 64,  64, 1 };
+        s_wg_denoms = { 32,  32, 1 };
+
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id);
+    }
+#undef CREATE_MM
 
     // mul mat vec
 
@@ -2266,6 +2359,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f32_f32_len, mul_mat_vec_bf16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
@@ -2288,6 +2382,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f16_f32_len, mul_mat_vec_bf16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
@@ -2311,6 +2406,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_BF16], "mul_mat_vec_id_bf16_f32", mul_mat_vec_id_bf16_f32_len, mul_mat_vec_id_bf16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
@@ -2356,6 +2452,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     // get_rows
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F32 ], "get_rows_f32",  get_rows_f32_len,  get_rows_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F16 ], "get_rows_f16",  get_rows_f16_len,  get_rows_f16_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_BF16], "get_rows_bf16", get_rows_bf16_len, get_rows_bf16_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_0], "get_rows_q4_0", get_rows_q4_0_len, get_rows_q4_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_1], "get_rows_q4_1", get_rows_q4_1_len, get_rows_q4_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_0], "get_rows_q5_0", get_rows_q5_0_len, get_rows_q5_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
@@ -2373,6 +2470,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F32 ], "get_rows_f32_f32",  get_rows_f32_f32_len,  get_rows_f32_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F16 ], "get_rows_f16_f32",  get_rows_f16_f32_len,  get_rows_f16_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_BF16], "get_rows_bf16_f32", get_rows_bf16_f32_len, get_rows_bf16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_0], "get_rows_q4_0_f32", get_rows_q4_0_f32_len, get_rows_q4_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_1], "get_rows_q4_1_f32", get_rows_q4_1_f32_len, get_rows_q4_1_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_0], "get_rows_q5_0_f32", get_rows_q5_0_f32_len, get_rows_q5_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
@@ -2399,7 +2497,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len,              mul_mat_vec_p021_f16_f32_data,              "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true);
         }
     }
-    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 9 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
@@ -2410,10 +2508,13 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_bf16,"cpy_f32_bf16",cpy_f32_bf16_len,cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f32, "contig_cpy_f32_f32", contig_cpy_f32_f32_len, contig_cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_bf16,"contig_cpy_f32_bf16",contig_cpy_f32_bf16_len,contig_cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
     if (device->float_controls_rte_fp16) {
         ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
@@ -2529,6 +2630,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_cwhn_f32, "conv2d_dw_cwhn_f32", conv2d_dw_cwhn_f32_len, conv2d_dw_cwhn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
+
     for (auto &c : compiles) {
         c.wait();
     }
@@ -2578,6 +2682,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         bool coopmat2_support = false;
         device->coopmat_support = false;
         device->integer_dot_product = false;
+        bool bfloat16_support = false;
 
         for (const auto& properties : ext_props) {
             if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
@@ -2608,6 +2713,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
                        !getenv("GGML_VK_DISABLE_INTEGER_DOT_PRODUCT")) {
                 device->integer_dot_product = true;
 #endif
+            } else if (strcmp("VK_KHR_shader_bfloat16", properties.extensionName) == 0 &&
+                       !getenv("GGML_VK_DISABLE_BFLOAT16")) {
+                bfloat16_support = true;
             }
         }
 
@@ -2794,6 +2902,17 @@ static vk_device ggml_vk_get_device(size_t idx) {
         }
 #endif
 
+#if defined(VK_KHR_shader_bfloat16)
+        VkPhysicalDeviceShaderBfloat16FeaturesKHR bfloat16_features {};
+        bfloat16_features.pNext = nullptr;
+        bfloat16_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR;
+        if (bfloat16_support) {
+            last_struct->pNext = (VkBaseOutStructure *)&bfloat16_features;
+            last_struct = (VkBaseOutStructure *)&bfloat16_features;
+            device_extensions.push_back("VK_KHR_shader_bfloat16");
+        }
+#endif
+
         VkPhysicalDeviceMaintenance4Features maint4_features {};
         maint4_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_4_FEATURES;
         if (maintenance4_support) {
@@ -2991,6 +3110,25 @@ static vk_device ggml_vk_get_device(size_t idx) {
                     device->coopmat_int_n = prop.NSize;
                     device->coopmat_int_k = prop.KSize;
                 }
+#if defined(VK_KHR_shader_bfloat16) && defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+                if (prop.AType == VK_COMPONENT_TYPE_BFLOAT16_KHR &&
+                    prop.BType == VK_COMPONENT_TYPE_BFLOAT16_KHR &&
+                    prop.CType == VK_COMPONENT_TYPE_FLOAT32_KHR &&
+                    prop.ResultType == VK_COMPONENT_TYPE_FLOAT32_KHR &&
+                    (vk::ScopeKHR)prop.scope == vk::ScopeKHR::eSubgroup
+                ) {
+                    // coopmat sizes not set yet
+                    if (device->coopmat_m == 0) {
+                        device->coopmat_bf16_support = true;
+                        device->coopmat_m = prop.MSize;
+                        device->coopmat_n = prop.NSize;
+                        device->coopmat_k = prop.KSize;
+                    } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) {
+                        // Only enable if shape is identical
+                        device->coopmat_bf16_support = true;
+                    }
+                }
+#endif
             }
 
             if (device->coopmat_m == 0 || !device->coopmat_acc_f32_support) {
@@ -2998,11 +3136,19 @@ static vk_device ggml_vk_get_device(size_t idx) {
                 GGML_LOG_DEBUG("ggml_vulkan: WARNING: No suitable matrix core mode found. Disabling matrix cores.\n");
                 device->coopmat_support = false;
             }
+            if (getenv("GGML_VK_DISABLE_BFLOAT16")) {
+                device->coopmat_bf16_support = false;
+            }
         }
 
         if (device->coopmat_support) {
             device_extensions.push_back("VK_KHR_cooperative_matrix");
         }
+#if defined(VK_KHR_shader_bfloat16)
+        if (device->coopmat_bf16_support) {
+            device_extensions.push_back("VK_KHR_shader_bfloat16");
+        }
+#endif
 #endif
         device->name = GGML_VK_NAME + std::to_string(idx);
 
@@ -3459,6 +3605,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
         return ctx->device->pipeline_matmul_f32_f16;
     }
+    if (src0_type == GGML_TYPE_BF16 && src1_type == GGML_TYPE_BF16) {
+        return ctx->device->pipeline_matmul_bf16;
+    }
     if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) {
         if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
             return ctx->device->pipeline_matmul_f16_f32.f16acc;
@@ -3530,6 +3679,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
     switch (a_type) {
         case GGML_TYPE_F32:
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -3562,6 +3712,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
         return ctx->device->pipeline_matmul_id_f32;
     }
+    if (src0_type == GGML_TYPE_BF16 && src1_type == GGML_TYPE_BF16) {
+        return ctx->device->pipeline_matmul_id_bf16;
+    }
     if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) {
         if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
             return ctx->device->pipeline_matmul_id_f16_f32.f16acc;
@@ -3615,6 +3768,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context
     switch (a_type) {
         case GGML_TYPE_F32:
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -4350,6 +4504,13 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_cpy_f16_f16;
         }
     }
+    if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_BF16) {
+        if (contig) {
+            return ctx->device->pipeline_contig_cpy_f32_bf16;
+        } else {
+            return ctx->device->pipeline_cpy_f32_bf16;
+        }
+    }
     if (src->type == GGML_TYPE_F32) {
         switch (to) {
         case GGML_TYPE_Q4_0:
@@ -4477,8 +4638,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) ||
                               !ggml_vk_dim01_contiguous(src0);
     const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) ||
+                              (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) ||
                               !ggml_vk_dim01_contiguous(src1);
 
+    // If src0 is BF16, try to use a BF16 x BF16 multiply
+    ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16;
+
     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
 
     bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0;
@@ -4488,25 +4653,25 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
 
     if (mmp == nullptr) {
         // Fall back to f16 dequant mul mat
-        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]);
+        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]);
         quantize_y = false;
     }
 
     const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
-    const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig);
+    const bool qy_needs_dequant = !quantize_y && ((src1->type != f16_type && !y_f32_kernel) || y_non_contig);
 
     if (qx_needs_dequant) {
         // Fall back to dequant + f16 mulmat
-        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16, (ggml_prec)dst->op_params[0]);
+        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, f16_type, y_f32_kernel ? GGML_TYPE_F32 : f16_type, (ggml_prec)dst->op_params[0]);
     }
 
     // Not implemented
     GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
 
-    const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? GGML_TYPE_F16 : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type)));
+    const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type)));
     const bool aligned = !quantize_y && ne10 == kpad && ne01 > 8 && ne11 > 8;
 
-    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type));
+    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type));
 
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) : ne11;
@@ -4527,12 +4692,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     vk_pipeline to_q8_1 = nullptr;
 
     if (x_non_contig) {
-        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16);
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type);
     } else {
         to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
     }
     if (y_non_contig) {
-        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16);
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type);
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
@@ -4949,6 +5114,8 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint64_t nb01 = src0->nb[1];
     const uint64_t nb02 = src0->nb[2];
 
+    const uint64_t nb12 = src1->nb[2];
+
     // const uint64_t ne10 = src1->ne[0];
     const uint64_t ne11 = src1->ne[1];
     const uint64_t ne12 = src1->ne[2];
@@ -4974,6 +5141,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
 
     const uint32_t row_stride_x = nb01 / sizeof(ggml_fp16_t);
     const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t);
+    const uint32_t channel_stride_y = nb12 / sizeof(float);
 
     const uint64_t qx_sz = ggml_nbytes(src0);
     const uint64_t qy_sz = ggml_nbytes(src1);
@@ -5004,7 +5172,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset;
 
     // compute
-    const std::array<uint32_t, 7> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
+    const std::array<uint32_t, 9> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
         { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
@@ -5029,7 +5197,7 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, c
     // mul_mat_vec supports batching ne12*ne13 when ne11==1, or treating ne11 as the batch size (up to four)
     // when ne12 and ne13 are one.
     } else if ((dst->ne[1] == 1 || (dst->ne[1] <= mul_mat_vec_max_cols && src1->ne[2] * src1->ne[3] == 1)) &&
-               (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
+               (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || ggml_is_quantized(src0->type))) {
         ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst, dryrun);
     } else {
         ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst, dryrun);
@@ -5097,27 +5265,31 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) ||
                               !ggml_vk_dim01_contiguous(src0);
     const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) ||
+                              (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) ||
                               !ggml_vk_dim01_contiguous(src1);
 
+    // If src0 is BF16, try to use a BF16 x BF16 multiply
+    ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16;
+
     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
 
-    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]);
+    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]);
 
     const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
-    const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
+    const bool qy_needs_dequant = (src1->type != f16_type && !y_f32_kernel) || y_non_contig;
 
     if (qx_needs_dequant) {
         // Fall back to dequant + f16 mulmat
-        mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16, (ggml_prec)dst->op_params[0]);
+        mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, f16_type, y_f32_kernel ? GGML_TYPE_F32 : f16_type, (ggml_prec)dst->op_params[0]);
     }
 
     // Not implemented
     GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
 
-    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? GGML_TYPE_F16 : src0->type));
+    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? f16_type : src0->type));
     const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8;
 
-    vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type);
+    vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? f16_type : src0->type);
 
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11;
@@ -5136,12 +5308,12 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     vk_pipeline to_fp16_vk_1 = nullptr;
 
     if (x_non_contig) {
-        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16);
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type);
     } else {
         to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
     }
     if (y_non_contig) {
-        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16);
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type);
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
@@ -5988,6 +6160,15 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_leaky_relu_f32;
         }
         return nullptr;
+    case GGML_OP_CONV_2D_DW:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            if (ggml_is_contiguous(src1)) {
+                return ctx->device->pipeline_conv2d_dw_whcn_f32;
+            } else if (ggml_is_contiguous_channels(src1)) {
+                return ctx->device->pipeline_conv2d_dw_cwhn_f32;
+            }
+        }
+        return nullptr;
     default:
         return nullptr;
     }
@@ -6014,6 +6195,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_ROPE:
     case GGML_OP_RMS_NORM:
+    case GGML_OP_CONV_2D_DW:
         return true;
     default:
         return false;
@@ -6310,6 +6492,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_CONCAT:
     case GGML_OP_UPSCALE:
     case GGML_OP_UNARY:
+    case GGML_OP_CONV_2D_DW:
         {
             const uint32_t ne = ggml_nelements(dst);
             if (ne > 262144) {
@@ -7096,6 +7279,30 @@ static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, c
     }, dryrun);
 }
 
+static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    vk_op_conv2d_dw_push_constants p{};
+    p.ne = ggml_nelements(dst);
+    p.channels = dst->ne[2];
+    p.batches = dst->ne[3];
+    p.dst_w = dst->ne[0];
+    p.dst_h = dst->ne[1];
+    p.src_w = src1->ne[0];
+    p.src_h = src1->ne[1];
+    p.knl_w = src0->ne[0];
+    p.knl_h = src0->ne[1];
+    p.stride_x = dst->op_params[0];
+    p.stride_y = dst->op_params[1];
+    p.pad_x = dst->op_params[2];
+    p.pad_y = dst->op_params[3];
+    p.dilation_x = dst->op_params[4];
+    p.dilation_y = dst->op_params[5];
+
+    GGML_ASSERT(src0->ne[3] == p.channels);
+    GGML_ASSERT(src1->ne[3] == p.batches);
+
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_2D_DW, std::move(p), dryrun);
+}
+
 static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     const float * op_params = (const float *)dst->op_params;
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun);
@@ -8116,6 +8323,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
     case GGML_OP_POOL_2D:
+    case GGML_OP_CONV_2D_DW:
     case GGML_OP_RWKV_WKV6:
     case GGML_OP_RWKV_WKV7:
     case GGML_OP_LEAKY_RELU:
@@ -8179,6 +8387,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         case GGML_OP_IM2COL:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_POOL_2D:
+        case GGML_OP_CONV_2D_DW:
         case GGML_OP_LEAKY_RELU:
             {
                 // These operations all go through ggml_vk_op_f32, so short-circuit and
@@ -8352,6 +8561,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_POOL_2D:
         ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_CONV_2D_DW:
+        ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node, dryrun);
+
         break;
     case GGML_OP_LEAKY_RELU:
         ggml_vk_leaky_relu(ctx, compute_ctx, src0, node, dryrun);
@@ -8473,6 +8686,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
     case GGML_OP_POOL_2D:
+    case GGML_OP_CONV_2D_DW:
     case GGML_OP_RWKV_WKV6:
     case GGML_OP_RWKV_WKV7:
     case GGML_OP_LEAKY_RELU:
@@ -9227,6 +9441,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 switch (src0_type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -9262,10 +9477,15 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (a->ne[3] != b->ne[3]) {
                     return false;
                 }
-                if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) ||
+                if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_BF16) ||
                     !(ggml_vk_dim01_contiguous(op->src[1]) || op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16)) {
                     return false;
                 }
+                if (op->src[0]->type == GGML_TYPE_BF16 && op->src[1]->type == GGML_TYPE_F16) {
+                    // We currently don't have a bf16 x f16 shader, or an fp16->bf16 copy shader.
+                    // So don't support this combination for now.
+                    return false;
+                }
 
                 return true;
             } break;
@@ -9338,6 +9558,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 switch (op->src[0]->type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -9368,6 +9589,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     switch (src1_type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -9442,6 +9664,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_COUNT_EQUAL:
         case GGML_OP_IM2COL:
         case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_CONV_2D_DW:
         case GGML_OP_POOL_2D:
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:

ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt

@@ -12,6 +12,9 @@ endif()
 if (GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
     add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
 endif()
+if (GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+    add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+endif()
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)

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

@@ -18,7 +18,11 @@ void main() {
     // fast path for when all four iterations are in-bounds
     if (idx + (num_iter-1)*num_threads < p.ne) {
         [[unroll]] for (uint i = 0; i < num_iter; ++i) {
-#ifndef OPTIMIZATION_ERROR_WORKAROUND
+
+#if defined(DATA_D_BF16)
+            float f = float(data_a[get_aoffset() + idx]);
+            data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f));
+#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
             data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
 #else
             data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
@@ -31,7 +35,10 @@ void main() {
                 continue;
             }
 
-#ifndef OPTIMIZATION_ERROR_WORKAROUND
+#if defined(DATA_D_BF16)
+            float f = float(data_a[get_aoffset() + idx]);
+            data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f));
+#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
             data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
 #else
             data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];

ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp

@@ -0,0 +1,105 @@
+#version 450
+
+#include "types.comp"
+
+layout (push_constant) uniform parameter
+{
+    uint ne;
+    uint batches;
+    uint channels;
+    uint dst_w;
+    uint dst_h;
+    uint src_w;
+    uint src_h;
+    uint knl_w;
+    uint knl_h;
+    int stride_x;
+    int stride_y;
+    int pad_x;
+    int pad_y;
+    int dilation_x;
+    int dilation_y;
+} p;
+
+layout (binding = 0) readonly buffer A {A_TYPE knl_data[];};
+layout (binding = 1) readonly buffer B {B_TYPE src_data[];};
+layout (binding = 2) writeonly buffer D {D_TYPE dst_data[];};
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE conv_2d_dw_whcn(uint idx) {
+    uint i0 = idx / p.dst_w;
+    uint dst_x = idx - i0 * p.dst_w;
+    uint i1 = i0 / p.dst_h;
+    uint dst_y = i0 - i1 * p.dst_h;
+    uint n = i1 / p.channels;
+    uint c = i1 - n * p.channels;
+
+    uint src_i = n * p.channels * p.src_h * p.src_w + c * p.src_h * p.src_w;
+    uint knl_i = c * p.knl_h * p.knl_w;
+
+    FLOAT_TYPE sum = 0.0;
+    for (uint knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+        uint src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y;
+        if (src_y >= p.src_h) { // src_y < 0 will wrap to a large unsigned int
+            continue;
+        }
+        for (uint knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+            uint src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x;
+            if (src_x >= p.src_w) { // src_x < 0 will wrap to a large unsigned int
+                continue;
+            }
+            FLOAT_TYPE v = FLOAT_TYPE(src_data[src_i + src_y * p.src_w + src_x]);
+            FLOAT_TYPE k = FLOAT_TYPE(knl_data[knl_i + knl_y * p.knl_w + knl_x]);
+            sum = fma(v, k, sum);
+        }
+    }
+    return sum;
+}
+
+FLOAT_TYPE conv_2d_dw_cwhn(uint idx) {
+    uint i0 = idx / p.channels;
+    uint c = idx - i0 * p.channels;
+    uint i1 = i0 / p.dst_w;
+    uint dst_x = i0 - i1 * p.dst_w;
+    uint n = i1 / p.dst_h;
+    uint dst_y = i1 - n * p.dst_h;
+
+    uint src_i = n * p.channels * p.src_h * p.src_w;
+    uint src_row = p.src_w * p.channels;
+    uint knl_row = p.knl_w * p.channels;
+
+    FLOAT_TYPE sum = 0.0;
+    for (uint knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+        uint src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y;
+        if (src_y >= p.src_h) { // src_y < 0 will wrap to a large unsigned int
+            continue;
+        }
+        for (uint knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+            uint src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x;
+            if (src_x >= p.src_w) { // src_x < 0 will wrap to a large unsigned int
+                continue;
+            }
+            FLOAT_TYPE v = FLOAT_TYPE(src_data[src_i + src_y * src_row + src_x * p.channels + c]);
+            FLOAT_TYPE k = FLOAT_TYPE(knl_data[        knl_y * knl_row + knl_x * p.channels + c]);
+            sum = fma(v, k, sum);
+        }
+    }
+    return sum;
+}
+
+void main() {
+    uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+    if (idx >= p.ne) {
+        return;
+    }
+
+    FLOAT_TYPE result =
+#ifdef WHCN
+        conv_2d_dw_whcn(idx);
+#else
+        conv_2d_dw_cwhn(idx);
+#endif
+    dst_data[idx] = D_TYPE(result);
+}
+

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

@@ -12,7 +12,10 @@ void main() {
         return;
     }
 
-#ifndef OPTIMIZATION_ERROR_WORKAROUND
+#if defined(DATA_D_BF16)
+    float f = float(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(fp32_to_bf16(f));
+#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
     data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
 #else
     data_d[get_doffset() + dst_idx(idx)] = data_a[get_aoffset() + src0_idx(idx)];

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

@@ -23,6 +23,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
 }
 #endif
 
+#if defined(DATA_A_BF16)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    return vec2(bf16_to_fp32(data_a[a_offset + ib]), bf16_to_fp32(data_a[a_offset + ib + 1]));
+}
+#endif
+
 #if defined(DATA_A_Q4_0)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
@@ -428,7 +434,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
 }
 #endif
 
-#if defined(DATA_A_F32) || defined(DATA_A_F16)
+#if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(0, 0);
 }

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp

@@ -482,7 +482,7 @@ float16_t dequantFuncIQ2_XXS(const in decodeBufIQ2_XXS bl, const in uint blockCo
     const uint ib8 = (idx & 0x18) >> 3;  // 0..3
     const uint iqs = 8 * ib32 + ib8;
 
-    const uint8_t qs = bl.block.qs[iqs];
+    const uint qs = bl.block.qs[iqs];
     const uint signscale = pack32(u16vec2(bl16.block.qs[4*ib32+2], bl16.block.qs[4*ib32+3]));
 
     const float dscale = float(bl.block.d) * 0.25 * (0.5 + float(signscale >> 28));

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

@@ -20,9 +20,14 @@ void main() {
     const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
     const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
 
-#ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]);
+#if defined(DATA_A_BF16)
+    FLOAT_TYPE v = FLOAT_TYPE(bf16_to_fp32(data_a[a_offset + i00]));
 #else
-    data_d[d_offset + i00] = data_a[a_offset + i00];
+    FLOAT_TYPE v = FLOAT_TYPE(data_a[a_offset + i00]);
+#endif
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[d_offset + i00] = D_TYPE(v);
+#else
+    data_d[d_offset + i00] = D_TYPE(v);
 #endif
 }

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

@@ -6,7 +6,7 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
+#if !defined(DATA_A_F32) && !defined(DATA_A_F16) && !defined(DATA_A_BF16)
 #define K_PER_ITER 8
 #else
 #define K_PER_ITER 2

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

@@ -21,7 +21,9 @@ layout (push_constant) uniform parameter
     uint nrows_x;
     uint row_stride_x;
     uint channel_stride_x;
+    uint channel_stride_y;
     uint channel_x_divisor;
+    uint ne12;
     uint b_offset;
     uint d_offset;
 } p;
@@ -33,6 +35,7 @@ void main() {
     const uint row_x     = gl_GlobalInvocationID.y;
     const uint channel   = gl_GlobalInvocationID.z;
     const uint channel_x = channel / p.channel_x_divisor;
+    const uint channel_y = channel % p.ne12;
 
     const uint nrows_y   = p.ncols_x;
     const uint nrows_dst = p.nrows_x;
@@ -56,7 +59,7 @@ void main() {
                 const uint row_y = col_x;
 
                 const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-                const uint iy = channel*nrows_y + row_y;
+                const uint iy = channel_y*p.channel_stride_y + row_y;
 
                 const vec4 av4 = vec4(data_a_v4[ix / 4]);
                 const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -72,7 +75,7 @@ void main() {
             const uint row_y = col_x;
 
             const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel*nrows_y + row_y;
+            const uint iy = channel_y*p.channel_stride_y + row_y;
 
             const vec4 av4 = vec4(data_a_v4[ix / 4]);
             const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -89,7 +92,7 @@ void main() {
             const uint row_y = col_x;
 
             const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel*nrows_y + row_y;
+            const uint iy = channel_y*p.channel_stride_y + row_y;
 
             const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
 

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

@@ -10,6 +10,10 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
 #endif
 
+#if defined(DATA_A_BF16) && defined(COOPMAT)
+#extension GL_EXT_bfloat16 : enable
+#endif
+
 #ifdef COOPMAT
 #extension GL_KHR_cooperative_matrix : enable
 #extension GL_KHR_memory_scope_semantics : enable
@@ -29,6 +33,10 @@
 #define LOAD_VEC_B 1
 #endif
 
+#if !defined(TO_FLOAT_TYPE)
+#define TO_FLOAT_TYPE FLOAT_TYPE
+#endif
+
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -202,8 +210,8 @@ void main() {
 #endif
 
 #ifdef COOPMAT
-    coopmat<float16_t, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a;
-    coopmat<float16_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
+    coopmat<FLOAT_TYPE, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a;
+    coopmat<FLOAT_TYPE, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
     coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col];
 
     [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) {
@@ -248,6 +256,21 @@ void main() {
                 buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = FLOAT_TYPE(0.0f);
             }
 #endif
+#elif defined(DATA_A_BF16)
+#if LOAD_VEC_A == 4
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
+            buf_a[buf_idx    ] = TO_FLOAT_TYPE(data_a[idx].x);
+            buf_a[buf_idx + 1] = TO_FLOAT_TYPE(data_a[idx].y);
+            buf_a[buf_idx + 2] = TO_FLOAT_TYPE(data_a[idx].z);
+            buf_a[buf_idx + 3] = TO_FLOAT_TYPE(data_a[idx].w);
+#else
+            if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) {
+                buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = TO_FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]);
+            } else {
+                buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = TO_FLOAT_TYPE(uint16_t(0));
+            }
+#endif
 #elif defined(DATA_A_Q4_0)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
             const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 4 * loadr_a;
@@ -695,13 +718,13 @@ void main() {
             const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
 #endif
             const uint buf_idx = (loadc_b + l) * SHMEM_STRIDE + loadr_b * LOAD_VEC_B;
-            buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx].x);
-            buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx].y);
-            buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx].z);
-            buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx].w);
+            buf_b[buf_idx + 0] = TO_FLOAT_TYPE(data_b[idx].x);
+            buf_b[buf_idx + 1] = TO_FLOAT_TYPE(data_b[idx].y);
+            buf_b[buf_idx + 2] = TO_FLOAT_TYPE(data_b[idx].z);
+            buf_b[buf_idx + 3] = TO_FLOAT_TYPE(data_b[idx].w);
 #elif !MUL_MAT_ID
             if (ic * BN + loadc_b + l < p.N && block + loadr_b < end_k) {
-                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]);
+                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]);
             } else {
                 buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f);
             }
@@ -709,7 +732,7 @@ void main() {
             const uint row_i = ic * BN + loadc_b + l;
             if (row_i < _ne1) {
                 const u16vec2 row_idx = row_ids[row_i];
-                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
+                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
             } else {
                 buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f);
             }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -14,6 +14,9 @@
 #extension GL_EXT_buffer_reference : enable
 #extension GL_KHR_shader_subgroup_ballot : enable
 #extension GL_KHR_shader_subgroup_vote : enable
+#ifdef DATA_A_BF16
+#extension GL_EXT_bfloat16 : enable
+#endif
 
 #include "types.comp"
 
@@ -80,6 +83,12 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 #define store_scales(a)
 #endif
 
+#if defined(DATA_A_BF16)
+#define MAT_TYPE bfloat16_t
+#else
+#define MAT_TYPE FLOAT_TYPE
+#endif
+
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};
 
@@ -271,8 +280,8 @@ void main() {
 
                 // Manually partial unroll
                 [[unroll]] for (uint j = 0; j < unroll_count; ++j) {
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                     coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose);
@@ -286,8 +295,8 @@ void main() {
                 store_scales(tid);
             }
             while (block_k < end_k) {
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose);
@@ -310,8 +319,8 @@ void main() {
 
                 // Manually partial unroll
                 [[unroll]] for (uint j = 0; j < unroll_count; ++j) {
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                     coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose);
@@ -325,8 +334,8 @@ void main() {
                 store_scales(tid);
             }
             while (block_k < end_k) {
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose);
@@ -350,8 +359,8 @@ void main() {
 
                 // Manually partial unroll
                 [[unroll]] for (uint j = 0; j < unroll_count; ++j) {
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                     coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose);
@@ -365,8 +374,8 @@ void main() {
                 store_scales(tid);
             }
             while (block_k < end_k) {
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose);
@@ -405,8 +414,8 @@ void main() {
                 fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false);
             }
 
-            coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-            coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+            coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+            coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
 
             coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
 #ifdef MUL_MAT_ID

ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp

@@ -0,0 +1,7 @@
+#version 460
+
+#extension GL_EXT_bfloat16 : require
+
+void main()
+{
+}

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

@@ -33,6 +33,19 @@
 #endif
 #endif
 
+#if defined(DATA_A_BF16)
+#define QUANT_K 1
+#define QUANT_R 1
+
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
+#define A_TYPE uint16_t
+#elif LOAD_VEC_A == 4
+#define A_TYPE u16vec4
+#elif LOAD_VEC_A == 8
+#error unsupported
+#endif
+#endif
+
 #define QUANT_K_Q4_0 32
 #define QUANT_R_Q4_0 2
 
@@ -1343,4 +1356,18 @@ void init_iq_shmem(uvec3 wgsize)
 }
 #endif
 
+// returns the bfloat value in the low 16b.
+// See ggml_compute_fp32_to_bf16
+uint32_t fp32_to_bf16(float f)
+{
+    uint32_t u = floatBitsToUint(f);
+    u = (u + (0x7fff + ((u >> 16) & 1))) >> 16;
+    return u;
+}
+
+float bf16_to_fp32(uint32_t u)
+{
+    return uintBitsToFloat(u << 16);
+}
+
 #endif // !defined(GGML_TYPES_COMP)

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -63,7 +63,8 @@ const std::vector<std::string> type_names = {
     "iq3_xxs",
     "iq3_s",
     "iq4_xs",
-    "iq4_nl"
+    "iq4_nl",
+    "bf16",
 };
 
 namespace {
@@ -296,7 +297,6 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
     std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4";
 
     std::map<std::string, std::string> base_dict = {
-        {"FLOAT_TYPE", (coopmat2 || fp16) ? "float16_t" : "float"},
         {"FLOAT_TYPE_VEC2", (coopmat2 || fp16) ? "f16vec2" : "vec2"},
     };
     std::string shader_name = "matmul";
@@ -318,12 +318,45 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
 
     const std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp";
 
-    // Shaders with f16 B_TYPE
-    string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
-    string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    auto const &FLOAT_TYPE = [&](const std::string &t) -> std::string {
+        if (t == "bf16") {
+            // scalar path promotes to float
+            if (!coopmat && !coopmat2) {
+                return "float";
+            }
+            return "bfloat16_t";
+        }
+        if (coopmat2 || fp16) {
+            return "float16_t";
+        }
+        return "float";
+    };
 
-    string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
-    string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+    // Shaders with f16 B_TYPE
+    string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+
+    string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+
+    // bf16
+    {
+        std::string load_vec_a_unaligned = "1";
+        // For aligned matmul loads
+        std::string load_vec_a = coopmat2 ? "1" : "4";
+
+        // scalar path promotes to float
+        std::string to_float_type = (coopmat || coopmat2) ? "uintBitsToBFloat16EXT" : "bf16_to_fp32";
+
+        // If bfloat16 is not supported, then only compile the scalar (promote to fp32) shader
+#if !defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (!(coopmat || coopmat2))
+#endif
+        {
+            string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"},   {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_bf16",         source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                      {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"},                          {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}),                   fp16, coopmat, coopmat2, f16acc);
+        }
+    }
 
     for (const auto& tname : type_names) {
         std::string load_vec_quant = "2";
@@ -332,26 +365,30 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
         else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq4_nl"))
             load_vec_quant = "4";
 
+        if (tname == "bf16") {
+            continue;
+        }
+
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
         // For unaligned, load one at a time for f32/f16, or two at a time for quants
-        std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16") ? "1" : load_vec_quant;
+        std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? "1" : load_vec_quant;
         // For aligned matmul loads
-        std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16") ? load_vec : load_vec_quant;
+        std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? load_vec : load_vec_quant;
 
         // don't generate f32 variants for coopmat2
         if (!coopmat2) {
-            string_to_spv(shader_name + "_" + tname + "_f32",         source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float"},            {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f32",         source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float"},            {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
         }
 
         if (tname != "f16" && tname != "f32") {
-            string_to_spv(shader_name + "_" + tname + "_f16",         source_name,  merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float16_t"},        {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name,  merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f16",         source_name,  merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float16_t"},        {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name,  merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
         }
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
         if (!coopmat && !coopmat2 && !matmul_id && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) {
-            string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
         }
 #endif
     }
@@ -393,6 +430,7 @@ void process_shaders() {
             if (tname == "f32") {
                 continue;
             }
+            if (tname == "bf16") continue;
 
             if (tname == "f16") {
                 string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
@@ -417,12 +455,12 @@ void process_shaders() {
         string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
 
         // Dequant shaders
-        if (tname != "f16") {
+        if (tname != "f16" && tname != "bf16") {
             string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}}));
         }
 
         if (!string_ends_with(tname, "_k")) {
-            shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp";
+            shader = (tname == "f32" || tname == "f16" || tname == "bf16") ? "get_rows.comp" : "get_rows_quant.comp";
 
             if (tname == "f16") {
                 string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}));
@@ -447,9 +485,11 @@ void process_shaders() {
     string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
     string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+    string_to_spv("cpy_f32_bf16","copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
     string_to_spv("contig_cpy_f32_f32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("contig_cpy_f32_f16", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
     string_to_spv("contig_cpy_f16_f16", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+    string_to_spv("contig_cpy_f32_bf16","contig_copy.comp",{{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
 
     for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
         string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
@@ -544,6 +584,9 @@ void process_shaders() {
 
     string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
+    string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
+    string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));
+
     for (auto &c : compiles) {
         c.wait();
     }

gguf-py/gguf/constants.py

@@ -104,6 +104,7 @@ class Keys:
         EXPERT_WEIGHTS_SCALE              = "{arch}.expert_weights_scale"
         EXPERT_WEIGHTS_NORM               = "{arch}.expert_weights_norm"
         EXPERT_GATING_FUNC                = "{arch}.expert_gating_func"
+        MOE_EVERY_N_LAYERS                = "{arch}.moe_every_n_layers"
         POOLING_TYPE                      = "{arch}.pooling_type"
         LOGIT_SCALE                       = "{arch}.logit_scale"
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
@@ -230,8 +231,10 @@ class Keys:
         BLOCK_COUNT         = "clip.vision.block_count"
         IMAGE_MEAN          = "clip.vision.image_mean"
         IMAGE_STD           = "clip.vision.image_std"
+        SPATIAL_MERGE_SIZE  = "clip.vision.spatial_merge_size"
         USE_GELU            = "clip.use_gelu"
         USE_SILU            = "clip.use_silu"
+        N_WA_PATTERN        = "clip.vision.n_wa_pattern" # used by qwen2.5vl
 
         class Attention:
             HEAD_COUNT      = "clip.vision.attention.head_count"
@@ -267,6 +270,7 @@ class MODEL_ARCH(IntEnum):
     REFACT           = auto()
     BERT             = auto()
     NOMIC_BERT       = auto()
+    NOMIC_BERT_MOE   = auto()
     JINA_BERT_V2     = auto()
     BLOOM            = auto()
     STABLELM         = auto()
@@ -489,6 +493,7 @@ class MODEL_TENSOR(IntEnum):
     V_ENC_FFN_DOWN       = auto()
     V_PRE_NORM           = auto()
     V_POST_NORM          = auto()
+    V_MM_INP_NORM        = auto()
     V_MM_INP_PROJ        = auto() # gemma3
     V_MM_SOFT_EMB_NORM   = auto() # gemma3
     V_RESMPL_POS_EMBD_K  = auto() # minicpmv
@@ -503,6 +508,7 @@ class MODEL_TENSOR(IntEnum):
     V_RESMPL_PROJ        = auto() # minicpmv
     V_RESMPL_QUERY       = auto() # minicpmv
     V_TOK_EMBD_IMG_BREAK = auto() # pixtral
+    V_MM_PATCH_MERGER    = auto() # mistral small 3.1
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -521,6 +527,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.REFACT:           "refact",
     MODEL_ARCH.BERT:             "bert",
     MODEL_ARCH.NOMIC_BERT:       "nomic-bert",
+    MODEL_ARCH.NOMIC_BERT_MOE:   "nomic-bert-moe",
     MODEL_ARCH.JINA_BERT_V2:     "jina-bert-v2",
     MODEL_ARCH.BLOOM:            "bloom",
     MODEL_ARCH.STABLELM:         "stablelm",
@@ -744,6 +751,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_PRE_NORM:                "v.pre_ln",
     MODEL_TENSOR.V_POST_NORM:               "v.post_ln",
     MODEL_TENSOR.V_MM_INP_PROJ:             "mm.input_projection",
+    MODEL_TENSOR.V_MM_INP_NORM:             "mm.input_norm",
     MODEL_TENSOR.V_MM_SOFT_EMB_NORM:        "mm.soft_emb_norm",
     MODEL_TENSOR.V_RESMPL_POS_EMBD_K:       "resampler.pos_embd_k",
     MODEL_TENSOR.V_RESMPL_ATTN_Q:           "resampler.attn.q",
@@ -757,6 +765,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_RESMPL_PROJ:             "resampler.proj",
     MODEL_TENSOR.V_RESMPL_QUERY:            "resampler.query",
     MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK:      "v.token_embd.img_break", # pixtral
+    MODEL_TENSOR.V_MM_PATCH_MERGER:         "mm.patch_merger", # mistral small 3.1
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -780,6 +789,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_PRE_NORM,
         MODEL_TENSOR.V_POST_NORM,
         MODEL_TENSOR.V_MM_INP_PROJ,
+        MODEL_TENSOR.V_MM_INP_NORM,
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM,
         MODEL_TENSOR.V_RESMPL_POS_EMBD_K,
         MODEL_TENSOR.V_RESMPL_ATTN_Q,
@@ -793,6 +803,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_RESMPL_PROJ,
         MODEL_TENSOR.V_RESMPL_QUERY,
         MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK,
+        MODEL_TENSOR.V_MM_PATCH_MERGER,
     ],
     MODEL_ARCH.LLAMA: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -960,6 +971,22 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP,
         MODEL_TENSOR.LAYER_OUT_NORM,
     ],
+    MODEL_ARCH.NOMIC_BERT_MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.TOKEN_TYPES,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.LAYER_OUT_NORM,
+    ],
     MODEL_ARCH.JINA_BERT_V2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.TOKEN_EMBD_NORM,
@@ -2006,6 +2033,8 @@ class PoolingType(IntEnum):
     NONE = 0
     MEAN = 1
     CLS  = 2
+    LAST = 3
+    RANK = 4
 
 
 class GGMLQuantizationType(IntEnum):
@@ -2136,6 +2165,8 @@ class VisionProjectorType:
     GEMMA3 = "gemma3"
     IDEFICS3 = "idefics3"
     PIXTRAL = "pixtral"
+    QWEN2VL = "qwen2vl_merger"
+    QWEN25VL = "qwen2.5vl_merger"
 
 
 # Items here are (block size, type size)

gguf-py/gguf/gguf_writer.py

@@ -728,6 +728,9 @@ class GGUFWriter:
     def add_expert_gating_func(self, value: ExpertGatingFuncType) -> None:
         self.add_uint32(Keys.LLM.EXPERT_GATING_FUNC.format(arch=self.arch), value.value)
 
+    def add_moe_every_n_layers(self, value: int) -> None:
+        self.add_uint32(Keys.LLM.MOE_EVERY_N_LAYERS.format(arch=self.arch), value)
+
     def add_swin_norm(self, value: bool) -> None:
         self.add_bool(Keys.LLM.SWIN_NORM.format(arch=self.arch), value)
 
@@ -969,6 +972,9 @@ class GGUFWriter:
     def add_vision_image_std(self, values: Sequence[float]) -> None:
         self.add_array(Keys.ClipVision.IMAGE_STD, values)
 
+    def add_vision_spatial_merge_size(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.SPATIAL_MERGE_SIZE, value)
+
     def add_vision_use_gelu(self, value: bool) -> None:
         self.add_bool(Keys.ClipVision.USE_GELU, value)
 
@@ -978,6 +984,9 @@ class GGUFWriter:
     def add_vision_projector_scale_factor(self, value: int) -> None:
         self.add_uint32(Keys.ClipVision.Projector.SCALE_FACTOR, value)
 
+    def add_vision_n_wa_pattern(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.N_WA_PATTERN, value)
+
     def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
         pack_prefix = ''
         if not skip_pack_prefix:

gguf-py/gguf/tensor_mapping.py

@@ -290,6 +290,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.ffn.router.layer",        # dbrx
             "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
             "language_model.model.layers.{bid}.feed_forward.router", # llama4
+            "encoder.layers.{bid}.mlp.router.layer",            # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
@@ -322,6 +323,7 @@ class TensorNameMap:
             "model.layers.layers.{bid}.mlp.up_proj",                  # plamo
             "model.layers.{bid}.feed_forward.w3",                     # internlm2
             "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
+            "encoder.layers.{bid}.mlp.fc1",                           # nomic-bert-moe
             "model.layers.{bid}.mlp.c_fc",                            # starcoder2
             "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2
             "model.layers.{bid}.residual_mlp.w3",                     # arctic
@@ -337,6 +339,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.experts.up_proj",         # qwen2moe olmoe (merged)
             "model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged)
             "language_model.model.layers.{bid}.feed_forward.experts.up_proj", # llama4
+            "encoder.layers.{bid}.mlp.experts.mlp.w1",        # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_UP_SHEXP: (
@@ -418,6 +421,7 @@ class TensorNameMap:
             "model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe
             "model.layers.{bid}.block_sparse_moe.experts.w2",    # phimoe (merged)
             "language_model.model.layers.{bid}.feed_forward.experts.down_proj", # llama4
+            "encoder.layers.{bid}.mlp.experts.mlp.w2",           # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_DOWN_SHEXP: (
@@ -892,6 +896,7 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_MMPROJ: (
             "multi_modal_projector.linear_{bid}",
+            "visual.merger.mlp.{bid}", # qwen2vl
         ),
 
         MODEL_TENSOR.V_MMPROJ_FC: (
@@ -915,6 +920,7 @@ class TensorNameMap:
             "vpm.embeddings.patch_embedding",
             "model.vision_model.embeddings.patch_embedding", # SmolVLM
             "vision_tower.patch_conv", # pixtral
+            "visual.patch_embed.proj", # qwen2vl
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_POS: (
@@ -928,6 +934,7 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.self_attn.q_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.q_proj", # SmolVLM
             "vision_tower.transformer.layers.{bid}.attention.q_proj", # pixtral
+            "visual.blocks.{bid}.attn.q", # qwen2vl, generated
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_K: (
@@ -935,6 +942,7 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.self_attn.k_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.k_proj", # SmolVLM
             "vision_tower.transformer.layers.{bid}.attention.k_proj", # pixtral
+            "visual.blocks.{bid}.attn.k", # qwen2vl, generated
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_V: (
@@ -942,6 +950,7 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.self_attn.v_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.v_proj", # SmolVLM
             "vision_tower.transformer.layers.{bid}.attention.v_proj", # pixtral
+            "visual.blocks.{bid}.attn.v", # qwen2vl, generated
         ),
 
         MODEL_TENSOR.V_ENC_INPUT_NORM: (
@@ -949,6 +958,7 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.layer_norm1",
             "model.vision_model.encoder.layers.{bid}.layer_norm1", # SmolVLM
             "vision_tower.transformer.layers.{bid}.attention_norm", # pixtral
+            "visual.blocks.{bid}.norm1", # qwen2vl
         ),
 
         MODEL_TENSOR.V_ENC_OUTPUT: (
@@ -956,6 +966,7 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.self_attn.out_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.out_proj", # SmolVLM
             "vision_tower.transformer.layers.{bid}.attention.o_proj", # pixtral
+            "visual.blocks.{bid}.attn.proj", # qwen2vl
         ),
 
         MODEL_TENSOR.V_ENC_OUTPUT_NORM: (
@@ -963,17 +974,24 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.layer_norm2",
             "model.vision_model.encoder.layers.{bid}.layer_norm2", # SmolVLM
             "vision_tower.transformer.layers.{bid}.ffn_norm", # pixtral
+            "visual.blocks.{bid}.norm2", # qwen2vl
         ),
 
+        # some namings are messed up because the original llava code swapped fc1 and fc2
+        # we have no better way to fix it, just be careful
+        # new models like pixtral use the correct naming
         MODEL_TENSOR.V_ENC_FFN_UP: (
             "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc1",
             "vpm.encoder.layers.{bid}.mlp.fc1",
             "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3 (note: name is swapped)
             "vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral
+            "visual.blocks.{bid}.mlp.fc2", # qwen2vl
+            "visual.blocks.{bid}.mlp.up_proj", # qwen2.5vl
         ),
 
         MODEL_TENSOR.V_ENC_FFN_GATE: (
             "vision_tower.transformer.layers.{bid}.feed_forward.gate_proj", # pixtral
+            "visual.blocks.{bid}.mlp.gate_proj", # qwen2.5vl
         ),
 
         MODEL_TENSOR.V_ENC_FFN_DOWN: (
@@ -981,6 +999,8 @@ class TensorNameMap:
             "vpm.encoder.layers.{bid}.mlp.fc2",
             "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3 (note: name is swapped)
             "vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral
+            "visual.blocks.{bid}.mlp.fc1", # qwen2vl
+            "visual.blocks.{bid}.mlp.down_proj", # qwen2.5vl
         ),
 
         MODEL_TENSOR.V_PRE_NORM: (
@@ -991,12 +1011,17 @@ class TensorNameMap:
         MODEL_TENSOR.V_POST_NORM: (
             "vision_tower.vision_model.post_layernorm",
             "model.vision_model.post_layernorm", # SmolVLM
+            "visual.merger.ln_q", # qwen2vl
         ),
 
         MODEL_TENSOR.V_MM_INP_PROJ: (
             "multi_modal_projector.mm_input_projection",
         ),
 
+        MODEL_TENSOR.V_MM_INP_NORM: (
+            "multi_modal_projector.norm",
+        ),
+
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM: (
             "multi_modal_projector.mm_soft_emb_norm",
         ),
@@ -1048,6 +1073,10 @@ class TensorNameMap:
         MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK: (
             "v.token_embd.img_break", # for pixtral, this is a generated vector
         ),
+
+        MODEL_TENSOR.V_MM_PATCH_MERGER: (
+            "multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1
+        ),
     }
 
     # architecture-specific block mappings

grammars/README.md

@@ -1,6 +1,6 @@
 # GBNF Guide
 
-GBNF (GGML BNF) is a format for defining [formal grammars](https://en.wikipedia.org/wiki/Formal_grammar) to constrain model outputs in `llama.cpp`. For example, you can use it to force the model to generate valid JSON, or speak only in emojis. GBNF grammars are supported in various ways in `examples/main` and `examples/server`.
+GBNF (GGML BNF) is a format for defining [formal grammars](https://en.wikipedia.org/wiki/Formal_grammar) to constrain model outputs in `llama.cpp`. For example, you can use it to force the model to generate valid JSON, or speak only in emojis. GBNF grammars are supported in various ways in `tools/main` and `tools/server`.
 
 ## Background
 
@@ -110,21 +110,21 @@ While semantically correct, the syntax `x? x? x?.... x?` (with N repetitions) ma
 
 You can use GBNF grammars:
 
-- In [llama-server](../examples/server)'s completion endpoints, passed as the `grammar` body field
-- In [llama-cli](../examples/main), passed as the `--grammar` & `--grammar-file` flags
+- In [llama-server](../tools/server)'s completion endpoints, passed as the `grammar` body field
+- In [llama-cli](../tools/main), passed as the `--grammar` & `--grammar-file` flags
 - With [test-gbnf-validator](../tests/test-gbnf-validator.cpp), to test them against strings.
 
 ## JSON Schemas → GBNF
 
 `llama.cpp` supports converting a subset of https://json-schema.org/ to GBNF grammars:
 
-- In [llama-server](../examples/server):
+- In [llama-server](../tools/server):
     - For any completion endpoints, passed as the `json_schema` body field
     - For the `/chat/completions` endpoint, passed inside the `response_format` body field (e.g. `{"type", "json_object", "schema": {"items": {}}}` or `{ type: "json_schema", json_schema: {"schema": ...} }`)
-- In [llama-cli](../examples/main), passed as the `--json` / `-j` flag
+- In [llama-cli](../tools/main), passed as the `--json` / `-j` flag
 - To convert to a grammar ahead of time:
     - in CLI, with [examples/json_schema_to_grammar.py](../examples/json_schema_to_grammar.py)
-    - in JavaScript with [json-schema-to-grammar.mjs](../examples/server/public_legacy/json-schema-to-grammar.mjs) (this is used by the [server](../examples/server)'s Web UI)
+    - in JavaScript with [json-schema-to-grammar.mjs](../tools/server/public_legacy/json-schema-to-grammar.mjs) (this is used by the [server](../tools/server)'s Web UI)
 
 Take a look at [tests](../tests/test-json-schema-to-grammar.cpp) to see which features are likely supported (you'll also find usage examples in https://github.com/ggml-org/llama.cpp/pull/5978, https://github.com/ggml-org/llama.cpp/pull/6659 & https://github.com/ggml-org/llama.cpp/pull/6555).
 

include/llama.h

@@ -1232,6 +1232,7 @@ extern "C" {
         "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)");
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    /// Setting k <= 0 makes this a noop
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751

pyrightconfig.json

@@ -15,7 +15,7 @@
     },
     {
       // uses match expressions in steps.py
-      "root": "examples/server/tests",
+      "root": "tools/server/tests",
       "pythonVersion": "3.10",
     },
   ],

requirements/requirements-all.txt

@@ -1,6 +1,6 @@
--r ../examples/llava/requirements.txt
--r ../examples/server/bench/requirements.txt
--r ../examples/server/tests/requirements.txt
+-r ../tools/llava/requirements.txt
+-r ../tools/server/bench/requirements.txt
+-r ../tools/server/tests/requirements.txt
 
 -r ./requirements-compare-llama-bench.txt
 -r ./requirements-pydantic.txt

scripts/compare-llama-bench.py

@@ -19,9 +19,9 @@ logger = logging.getLogger("compare-llama-bench")
 
 # Properties by which to differentiate results per commit:
 KEY_PROPERTIES = [
-    "cpu_info", "gpu_info", "backends", "n_gpu_layers", "model_filename", "model_type", "n_batch", "n_ubatch",
-    "embeddings", "cpu_mask", "cpu_strict", "poll", "n_threads", "type_k", "type_v", "use_mmap", "no_kv_offload",
-    "split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen"
+    "cpu_info", "gpu_info", "backends", "n_gpu_layers", "tensor_buft_overrides", "model_filename", "model_type",
+    "n_batch", "n_ubatch", "embeddings", "cpu_mask", "cpu_strict", "poll", "n_threads", "type_k", "type_v",
+    "use_mmap", "no_kv_offload", "split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen", "n_depth"
 ]
 
 # Properties that are boolean and are converted to Yes/No for the table:
@@ -30,11 +30,11 @@ BOOL_PROPERTIES = ["embeddings", "cpu_strict", "use_mmap", "no_kv_offload", "fla
 # Header names for the table:
 PRETTY_NAMES = {
     "cpu_info": "CPU", "gpu_info": "GPU", "backends": "Backends", "n_gpu_layers": "GPU layers",
-    "model_filename": "File", "model_type": "Model", "model_size": "Model size [GiB]",
-    "model_n_params": "Num. of par.", "n_batch": "Batch size", "n_ubatch": "Microbatch size",
-    "embeddings": "Embeddings", "cpu_mask": "CPU mask", "cpu_strict": "CPU strict", "poll": "Poll",
-    "n_threads": "Threads", "type_k": "K type", "type_v": "V type", "split_mode": "Split mode", "main_gpu": "Main GPU",
-    "no_kv_offload": "NKVO", "flash_attn": "FlashAttention", "tensor_split": "Tensor split", "use_mmap": "Use mmap",
+    "tensor_buft_overrides": "Tensor overrides", "model_filename": "File", "model_type": "Model", "model_size": "Model size [GiB]",
+    "model_n_params": "Num. of par.", "n_batch": "Batch size", "n_ubatch": "Microbatch size", "embeddings": "Embeddings",
+    "cpu_mask": "CPU mask", "cpu_strict": "CPU strict", "poll": "Poll", "n_threads": "Threads", "type_k": "K type", "type_v": "V type",
+    "use_mmap": "Use mmap", "no_kv_offload": "NKVO", "split_mode": "Split mode", "main_gpu": "Main GPU", "tensor_split": "Tensor split",
+    "flash_attn": "FlashAttention",
 }
 
 DEFAULT_SHOW = ["model_type"]  # Always show these properties by default.
@@ -281,12 +281,12 @@ def get_rows(properties):
     The returned rows are unique in terms of property combinations.
     """
     select_string = ", ".join(
-        [f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"])
+        [f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "tb.n_depth", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"])
     equal_string = " AND ".join(
         [f"tb.{p} = tc.{p}" for p in KEY_PROPERTIES] + [
             f"tb.build_commit = '{hexsha8_baseline}'", f"tc.build_commit = '{hexsha8_compare}'"]
     )
-    group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt"])
+    group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt", "tb.n_depth"])
     query = (f"SELECT {select_string} FROM test tb JOIN test tc ON {equal_string} "
              f"GROUP BY {group_order_string} ORDER BY {group_order_string};")
     return cursor.execute(query).fetchall()
@@ -309,7 +309,7 @@ else:
     rows_full = get_rows(KEY_PROPERTIES)
     properties_different = []
     for i, kp_i in enumerate(KEY_PROPERTIES):
-        if kp_i in DEFAULT_SHOW or kp_i == "n_prompt" or kp_i == "n_gen":
+        if kp_i in DEFAULT_SHOW or kp_i in ["n_prompt", "n_gen", "n_depth"]:
             continue
         for row_full in rows_full:
             if row_full[i] != rows_full[0][i]:
@@ -340,17 +340,20 @@ else:
 
 table = []
 for row in rows_show:
-    n_prompt = int(row[-4])
-    n_gen    = int(row[-3])
+    n_prompt = int(row[-5])
+    n_gen    = int(row[-4])
+    n_depth  = int(row[-3])
     if n_prompt != 0 and n_gen == 0:
         test_name = f"pp{n_prompt}"
     elif n_prompt == 0 and n_gen != 0:
         test_name = f"tg{n_gen}"
     else:
         test_name = f"pp{n_prompt}+tg{n_gen}"
+    if n_depth != 0:
+        test_name = f"{test_name}@d{n_depth}"
     #           Regular columns    test name    avg t/s values              Speedup
     #            VVVVVVVVVVVVV     VVVVVVVVV    VVVVVVVVVVVVVV              VVVVVVV
-    table.append(list(row[:-4]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])
+    table.append(list(row[:-5]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])
 
 # Some a-posteriori fixes to make the table contents prettier:
 for bool_property in BOOL_PROPERTIES:
@@ -376,7 +379,7 @@ if "gpu_info" in show:
         for gns in GPU_NAME_STRIP:
             row_table[ip] = row_table[ip].replace(gns, "")
 
-        gpu_names = row_table[ip].split("/")
+        gpu_names = row_table[ip].split(", ")
         num_gpus = len(gpu_names)
         all_names_the_same = len(set(gpu_names)) == 1
         if len(gpu_names) >= 2 and all_names_the_same:

scripts/fetch_server_test_models.py

@@ -8,7 +8,7 @@
 
     Example:
         python scripts/fetch_server_test_models.py
-        ( cd examples/server/tests && ./tests.sh -v -x -m slow )
+        ( cd tools/server/tests && ./tests.sh -v -x -m slow )
 '''
 import ast
 import glob
@@ -66,7 +66,7 @@ if __name__ == '__main__':
 
     models = sorted(list(set([
         model
-        for test_file in glob.glob('examples/server/tests/unit/test_*.py')
+        for test_file in glob.glob('tools/server/tests/unit/test_*.py')
         for model in collect_hf_model_test_parameters(test_file)
     ])), key=lambda m: (m.hf_repo, m.hf_file))
 

scripts/sync-ggml.last

@@ -1 +1 @@
-13bcf9ce50651a8b4238ec6d136f46f2c1b23b6f
+0482de9c63b9134eb462c7732888c0ee0dbc2755

scripts/tool_bench.py

@@ -2,7 +2,7 @@
 '''
     Simplistic tool call benchmarks for llama-server and ollama.
 
-    Essentially runs the tests at server/examples/server/tests/unit/test_tool_call.py N times, at different temperatures and on different backends (current llama-server, baseline llama-server and ollama),
+    Essentially runs the tests at server/tools/server/tests/unit/test_tool_call.py N times, at different temperatures and on different backends (current llama-server, baseline llama-server and ollama),
     and plots the results of multiple runs (from same .jsonl file or multiple ones) as a success rate heatmap.
 
     Simple usage example:
@@ -51,8 +51,8 @@ import typer
 
 sys.path.insert(0, Path(__file__).parent.parent.as_posix())
 if True:
-    from examples.server.tests.utils import ServerProcess
-    from examples.server.tests.unit.test_tool_call import TIMEOUT_SERVER_START, do_test_calc_result, do_test_hello_world, do_test_weather
+    from tools.server.tests.utils import ServerProcess
+    from tools.server.tests.unit.test_tool_call import TIMEOUT_SERVER_START, do_test_calc_result, do_test_hello_world, do_test_weather
 
 
 @contextmanager

scripts/xxd.cmake

@@ -1,5 +1,5 @@
 # CMake equivalent of `xxd -i ${INPUT} ${OUTPUT}`
-# Usage: cmake -DINPUT=examples/server/public/index.html -DOUTPUT=examples/server/index.html.hpp -P scripts/xxd.cmake
+# Usage: cmake -DINPUT=tools/server/public/index.html -DOUTPUT=tools/server/index.html.hpp -P scripts/xxd.cmake
 
 SET(INPUT "" CACHE STRING "Input File")
 SET(OUTPUT "" CACHE STRING "Output File")

src/llama-arch.cpp

@@ -19,6 +19,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_REFACT,           "refact"           },
     { LLM_ARCH_BERT,             "bert"             },
     { LLM_ARCH_NOMIC_BERT,       "nomic-bert"       },
+    { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
     { LLM_ARCH_JINA_BERT_V2,     "jina-bert-v2"     },
     { LLM_ARCH_BLOOM,            "bloom"            },
     { LLM_ARCH_STABLELM,         "stablelm"         },
@@ -106,6 +107,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_WEIGHTS_SCALE,              "%s.expert_weights_scale"              },
     { LLM_KV_EXPERT_WEIGHTS_NORM,               "%s.expert_weights_norm"               },
     { LLM_KV_EXPERT_GATING_FUNC,                "%s.expert_gating_func"                },
+    { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
     { LLM_KV_LOGIT_SCALE,                       "%s.logit_scale"                       },
     { LLM_KV_DECODER_START_TOKEN_ID,            "%s.decoder_start_token_id"            },
@@ -472,6 +474,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_NOMIC_BERT_MOE,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
+            { LLM_TENSOR_TOKEN_TYPES,     "token_types" },
+            { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_JINA_BERT_V2,
         {

src/llama-arch.h

@@ -23,6 +23,7 @@ enum llm_arch {
     LLM_ARCH_REFACT,
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
+    LLM_ARCH_NOMIC_BERT_MOE,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
@@ -110,6 +111,7 @@ enum llm_kv {
     LLM_KV_EXPERT_WEIGHTS_SCALE,
     LLM_KV_EXPERT_WEIGHTS_NORM,
     LLM_KV_EXPERT_GATING_FUNC,
+    LLM_KV_MOE_EVERY_N_LAYERS,
     LLM_KV_POOLING_TYPE,
     LLM_KV_LOGIT_SCALE,
     LLM_KV_DECODER_START_TOKEN_ID,

src/llama-batch.cpp

@@ -189,7 +189,7 @@ llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) {
     return ubatch;
 }
 
-void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
+llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
     GGML_ASSERT(batch.n_tokens >= 0);
     this->batch = &batch;
     this->n_embd = n_embd;
@@ -203,6 +203,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
     for (size_t i = 0; i < n_tokens; ++i) {
         ids[i] = i;
     }
+
     if (simple_split) {
         seq.resize(1);
         llama_sbatch_seq & s = seq[0];
@@ -212,6 +213,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
         s.length = n_tokens;
         return;
     }
+
     std::sort(ids.begin(), ids.end(),
             [&batch](size_t a, size_t b) {
                 int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1;
@@ -239,6 +241,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
                 return n_seq_a > n_seq_b;
             }
     );
+
     // init seq
     llama_sbatch_seq * last_seq = nullptr;
 
@@ -262,6 +265,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim
         seq.push_back(new_seq);
         last_seq = &seq.back();
     }
+
     // keep shared prompts first at the end, then sort by length descending.
     std::sort(seq.begin(), seq.end(),
             [](llama_sbatch_seq & a, llama_sbatch_seq & b) {

src/llama-batch.h

@@ -70,7 +70,8 @@ struct llama_sbatch {
     // sequence-wise split
     llama_ubatch split_seq(size_t n_ubatch);
 
-    void from_batch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
+    llama_sbatch() = default;
+    llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
 };
 
 // temporary allocate memory for the input batch if needed

src/llama-chat.cpp

@@ -447,8 +447,16 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|assistant|>";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4 || tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4) {
         ss << "[gMASK]" << "<sop>";
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>" << "\n" << message->content;
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
         for (auto message : chat) {
             std::string role(message->role);
             ss << "<|" << role << "|>" << "\n" << message->content;

src/llama-context.cpp

@@ -6,11 +6,9 @@
 #include "llama-model.h"
 #include "llama-kv-cache.h"
 
-#include <cassert>
 #include <cstring>
 #include <stdexcept>
 #include <cinttypes>
-#include <cmath>
 
 //
 // llama_context
@@ -114,7 +112,7 @@ llama_context::llama_context(
     }
 
     if (n_ctx_per_seq > hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_pre_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
+        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
                 __func__, n_ctx_per_seq, hparams.n_ctx_train);
     }
 
@@ -177,44 +175,13 @@ llama_context::llama_context(
     }
 
     // init the memory module
-    // TODO: for now, always create a unified KV cache
     if (!hparams.vocab_only) {
-        kv_self.reset(static_cast<llama_kv_cache_unified *>(model.create_memory()));
+        llama_memory_params params_mem = {
+            /*.type_k =*/ params.type_k,
+            /*.type_v =*/ params.type_v,
+        };
 
-        LLAMA_LOG_DEBUG("%s: n_ctx = %u\n", __func__, cparams.n_ctx);
-
-        cparams.n_ctx = GGML_PAD(cparams.n_ctx, kv_self->get_padding(cparams));
-
-        LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
-
-        uint32_t kv_size = cparams.n_ctx;
-        ggml_type type_k = params.type_k;
-        ggml_type type_v = params.type_v;
-
-        if (llama_model_is_recurrent(&model)) {
-            // Mamba needs at least as many KV cells as there are sequences kept at any time
-            kv_size = std::max((uint32_t) 1, params.n_seq_max);
-            // it's probably best to keep as much precision as possible for the states
-            type_k = GGML_TYPE_F32; // required by ggml_ssm_conv for Mamba's conv_states
-            type_v = GGML_TYPE_F32; // required by ggml_ssm_scan for Mamba's ssm_states
-        }
-
-        GGML_ASSERT(hparams.n_embd_head_k % ggml_blck_size(type_k) == 0);
-        GGML_ASSERT(hparams.n_embd_head_v % ggml_blck_size(type_v) == 0);
-
-        if (!kv_self->init(model, cparams, type_k, type_v, kv_size, cparams.offload_kqv)) {
-            throw std::runtime_error("failed to initialize self-attention cache");
-        }
-
-        {
-            const size_t memory_size_k = kv_self->size_k_bytes();
-            const size_t memory_size_v = kv_self->size_v_bytes();
-
-            LLAMA_LOG_INFO("%s: KV self size  = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
-                    (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f),
-                    ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
-                    ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
-        }
+        memory.reset(model.create_memory(params_mem, cparams));
     }
 
     // init backends
@@ -305,7 +272,9 @@ llama_context::llama_context(
         int n_nodes_tg  = -1;
 
         // simulate full KV cache
-        kv_self->n = kv_self->size;
+        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+
+        kv_self->set_full();
 
         cross.v_embd.clear();
 
@@ -427,6 +396,18 @@ const llama_model & llama_context::get_model() const {
     return model;
 }
 
+const llama_cparams & llama_context::get_cparams() const {
+    return cparams;
+}
+
+ggml_backend_sched_t llama_context::get_sched() const {
+    return sched.get();
+}
+
+ggml_context * llama_context::get_ctx_compute() const {
+    return ctx_compute.get();
+}
+
 uint32_t llama_context::n_ctx() const {
     return cparams.n_ctx;
 }
@@ -456,337 +437,21 @@ uint32_t llama_context::n_threads_batch() const {
 }
 
 llama_kv_cache * llama_context::get_kv_self() {
-    return kv_self.get();
+    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+    return kv_self;
 }
 
 const llama_kv_cache * llama_context::get_kv_self() const {
-    return kv_self.get();
-}
-
-ggml_tensor * llama_context::build_rope_shift(
-        ggml_context * ctx0,
-        ggml_tensor * cur,
-        ggml_tensor * shift,
-        ggml_tensor * factors,
-              float   freq_base,
-              float   freq_scale) const {
-    const auto & n_ctx_orig = cparams.n_ctx_orig_yarn;
-
-    const auto & yarn_ext_factor  = cparams.yarn_ext_factor;
-    const auto & yarn_beta_fast   = cparams.yarn_beta_fast;
-    const auto & yarn_beta_slow   = cparams.yarn_beta_slow;
-
-    const auto & hparams = model.hparams;
-
-    const auto & n_rot     = hparams.n_rot;
-    const auto & rope_type = hparams.rope_type;
-
-    // See llm_build_deepseek2() for why attn_factor has to be scaled for YaRN RoPE to work correctly.
-    // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
-    const float yarn_attn_factor = model.arch == LLM_ARCH_DEEPSEEK2 ? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)) : cparams.yarn_attn_factor;
-
-    ggml_tensor * tmp;
-
-    if (ggml_is_quantized(cur->type)) {
-        // dequantize to f32 -> RoPE -> quantize back
-        tmp = ggml_cast(ctx0, cur, GGML_TYPE_F32);
-
-        tmp = ggml_rope_ext(ctx0, tmp,
-                shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow);
-
-        tmp = ggml_cpy(ctx0, tmp, cur);
-    } else {
-        // we rotate only the first n_rot dimensions
-        tmp = ggml_rope_ext_inplace(ctx0, cur,
-                shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow);
-    }
-
-    return tmp;
-}
-
-class llm_graph_input_k_shift : public llm_graph_input_i {
-public:
-    llm_graph_input_k_shift(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
-    virtual ~llm_graph_input_k_shift() = default;
-
-    void set_input(const llama_ubatch * ubatch) override;
-
-    ggml_tensor * k_shift; // I32 [kv_size]
-
-    const llama_kv_cache_unified * kv_self;
-};
-
-void llm_graph_input_k_shift::set_input(const llama_ubatch * ubatch) {
-    GGML_UNUSED(ubatch);
-
-    if (k_shift) {
-        assert(ggml_backend_buffer_is_host(k_shift->buffer));
-
-        int32_t * data = (int32_t *) k_shift->data;
-
-        for (uint32_t i = 0; i < kv_self->size; ++i) {
-            data[i] = kv_self->cells[i].delta;
-        }
-    }
-}
-
-llm_graph_result_ptr llama_context::build_kv_self_shift(
-        ggml_context * ctx0,
-        ggml_cgraph * gf) const {
-    auto res = std::make_unique<llm_graph_result>();
-
-    const auto & hparams = model.hparams;
-
-    const auto & n_layer = hparams.n_layer;
-
-    const auto & n_embd_head_k = hparams.n_embd_head_k;
-  //const auto & n_embd_head_v = hparams.n_embd_head_v;
-
-    //GGML_ASSERT(kv_self->size == n_ctx);
-
-    auto inp = std::make_unique<llm_graph_input_k_shift>(kv_self.get());
-
-    inp->k_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, cparams.n_ctx);
-    ggml_set_input(inp->k_shift);
-
-    for (uint32_t il = 0; il < n_layer; ++il) {
-        const int64_t n_head_kv    = hparams.n_head_kv(il);
-        const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
-
-        const bool is_swa = hparams.is_swa(il);
-
-        // note: the swa rope params could become part of the cparams in the future
-        //       if we decide to make them configurable, like the non-sliding ones
-        const float freq_base_l  = is_swa ? hparams.rope_freq_base_train_swa  : cparams.rope_freq_base;
-        const float freq_scale_l = is_swa ? hparams.rope_freq_scale_train_swa : cparams.rope_freq_scale;
-
-        ggml_tensor * rope_factors = kv_self->cbs.get_rope_factors(n_ctx_per_seq(), il);
-
-        ggml_tensor * k =
-            ggml_view_3d(ctx0, kv_self->k_l[il],
-                n_embd_head_k, n_head_kv, kv_self->size,
-                ggml_row_size(kv_self->k_l[il]->type, n_embd_head_k),
-                ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
-                0);
-
-        ggml_tensor * cur = build_rope_shift(ctx0, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l);
-
-        ggml_build_forward_expand(gf, cur);
-    }
-
-    res->add_input(std::move(inp));
-
-    return res;
-}
-
-llm_graph_result_ptr llama_context::build_kv_self_defrag(
-        ggml_context * ctx0,
-        ggml_cgraph * gf) const {
-    auto res = std::make_unique<llm_graph_result>();
-
-    const auto & hparams = model.hparams;
-
-    const auto & ids = kv_self->defrag_info.ids;
-
-#if 0
-    // CPU defrag
-    //
-    // TODO: optimizations are possible:
-    //       - multiple threads
-    //       - avoid copying to the host memory when already there
-    //
-    // likely not worth the effort, as we have ggml_graph based defrag
-    //
-
-    const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa();
-    const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa();
-
-    const uint32_t kv_size = size;
-
-    std::vector<uint8_t> buf_k;
-    std::vector<uint8_t> buf_v;
-
-    for (uint32_t il = 0; il < n_layer; ++il) {
-        const size_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa);
-        const size_t k_size     = ggml_row_size(k_l[il]->type, n_embd_k_gqa*kv_size);
-
-        const size_t v_size_el = ggml_type_size(v_l[il]->type);
-        const size_t v_size    = ggml_row_size (v_l[il]->type, n_embd_v_gqa*kv_size);
-
-        buf_k.resize(k_size);
-        buf_v.resize(v_size);
-
-        ggml_backend_tensor_get(k_l[il], buf_k.data(), 0, buf_k.size());
-        ggml_backend_tensor_get(v_l[il], buf_v.data(), 0, buf_v.size());
-
-        // batch move [i, i+nm) to [id, id+nm)
-        // note: cells can move only to a lower index
-        for (uint32_t i = 0; i < n_kv; ++i) {
-            const uint32_t id = ids[i];
-
-            if (i == id || id == n_kv) {
-                continue;
-            }
-
-            uint32_t nm = 1;
-
-            while (i + nm < n_kv && ids[i + nm] == id + nm) {
-                nm++;
-            }
-
-            // move keys
-            {
-                const int64_t os =  i*k_size_row;
-                const int64_t od = id*k_size_row;
-
-                memcpy(buf_k.data() + od, buf_k.data() + os, nm*k_size_row);
-            }
-
-            // move values (note: they are transposed)
-            {
-                const int64_t os =  i;
-                const int64_t od = id;
-
-                for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                    memcpy(buf_v.data() + (od + j*kv_size)*v_size_el, buf_v.data() + (os + j*kv_size)*v_size_el, nm*v_size_el);
-                }
-            }
-
-            i += nm - 1;
-        }
-
-        ggml_backend_tensor_set(k_l[il], buf_k.data(), 0, buf_k.size());
-        ggml_backend_tensor_set(v_l[il], buf_v.data(), 0, buf_v.size());
-    }
-#else
-    for (uint32_t i = 0; i < ids.size(); ++i) {
-        const uint32_t id = ids[i];
-
-        if (i == id || id == ids.size()) {
-            continue;
-        }
-
-        uint32_t nm = 1;
-
-        while (i + nm < ids.size() && ids[i + nm] == id + nm) {
-            nm++;
-        }
-
-        for (uint32_t il = 0; il < hparams.n_layer; ++il) { // NOLINT
-            const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
-            const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
-
-            ggml_tensor * view_k_src = ggml_view_2d(ctx0, kv_self->k_l[il],
-                    n_embd_k_gqa, nm,
-                    ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
-                    ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*i));
-
-            ggml_tensor * view_k_dst = ggml_view_2d(ctx0, kv_self->k_l[il],
-                    n_embd_k_gqa, nm,
-                    ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
-                    ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*id));
-
-            ggml_tensor * view_v_src;
-            ggml_tensor * view_v_dst;
-
-            if (cparams.flash_attn) {
-                // NOTE: the V cache is not transposed when using flash attention
-                view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il],
-                        n_embd_v_gqa, nm,
-                        ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa),
-                        ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*i));
-
-                view_v_dst = ggml_view_2d(ctx0, kv_self->v_l[il],
-                        n_embd_v_gqa, nm,
-                        ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa),
-                        ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*id));
-            } else {
-                view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il],
-                        nm, n_embd_v_gqa,
-                        ggml_row_size(kv_self->v_l[il]->type, kv_self->size),
-                        ggml_row_size(kv_self->v_l[il]->type, i));
-
-                view_v_dst = ggml_view_2d(ctx0, kv_self->v_l[il],
-                        nm, n_embd_v_gqa,
-                        ggml_row_size(kv_self->v_l[il]->type, kv_self->size),
-                        ggml_row_size(kv_self->v_l[il]->type, id));
-            }
-
-            ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_k_src, view_k_dst));
-            ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_v_src, view_v_dst));
-        }
-
-        i += nm - 1;
-    }
-
-    //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes);
-#endif
-
-    return res;
+    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+    return kv_self;
 }
 
 void llama_context::kv_self_update() {
-    auto & kv = kv_self;
-
     bool need_reserve = false;
 
-    if (kv->has_shift) {
-        if (!kv->get_can_shift()) {
-            GGML_ABORT("The current context does not support K-shift");
-        }
+    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
 
-        LLAMA_LOG_DEBUG("%s: applying K-shift\n", __func__);
-
-        // apply K-shift if needed
-        if (model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE) {
-            ggml_backend_sched_reset(sched.get());
-
-            auto * gf = graph_init();
-
-            auto res = build_kv_self_shift(ctx_compute.get(), gf);
-
-            ggml_backend_sched_alloc_graph(sched.get(), gf);
-
-            res->set_inputs(nullptr);
-
-            graph_compute(gf, false);
-
-            need_reserve = true;
-        }
-
-        {
-            kv->has_shift = false;
-
-            for (uint32_t i = 0; i < kv->size; ++i) {
-                kv->cells[i].delta = 0;
-            }
-        }
-    }
-
-    // defragment the KV cache if needed
-    if (kv->do_defrag) {
-        LLAMA_LOG_DEBUG("%s: defragmenting KV cache\n", __func__);
-
-        if (kv->defrag_prepare(graph_max_nodes())) {
-            ggml_backend_sched_reset(sched.get());
-
-            auto * gf = graph_init();
-
-            auto res = build_kv_self_defrag(ctx_compute.get(), gf);
-
-            ggml_backend_sched_alloc_graph(sched.get(), gf);
-
-            res->set_inputs(nullptr);
-
-            graph_compute(gf, false);
-
-            need_reserve = true;
-        }
-
-        kv->do_defrag = false;
-    }
+    need_reserve = kv_self->update(*this);
 
     // reserve a worst case graph if needed
     if (need_reserve) {
@@ -797,7 +462,7 @@ void llama_context::kv_self_update() {
         uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
         // simulate full KV cache
-        kv_self->n = kv_self->size;
+        kv_self->set_full();
 
         llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
         llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};
@@ -818,9 +483,6 @@ enum llama_pooling_type llama_context::pooling_type() const {
 }
 
 float * llama_context::get_logits() {
-    // reorder logits for backward compatibility
-    output_reorder();
-
     return logits;
 }
 
@@ -863,9 +525,6 @@ float * llama_context::get_logits_ith(int32_t i) {
 }
 
 float * llama_context::get_embeddings() {
-    // reorder embeddings for backward compatibility
-    output_reorder();
-
     return embd;
 }
 
@@ -1017,8 +676,8 @@ int llama_context::encode(llama_batch & inp_batch) {
     }
 
     // temporary allocate memory for the input batch if needed
-    // TODO: this is incorrect for multiple sequences because pos_max() is the maximum across all sequences
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->pos_max() + 1);
+    // note: during encode, we always pass the full sequence starting from pos = 0
+    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : 0);
 
     const llama_batch & batch = batch_allocr.batch;
     const int32_t n_tokens = batch.n_tokens;
@@ -1047,7 +706,7 @@ int llama_context::encode(llama_batch & inp_batch) {
 
     const int64_t n_embd = hparams.n_embd;
 
-    sbatch.from_batch(batch, n_embd, /* simple_split */ true, /* logits_all */ true);
+    llama_sbatch sbatch = llama_sbatch(batch, n_embd, /* simple_split */ true, /* logits_all */ true);
 
     const llama_ubatch ubatch = sbatch.split_simple(n_tokens);
 
@@ -1181,9 +840,11 @@ int llama_context::decode(llama_batch & inp_batch) {
         return -1;
     }
 
+    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+
     // temporary allocate memory for the input batch if needed
-    // TODO: this is incorrect for multiple sequences because pos_max() is the maximum across all sequences
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->pos_max() + 1);
+    // TODO: this is incorrect for multiple sequences because get_pos_max() is the maximum across all sequences
+    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->get_pos_max() + 1);
 
     const llama_batch & batch = batch_allocr.batch;
 
@@ -1195,7 +856,7 @@ int llama_context::decode(llama_batch & inp_batch) {
     const int64_t n_tokens_all = batch.n_tokens;
     const int64_t n_embd       = hparams.n_embd;
 
-    llama_kv_cache_guard kv_guard(kv_self.get());
+    llama_kv_cache_guard kv_guard(kv_self);
 
     GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
 
@@ -1236,11 +897,7 @@ int llama_context::decode(llama_batch & inp_batch) {
         n_outputs_all = 1;
     }
 
-    const bool logits_all = n_outputs_all == n_tokens_all;
-
-    sbatch.from_batch(batch, n_embd,
-            /* simple_split */ !kv_self->recurrent,
-            /* logits_all   */ logits_all);
+    llama_sbatch sbatch = kv_self->sbatch_init(batch, /* logits_all */ n_outputs_all == n_tokens_all);
 
     // reserve output buffer
     if (output_reserve(n_outputs_all) < n_outputs_all) {
@@ -1254,22 +911,7 @@ int llama_context::decode(llama_batch & inp_batch) {
     int64_t n_outputs_prev = 0;
 
     while (sbatch.n_tokens > 0) {
-        llama_ubatch ubatch = llama_ubatch();
-
-        const auto & n_ubatch = cparams.n_ubatch;
-
-        if (kv_self->recurrent) {
-            if (embd_pooled) {
-                // Pooled embeddings cannot be split across ubatches (yet)
-                ubatch = sbatch.split_seq(cparams.n_ubatch);
-            } else {
-                // recurrent model architectures are easier to implement
-                // with equal-length sequences
-                ubatch = sbatch.split_equal(cparams.n_ubatch);
-            }
-        } else {
-            ubatch = sbatch.split_simple(n_ubatch);
-        }
+        llama_ubatch ubatch = kv_self->ubatch_next(sbatch, cparams.n_ubatch, embd_pooled);
 
         // count the outputs in this u_batch
         {
@@ -1289,24 +931,12 @@ int llama_context::decode(llama_batch & inp_batch) {
         }
 
         // find KV slot
-        {
-            if (!kv_self->find_slot(ubatch)) {
-                LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
+        if (!kv_self->find_slot(ubatch)) {
+            LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
 
-                return 1;
-            }
-
-            if (!kv_self->recurrent) {
-                // a heuristic, to avoid attending the full cache if it is not yet utilized
-                // after enough generations, the benefit from this heuristic disappears
-                // if we start defragmenting the cache, the benefit from this will be more important
-                const uint32_t pad = kv_self->get_padding(cparams);
-                kv_self->n = std::min(kv_self->size, std::max(pad, GGML_PAD(kv_self->cell_max(), pad)));
-            }
+            return 1;
         }
 
-        //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self->n, kv_self->used, kv_self->head);
-
         ggml_backend_sched_reset(sched.get());
         ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data);
 
@@ -1420,43 +1050,68 @@ int llama_context::decode(llama_batch & inp_batch) {
     // finalize the batch processing
     kv_guard.commit();
 
+    // set to total number of outputs in the batch, for use in llama_get_logits_ith
+    n_outputs = n_outputs_all;
+
     // set output mappings
     {
         bool sorted_output = true;
 
-        GGML_ASSERT(sbatch.out_ids.size() == (size_t) n_outputs_all);
+        auto & out_ids = sbatch.out_ids;
+
+        GGML_ASSERT(out_ids.size() == (size_t) n_outputs_all);
 
         for (int64_t i = 0; i < n_outputs_all; ++i) {
-            int64_t out_id = sbatch.out_ids[i];
+            int64_t out_id = out_ids[i];
             output_ids[out_id] = i;
             if (out_id != i) {
                 sorted_output = false;
             }
         }
 
-        if (sorted_output) {
-            sbatch.out_ids.clear();
+        // make the outputs have the same order they had in the user-provided batch
+        // note: this is mostly relevant for recurrent models atm
+        if (!sorted_output) {
+            const uint32_t n_vocab = model.vocab.n_tokens();
+            const uint32_t n_embd  = model.hparams.n_embd;
+
+            GGML_ASSERT((size_t) n_outputs == out_ids.size());
+
+            // TODO: is there something more efficient which also minimizes swaps?
+            // selection sort, to minimize swaps (from https://en.wikipedia.org/wiki/Selection_sort)
+            for (int32_t i = 0; i < n_outputs - 1; ++i) {
+                int32_t j_min = i;
+                for (int32_t j = i + 1; j < n_outputs; ++j) {
+                    if (out_ids[j] < out_ids[j_min]) {
+                        j_min = j;
+                    }
+                }
+                if (j_min == i) { continue; }
+                std::swap(out_ids[i], out_ids[j_min]);
+                if (logits_size > 0) {
+                    for (uint32_t k = 0; k < n_vocab; k++) {
+                        std::swap(logits[i*n_vocab + k], logits[j_min*n_vocab + k]);
+                    }
+                }
+                if (embd_size > 0) {
+                    for (uint32_t k = 0; k < n_embd; k++) {
+                        std::swap(embd[i*n_embd + k], embd[j_min*n_embd + k]);
+                    }
+                }
+            }
+            std::fill(output_ids.begin(), output_ids.end(), -1);
+            for (int32_t i = 0; i < n_outputs; ++i) {
+                output_ids[out_ids[i]] = i;
+            }
         }
     }
 
-    // set to total number of outputs in the batch, for use in llama_get_logits_ith
-    n_outputs = n_outputs_all;
-
     // wait for the computation to finish (automatically done when obtaining the model output)
     //synchronize();
 
     // decide if we need to defrag the kv cache
-    if (cparams.causal_attn && cparams.defrag_thold > 0.0f) {
-        // - do not defrag small contexts (i.e. < 2048 tokens)
-        // - count the padding towards the number of used tokens
-        const float fragmentation = kv_self->n >= 2048 ? std::max(0.0f, 1.0f - float(kv_self->used + kv_self->get_padding(cparams))/float(kv_self->n)) : 0.0f;
-
-        // queue defragmentation for next llama_kv_cache_update
-        if (fragmentation > cparams.defrag_thold) {
-            LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation);
-
-            kv_self->defrag();
-        }
+    if (cparams.defrag_thold > 0.0f) {
+        kv_self->defrag_sched(cparams.defrag_thold);
     }
 
     // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
@@ -1542,44 +1197,6 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
     return n_outputs_max;
 }
 
-void llama_context::output_reorder() {
-    auto & out_ids = sbatch.out_ids;
-    if (!out_ids.empty()) {
-        const uint32_t n_vocab = model.vocab.n_tokens();
-        const uint32_t n_embd  = model.hparams.n_embd;
-
-        GGML_ASSERT((size_t) n_outputs == out_ids.size());
-
-        // TODO: is there something more efficient which also minimizes swaps?
-        // selection sort, to minimize swaps (from https://en.wikipedia.org/wiki/Selection_sort)
-        for (int32_t i = 0; i < n_outputs - 1; ++i) {
-            int32_t j_min = i;
-            for (int32_t j = i + 1; j < n_outputs; ++j) {
-                if (out_ids[j] < out_ids[j_min]) {
-                    j_min = j;
-                }
-            }
-            if (j_min == i) { continue; }
-            std::swap(out_ids[i], out_ids[j_min]);
-            if (logits_size > 0) {
-                for (uint32_t k = 0; k < n_vocab; k++) {
-                    std::swap(logits[i*n_vocab + k], logits[j_min*n_vocab + k]);
-                }
-            }
-            if (embd_size > 0) {
-                for (uint32_t k = 0; k < n_embd; k++) {
-                    std::swap(embd[i*n_embd + k], embd[j_min*n_embd + k]);
-                }
-            }
-        }
-        std::fill(output_ids.begin(), output_ids.end(), -1);
-        for (int32_t i = 0; i < n_outputs; ++i) {
-            output_ids[out_ids[i]] = i;
-        }
-        out_ids.clear();
-    }
-}
-
 //
 // graph
 //
@@ -1616,7 +1233,7 @@ llm_graph_result_ptr llama_context::graph_build(
                 /*.backend_cpu =*/ backend_cpu,
                 /*.cvec        =*/ &cvec,
                 /*.loras       =*/ &loras,
-                /*.memory      =*/ kv_self.get(),
+                /*.memory      =*/ memory.get(),
                 /*.cross       =*/ &cross,
                 /*.n_outputs   =*/ n_outputs,
                 /*.cb          =*/ graph_get_cb(),
@@ -2020,8 +1637,6 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
     {
         LLAMA_LOG_DEBUG("%s: - writing output ids\n", __func__);
 
-        output_reorder();
-
         const auto n_outputs    = this->n_outputs;
         const auto & output_ids = this->output_ids;
 
@@ -2075,6 +1690,8 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
     }
 
     LLAMA_LOG_DEBUG("%s: - writing KV self\n", __func__);
+    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+
     kv_self->state_write(io);
 
     return io.n_bytes();
@@ -2159,6 +1776,8 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
     }
 
     LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__);
+    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+
     kv_self->state_read(io);
 
     return io.n_bytes();
@@ -2167,6 +1786,8 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
 size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id) {
     GGML_UNUSED(seq_id);
 
+    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+
     kv_self->state_write(io, seq_id);
 
     return io.n_bytes();
@@ -2175,6 +1796,8 @@ size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id s
 size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id) {
     GGML_UNUSED(seq_id);
 
+    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+
     kv_self->state_read(io, seq_id);
 
     return io.n_bytes();
@@ -2530,7 +2153,7 @@ void llama_kv_cache_seq_cp(
          llama_seq_id   seq_id_dst,
             llama_pos   p0,
             llama_pos   p1) {
-    return llama_kv_self_seq_cp(ctx, seq_id_src, seq_id_dst, p0, p1);
+    llama_kv_self_seq_cp(ctx, seq_id_src, seq_id_dst, p0, p1);
 }
 
 void llama_kv_self_seq_cp(
@@ -2544,14 +2167,14 @@ void llama_kv_self_seq_cp(
         return;
     }
 
-    return kv->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    kv->seq_cp(seq_id_src, seq_id_dst, p0, p1);
 }
 
 // deprecated
 void llama_kv_cache_seq_keep(
         llama_context * ctx,
          llama_seq_id   seq_id) {
-    return llama_kv_self_seq_keep(ctx, seq_id);
+    llama_kv_self_seq_keep(ctx, seq_id);
 }
 
 void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) {
@@ -2560,7 +2183,7 @@ void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) {
         return;
     }
 
-    return kv->seq_keep(seq_id);
+    kv->seq_keep(seq_id);
 }
 
 // deprecated
@@ -2570,7 +2193,7 @@ void llama_kv_cache_seq_add(
             llama_pos   p0,
             llama_pos   p1,
             llama_pos   delta) {
-    return llama_kv_self_seq_add(ctx, seq_id, p0, p1, delta);
+    llama_kv_self_seq_add(ctx, seq_id, p0, p1, delta);
 }
 
 void llama_kv_self_seq_add(
@@ -2584,7 +2207,7 @@ void llama_kv_self_seq_add(
         return;
     }
 
-    return kv->seq_add(seq_id, p0, p1, delta);
+    kv->seq_add(seq_id, p0, p1, delta);
 }
 
 // deprecated
@@ -2594,7 +2217,7 @@ void llama_kv_cache_seq_div(
             llama_pos   p0,
             llama_pos   p1,
                   int   d) {
-    return llama_kv_self_seq_div(ctx, seq_id, p0, p1, d);
+    llama_kv_self_seq_div(ctx, seq_id, p0, p1, d);
 }
 
 void llama_kv_self_seq_div(
@@ -2608,7 +2231,7 @@ void llama_kv_self_seq_div(
         return;
     }
 
-    return kv->seq_div(seq_id, p0, p1, d);
+    kv->seq_div(seq_id, p0, p1, d);
 }
 
 // deprecated
@@ -2627,7 +2250,7 @@ llama_pos llama_kv_self_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
 
 // deprecated
 void llama_kv_cache_defrag(llama_context * ctx) {
-    return llama_kv_self_defrag(ctx);
+    llama_kv_self_defrag(ctx);
 }
 
 void llama_kv_self_defrag(llama_context * ctx) {
@@ -2636,7 +2259,8 @@ void llama_kv_self_defrag(llama_context * ctx) {
         return;
     }
 
-    return kv->defrag();
+    // force defrag
+    kv->defrag_sched(-1.0f);
 }
 
 // deprecated

src/llama-context.h

@@ -27,7 +27,12 @@ struct llama_context {
 
     void synchronize();
 
-    const llama_model & get_model() const;
+    const llama_model   & get_model()   const;
+    const llama_cparams & get_cparams() const;
+
+    ggml_backend_sched_t get_sched() const;
+
+    ggml_context * get_ctx_compute() const;
 
     uint32_t n_ctx()         const;
     uint32_t n_ctx_per_seq() const;
@@ -137,49 +142,30 @@ private:
     // Returns max number of outputs for which space was reserved.
     int32_t output_reserve(int32_t n_outputs);
 
-    // make the outputs have the same order they had in the user-provided batch
-    // TODO: maybe remove this
-    void output_reorder();
-
     //
     // graph
     //
 
+public:
     int32_t graph_max_nodes() const;
 
     // zero-out inputs and create the ctx_compute for the compute graph
     ggml_cgraph * graph_init();
 
-    llm_graph_result_ptr graph_build(
-            ggml_context * ctx,
-             ggml_cgraph * gf,
-      const llama_ubatch & ubatch,
-          llm_graph_type   gtype);
-
     // returns the result of ggml_backend_sched_graph_compute_async execution
     ggml_status graph_compute(
             ggml_cgraph * gf,
                    bool   batched);
 
+private:
+    llm_graph_result_ptr graph_build(
+            ggml_context * ctx,
+             ggml_cgraph * gf,
+      const llama_ubatch & ubatch,
+          llm_graph_type   gtype);
+
     llm_graph_cb graph_get_cb() const;
 
-    // used by kv_self_update()
-    ggml_tensor * build_rope_shift(
-        ggml_context * ctx0,
-        ggml_tensor * cur,
-        ggml_tensor * shift,
-        ggml_tensor * factors,
-              float   freq_base,
-              float   freq_scale) const;
-
-    llm_graph_result_ptr build_kv_self_shift(
-            ggml_context * ctx0,
-            ggml_cgraph * gf) const;
-
-    llm_graph_result_ptr build_kv_self_defrag(
-            ggml_context * ctx0,
-            ggml_cgraph * gf) const;
-
     // TODO: read/write lora adapters and cvec
     size_t state_write_data(llama_io_write_i & io);
     size_t state_read_data (llama_io_read_i  & io);
@@ -196,11 +182,10 @@ private:
     llama_cparams       cparams;
     llama_adapter_cvec  cvec;
     llama_adapter_loras loras;
-    llama_sbatch        sbatch;
 
     llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably
 
-    std::unique_ptr<llama_kv_cache_unified> kv_self;
+    std::unique_ptr<llama_memory_i> memory;
 
     // TODO: remove
     bool logits_all = false;

src/llama-graph.cpp

@@ -55,13 +55,16 @@ void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
     if (ubatch->pos && pos) {
         const int64_t n_tokens = ubatch->n_tokens;
 
-        if (ubatch->token && n_pos_per_embd > 1) {
+        if (ubatch->token && n_pos_per_embd == 4) {
             // in case we're using M-RoPE with text tokens, convert the 1D positions to 4D
-            // the other dimensions are all 0, they are unused for text tokens
-            std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd, 0);
+            // the 3 first dims are the same, and 4th dim is all 0
+            std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd);
             // copy the first dimension
             for (int i = 0; i < n_tokens; ++i) {
-                pos_data[i] = ubatch->pos[i];
+                pos_data[               i] = ubatch->pos[i];
+                pos_data[    n_tokens + i] = ubatch->pos[i];
+                pos_data[2 * n_tokens + i] = ubatch->pos[i];
+                pos_data[3 * n_tokens + i] = 0; // 4th dim is 0
             }
             ggml_backend_tensor_set(pos, pos_data.data(), 0, pos_data.size()*ggml_element_size(pos));
         } else {
@@ -281,24 +284,7 @@ void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
 
         // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
         for (uint32_t i = 0; i < n_kv; ++i) {
-            const uint32_t  cell_id = i + kv_self->head;
-
-            //////////////////////////////////////////////
-            // TODO: this should not mutate the KV cache !
-            llama_kv_cell & kv_cell = const_cast<class llama_kv_cache_unified *>(kv_self)->cells[i];
-
-            // prevent out-of-bound sources
-            if (kv_cell.src < 0 || (uint32_t) kv_cell.src >= kv_self->size) {
-                kv_cell.src = cell_id;
-            }
-
-            data[i] = kv_cell.src;
-
-            // TODO: do not mutate the KV cache
-            // ensure copy only happens once
-            if (kv_cell.src != (int32_t) cell_id) {
-                kv_cell.src = cell_id;
-            }
+            data[i] = kv_self->s_copy(i);
         }
     }
 }
@@ -314,18 +300,7 @@ void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) {
 
         // clear unused states
         for (int i = 0; i < n_kv; ++i) {
-            const uint32_t  cell_id = i + kv_self->head;
-
-            //////////////////////////////////////////////
-            // TODO: this should not mutate the KV cache !
-            llama_kv_cell & kv_cell = const_cast<class llama_kv_cache_unified *>(kv_self)->cells[i];
-
-            data[i] = (float) (kv_cell.src >= 0);
-
-            // only clear once
-            if (kv_cell.src < 0) {
-                kv_cell.src = cell_id;
-            }
+            data[i] = kv_self->s_mask(i);
         }
     }
 }
@@ -925,28 +900,35 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
     cb(up, "ffn_moe_up", il);
 
-    ggml_tensor * gate = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
-    cb(gate, "ffn_moe_gate", il);
+    ggml_tensor * experts = nullptr;
+    if (gate_exps) {
+        cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
+        cb(cur, "ffn_moe_gate", il);
+    } else {
+        cur = up;
+    }
 
     switch (type_op) {
         case LLM_FFN_SILU:
             {
-                gate = ggml_silu(ctx0, gate);
-                cb(gate, "ffn_moe_silu", il);
+                cur = ggml_silu(ctx0, cur);
+                cb(cur, "ffn_moe_silu", il);
             } break;
         case LLM_FFN_GELU:
             {
-                gate = ggml_gelu(ctx0, gate);
-                cb(gate, "ffn_moe_gelu", il);
+                cur = ggml_gelu(ctx0, cur);
+                cb(cur, "ffn_moe_gelu", il);
             } break;
         default:
             GGML_ABORT("fatal error");
     }
 
-    ggml_tensor * par = ggml_mul(ctx0, up, gate); // [n_ff, n_expert_used, n_tokens]
-    cb(par, "ffn_moe_gate_par", il);
+    if (gate_exps) {
+        cur = ggml_mul(ctx0, cur, up); // [n_ff, n_expert_used, n_tokens]
+        cb(cur, "ffn_moe_gate_par", il);
+    }
 
-    ggml_tensor * experts = build_lora_mm_id(down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens]
+    experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
     cb(experts, "ffn_moe_down", il);
 
     if (!weight_before_ffn) {
@@ -1095,7 +1077,7 @@ ggml_tensor * llm_graph_context::build_inp_cls() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_s_copy() const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     auto inp = std::make_unique<llm_graph_input_s_copy>(kv_self);
 
@@ -1112,7 +1094,7 @@ ggml_tensor * llm_graph_context::build_inp_s_copy() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_s_mask() const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     auto inp = std::make_unique<llm_graph_input_s_mask>(kv_self);
 
@@ -1426,8 +1408,6 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // store to KV cache
     {
-        GGML_ASSERT(!kv_self->recurrent);
-
         const auto kv_head = kv_self->head;
 
         GGML_ASSERT(kv_self->size == n_ctx);
@@ -1577,7 +1557,7 @@ ggml_tensor * llm_graph_context::build_copy_mask_state(
          ggml_tensor * state_mask,
              int32_t   n_state,
              int32_t   n_seqs) const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     const auto n_kv    = kv_self->n;
     const auto kv_head = kv_self->head;
@@ -1609,7 +1589,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
          ggml_tensor * state_mask,
   const llama_ubatch & ubatch,
                  int   il) const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     const auto token_shift_count = hparams.token_shift_count;
 
@@ -1630,7 +1610,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
          ggml_tensor * token_shift,
   const llama_ubatch & ubatch,
                  int   il) const {
-    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
 
     const auto token_shift_count = hparams.token_shift_count;
     const auto n_embd = hparams.n_embd;

src/llama-graph.h

@@ -19,6 +19,7 @@ struct llama_cparams;
 
 class llama_memory_i;
 class llama_kv_cache_unified;
+class llama_kv_cache_recurrent;
 
 // certain models (typically multi-modal) can produce different types of graphs
 enum llm_graph_type {
@@ -186,26 +187,26 @@ public:
 
 class llm_graph_input_s_copy : public llm_graph_input_i {
 public:
-    llm_graph_input_s_copy(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
+    llm_graph_input_s_copy(const llama_kv_cache_recurrent * kv_self) : kv_self(kv_self) {}
     virtual ~llm_graph_input_s_copy() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * s_copy; // I32 [kv_size]
 
-    const llama_kv_cache_unified * kv_self;
+    const llama_kv_cache_recurrent * kv_self;
 };
 
 class llm_graph_input_s_mask : public llm_graph_input_i {
 public:
-    llm_graph_input_s_mask(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
+    llm_graph_input_s_mask(const llama_kv_cache_recurrent * kv_self) : kv_self(kv_self) {}
     virtual ~llm_graph_input_s_mask() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * s_mask; // F32 [1, n_kv]
 
-    const llama_kv_cache_unified * kv_self;
+    const llama_kv_cache_recurrent * kv_self;
 };
 
 class llm_graph_input_cross_embd : public llm_graph_input_i {
@@ -350,8 +351,8 @@ struct llm_graph_params {
     const llama_cparams & cparams;
     const llama_ubatch  & ubatch;
 
-    ggml_backend_sched * sched;
-    ggml_backend * backend_cpu;
+    ggml_backend_sched_t sched;
+    ggml_backend_t backend_cpu;
 
     const llama_adapter_cvec  * cvec;
     const llama_adapter_loras * loras;
@@ -402,9 +403,9 @@ struct llm_graph_context {
 
     ggml_context * ctx0 = nullptr;
 
-    ggml_backend_sched * sched;
+    ggml_backend_sched_t sched;
 
-    ggml_backend * backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?
+    ggml_backend_t backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?
 
     const llama_adapter_cvec  * cvec;
     const llama_adapter_loras * loras;

src/llama-hparams.h

@@ -66,6 +66,7 @@ struct llama_hparams {
     float    expert_weights_scale = 0.0;
     bool     expert_weights_norm  = false;
     uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
+    uint32_t moe_every_n_layers   = 0;
 
     float f_norm_eps;
     float f_norm_rms_eps;

src/llama-kv-cache.h

@@ -2,32 +2,72 @@
 
 #include "llama.h"
 #include "llama-io.h"
+#include "llama-graph.h"
 #include "llama-memory.h"
 
 #include "ggml-cpp.h"
 
-#include <functional>
 #include <set>
 #include <vector>
 
 struct llama_cparams;
 struct llama_hparams;
 struct llama_ubatch;
+struct llama_sbatch;
+struct llama_model;
+struct llama_context;
 
 struct llama_kv_cache : public llama_memory_i {
-    using llama_memory_i::llama_memory_i;
+    virtual ~llama_kv_cache() = default;
 
-    virtual void restore() = 0; // call if batch processing fails - restores the cache state
-    virtual void commit() = 0;  // call after successful batch processing - clears any pending state
+    // call if batch processing fails - restores the cache state
+    virtual void restore() = 0;
 
-    virtual int32_t get_n_tokens()   const = 0;
-    virtual int32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
+    // call after successful batch processing - clears any pending state
+    virtual void commit()  = 0;
 
-    virtual bool get_can_shift() const = 0;
+    // process any pending defrag/shift/etc. operations
+    // optionally call once before processing a new batch
+    virtual bool update(llama_context & lctx) = 0;
+
+    // schedule a defrag if the fragmentation threshold is exceeded. otherwise, do nothing
+    virtual void defrag_sched(float thold) = 0;
+
+    // simulate full cache, used for allocating worst-case compute buffers
+    virtual void set_full() = 0;
+
+    //
+    // batch processing
+    //
+
+    virtual llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) = 0;
+
+    // different KV caches require different batch splitting strategies
+    virtual llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const = 0;
+
+    // find an empty slot of size "n_tokens" in the cache
+    virtual bool find_slot(const llama_ubatch & batch) = 0;
+
+    // getters
+    virtual int32_t   get_n_tokens()   const = 0;
+    virtual int32_t   get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
+    virtual llama_pos get_pos_max()    const = 0;
+    virtual bool      get_can_shift()  const = 0;
 
     bool get_can_edit() const override { return get_can_shift(); }
+
+    //
+    // state write/read
+    //
+
+    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0;
+    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) = 0;
 };
 
+//
+// llama_kv_cache_guard
+//
+
 struct llama_kv_cache_guard {
     llama_kv_cache_guard(llama_kv_cache * kv) : kv(kv) {}
 
@@ -43,65 +83,50 @@ private:
     llama_kv_cache * kv;
 };
 
-struct llama_kv_cell {
-    llama_pos pos   = -1;
-    llama_pos delta =  0;
-    int32_t   src   = -1; // used by recurrent state models to copy states
-    int32_t   tail  = -1;
+//
+// llama_kv_cache_unified
+//
 
-    std::set<llama_seq_id> seq_id;
-
-    bool has_seq_id(const llama_seq_id & id) const {
-        return seq_id.find(id) != seq_id.end();
-    }
-
-    bool is_empty() const {
-        return seq_id.empty();
-    }
-
-    bool is_same_seq(const llama_kv_cell & other) const {
-        return seq_id == other.seq_id;
-    }
-};
-
-// ring-buffer of cached KV data
-// TODO: pimpl
 // TODO: add notion of max sequences
 class llama_kv_cache_unified : public llama_kv_cache {
 public:
-    // can be used to query data from the model if needed
-    struct callbacks {
-        std::function<ggml_tensor * (uint32_t n_ctx_per_seq, int il)> get_rope_factors;
+    struct kv_cell {
+        llama_pos pos   = -1;
+        llama_pos delta =  0;
+
+        std::set<llama_seq_id> seq_id;
+
+        bool has_seq_id(const llama_seq_id & id) const {
+            return seq_id.find(id) != seq_id.end();
+        }
+
+        bool is_empty() const {
+            return seq_id.empty();
+        }
+
+        bool is_same_seq(const kv_cell & other) const {
+            return seq_id == other.seq_id;
+        }
     };
 
+    static uint32_t get_padding(const llama_cparams & cparams);
+
     llama_kv_cache_unified(
-            const llama_hparams & hparams,
-            callbacks             cbs);
-
-    virtual ~llama_kv_cache_unified() = default;
-
-    // TODO: become constructor
-    bool init(
-            const llama_model & model,   // TODO: do not reference the model
-          const llama_cparams & cparams,
+            const llama_model & model,
                     ggml_type   type_k,
                     ggml_type   type_v,
+                         bool   v_trans,
+                         bool   offload,
                      uint32_t   kv_size,
-                         bool   offload);
+                     uint32_t   padding);
 
-    int32_t get_n_tokens()   const override;
-    int32_t get_used_cells() const override;
+    ~llama_kv_cache_unified() = default;
 
-    size_t total_size() const;
-
-    // TODO: better data structures to reduce the cost of this operation
-    llama_pos pos_max() const;
+    //
+    // llama_memory_i
+    //
 
     void clear() override;
-    void defrag() override;
-
-    virtual void restore() override;
-    virtual void commit() override;
 
     bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
     void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
@@ -111,63 +136,40 @@ public:
 
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
-    bool get_can_shift() const override;
+    //
+    // llama_kv_cache
+    //
+
+    void restore() override;
+    void commit()  override;
+
+    bool update(llama_context & ctx) override;
+
+    void defrag_sched(float thold) override;
+
+    void set_full() override;
+
+    llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
+
+    llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
 
-    // find an empty slot of size "n_tokens" in the cache
     // updates the cache head
     // Note: On success, it's important that cache.head points
     // to the first cell of the slot.
-    bool find_slot(const llama_ubatch & batch);
+    bool find_slot(const llama_ubatch & batch) override;
 
-    // TODO: maybe not needed
-    uint32_t get_padding(const llama_cparams & cparams) const;
+    int32_t get_n_tokens()   const override;
+    int32_t get_used_cells() const override;
 
-    // find how many cells are currently in use
-    uint32_t cell_max() const;
+    // TODO: better data structures to reduce the cost of this operation
+    llama_pos get_pos_max() const override;
 
-    size_t size_k_bytes() const;
-    size_t size_v_bytes() const;
-
-    // defrag
-
-    struct {
-        std::vector<uint32_t> ids;
-    } defrag_info;
-
-    // return true if cells have been moved
-    bool defrag_prepare(int32_t n_max_nodes);
-
-    // commit/restore cache
-
-    struct slot_range {
-        uint32_t c0 = 0; // note: these are cell indices, not sequence positions
-        uint32_t c1 = 0;
-    };
-
-    // pending cell updates that are not yet committed
-    struct {
-        std::vector<slot_range> ranges;
-    } pending;
+    bool get_can_shift() const override;
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1);
-
-    // members
-
-    const llama_hparams & hparams;
-
-    callbacks cbs;
-
-    bool has_shift = false;
-    bool do_defrag = false;
-
-    // TODO: remove this and implement llama_kv_cache_recurrent instead
-    bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token
-
-    bool v_trans   = true;  // the value tensor is transposed
-    bool can_shift = false;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
 
     // Note: The value of head isn't only used to optimize searching
     // for a free KV slot. llama_decode_impl also uses it, so it
@@ -179,18 +181,213 @@ public:
     // computed before each graph build
     uint32_t n = 0;
 
-    std::vector<llama_kv_cell> cells;
+    std::vector<kv_cell> cells;
 
     std::vector<ggml_tensor *> k_l; // per layer
     std::vector<ggml_tensor *> v_l;
 
 private:
+    const llama_model & model;
+    const llama_hparams & hparams;
+
+    bool has_shift = false;
+    bool do_defrag = false;
+
+    bool v_trans   = true;  // the value tensor is transposed
+    bool can_shift = false;
+
+    // required padding
+    uint32_t padding = 1;
+
     ggml_type type_k = GGML_TYPE_F16;
     ggml_type type_v = GGML_TYPE_F16;
 
     std::vector<ggml_context_ptr>        ctxs;
     std::vector<ggml_backend_buffer_ptr> bufs;
 
+    // defrag
+    struct {
+        std::vector<uint32_t> ids;
+    } defrag_info;
+
+    // return true if cells have been moved
+    bool defrag_prepare(int32_t n_max_nodes);
+
+    // commit/restore cache
+    struct slot_range {
+        uint32_t c0 = 0; // note: these are cell indices, not sequence positions
+        uint32_t c1 = 0;
+    };
+
+    // pending cell updates that are not yet committed
+    struct {
+        std::vector<slot_range> ranges;
+    } pending;
+
+    // find how many cells are currently in use
+    uint32_t cell_max() const;
+
+    size_t total_size() const;
+
+    size_t size_k_bytes() const;
+    size_t size_v_bytes() const;
+
+    ggml_tensor * build_rope_shift(
+            const llama_cparams & cparams,
+                   ggml_context * ctx,
+                    ggml_tensor * cur,
+                    ggml_tensor * shift,
+                    ggml_tensor * factors,
+                          float   freq_base,
+                          float   freq_scale) const;
+
+    llm_graph_result_ptr build_graph_shift(
+            const llama_cparams & cparams,
+                   ggml_context * ctx,
+                    ggml_cgraph * gf) const;
+
+    llm_graph_result_ptr build_graph_defrag(
+            const llama_cparams & cparams,
+                   ggml_context * ctx,
+                    ggml_cgraph * gf) const;
+
+    void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
+    void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
+
+    bool state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
+    bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
+};
+
+//
+// llama_kv_cache_recurrent
+//
+
+class llama_kv_cache_recurrent : public llama_kv_cache {
+public:
+    struct kv_cell {
+        llama_pos pos  = -1;
+        int32_t   src  = -1; // used to copy states
+        int32_t   tail = -1;
+
+        std::set<llama_seq_id> seq_id;
+
+        bool has_seq_id(const llama_seq_id & id) const {
+            return seq_id.find(id) != seq_id.end();
+        }
+
+        bool is_empty() const {
+            return seq_id.empty();
+        }
+
+        bool is_same_seq(const kv_cell & other) const {
+            return seq_id == other.seq_id;
+        }
+    };
+
+    llama_kv_cache_recurrent(
+            const llama_model & model,
+                    ggml_type   type_k,
+                    ggml_type   type_v,
+                         bool   offload,
+                     uint32_t   kv_size);
+
+    ~llama_kv_cache_recurrent() = default;
+
+    //
+    // llama_memory_i
+    //
+
+    void clear() override;
+
+    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
+    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
+    void seq_keep(llama_seq_id seq_id) override;
+    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos delta) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    //
+    // llama_kv_cache
+    //
+
+    void restore() override;
+    void commit()  override;
+
+    bool update(llama_context & lctx) override;
+
+    void defrag_sched(float thold) override;
+
+    void set_full() override;
+
+    llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
+
+    llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
+
+    bool find_slot(const llama_ubatch & batch) override;
+
+    int32_t get_n_tokens()   const override;
+    int32_t get_used_cells() const override;
+
+    // TODO: better data structures to reduce the cost of this operation
+    llama_pos get_pos_max() const override;
+
+    bool get_can_shift() const override;
+
+    // TODO: temporary methods - they are not really const as they do const_cast<>, fix this
+    int32_t s_copy(int i) const;
+    float   s_mask(int i) const;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
+
+    // Note: The value of head isn't only used to optimize searching
+    // for a free KV slot. llama_decode_impl also uses it, so it
+    // cannot be freely changed after a slot has been allocated.
+    uint32_t head = 0;
+    uint32_t size = 0;
+    uint32_t used = 0; // used cells (i.e. at least one seq_id)
+
+    // computed before each graph build
+    uint32_t n = 0;
+
+    std::vector<kv_cell> cells;
+
+    std::vector<ggml_tensor *> k_l; // per layer
+    std::vector<ggml_tensor *> v_l;
+
+private:
+    //const llama_model & model;
+    const llama_hparams & hparams;
+
+    // commit/restore cache
+    // TODO: rework for recurrent cache
+    struct slot_range {
+        uint32_t c0 = 0; // note: these are cell indices, not sequence positions
+        uint32_t c1 = 0;
+    };
+
+    // pending cell updates that are not yet committed
+    struct {
+        std::vector<slot_range> ranges;
+    } pending;
+
+    ggml_type type_k = GGML_TYPE_F16;
+    ggml_type type_v = GGML_TYPE_F16;
+
+    std::vector<ggml_context_ptr>        ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    // find how many cells are currently in use
+    uint32_t cell_max() const;
+
+    size_t total_size() const;
+
+    size_t size_k_bytes() const;
+    size_t size_v_bytes() const;
+
     void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
     void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
 
@@ -198,11 +395,6 @@ private:
     bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
 };
 
-// TODO: temporary reusing llama_kv_cache_unified -- implement recurrent cache and simplify llama_kv_cache_unified
-//class llama_kv_cache_recurrent : public llama_kv_cache_unified {
-//public:
-//    using llama_kv_cache_unified::llama_kv_cache_unified;
-//};
 
 //
 // kv cache view

src/llama-memory.h

@@ -2,12 +2,22 @@
 
 #include "llama.h"
 
+struct llama_memory_params {
+    // kv cache
+    ggml_type type_k;
+    ggml_type type_v;
+
+    // parameters for other types of memory
+    // ...
+};
+
 // general concept of LLM memory
 // the KV cache is a type of LLM memory, but there can be other types
 class llama_memory_i {
 public:
+    virtual ~llama_memory_i() = default;
+
     virtual void clear() = 0;
-    virtual void defrag() = 0;
 
     virtual bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) = 0;
     virtual void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) = 0;