server: Reorder methods in server-task.cpp (#19016)

* Move `task_result_state::update_chat_msg` to match with header

* Move `server_task_result_cmpl_partial::to_json_anthropic()` to match with header

---------

Co-authored-by: openingnow <>

.github/workflows/build-cache.yml

@@ -16,7 +16,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Get latest Vulkan SDK version
         id: vulkan_sdk_version
@@ -24,7 +24,7 @@ jobs:
           echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
 
       - name: Setup Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-sdk
         with:
           path: ./vulkan_sdk
@@ -47,10 +47,10 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-toolchain
         with:
           path: ./spacemit_toolchain
@@ -73,10 +73,10 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-rocm
         with:
           path: C:\Program Files\AMD\ROCm

.github/workflows/build-cmake-pkg.yml

@@ -7,7 +7,7 @@ jobs:
   linux:
     runs-on: ubuntu-24.04
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           fetch-depth: 0
 

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

@@ -8,7 +8,7 @@ jobs:
   #   runs-on: ubuntu-24.04
 
   #   steps:
-  #     - uses: actions/checkout@v4
+  #     - uses: actions/checkout@v6
   #     - name: Setup Riscv
   #       run: |
   #         sudo dpkg --add-architecture riscv64
@@ -52,7 +52,7 @@ jobs:
   #   runs-on: ubuntu-24.04
 
   #   steps:
-  #     - uses: actions/checkout@v4
+  #     - uses: actions/checkout@v6
   #     - name: Setup Riscv
   #       run: |
   #         sudo dpkg --add-architecture riscv64
@@ -99,7 +99,7 @@ jobs:
   #   runs-on: ubuntu-24.04
 
   #   steps:
-  #     - uses: actions/checkout@v4
+  #     - uses: actions/checkout@v6
   #     - name: Setup Arm64
   #       run: |
   #         sudo dpkg --add-architecture arm64
@@ -146,7 +146,7 @@ jobs:
     container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
       - name: Setup LoongArch
         run: |
           rm -f /etc/apt/sources.list.d/*
@@ -201,7 +201,7 @@ jobs:
     container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
       - name: Setup LoongArch
         run: |
           rm -f /etc/apt/sources.list.d/*
@@ -262,10 +262,10 @@ jobs:
       SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
 
       - name: Use SpacemiT Toolchain Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-toolchain
         with:
           path: ./spacemit_toolchain

.github/workflows/build.yml

@@ -63,7 +63,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -99,7 +99,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -135,7 +135,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -189,7 +189,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -269,7 +269,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -317,7 +317,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Dependencies
         id: depends
@@ -347,7 +347,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       # - name: ccache
       #   uses: ggml-org/ccache-action@v1.2.16
@@ -380,7 +380,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -414,7 +414,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -436,7 +436,7 @@ jobs:
           echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
 
       - name: Use Vulkan SDK Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-sdk
         with:
           path: ./vulkan_sdk
@@ -472,7 +472,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -494,7 +494,7 @@ jobs:
           echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
 
       - name: Use Vulkan SDK Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-sdk
         with:
           path: ./vulkan_sdk
@@ -543,7 +543,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -585,7 +585,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Dependencies
         id: depends
@@ -616,7 +616,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Dependencies
         id: depends
@@ -644,7 +644,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
 
       - name: add oneAPI to apt
         shell: bash
@@ -668,7 +668,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -693,7 +693,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
 
       - name: add oneAPI to apt
         shell: bash
@@ -717,7 +717,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -749,7 +749,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -781,7 +781,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -813,7 +813,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Build
         id: cmake_build
@@ -843,7 +843,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -853,7 +853,7 @@ jobs:
           save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Download xcframework artifact
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v7
         with:
           name: llama-xcframework
           path: build-apple/llama.xcframework/
@@ -885,7 +885,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -954,7 +954,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1053,7 +1053,7 @@ jobs:
     steps:
         - name: Clone
           id: checkout
-          uses: actions/checkout@v4
+          uses: actions/checkout@v6
 
         - name: Install dependencies
           env:
@@ -1092,7 +1092,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Install ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1145,7 +1145,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1177,7 +1177,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Grab rocWMMA package
         id: grab_rocwmma
@@ -1187,7 +1187,7 @@ jobs:
           7z x data.tar
 
       - name: Use ROCm Installation Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-rocm
         with:
           path: C:\Program Files\AMD\ROCm
@@ -1239,7 +1239,7 @@ jobs:
 
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup Xcode
         uses: maxim-lobanov/setup-xcode@v1
@@ -1269,7 +1269,7 @@ jobs:
           ./build-xcframework.sh
 
       - name: Upload xcframework artifact
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           name: llama-xcframework
           path: build-apple/llama.xcframework/
@@ -1285,7 +1285,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       # Disabled due to size (400MB) and always 0 cache hits
       # - name: ccache
@@ -1295,7 +1295,7 @@ jobs:
       #     evict-old-files: 1d
 
       - name: Set up JDK
-        uses: actions/setup-java@v3
+        uses: actions/setup-java@v5
         with:
           java-version: 17
           distribution: zulu
@@ -1327,7 +1327,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Install OpenCL Headers and Libs
         id: install_opencl
@@ -1402,7 +1402,7 @@ jobs:
     runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
     steps:
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -1460,7 +1460,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1486,7 +1486,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1512,7 +1512,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1538,7 +1538,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1564,7 +1564,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1590,7 +1590,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Test
         id: ggml-ci
@@ -1604,7 +1604,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Test
         id: ggml-ci
@@ -1618,7 +1618,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Test
         id: ggml-ci
@@ -1632,7 +1632,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Test
         id: ggml-ci
@@ -1645,7 +1645,7 @@ jobs:
   #   steps:
   #     - name: Clone
   #       id: checkout
-  #       uses: actions/checkout@v4
+  #       uses: actions/checkout@v6
 
   #     - name: Test
   #       id: ggml-ci
@@ -1659,7 +1659,7 @@ jobs:
   #   steps:
   #     - name: Clone
   #       id: checkout
-  #       uses: actions/checkout@v4
+  #       uses: actions/checkout@v6
 
   #     - name: Test
   #       id: ggml-ci
@@ -1673,7 +1673,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Test
         id: ggml-ci
@@ -1686,7 +1686,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Dawn Dependency
         id: dawn-depends
@@ -1714,7 +1714,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Test
         id: ggml-ci
@@ -1728,7 +1728,7 @@ jobs:
      steps:
        - name: Clone
          id: checkout
-         uses: actions/checkout@v4
+         uses: actions/checkout@v6
 
        - name: ccache
          uses: ggml-org/ccache-action@v1.2.16
@@ -1773,7 +1773,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Check environment
         run: |
@@ -1875,7 +1875,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup ccache
         run: |
@@ -1969,7 +1969,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup ccache
         run: |
@@ -2043,7 +2043,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup ccache
         run: |
@@ -2089,7 +2089,7 @@ jobs:
      steps:
        - name: Clone
          id: checkout
-         uses: actions/checkout@v4
+         uses: actions/checkout@v6
 
        - name: Dependencies
          id: depends

.github/workflows/check-vendor.yml

@@ -23,12 +23,12 @@ jobs:
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
       - name: Setup Python
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v6
         with:
           python-version: '3.x'
 

.github/workflows/close-issue.yml

@@ -15,7 +15,7 @@ jobs:
       issues: write
       pull-requests: write
     steps:
-      - uses: actions/stale@v5
+      - uses: actions/stale@v10
         with:
           exempt-issue-labels: "refactoring,help wanted,good first issue,research 🔬,bug,roadmap"
           days-before-issue-stale: 30

.github/workflows/copilot-setup-steps.yml

@@ -26,7 +26,7 @@ jobs:
     # If you do not check out your code, Copilot will do this for you.
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -45,7 +45,7 @@ jobs:
           sudo chmod +x /usr/local/bin/git-clang-format
 
       - name: Set up Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: '3.11'
 

.github/workflows/docker.yml

@@ -49,7 +49,7 @@ jobs:
           - { tag: "rocm",   dockerfile: ".devops/rocm.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
     steps:
       - name: Check out the repo
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0 # preserve git history, so we can determine the build number
 
@@ -63,7 +63,7 @@ jobs:
         uses: docker/setup-buildx-action@v3
 
       - name: Log in to Docker Hub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           registry: ghcr.io
           username: ${{ github.repository_owner }}
@@ -208,7 +208,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 

.github/workflows/editorconfig.yml

@@ -22,7 +22,7 @@ jobs:
   editorconfig:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
       - uses: editorconfig-checker/action-editorconfig-checker@v2
         with:
           version: v3.0.3

.github/workflows/gguf-publish.yml

@@ -24,9 +24,9 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v6
     - name: Set up Python
-      uses: actions/setup-python@v5
+      uses: actions/setup-python@v6
       with:
         python-version: '3.9.x'
     - name: Install dependencies

.github/workflows/labeler.yml

@@ -9,9 +9,9 @@ jobs:
       pull-requests: write
     runs-on: ubuntu-latest
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v6
       with:
         repository: "ggml-org/llama.cpp"
-    - uses: actions/labeler@v5
+    - uses: actions/labeler@v6
       with:
         configuration-path: '.github/labeler.yml'

.github/workflows/pre-tokenizer-hashes.yml

@@ -16,10 +16,10 @@ jobs:
 
         steps:
         - name: Checkout repository
-          uses: actions/checkout@v4
+          uses: actions/checkout@v6
 
         - name: Set up Python
-          uses: actions/setup-python@v5
+          uses: actions/setup-python@v6
           with:
               python-version: '3.11'
 

.github/workflows/python-check-requirements.yml

@@ -24,9 +24,9 @@ jobs:
     name: check-requirements
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
       - name: Set up Python environment
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: "3.11"
       - name: Run check-requirements.sh script

.github/workflows/python-lint.yml

@@ -19,9 +19,9 @@ jobs:
     name: Lint
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
       - name: Set up Python environment
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: "3.11"
       - name: flake8 Lint

.github/workflows/python-type-check.yml

@@ -24,9 +24,9 @@ jobs:
     name: pyright type-check
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
       - name: Set up Python environment
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: "3.11"
       - name: Install Python dependencies

.github/workflows/release.yml

@@ -27,7 +27,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -63,7 +63,7 @@ jobs:
           tar -czvf llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz -s ",./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz
           name: llama-bin-macos-arm64.tar.gz
@@ -74,7 +74,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -111,7 +111,7 @@ jobs:
           tar -czvf llama-${{ steps.tag.outputs.name }}-bin-macos-x64.tar.gz -s ",./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.tar.gz
           name: llama-bin-macos-x64.tar.gz
@@ -133,7 +133,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -173,7 +173,7 @@ jobs:
           tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-${{ matrix.build }}.tar.gz
@@ -184,7 +184,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -226,7 +226,7 @@ jobs:
           tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz
           name: llama-bin-ubuntu-vulkan-x64.tar.gz
@@ -242,7 +242,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -278,7 +278,7 @@ jobs:
           7z a -snl llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-bin-win-cpu-${{ matrix.arch }}.zip
           name: llama-bin-win-cpu-${{ matrix.arch }}.zip
@@ -305,7 +305,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -360,7 +360,7 @@ jobs:
           7z a -snl llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip .\build\bin\Release\${{ matrix.target }}.dll
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
           name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
@@ -375,7 +375,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Install ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -416,7 +416,7 @@ jobs:
           7z a -snl llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
           name: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
@@ -431,7 +431,7 @@ jobs:
           7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
           name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
@@ -451,7 +451,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -511,7 +511,7 @@ jobs:
           7z a -snl llama-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload the release package
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
@@ -531,7 +531,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Grab rocWMMA package
         id: grab_rocwmma
@@ -542,7 +542,7 @@ jobs:
 
       - name: Cache ROCm Installation
         id: cache-rocm
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         with:
           path: C:\Program Files\AMD\ROCm
           key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
@@ -617,7 +617,7 @@ jobs:
           7z a -snl llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-bin-win-hip-${{ matrix.name }}-x64.zip
           name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
@@ -627,7 +627,7 @@ jobs:
 
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -672,7 +672,7 @@ jobs:
           zip -r -y llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-${{ steps.tag.outputs.name }}-xcframework.zip
           name: llama-${{ steps.tag.outputs.name }}-xcframework.zip
@@ -703,7 +703,7 @@ jobs:
     runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
     steps:
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -763,7 +763,7 @@ jobs:
           tar -czvf llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz
           name: llama-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz
@@ -794,7 +794,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -804,7 +804,7 @@ jobs:
 
       - name: Download artifacts
         id: download-artifact
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v7
         with:
           path: ./artifact
           merge-multiple: true
@@ -887,7 +887,7 @@ jobs:
 
       - name: Upload release
         id: upload_release
-        uses: actions/github-script@v3
+        uses: actions/github-script@v8
         with:
           github-token: ${{secrets.GITHUB_TOKEN}}
           script: |

.github/workflows/server-webui.yml

@@ -37,14 +37,14 @@ jobs:
     continue-on-error: true
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
       - name: Setup Node.js
         id: node
-        uses: actions/setup-node@v4
+        uses: actions/setup-node@v6
         with:
           node-version: "22"
           cache: "npm"
@@ -131,14 +131,14 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
       - name: Python setup
         id: setup_python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: '3.11'
 
@@ -148,7 +148,7 @@ jobs:
           pip install -r tools/server/tests/requirements.txt
 
       - name: Setup Node.js for WebUI
-        uses: actions/setup-node@v4
+        uses: actions/setup-node@v6
         with:
           node-version: "22"
           cache: "npm"

.github/workflows/server.yml

@@ -64,7 +64,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
@@ -77,7 +77,7 @@ jobs:
 
       - name: Python setup
         id: setup_python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: '3.11'
 
@@ -100,7 +100,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
@@ -113,7 +113,7 @@ jobs:
 
       - name: Python setup
         id: setup_python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: '3.11'
 

.github/workflows/update-ops-docs.yml

@@ -18,10 +18,10 @@ jobs:
 
         steps:
         - name: Checkout repository
-          uses: actions/checkout@v4
+          uses: actions/checkout@v6
 
         - name: Set up Python
-          uses: actions/setup-python@v5
+          uses: actions/setup-python@v6
           with:
               python-version: '3.x'
 

.github/workflows/winget.yml

@@ -21,7 +21,7 @@ jobs:
 
       - name: Find latest release
         id: find_latest_release
-        uses: actions/github-script@v6
+        uses: actions/github-script@v8
         with:
           script: |
             const { data: releases } = await github.rest.repos.listReleases({

ci/run.sh

@@ -254,7 +254,7 @@ function gg_run_ctest_release {
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
-        (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure -L 'main|python' ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     else
         (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     fi

common/chat-parser.cpp

@@ -129,7 +129,7 @@ static void parse_json_tool_calls(
     }
 }
 
-common_chat_msg_parser::common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax)
+common_chat_msg_parser::common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_parser_params & syntax)
     : input_(input), is_partial_(is_partial), syntax_(syntax)
 {
     result_.role = "assistant";
@@ -1611,7 +1611,7 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
     builder.finish();
 }
 
-common_chat_msg common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax) {
+common_chat_msg common_chat_parse(const std::string & input, bool is_partial, const common_chat_parser_params & syntax) {
     if (syntax.format == COMMON_CHAT_FORMAT_PEG_SIMPLE ||
         syntax.format == COMMON_CHAT_FORMAT_PEG_NATIVE ||
         syntax.format == COMMON_CHAT_FORMAT_PEG_CONSTRUCTED) {
@@ -1635,7 +1635,7 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
     return msg;
 }
 
-common_chat_msg common_chat_peg_parse(const common_peg_arena & parser, const std::string & input, bool is_partial, const common_chat_syntax & syntax) {
+common_chat_msg common_chat_peg_parse(const common_peg_arena & parser, const std::string & input, bool is_partial, const common_chat_parser_params & syntax) {
     if (parser.empty()) {
         throw std::runtime_error("Failed to parse due to missing parser definition.");
     }

common/chat-parser.h

@@ -5,7 +5,7 @@
 #include "json-partial.h"
 #include "regex-partial.h"
 
-#include <nlohmann/json.hpp>
+#include <nlohmann/json_fwd.hpp>
 
 #include <optional>
 #include <string>
@@ -19,20 +19,20 @@ class common_chat_msg_partial_exception : public std::runtime_error {
 class common_chat_msg_parser {
     std::string input_;
     bool is_partial_;
-    common_chat_syntax syntax_;
+    common_chat_parser_params syntax_; // TODO: rename to params
     std::string healing_marker_;
 
     size_t pos_ = 0;
     common_chat_msg result_;
 
   public:
-    common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
+    common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_parser_params & syntax);
     const std::string & input() const { return input_; }
     size_t pos() const { return pos_; }
     const std::string & healing_marker() const { return healing_marker_; }
     const bool & is_partial() const { return is_partial_; }
     const common_chat_msg & result() const { return result_; }
-    const common_chat_syntax & syntax() const { return syntax_; }
+    const common_chat_parser_params & syntax() const { return syntax_; }
 
     void move_to(size_t pos) {
         if (pos > input_.size()) {

common/chat.cpp

@@ -601,18 +601,18 @@ bool common_chat_templates_was_explicit(const struct common_chat_templates * tmp
     return tmpls->has_explicit_template;
 }
 
-const char * common_chat_templates_source(const struct common_chat_templates * tmpls, const char * variant) {
-    if (variant != nullptr) {
-        if (strcmp(variant, "tool_use") == 0) {
+std::string common_chat_templates_source(const struct common_chat_templates * tmpls, const std::string & variant) {
+    if (!variant.empty()) {
+        if (variant == "tool_use") {
             if (tmpls->template_tool_use) {
-                return tmpls->template_tool_use->source().c_str();
+                return tmpls->template_tool_use->source();
             }
-            return nullptr;
+            return "";
         } else {
-            LOG_DBG("%s: unknown template variant: %s\n", __func__, variant);
+            LOG_DBG("%s: unknown template variant: %s\n", __func__, variant.c_str());
         }
     }
-    return tmpls->template_default->source().c_str();
+    return tmpls->template_default->source();
 }
 
 common_chat_templates_ptr common_chat_templates_init(

common/chat.h

@@ -145,7 +145,7 @@ struct common_chat_templates_inputs {
     std::vector<common_chat_tool> tools;
     common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
     bool parallel_tool_calls = false;
-    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE; // TODO: refactor this to "bool enable_thinking"
     bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
     std::map<std::string, std::string> chat_template_kwargs;
@@ -165,14 +165,21 @@ struct common_chat_params {
     std::string                         parser;
 };
 
-struct common_chat_syntax {
+// per-message parsing syntax
+// should be derived from common_chat_params
+struct common_chat_parser_params {
     common_chat_format       format                = COMMON_CHAT_FORMAT_CONTENT_ONLY;
-    common_reasoning_format  reasoning_format      = COMMON_REASONING_FORMAT_NONE;
+    common_reasoning_format  reasoning_format      = COMMON_REASONING_FORMAT_NONE; // TODO: refactor this to "bool parse_reasoning"
     // Whether reasoning_content should be inlined in the content (e.g. for reasoning_format=deepseek in stream mode)
     bool                     reasoning_in_content  = false;
     bool                     thinking_forced_open  = false;
     bool                     parse_tool_calls      = true;
     common_peg_arena         parser                = {};
+    common_chat_parser_params() = default;
+    common_chat_parser_params(const common_chat_params & chat_params) {
+        format               = chat_params.format;
+        thinking_forced_open = chat_params.thinking_forced_open;
+    }
 };
 
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
@@ -191,7 +198,7 @@ common_chat_templates_ptr common_chat_templates_init(
                                            const std::string & eos_token_override = "");
 
 bool         common_chat_templates_was_explicit(const struct common_chat_templates * tmpls);
-const char * common_chat_templates_source(const struct common_chat_templates * tmpls, const char * variant = nullptr);
+std::string  common_chat_templates_source(const struct common_chat_templates * tmpls, const std::string & variant = "");
 
 
 struct common_chat_params      common_chat_templates_apply(
@@ -213,10 +220,12 @@ std::string common_chat_format_example(
     const std::map<std::string, std::string> & chat_template_kwargs);
 
 const char*               common_chat_format_name(common_chat_format format);
-const char*               common_reasoning_format_name(common_reasoning_format format);
-common_reasoning_format   common_reasoning_format_from_name(const std::string & format);
-common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
-common_chat_msg           common_chat_peg_parse(const common_peg_arena & parser, const std::string & input, bool is_partial, const common_chat_syntax & syntax);
+common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_parser_params & syntax);
+common_chat_msg           common_chat_peg_parse(const common_peg_arena & parser, const std::string & input, bool is_partial, const common_chat_parser_params & syntax);
+
+// used by arg and server
+const char *             common_reasoning_format_name(common_reasoning_format format);
+common_reasoning_format  common_reasoning_format_from_name(const std::string & format);
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
 

common/common.h

@@ -57,6 +57,8 @@ extern const char * LLAMA_COMMIT;
 extern const char * LLAMA_COMPILER;
 extern const char * LLAMA_BUILD_TARGET;
 
+const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
+
 struct common_control_vector_load_info;
 
 //
@@ -284,6 +286,7 @@ struct common_params_diffusion {
 };
 
 // reasoning API response format (not to be confused as chat template's reasoning format)
+// only used by server
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
     COMMON_REASONING_FORMAT_AUTO,            // Same as deepseek, using `message.reasoning_content`

common/download.cpp

@@ -314,23 +314,26 @@ static bool common_pull_file(httplib::Client & cli,
 
 // download one single file from remote URL to local path
 // returns status code or -1 on error
-static int common_download_file_single_online(const std::string & url,
-                                               const std::string & path,
-                                               const std::string & bearer_token,
-                                               const common_header_list & custom_headers) {
+static int common_download_file_single_online(const std::string        & url,
+                                              const std::string        & path,
+                                              const std::string        & bearer_token,
+                                              const common_header_list & custom_headers) {
     static const int max_attempts        = 3;
     static const int retry_delay_seconds = 2;
 
     auto [cli, parts] = common_http_client(url);
 
-    httplib::Headers default_headers = {{"User-Agent", "llama-cpp"}};
-    if (!bearer_token.empty()) {
-        default_headers.insert({"Authorization", "Bearer " + bearer_token});
-    }
+    httplib::Headers headers;
     for (const auto & h : custom_headers) {
-        default_headers.emplace(h.first, h.second);
+        headers.emplace(h.first, h.second);
     }
-    cli.set_default_headers(default_headers);
+    if (headers.find("User-Agent") == headers.end()) {
+        headers.emplace("User-Agent", "llama-cpp/" + build_info);
+    }
+    if (!bearer_token.empty()) {
+        headers.emplace("Authorization", "Bearer " + bearer_token);
+    }
+    cli.set_default_headers(headers);
 
     const bool file_exists = std::filesystem::exists(path);
 
@@ -437,10 +440,12 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string
                                                              const common_remote_params & params) {
     auto [cli, parts] = common_http_client(url);
 
-    httplib::Headers headers = {{"User-Agent", "llama-cpp"}};
-
-    for (const auto & header : params.headers) {
-        headers.emplace(header.first, header.second);
+    httplib::Headers headers;
+    for (const auto & h : params.headers) {
+        headers.emplace(h.first, h.second);
+    }
+    if (headers.find("User-Agent") == headers.end()) {
+        headers.emplace("User-Agent", "llama-cpp/" + build_info);
     }
 
     if (params.timeout > 0) {

common/http.h

@@ -57,6 +57,17 @@ static std::pair<httplib::Client, common_http_url> common_http_client(const std:
         throw std::runtime_error("error: invalid URL format");
     }
 
+#ifndef CPPHTTPLIB_OPENSSL_SUPPORT
+    if (parts.scheme == "https") {
+        throw std::runtime_error(
+            "HTTPS is not supported. Please rebuild with:\n"
+            "  -DLLAMA_BUILD_BORINGSSL=ON\n"
+            "  -DLLAMA_BUILD_LIBRESSL=ON\n"
+            "or ensure dev files of an OpenSSL-compatible library are available when building."
+        );
+    }
+#endif
+
     httplib::Client cli(parts.scheme + "://" + parts.host);
 
     if (!parts.user.empty()) {

common/jinja/runtime.cpp

@@ -805,7 +805,7 @@ value member_expression::execute_impl(context & ctx) {
         } else if (is_val<value_string>(property)) {
             auto key = property->as_string().str();
             JJ_DEBUG("Accessing %s built-in '%s'", is_val<value_array>(object) ? "array" : "string", key.c_str());
-            val = try_builtin_func(ctx, key, object);
+            val = try_builtin_func(ctx, key, object, true);
         } else {
             throw std::runtime_error("Cannot access property with non-string/non-number: got " + property->type());
         }
@@ -814,7 +814,7 @@ value member_expression::execute_impl(context & ctx) {
             throw std::runtime_error("Cannot access property with non-string: got " + property->type());
         }
         auto key = property->as_string().str();
-        val = try_builtin_func(ctx, key, object);
+        val = try_builtin_func(ctx, key, object, true);
     }
 
     if (ctx.is_get_stats && val && object && property) {

common/jinja/value.cpp

@@ -1005,6 +1005,7 @@ const func_builtins & value_none_t::get_builtins() const {
     static const func_builtins builtins = {
         {"default", default_value},
         {"tojson", tojson},
+        {"string", [](const func_args &) -> value { return mk_val<value_string>("None"); }}
     };
     return builtins;
 }

common/jinja/value.h

@@ -203,6 +203,9 @@ struct value_int_t : public value_t {
     virtual int64_t as_int() const override { return val_int; }
     virtual double as_float() const override { return static_cast<double>(val_int); }
     virtual string as_string() const override { return std::to_string(val_int); }
+    virtual bool as_bool() const override {
+        return val_int != 0;
+    }
     virtual const func_builtins & get_builtins() const override;
 };
 using value_int = std::shared_ptr<value_int_t>;
@@ -219,6 +222,9 @@ struct value_float_t : public value_t {
         if (out.back() == '.') out.push_back('0'); // leave one zero if no decimals
         return out;
     }
+    virtual bool as_bool() const override {
+        return val_flt != 0.0;
+    }
     virtual const func_builtins & get_builtins() const override;
 };
 using value_float = std::shared_ptr<value_float_t>;
@@ -336,12 +342,12 @@ struct value_none_t : public value_t {
     virtual std::string type() const override { return "None"; }
     virtual bool is_none() const override { return true; }
     virtual bool as_bool() const override { return false; }
+    virtual string as_string() const override { return string("None"); }
     virtual std::string as_repr() const override { return type(); }
     virtual const func_builtins & get_builtins() const override;
 };
 using value_none = std::shared_ptr<value_none_t>;
 
-
 struct value_undefined_t : public value_t {
     std::string hint; // for debugging, to indicate where undefined came from
     value_undefined_t(const std::string & h = "") : hint(h) {}

common/json-partial.h

@@ -1,5 +1,6 @@
 #pragma once
 
+// TODO: use json_fwd.hpp when possible
 #include <nlohmann/json.hpp>
 
 // Healing marker (empty if the JSON was fully parsed / wasn't healed).

convert_hf_to_gguf.py

@@ -1078,6 +1078,9 @@ class TextModel(ModelBase):
         if chkhsh == "b3d1dd861f1d4c5c0d2569ce36baf3f90fe8a102db3de50dd71ff860d91be3df":
             # ref: https://huggingface.co/aari1995/German_Semantic_V3
             res = "jina-v2-de"
+        if chkhsh == "cdf5f35325780597efd76153d4d1c16778f766173908894c04afc20108536267":
+            # ref: https://huggingface.co/zai-org/GLM-4.7-Flash
+            res = "glm4"
         if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
             # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
             res = "llama-bpe"
@@ -2976,7 +2979,10 @@ class Llama4VisionModel(MmprojModel):
         return []
 
 
-@ModelBase.register("Mistral3ForConditionalGeneration")
+@ModelBase.register(
+    "Mistral3ForConditionalGeneration",
+    "Ministral3ForCausalLM",
+)
 class Mistral3Model(LlamaModel):
     model_arch = gguf.MODEL_ARCH.MISTRAL3
 
@@ -7458,7 +7464,7 @@ class DeepseekModel(TextModel):
     "DeepseekV3ForCausalLM",
     "KimiVLForConditionalGeneration",
     "YoutuForCausalLM",
-    "YoutuVLForConditionalGeneration"
+    "YoutuVLForConditionalGeneration",
 )
 class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
@@ -8446,6 +8452,32 @@ class Glm4MoeModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("Glm4MoeLiteForCausalLM")
+class Glm4MoeLiteModel(DeepseekV2Model):
+    model_arch = gguf.MODEL_ARCH.DEEPSEEK2
+
+    # copied from Glm4MoeModel
+    def set_vocab(self):
+        from transformers import AutoTokenizer
+
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        # Special tokens
+        # Note: Using <|endoftext|> (151329) for eot causes endless generation
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"])  # 151331
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # 151336
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"]) # 151329
+        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # 151338
+
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+
 @ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
 class ChatGLMModel(TextModel):
     model_arch = gguf.MODEL_ARCH.CHATGLM
@@ -9183,7 +9215,7 @@ class NemotronHModel(GraniteHybridModel):
                 return [(mapped_name, reshaped_data)]
 
             if name.endswith("mixer.norm.weight"):
-                reshaped_data = data_torch.reshape(8, 512)
+                reshaped_data = data_torch.reshape(self.n_group, -1)
                 mapped_name = self.map_tensor_name(name)
                 return [(mapped_name, reshaped_data)]
 

convert_hf_to_gguf_update.py

@@ -170,6 +170,7 @@ pre_computed_hashes = [
     {"name": "grok-2",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/alvarobartt/grok-2-tokenizer", "chkhsh": "66b8d4e19ab16c3bfd89bce5d785fb7e0155e8648708a1f42077cb9fe002c273"},
     # jina-v2-de variants
     {"name": "jina-v2-de", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/aari1995/German_Semantic_V3", "chkhsh": "b3d1dd861f1d4c5c0d2569ce36baf3f90fe8a102db3de50dd71ff860d91be3df"},
+    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.7-Flash", "chkhsh": "cdf5f35325780597efd76153d4d1c16778f766173908894c04afc20108536267"},
 ]
 
 

docs/backend/OPENCL.md

@@ -8,6 +8,7 @@
 - [CMake Options](#cmake-options)
 - [Android](#android)
 - [Windows 11 Arm64](#windows-11-arm64)
+- [Linux](#Linux)
 - [Known Issue](#known-issues)
 - [TODO](#todo)
 

examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh

@@ -4,6 +4,7 @@ set -e
 
 # First try command line argument, then environment variable, then file
 CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+BUILD_DIR="${2:-"$BUILD_DIR"}"
 
 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
@@ -13,6 +14,10 @@ if [ -z "$CONVERTED_MODEL" ]; then
     exit 1
 fi
 
-cmake --build ../../build --target llama-debug -j8
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
 
-../../build/bin/llama-debug -m $CONVERTED_MODEL --embedding -p "Hello world today" --save-logits
+cmake --build ${BUILD_DIR} --target llama-debug -j8
+
+${BUILD_DIR}/bin/llama-debug -m $CONVERTED_MODEL --embedding -p "Hello world today" --save-logits

examples/model-conversion/scripts/causal/run-converted-model.sh

@@ -5,11 +5,16 @@ set -e
 # First try command line argument, then environment variable, then file
 CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
 MODEL_TESTING_PROMPT="${2:-"$MODEL_TESTING_PROMPT"}"
+BUILD_DIR="${3:-"$BUILD_DIR"}"
 
-if [ -z "$MODEL_TESTING_PROMPT"]; then
+if [ -z "$MODEL_TESTING_PROMPT" ]; then
     MODEL_TESTING_PROMPT="Hello, my name is"
 fi
 
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
+
 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
     echo "Error: Model path must be provided either as:" >&2
@@ -21,6 +26,6 @@ fi
 echo $CONVERTED_MODEL
 echo $MODEL_TESTING_PROMPT
 
-cmake --build ../../build --target llama-debug -j8
+cmake --build ${BUILD_DIR} --target llama-debug -j8
 
-../../build/bin/llama-debug -m "$CONVERTED_MODEL" -p "$MODEL_TESTING_PROMPT" --save-logits
+${BUILD_DIR}/bin/llama-debug -m "$CONVERTED_MODEL" -p "$MODEL_TESTING_PROMPT" --save-logits

examples/model-conversion/scripts/embedding/run-converted-model.sh

@@ -28,6 +28,7 @@ done
 
 # First try command line argument, then environment variable
 CONVERTED_MODEL="${CONVERTED_MODEL:-"$CONVERTED_EMBEDDING_MODEL"}"
+BUILD_DIR="${BUILD_DIR:-"../../build"}"
 
 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
@@ -50,5 +51,5 @@ fi
 
 echo $CONVERTED_MODEL
 
-cmake --build ../../build --target llama-debug -j8
-../../build/bin/llama-debug -m "$CONVERTED_MODEL" --embedding -p "$PROMPT" --save-logits --embd-normalize $EMBD_NORMALIZE
+cmake --build ${BUILD_DIR} --target llama-debug -j8
+${BUILD_DIR}/bin/llama-debug -m "$CONVERTED_MODEL" --embedding -p "$PROMPT" --save-logits --embd-normalize $EMBD_NORMALIZE

ggml/include/ggml.h

@@ -630,10 +630,11 @@ extern "C" {
 
     // this tensor...
     enum ggml_tensor_flag {
-        GGML_TENSOR_FLAG_INPUT  =  1, // ...is an input for the GGML compute graph
-        GGML_TENSOR_FLAG_OUTPUT =  2, // ...is an output for the GGML compute graph
-        GGML_TENSOR_FLAG_PARAM  =  4, // ...contains trainable parameters
-        GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
+        GGML_TENSOR_FLAG_INPUT   =  1, // ...is an input for the GGML compute graph
+        GGML_TENSOR_FLAG_OUTPUT  =  2, // ...is an output for the GGML compute graph
+        GGML_TENSOR_FLAG_PARAM   =  4, // ...contains trainable parameters
+        GGML_TENSOR_FLAG_LOSS    =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
+        GGML_TENSOR_FLAG_COMPUTE = 16, // ...must be computed
     };
 
     enum ggml_tri_type {
@@ -2577,11 +2578,42 @@ extern "C" {
         struct ggml_tensor *  grad,
         struct ggml_tensor *  sgd_params); // alpha, weight decay
 
+    // build forward mutiple tensors and select one of them for computing
+    // this is useful for creating graphs that have constant topology but compute different things based on the input
+    // ref: https://github.com/ggml-org/llama.cpp/pull/18550
     //
-    // automatic differentiation
+    // nodes:
+    //   | - build forward into the graph but do not compute
+    //   c - build forward into the graph and compute
     //
+    //    |  |  ...  c  ...  |
+    //    |  |  ...  c  ...  |
+    //    |  |  ...  c  ...  |
+    //   [0  1  ... idx ...  n-1]        <-- ggml_build_forward_select(..., n, idx)
+    //               c
+    //               c
+    //
+    // example:
+    //   struct ggml_tensor * curs[3];
+    //
+    //   curs[0]  = compute0(...);
+    //   curs[1]  = compute1(...);
+    //   curs[2]  = compute2(...);
+    //
+    //   int idx = select_branch(some_input);
+    //
+    //   struct ggml_tensor * out = ggml_build_forward_select(cgraph, curs, 3, idx);
+    //
+    GGML_API struct ggml_tensor * ggml_build_forward_select(
+            struct ggml_cgraph  * cgraph,
+            struct ggml_tensor ** tensors,
+            int                   n_tensors,
+            int                   idx);
+
+    GGML_API void ggml_build_forward_expand(
+            struct ggml_cgraph * cgraph,
+            struct ggml_tensor * tensor);
 
-    GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
     GGML_API void ggml_build_backward_expand(
         struct ggml_context *  ctx,        // context for gradient computation
         struct ggml_cgraph  *  cgraph,
@@ -2613,7 +2645,7 @@ extern "C" {
     GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
 
     // dump the graph into a file using the dot format
-    GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
+    GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * cgraph, const char * filename);
 
     // TODO these functions were sandwiched in the old optimization interface, is there a better place for them?
     typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);

ggml/src/ggml-backend-reg.cpp

@@ -77,39 +77,23 @@
 #include "ggml-zendnn.h"
 #endif
 
-// disable C++17 deprecation warning for std::codecvt_utf8
-#if defined(__clang__)
-#    pragma clang diagnostic push
-#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
-#endif
-
 namespace fs = std::filesystem;
 
 static std::string path_str(const fs::path & path) {
-    std::string u8path;
     try {
 #if defined(__cpp_lib_char8_t)
         // C++20 and later: u8string() returns std::u8string
-        std::u8string u8str = path.u8string();
-        u8path = std::string(reinterpret_cast<const char*>(u8str.c_str()));
+        const std::u8string u8str = path.u8string();
+        return std::string(reinterpret_cast<const char *>(u8str.data()), u8str.size());
 #else
         // C++17: u8string() returns std::string
-        u8path = path.u8string();
+        return path.u8string();
 #endif
     } catch (...) {
+        return std::string();
     }
-    return u8path;
 }
 
-#if defined(__clang__)
-#    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
-#endif
-
 #ifdef _WIN32
 
 using dl_handle = std::remove_pointer_t<HMODULE>;

ggml/src/ggml-backend.cpp

@@ -874,9 +874,9 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
         }
         if (sched->debug > 1) {
             ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
-            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s] use=%d:", i, ggml_op_name(node->op), node->name,
+            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s] use=%d,c=%d:", i, ggml_op_name(node->op), node->name,
                 fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node),
-                graph->use_counts[ggml_hash_find(&graph->visited_hash_set, node)]);
+                graph->use_counts[ggml_hash_find(&graph->visited_hash_set, node)], node->flags & GGML_TENSOR_FLAG_COMPUTE ? 1 : 0);
             for (int j = 0; j < GGML_MAX_SRC; j++) {
                 struct ggml_tensor * src = node->src[j];
                 if (src == NULL) {
@@ -1922,6 +1922,7 @@ static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set,
         dst->view_offs = src->view_offs;
     }
     dst->op = src->op;
+    dst->flags = src->flags;
     memcpy(dst->op_params, src->op_params, sizeof(dst->op_params));
     ggml_set_name(dst, src->name);
 

ggml/src/ggml-blas/ggml-blas.cpp

@@ -226,6 +226,10 @@ static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend,
     for (int i = 0; i < cgraph->n_nodes; i++) {
         struct ggml_tensor * node = cgraph->nodes[i];
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         switch (node->op) {
             case GGML_OP_MUL_MAT:
                 ggml_backend_blas_mul_mat(ctx, node);

ggml/src/ggml-cann/ggml-cann.cpp

@@ -2146,6 +2146,10 @@ static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx
                 continue;
             }
 
+            if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+                continue;
+            }
+
             bool ok = ggml_cann_compute_forward(*cann_ctx, node);
             if (!ok) {
                 GGML_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));

ggml/src/ggml-cpu/ggml-cpu.c

@@ -2943,6 +2943,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
             continue;
         }
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         ggml_compute_forward(&params, node);
 
         if (state->ith == 0 && cplan->abort_callback &&

ggml/src/ggml-cuda/argsort.cu

@@ -2,6 +2,9 @@
 
 #ifdef GGML_CUDA_USE_CUB
 #    include <cub/cub.cuh>
+#    if (CCCL_MAJOR_VERSION >= 3 && CCCL_MINOR_VERSION >= 1)
+#        define STRIDED_ITERATOR_AVAILABLE
+#    endif
 using namespace cub;
 #endif  // GGML_CUDA_USE_CUB
 
@@ -14,12 +17,14 @@ static __global__ void init_indices(int * indices, const int ncols, const int nr
     }
 }
 
+#ifndef STRIDED_ITERATOR_AVAILABLE
 static __global__ void init_offsets(int * offsets, const int ncols, const int nrows) {
     const int idx = blockIdx.x * blockDim.x + threadIdx.x;
     if (idx <= nrows) {
         offsets[idx] = idx * ncols;
     }
 }
+#endif  // STRIDED_ITERATOR_AVAILABLE
 
 #ifdef GGML_CUDA_USE_CUB
 void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
@@ -31,19 +36,22 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
                               cudaStream_t     stream) {
     ggml_cuda_pool_alloc<int>   temp_indices_alloc(pool, ncols * nrows);
     ggml_cuda_pool_alloc<float> temp_keys_alloc(pool, ncols * nrows);
-    ggml_cuda_pool_alloc<int>   offsets_alloc(pool, nrows + 1);
 
     int *   temp_indices = temp_indices_alloc.get();
     float * temp_keys    = temp_keys_alloc.get();
-    int *   d_offsets    = offsets_alloc.get();
 
     static const int block_size = 256;
     const dim3 grid_size((ncols + block_size - 1) / block_size, nrows);
     init_indices<<<grid_size, block_size, 0, stream>>>(temp_indices, ncols, nrows);
 
-    const dim3 offset_grid((nrows + block_size - 1) / block_size);
-    init_offsets<<<offset_grid, block_size, 0, stream>>>(d_offsets, ncols, nrows);
-
+#ifdef STRIDED_ITERATOR_AVAILABLE
+    auto offset_iterator = cuda::make_strided_iterator(cuda::make_counting_iterator(0), ncols);
+#else
+    ggml_cuda_pool_alloc<int> offsets_alloc(pool, nrows + 1);
+    int *                     offset_iterator = offsets_alloc.get();
+    const dim3                offset_grid((nrows + block_size - 1) / block_size);
+    init_offsets<<<offset_grid, block_size, 0, stream>>>(offset_iterator, ncols, nrows);
+#endif
     CUDA_CHECK(cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream));
 
     size_t temp_storage_bytes = 0;
@@ -57,7 +65,7 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
             DeviceSegmentedSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
                                            temp_indices, dst,                                  // values (indices)
                                            ncols * nrows, nrows,  // num items, num segments
-                                           d_offsets, d_offsets + 1, stream);
+                                           offset_iterator, offset_iterator + 1, stream);
         }
     } else {
         if (nrows == 1) {
@@ -66,7 +74,8 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
                                                  ncols, 0, sizeof(float) * 8, stream);
         } else {
             DeviceSegmentedSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys, temp_indices,
-                                                     dst, ncols * nrows, nrows, d_offsets, d_offsets + 1, stream);
+                                                     dst, ncols * nrows, nrows, offset_iterator, offset_iterator + 1,
+                                                     stream);
         }
     }
 
@@ -80,7 +89,7 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
                                        ncols, 0, sizeof(float) * 8, stream);
         } else {
             DeviceSegmentedSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst,
-                                           ncols * nrows, nrows, d_offsets, d_offsets + 1, stream);
+                                           ncols * nrows, nrows, offset_iterator, offset_iterator + 1, stream);
         }
     } else {
         if (nrows == 1) {
@@ -89,8 +98,8 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
                                                  ncols, 0, sizeof(float) * 8, stream);
         } else {
             DeviceSegmentedSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
-                                                     temp_indices, dst, ncols * nrows, nrows, d_offsets, d_offsets + 1,
-                                                     stream);
+                                                     temp_indices, dst, ncols * nrows, nrows, offset_iterator,
+                                                     offset_iterator + 1, stream);
         }
     }
 }

ggml/src/ggml-cuda/common.cuh

@@ -1123,6 +1123,7 @@ struct ggml_tensor_extra_gpu {
 struct ggml_cuda_graph_node_properties {
     void * node_address;
     ggml_op node_op;
+    int32_t flags;
     int64_t ne[GGML_MAX_DIMS];
     size_t nb[GGML_MAX_DIMS];
     void * src_address[GGML_MAX_SRC];

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -432,7 +432,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     constexpr int  ncols           = ncols1 * ncols2;
     constexpr int  cols_per_warp   = T_B_KQ::I;
     constexpr int  cols_per_thread = get_cols_per_thread();
-    constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
+    constexpr int  np              = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa(DKQ, DV, ncols);
     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2(DKQ, DV, ncols);
     constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2(DKQ, DV, ncols);
@@ -510,7 +510,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 }
             }
         } else {
-            static_assert(cols_per_warp != 8, "cols_per_warp == 8 not implemented");
 #pragma unroll
             for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += T_A_KQ::J) {
                 load_ldmatrix(Q_B[0], tile_Q + (threadIdx.y / np)*(T_B_KQ::I*stride_tile_Q) + k_KQ_0, stride_tile_Q);
@@ -522,14 +521,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                     T_A_KQ K_A;
                     load_ldmatrix(K_A, tile_K + i_KQ_0*stride_tile_K + (k_KQ_0 - k0_start), stride_tile_K);
 
-                    // Wide version of KQ_C is column-major
+                    if constexpr (cols_per_warp == 8) {
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
+                    } else {
+                        // Wide version of KQ_C is column-major
 #if defined(AMD_WMMA_AVAILABLE)
-                    // RDNA matrix C is column-major.
-                    mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
+                        // RDNA matrix C is column-major.
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
 #else
-                    // swap A and B for CUDA.
-                    mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
+                        // swap A and B for CUDA.
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
 #endif // defined(AMD_WMMA_AVAILABLE)
+                    }
                 }
             }
         }
@@ -953,7 +956,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
 
     constexpr int  cols_per_warp   = T_B_KQ::I;
     constexpr int  cols_per_thread = get_cols_per_thread();
-    constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
+    constexpr int  np              = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa     (DKQ, DV, ncols);
     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2     (DKQ, DV, ncols);
     constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2     (DKQ, DV, ncols);
@@ -1484,6 +1487,13 @@ static __global__ void flash_attn_ext_f16(
         NO_DEVICE_CODE;
         return;
     }
+#ifdef VOLTA_MMA_AVAILABLE
+    if (ncols1*ncols2 < 32) {
+        NO_DEVICE_CODE;
+        return;
+    }
+#endif // VOLTA_MMA_AVAILABLE
+
 #if __CUDA_ARCH__ == GGML_CUDA_CC_TURING
     if (ncols1*ncols2 > 32) {
         NO_DEVICE_CODE;
@@ -1728,3 +1738,8 @@ DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256,  64)
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16);
+
+// For GLM 4.7 Flash
+extern DECL_FATTN_MMA_F16_CASE(576, 512,  4,  4);
+extern DECL_FATTN_MMA_F16_CASE(576, 512,  8,  4);
+extern DECL_FATTN_MMA_F16_CASE(576, 512, 16,  4);

ggml/src/ggml-cuda/fattn-tile.cuh

@@ -68,6 +68,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
 
     return 0;
@@ -122,6 +124,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  32,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  32,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  32,  64)
 
     return 0;
@@ -183,6 +187,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32, 128)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128,  64)
 
@@ -245,6 +251,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5,  32, 256)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3,  64, 128)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128,  64)
 
@@ -1187,6 +1195,10 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
             launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
             return;
         }
+        if (use_gqa_opt && gqa_ratio % 4 == 0) {
+            launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
+            return;
+        }
     }
 
     if constexpr (DV <= 256) {

ggml/src/ggml-cuda/fattn.cu

@@ -121,8 +121,12 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
 
             GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
             const int gqa_ratio = Q->ne[2] / K->ne[2];
-            GGML_ASSERT(gqa_ratio % 16 == 0);
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+            GGML_ASSERT(gqa_ratio % 4 == 0);
+            if (gqa_ratio % 16 == 0) {
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+            } else {
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512,  4>(ctx, dst);
+            }
         } break;
         default:
             GGML_ABORT("fatal error");
@@ -262,7 +266,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
             if (V->ne[0] != 512) {
                 return BEST_FATTN_KERNEL_NONE;
             }
-            if (!gqa_opt_applies || gqa_ratio % 16 != 0) {
+            if (!gqa_opt_applies || gqa_ratio % 4 != 0) {
                 return BEST_FATTN_KERNEL_NONE;
             }
             break;

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

@@ -2918,6 +2918,7 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
 static void ggml_cuda_graph_node_set_properties(ggml_cuda_graph_node_properties * props, ggml_tensor * node) {
     props->node_address = node->data;
     props->node_op = node->op;
+    props->flags = node->flags;
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
         props->ne[i] = node->ne[i];
         props->nb[i] = node->nb[i];
@@ -2961,6 +2962,10 @@ static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_
         return false;
     }
 
+    if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) != (props->flags & GGML_TENSOR_FLAG_COMPUTE)) {
+        return false;
+    }
+
     return true;
 }
 
@@ -3378,6 +3383,9 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                     continue;
                 }
 
+                if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+                    continue;
+                }
 
                 // start of fusion operations
                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 16, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 2, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 4, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 8, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 8, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 8, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 8, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 8, 4);

ggml/src/ggml-cuda/template-instances/generate_cu_files.py

@@ -85,7 +85,7 @@ for ncols in [8, 16, 32, 64]:
                     continue
                 if head_size_kq != 576 and ncols2 == 16:
                     continue
-                if head_size_kq == 576 and ncols2 != 16:
+                if head_size_kq == 576 and ncols2 not in (4, 16):
                     continue
                 head_size_v = head_size_kq if head_size_kq != 576 else 512
                 f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size_kq=head_size_kq, head_size_v=head_size_v))

ggml/src/ggml-cuda/top-k.cu

@@ -4,7 +4,6 @@
 #ifdef GGML_CUDA_USE_CUB
 #    include <cub/cub.cuh>
 #    if (CCCL_MAJOR_VERSION >= 3 && CCCL_MINOR_VERSION >= 2)
-#        include <cuda/iterator>
 #        define CUB_TOP_K_AVAILABLE
 using namespace cub;
 #    endif  // CCCL_MAJOR_VERSION >= 3 && CCCL_MINOR_VERSION >= 2

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -2497,6 +2497,10 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
             continue;
         }
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         uint32_t flags = 0;
 
         // skip quantizer if src1 is reused

ggml/src/ggml-impl.h

@@ -611,6 +611,9 @@ static inline bool ggml_can_fuse_ext(const struct ggml_cgraph * cgraph, const in
         if (node->op != ops[i]) {
             return false;
         }
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            return false;
+        }
         if (i < num_ops - 1 && !ggml_node_has_n_uses(cgraph, node_idxs[i], 1)) {
             return false;
         }

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

@@ -1078,12 +1078,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 op->src[0]->ne[0] != 112 &&
                 op->src[0]->ne[0] != 128 &&
                 op->src[0]->ne[0] != 192 &&
-                op->src[0]->ne[0] != 256) {
-                return false;
-            }
-            if (op->src[0]->ne[0] == 576) {
-                // DeepSeek sizes
-                // TODO: disabled for now, until optmized
+                op->src[0]->ne[0] != 256 &&
+                op->src[0]->ne[0] != 576) {
                 return false;
             }
             if (op->src[1]->type != op->src[2]->type) {

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -203,6 +203,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
         GGML_ABORT("unsupported op");
     }
 
+    if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+        return 1;
+    }
+
     int n_fuse = 1;
 
     // check if the current node can run concurrently with other nodes before it
@@ -2516,7 +2520,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
         // simdgroups per threadgroup (a.k.a. warps)
         //nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4;
-        int32_t nsg = 4;
+        int32_t nsg = ne00 >= 512 ? 8 : 4;
 
         const size_t smem = FATTN_SMEM(nsg);
 

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

@@ -5552,9 +5552,7 @@ void kernel_flash_attn_ext_impl(
 
                 constexpr short NC = (C/8)/NSG;
 
-                // note: do not unroll for large heads
-                #pragma unroll (DK <= 64 ? NC : 1)
-                for (short cc = 0; cc < NC; ++cc) {
+                FOR_UNROLL (short cc = 0; cc < NC; ++cc) {
                     qk8x8_t mqk = make_filled_simdgroup_matrix<qk_t, 8>((qk_t) 0.0f);
 
                     if (DK % 16 != 0) {
@@ -5575,7 +5573,9 @@ void kernel_flash_attn_ext_impl(
                         k8x8_t mk[2];
                         q8x8_t mq[2];
 
-                        FOR_UNROLL (short i = 0; i < DK8/2; ++i) {
+                        // note: too much unroll can tank the performance for large heads
+                        #pragma unroll (MIN(DK8/2, 4*NSG))
+                        for (short i = 0; i < DK8/2; ++i) {
                             simdgroup_barrier(mem_flags::mem_none);
 
                             simdgroup_load(mq[0], pq + 0*8 + 16*i, DK);
@@ -5749,7 +5749,9 @@ void kernel_flash_attn_ext_impl(
                                 pv  += 8*NS20;
                             }
                         } else {
-                            FOR_UNROLL (short cc = 0; cc < (C/8)/2; ++cc) {
+                            constexpr short NC = (C/8)/2;
+
+                            FOR_UNROLL (short cc = 0; cc < NC; ++cc) {
                                 s8x8_t vs[2];
 
                                 simdgroup_load(vs[0], ss + 16*cc + 0, SH, 0, false);
@@ -5952,6 +5954,7 @@ kernel void kernel_flash_attn_ext(
       //case 1: kernel_flash_attn_ext_impl<FWD_TMPL, 1>(FWD_ARGS); break;
       //case 2: kernel_flash_attn_ext_impl<FWD_TMPL, 2>(FWD_ARGS); break;
         case 4: kernel_flash_attn_ext_impl<FWD_TMPL, 4>(FWD_ARGS); break;
+        case 8: kernel_flash_attn_ext_impl<FWD_TMPL, 8>(FWD_ARGS); break;
     }
 #undef FWD_TMPL
 #undef FWD_ARGS

ggml/src/ggml-opencl/CMakeLists.txt

@@ -57,6 +57,7 @@ set(GGML_OPENCL_KERNELS
     add
     add_id
     argsort
+    tri
     fill
     clamp
     cpy

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -489,6 +489,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_gelu_quick, kernel_gelu_quick_4;
     cl_kernel kernel_relu;
     cl_kernel kernel_sigmoid_f32, kernel_sigmoid_f16;
+    cl_kernel kernel_tri;
     cl_kernel kernel_fill;
     cl_kernel kernel_clamp;
     cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu, kernel_swiglu_oai, kernel_geglu_erf, kernel_geglu_quick,
@@ -793,6 +794,24 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // tri
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tri.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("tri.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_tri = clCreateKernel(prog, "kernel_tri_f32", &err), err));
+        GGML_LOG_CONT(".");
+
+        CL_CHECK(clReleaseProgram(prog));
+    }
+
     // fill
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -3058,6 +3077,10 @@ static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggm
             continue;
         }
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         if (!backend_ctx->disable_fusion && ggml_opencl_can_fuse(cgraph, i, { GGML_OP_NORM, GGML_OP_MUL, GGML_OP_ADD })) {
             ggml_opencl_op_norm_fused(backend, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
             i += 2;
@@ -3201,6 +3224,8 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 default:
                     return false;
             }
+        case GGML_OP_TRI:
+            return op->type == GGML_TYPE_F32 && ggml_is_contiguous(op);
         case GGML_OP_FILL:
             return op->type == GGML_TYPE_F32 && ggml_is_contiguous(op);
         case GGML_OP_CLAMP:
@@ -5961,6 +5986,44 @@ static void ggml_cl_sigmoid(ggml_backend_t backend, const ggml_tensor * src0, co
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
 }
 
+static void ggml_cl_tri(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int tri_type = ggml_get_op_params_i32(dst, 0);
+    const int64_t n = ggml_nelements(dst);
+    const int     ne0  = dst->ne[0];
+    const int     ne1  = dst->ne[1];
+
+    cl_kernel kernel = backend_ctx->kernel_tri;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &n));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne1));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &tri_type));
+
+    size_t local_work_size[1] = { 256 };
+    size_t global_work_size[1] = { ((size_t)n + local_work_size[0] - 1) / local_work_size[0] * local_work_size[0] };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_fill(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
@@ -10008,6 +10071,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_glu;
             break;
+        case GGML_OP_TRI:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_tri;
+            break;
         case GGML_OP_FILL:
             if (!any_on_device) {
                 return false;

ggml/src/ggml-opencl/kernels/tri.cl

@@ -0,0 +1,32 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// tri
+//------------------------------------------------------------------------------
+__kernel void kernel_tri_f32(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int n,
+        int ne0,
+        int ne1,
+        int tri_type
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int idx = get_global_id(0);
+    if (idx >= n) return;
+
+    int i0 = idx % ne0;
+    int i1 = (idx / ne0) % ne1;
+
+    int keep = 0;
+    if (tri_type == 0) keep = (i0 >= i1);
+    else if (tri_type == 1) keep = (i0 >  i1);
+    else if (tri_type == 2) keep = (i0 <= i1);
+    else                    keep = (i0 <  i1);
+
+    dst[idx] = keep ? src0[idx] : 0.0f;
+}

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -4109,6 +4109,9 @@ static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * syc
         if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
             continue;
         }
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
 #ifndef NDEBUG
         assert(node->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device));
         for (int j = 0; j < GGML_MAX_SRC; j++) {

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

@@ -991,6 +991,8 @@ struct vk_mat_vec_id_push_constants {
     uint32_t fusion_flags;
     uint32_t nei0;
     uint32_t ne11;
+    uint32_t expert_i1;
+    uint32_t nbi1;
 };
 
 struct vk_flash_attn_push_constants {
@@ -1516,6 +1518,15 @@ struct vk_quantize_q8_1_push_constants {
     uint32_t num_blocks;
 };
 
+struct vk_op_flash_attn_split_k_reduce_push_constants {
+    uint32_t D;
+    uint32_t ne1;
+    uint32_t ne2;
+    uint32_t ne3;
+    uint32_t k_num;
+    uint32_t sinks;
+};
+
 // Allow pre-recording command buffers
 struct vk_staging_memcpy {
     vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
@@ -1802,7 +1813,6 @@ struct ggml_backend_vk_context {
     bool prealloc_x_need_sync, prealloc_y_need_sync, prealloc_split_k_need_sync;
 
     vk_context_ref compute_ctx;
-    vk_context_ref transfer_ctx;
 
     std::vector<vk_context_ref> tensor_ctxs;
 
@@ -1812,7 +1822,6 @@ struct ggml_backend_vk_context {
     uint32_t pipeline_descriptor_set_requirements {};
 
     vk_command_pool compute_cmd_pool;
-    vk_command_pool transfer_cmd_pool;
 
     // number of additional consecutive nodes that are being fused with the
     // node currently being processed
@@ -3178,15 +3187,15 @@ static void ggml_vk_load_shaders(vk_device& device) {
             if (path == FAPATH) { \
                 if (aligned) { \
                     if (f32acc) { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
                     } else { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
                     } \
                 } else { \
                     if (f32acc) { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1,                                        true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1,                                        true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
                     } else { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1,                                        true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1,                                        true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
                     } \
                 } \
             } \
@@ -3980,7 +3989,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_MXFP4],   "get_rows_mxfp4_f32",   get_rows_mxfp4_f32_len,   get_rows_mxfp4_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, 5 * sizeof(uint32_t), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, sizeof(vk_op_flash_attn_split_k_reduce_push_constants), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
 
     if (device->subgroup_clustered && device->subgroup_require_full_support) {
         ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_subgroup_len, quantize_q8_1_x4_subgroup_data, "main", 2, sizeof(vk_quantize_q8_1_push_constants), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
@@ -5647,7 +5656,6 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     ctx->almost_ready_fence = ctx->device->device.createFence({});
 
     ctx->compute_cmd_pool.init(ctx->device, &ctx->device->compute_queue);
-    ctx->transfer_cmd_pool.init(ctx->device, &ctx->device->transfer_queue);
 
     if (vk_perf_logger_enabled) {
         ctx->perf_logger = std::unique_ptr<vk_perf_logger>(new vk_perf_logger());
@@ -8083,8 +8091,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
 
     const uint64_t nei0 = ids->ne[0];
     const uint64_t nei1 = ids->ne[1];
-
-    GGML_ASSERT(nei1 == 1);
+    const uint32_t nbi1 = (uint32_t)(ids->nb[1] / sizeof(int));
 
     const uint64_t ne20 = dst->ne[0];
     const uint64_t ne21 = dst->ne[1];
@@ -8168,7 +8175,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
         if (quantize_y) {
             ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
         }
-        ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, dmmv, nei1);
     }
 
     vk_subbuffer d_D = ggml_vk_tensor_subbuffer(ctx, cgraph->nodes[node_idx + ctx->num_additional_fused_ops]);
@@ -8226,7 +8233,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     uint32_t stride_batch_y = ne10*ne11;
 
     if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant) {
-        stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
+        stride_batch_y = src1->nb[2] / ggml_type_size(src1->type);
     }
 
     const uint32_t max_groups_x = ctx->device->properties.limits.maxComputeWorkGroupCount[0];
@@ -8262,23 +8269,25 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
         fusion_flags |= MAT_VEC_FUSION_FLAGS_SCALE1;
     }
 
-    // compute
-    const vk_mat_vec_id_push_constants pc = {
-        (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
-        (uint32_t)(ne00 * ne01), stride_batch_y, (uint32_t)(ne20 * ne21),
-        fusion_flags,
-        (uint32_t)nei0, (uint32_t)ne11,
-    };
-    ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
-        {
-            d_X,
-            d_Y,
-            d_D,
-            d_F0,
-            d_F1,
-            d_ids,
-        },
-        pc, { groups_x, (uint32_t)nei0, groups_z });
+    // Loop over the batch dimension
+    for (uint32_t expert_i1 = 0; expert_i1 < nei1; ++expert_i1) {
+        const vk_mat_vec_id_push_constants pc = {
+            (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
+            (uint32_t)(ne00 * ne01), stride_batch_y, (uint32_t)(ne20 * ne21),
+            fusion_flags,
+            (uint32_t)nei0, (uint32_t)ne11, expert_i1, nbi1
+        };
+        ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
+            {
+                d_X,
+                d_Y,
+                d_D,
+                d_F0,
+                d_F1,
+                d_ids,
+            },
+            pc, { groups_x, (uint32_t)nei0, groups_z });
+    }
 
     if (x_non_contig) {
         ctx->prealloc_x_need_sync = true;
@@ -8292,7 +8301,7 @@ static bool ggml_vk_use_mul_mat_vec_id(const struct ggml_cgraph * cgraph, int no
     ggml_tensor * dst = cgraph->nodes[node_idx];
     ggml_tensor * src0 = dst->src[0];
     ggml_tensor * src2 = dst->src[2];
-    return src2->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type));
+    return (src2->ne[1] <= 8) && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type));
 }
 
 static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
@@ -8454,14 +8463,14 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         GGML_ASSERT(0);
     }
 
-    if (N == 1 && qk_ratio > 1 && qk_ratio <= max_gqa &&
+    if (N <= 8 && qk_ratio > 1 && qk_ratio <= max_gqa &&
         qk_ratio * nek2 == neq2 && nek2 == nev2 && nem2 <= 1) {
         // grouped query attention - make the N dimension equal to gqa_ratio, reduce
         // workgroups proportionally in y dimension. The shader will detect gqa_ratio > 1
         // and change addressing calculations to index Q's dimension 2.
         gqa_ratio = qk_ratio;
         N = gqa_ratio;
-        workgroups_y /= N;
+        workgroups_y /= gqa_ratio;
     }
 
     bool small_rows = N <= get_fa_num_small_rows(path);
@@ -8523,6 +8532,8 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     }
 
     assert(pipeline);
+    // Compile early to initialize wg_denoms.
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
 
     uint32_t split_kv = KV;
     uint32_t split_k = 1;
@@ -8530,22 +8541,24 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     // Use a placeholder core count if one isn't available. split_k is a big help for perf.
     const uint32_t shader_core_count = ctx->device->shader_core_count ? ctx->device->shader_core_count : 16;
 
-    // Try to use split_k when KV is large enough to be worth the overhead
-    if (workgroups_x == 1 && shader_core_count > 0) {
+    // Try to use split_k when KV is large enough to be worth the overhead.
+    // Must either be a single batch or be using gqa, we can't mix the two.
+    if (workgroups_x <= pipeline->wg_denoms[0] && (workgroups_x == 1 || gqa_ratio > 1)) {
         // Try to run two workgroups per SM.
-        split_k = shader_core_count * 2 / (workgroups_y * workgroups_z);
+        split_k = shader_core_count * 2 / (workgroups_x * workgroups_y * workgroups_z);
         if (split_k > 1) {
             // Try to evenly split KV into split_k chunks, but it needs to be a multiple
             // of "align", so recompute split_k based on that.
             split_kv = ROUNDUP_POW2(std::max(1u, KV / split_k), alignment);
             split_k = CEIL_DIV(KV, split_kv);
-            workgroups_x = split_k;
         }
     }
 
     // Reserve space for split_k temporaries. For each split x batch, we need to store the O matrix (D x ne1)
     // and the per-row m and L values (ne1 rows). We store all the matrices first, followed by the rows.
-    const uint64_t split_k_size = split_k > 1 ? (HSV * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k * ne3 : 0;
+    // For matrices, the order is (inner to outer) [HSV, ne1, k, ne2, ne3].
+    // For L/M, the order is (inner to outer) [ne1, k, ne2, ne3].
+    const uint64_t split_k_size = split_k > 1 ? (HSV * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k * ne2 * ne3 : 0;
     if (split_k_size > ctx->device->properties.limits.maxStorageBufferRange) {
         GGML_ABORT("Requested preallocation size is too large");
     }
@@ -8556,7 +8569,6 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
 
     {
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         if (split_k > 1) {
             ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_flash_attn_split_k_reduce, 1);
         }
@@ -8605,7 +8617,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         if (ctx->prealloc_split_k_need_sync) {
             ggml_vk_sync_buffers(ctx, subctx);
         }
-
+        workgroups_x *= pipeline->wg_denoms[0];
         vk_subbuffer split_k_buf = ggml_vk_subbuffer(ctx, ctx->prealloc_split_k, 0);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                     {q_buf, k_buf, v_buf, mask_buf, sinks_buf, split_k_buf},
@@ -8613,15 +8625,19 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                     // there's no more than one tile of rows (i.e. workgroups_x would have been
                                     // one). We reuse workgroups_x to mean the number of splits, so we need to
                                     // cancel out the divide by wg_denoms[0].
-                                    pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
+                                    pc, { split_k * workgroups_x, workgroups_y, workgroups_z });
 
         ggml_vk_sync_buffers(ctx, subctx);
-        const std::array<uint32_t, 5> pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne3, split_k, (sinks != nullptr) };
+        const vk_op_flash_attn_split_k_reduce_push_constants pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3, split_k, (sinks != nullptr) };
         ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_flash_attn_split_k_reduce,
                                     {split_k_buf, sinks_buf, dst_buf},
-                                    pc2, { (uint32_t)ne1, HSV, (uint32_t)ne3 });
+                                    pc2, { (uint32_t)ne1, HSV, (uint32_t)(ne2 * ne3) });
         ctx->prealloc_split_k_need_sync = true;
     } else {
+        if (gqa_ratio > 1) {
+            // When using gqa, we want one actual workgroup per batch, so cancel out wg_denoms
+            workgroups_x *= pipeline->wg_denoms[0];
+        }
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                     {q_buf, k_buf, v_buf, mask_buf, sinks_buf, dst_buf},
                                     pc, { workgroups_x, workgroups_y, workgroups_z });
@@ -11560,7 +11576,6 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
     free(d_chk);
 
     ggml_vk_command_pool_cleanup(ctx->device, ctx->compute_cmd_pool);
-    ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool);
 
     ggml_vk_destroy_buffer(d_X);
     ggml_vk_destroy_buffer(d_Y);
@@ -12145,7 +12160,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_contex
         ggml_vk_submit(subctx, {});
         ctx->submit_pending = true;
         ggml_vk_synchronize(ctx);
+        GGML_ASSERT(ctx->compute_ctx.expired());
         ggml_vk_ctx_begin(ctx->device, subctx);
+        ctx->compute_ctx = subctx;
     }
 
     if (ctx->prealloc_x == nullptr || (ctx->prealloc_size_x > 0 && ctx->prealloc_x->size < ctx->prealloc_size_x)) {
@@ -12163,6 +12180,7 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_contex
             ggml_vk_destroy_buffer(ctx->prealloc_y);
         }
         ctx->prealloc_y = ggml_vk_create_buffer_device(ctx->device, ctx->prealloc_size_y);
+        ctx->prealloc_y_last_tensor_used = nullptr;
     }
     if (ctx->prealloc_split_k == nullptr || (ctx->prealloc_size_split_k > 0 && ctx->prealloc_split_k->size < ctx->prealloc_size_split_k)) {
         VK_LOG_MEMORY("ggml_vk_preallocate_buffers(split_k_size: " << ctx->prealloc_size_split_k << ")");
@@ -12191,6 +12209,9 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     if (ggml_is_empty(node) || ggml_op_is_empty(node->op) || !node->buffer) {
         return false;
     }
+    if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+        return false;
+    }
 
     VK_LOG_DEBUG("ggml_vk_build_graph(" << node << ", " << ggml_op_name(node->op) << ")");
     ctx->semaphore_idx = 0;
@@ -12740,7 +12761,6 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
     ctx->prealloc_x_need_sync = ctx->prealloc_y_need_sync = ctx->prealloc_split_k_need_sync = false;
 
     ggml_vk_command_pool_cleanup(ctx->device, ctx->compute_cmd_pool);
-    ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool);
 
     for (size_t i = 0; i < ctx->gc.semaphores.size(); i++) {
         ctx->device->device.destroySemaphore({ ctx->gc.semaphores[i].s });
@@ -12769,7 +12789,7 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
 static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
     VK_LOG_DEBUG("ggml_vk_cleanup(" << ctx->name << ")");
     // discard any unsubmitted command buffers
-    ctx->transfer_ctx.reset();
+    ctx->compute_ctx.reset();
     // wait for any pending command buffers to finish
     ggml_vk_synchronize(ctx);
 
@@ -12802,7 +12822,6 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
     ctx->descriptor_sets.clear();
 
     ctx->compute_cmd_pool.destroy(ctx->device->device);
-    ctx->transfer_cmd_pool.destroy(ctx->device->device);
     if (vk_perf_logger_enabled) {
         ctx->perf_logger->print_timings(true);
     }
@@ -13074,34 +13093,34 @@ static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, ggml_tensor
 
     ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context;
 
-    vk_context transfer_ctx;
+    vk_context compute_ctx;
 
-    if (ctx->transfer_ctx.expired()) {
+    if (ctx->compute_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
-        ctx->transfer_ctx = transfer_ctx;
-        ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+        compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+        ctx->compute_ctx = compute_ctx;
+        ggml_vk_ctx_begin(ctx->device, compute_ctx);
     } else {
-        transfer_ctx = ctx->transfer_ctx.lock();
+        compute_ctx = ctx->compute_ctx.lock();
     }
 
     vk_buffer buf = buf_ctx->dev_buffer;
 
     auto dst_offset = vk_tensor_offset(tensor) + tensor->view_offs + offset;
 
-    bool ret = ggml_vk_buffer_write_async(transfer_ctx, buf, dst_offset, data, size);
+    bool ret = ggml_vk_buffer_write_async(compute_ctx, buf, dst_offset, data, size);
 
     if (!ret) {
         ggml_vk_ensure_sync_staging_buffer(ctx, size);
-        ggml_vk_sync_buffers(nullptr, transfer_ctx);
+        ggml_vk_sync_buffers(nullptr, compute_ctx);
 
         vk::BufferCopy buffer_cpy;
         buffer_cpy.srcOffset = 0;
         buffer_cpy.dstOffset = dst_offset;
         buffer_cpy.size = size;
 
-        transfer_ctx->s->buffer.copyBuffer(ctx->sync_staging->buffer, buf->buffer, { buffer_cpy });
-        deferred_memcpy(ctx->sync_staging->ptr, data, size, &transfer_ctx->in_memcpys);
+        compute_ctx->s->buffer.copyBuffer(ctx->sync_staging->buffer, buf->buffer, { buffer_cpy });
+        deferred_memcpy(ctx->sync_staging->ptr, data, size, &compute_ctx->in_memcpys);
         ggml_vk_synchronize(ctx);
     }
 }
@@ -13113,34 +13132,34 @@ static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_
 
     ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context;
 
-    vk_context transfer_ctx;
+    vk_context compute_ctx;
 
-    if (ctx->transfer_ctx.expired()) {
+    if (ctx->compute_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
-        ctx->transfer_ctx = transfer_ctx;
-        ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+        compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+        ctx->compute_ctx = compute_ctx;
+        ggml_vk_ctx_begin(ctx->device, compute_ctx);
     } else {
-        transfer_ctx = ctx->transfer_ctx.lock();
+        compute_ctx = ctx->compute_ctx.lock();
     }
 
     vk_buffer buf = buf_ctx->dev_buffer;
 
     auto src_offset = vk_tensor_offset(tensor) + tensor->view_offs + offset;
-    bool ret = ggml_vk_buffer_read_async(transfer_ctx, buf, src_offset, data, size);
+    bool ret = ggml_vk_buffer_read_async(compute_ctx, buf, src_offset, data, size);
 
     // If that failed, copy synchronously through a staging buffer
     if (!ret) {
         ggml_vk_ensure_sync_staging_buffer(ctx, size);
-        ggml_vk_sync_buffers(nullptr, transfer_ctx);
+        ggml_vk_sync_buffers(nullptr, compute_ctx);
 
         vk::BufferCopy buffer_cpy;
         buffer_cpy.srcOffset = src_offset;
         buffer_cpy.dstOffset = 0;
         buffer_cpy.size = size;
 
-        transfer_ctx->s->buffer.copyBuffer(buf->buffer, ctx->sync_staging->buffer, { buffer_cpy });
-        deferred_memcpy(data, ctx->sync_staging->ptr, size, &transfer_ctx->out_memcpys);
+        compute_ctx->s->buffer.copyBuffer(buf->buffer, ctx->sync_staging->buffer, { buffer_cpy });
+        deferred_memcpy(data, ctx->sync_staging->ptr, size, &compute_ctx->out_memcpys);
         ggml_vk_synchronize(ctx);
     }
 }
@@ -13152,21 +13171,21 @@ static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_
         ggml_backend_vk_buffer_context * src_buf_ctx = (ggml_backend_vk_buffer_context *)src->buffer->context;
         ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
 
-        vk_context transfer_ctx;
+        vk_context compute_ctx;
 
-        if (ctx->transfer_ctx.expired()) {
+        if (ctx->compute_ctx.expired()) {
             // Initialize new transfer context
-            transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
-            ctx->transfer_ctx = transfer_ctx;
-            ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+            compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+            ctx->compute_ctx = compute_ctx;
+            ggml_vk_ctx_begin(ctx->device, compute_ctx);
         } else {
-            transfer_ctx = ctx->transfer_ctx.lock();
+            compute_ctx = ctx->compute_ctx.lock();
         }
 
         vk_buffer src_buf = src_buf_ctx->dev_buffer;
         vk_buffer dst_buf = dst_buf_ctx->dev_buffer;
 
-        ggml_vk_buffer_copy_async(transfer_ctx, dst_buf, vk_tensor_offset(dst) + dst->view_offs, src_buf, vk_tensor_offset(src) + src->view_offs, ggml_nbytes(src));
+        ggml_vk_buffer_copy_async(compute_ctx, dst_buf, vk_tensor_offset(dst) + dst->view_offs, src_buf, vk_tensor_offset(src) + src->view_offs, ggml_nbytes(src));
         return true;
     }
 
@@ -13176,19 +13195,19 @@ static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_
 static void ggml_vk_synchronize(ggml_backend_vk_context * ctx) {
     VK_LOG_DEBUG("ggml_vk_synchronize()");
 
-    bool do_transfer = !ctx->transfer_ctx.expired();
+    bool do_transfer = !ctx->compute_ctx.expired();
 
-    vk_context transfer_ctx;
+    vk_context compute_ctx;
     if (do_transfer) {
-        transfer_ctx = ctx->transfer_ctx.lock();
+        compute_ctx = ctx->compute_ctx.lock();
 
-        ggml_vk_ctx_end(transfer_ctx);
+        ggml_vk_ctx_end(compute_ctx);
 
-        for (auto& cpy : transfer_ctx->in_memcpys) {
+        for (auto& cpy : compute_ctx->in_memcpys) {
             memcpy(cpy.dst, cpy.src, cpy.n);
         }
 
-        ggml_vk_submit(transfer_ctx, {});
+        ggml_vk_submit(compute_ctx, {});
         ctx->submit_pending = true;
     }
 
@@ -13202,10 +13221,10 @@ static void ggml_vk_synchronize(ggml_backend_vk_context * ctx) {
     }
 
     if (do_transfer) {
-        for (auto& cpy : transfer_ctx->out_memcpys) {
+        for (auto& cpy : compute_ctx->out_memcpys) {
             memcpy(cpy.dst, cpy.src, cpy.n);
         }
-        ctx->transfer_ctx.reset();
+        ctx->compute_ctx.reset();
     }
 }
 
@@ -13645,7 +13664,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     int last_node = cgraph->n_nodes - 1;
 
     // If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly
-    while (last_node > 0 && ggml_vk_is_empty(cgraph->nodes[last_node])) {
+    while (last_node > 0 && (ggml_vk_is_empty(cgraph->nodes[last_node]) || ((cgraph->nodes[last_node]->flags & GGML_TENSOR_FLAG_COMPUTE) == 0))) {
         last_node -= 1;
     }
 
@@ -13874,6 +13893,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         ggml_vk_submit(compute_ctx, ctx->device->fence);
         VK_CHECK(ctx->device->device.waitForFences({ ctx->device->fence }, true, UINT64_MAX), "GGML_VULKAN_PERF waitForFences");
         ctx->device->device.resetFences({ ctx->device->fence });
+        ctx->compute_ctx.reset();
 
         // Get the results and pass them to the logger
         std::vector<uint64_t> timestamps(cgraph->n_nodes + 1);
@@ -14160,15 +14180,15 @@ static void ggml_backend_vk_event_record(ggml_backend_t backend, ggml_backend_ev
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     vk_event *vkev = (vk_event *)event->context;
 
-    vk_context transfer_ctx;
+    vk_context compute_ctx;
 
-    if (ctx->transfer_ctx.expired()) {
+    if (ctx->compute_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
-        ctx->transfer_ctx = transfer_ctx;
-        ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+        compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+        ctx->compute_ctx = compute_ctx;
+        ggml_vk_ctx_begin(ctx->device, compute_ctx);
     } else {
-        transfer_ctx = ctx->transfer_ctx.lock();
+        compute_ctx = ctx->compute_ctx.lock();
     }
 
     // the backend interface doesn't have an explicit reset, so reset it here
@@ -14176,13 +14196,13 @@ static void ggml_backend_vk_event_record(ggml_backend_t backend, ggml_backend_ev
     ctx->device->device.resetEvent(vkev->event);
     ctx->device->device.resetFences({ vkev->fence });
 
-    ggml_vk_set_event(transfer_ctx, vkev->event);
+    ggml_vk_set_event(compute_ctx, vkev->event);
 
-    ggml_vk_ctx_end(transfer_ctx);
+    ggml_vk_ctx_end(compute_ctx);
 
-    ggml_vk_submit(transfer_ctx, {vkev->fence});
+    ggml_vk_submit(compute_ctx, {vkev->fence});
     ctx->submit_pending = true;
-    ctx->transfer_ctx.reset();
+    ctx->compute_ctx.reset();
 }
 
 static void ggml_backend_vk_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
@@ -14190,20 +14210,20 @@ static void ggml_backend_vk_event_wait(ggml_backend_t backend, ggml_backend_even
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     vk_event *vkev = (vk_event *)event->context;
 
-    vk_context transfer_ctx;
+    vk_context compute_ctx;
 
-    if (ctx->transfer_ctx.expired()) {
+    if (ctx->compute_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
-        ctx->transfer_ctx = transfer_ctx;
-        ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+        compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+        ctx->compute_ctx = compute_ctx;
+        ggml_vk_ctx_begin(ctx->device, compute_ctx);
     } else {
-        transfer_ctx = ctx->transfer_ctx.lock();
+        compute_ctx = ctx->compute_ctx.lock();
     }
 
-    ggml_vk_wait_events(transfer_ctx, {vkev->event});
-    ggml_vk_ctx_end(transfer_ctx);
-    ctx->transfer_ctx.reset();
+    ggml_vk_wait_events(compute_ctx, {vkev->event});
+    ggml_vk_ctx_end(compute_ctx);
+    ctx->compute_ctx.reset();
 }
 
 // TODO: enable async and synchronize

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

@@ -53,7 +53,7 @@ void main() {
     const uint32_t d_tid = gl_LocalInvocationIndex % D_split;
     const uint32_t col_tid = gl_LocalInvocationIndex / D_split;
 
-    uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
+    uint32_t q_offset = gqa_iq1*p.nb01 + (iq2*p.nb02 + iq3*p.nb03) / 4;
 
     [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
         uint32_t d = (idx + tid) % (HSK / 4);
@@ -101,9 +101,9 @@ void main() {
     uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
-    uint32_t m_offset = 0;
+    uint32_t m_offset = gqa_iq1*KV;
     if (p.nem2 != 1 || p.nem3 != 1) {
-        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
+        m_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
     }
 
     [[dont_unroll]]
@@ -320,7 +320,8 @@ void main() {
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
+        // note: O and Q have swapped coord 1,2.
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
 
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
@@ -332,7 +333,7 @@ void main() {
             }
         }
 
-        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
+        o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
                 perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -378,7 +379,7 @@ void main() {
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
+    uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl

@@ -165,7 +165,7 @@ ACC_TYPE perElemOpGetSink(const in uint32_t r, const in uint32_t c, const in ACC
 }
 
 uint32_t i, N, KV, split_k_index, Tr, start_j, end_j,
-         iq2, iq3, rk2, rk3, rv2, rv3, ik2, ik3, iv2, iv3,
+         gqa_iq1, iq2, iq3, rk2, rk3, rv2, rv3, ik2, ik3, iv2, iv3,
          q_stride, k_stride, v_stride, m_stride;
 
 void init_indices()
@@ -173,12 +173,19 @@ void init_indices()
     N = p.N;
     KV = p.KV;
 
-    i = gl_WorkGroupID.x;
-    split_k_index = 0;
-
     if (p.k_num > 1) {
         i = 0;
-        split_k_index = gl_WorkGroupID.x;
+        // batch and split_k share gl_WorkGroupID.x
+        gqa_iq1 = gl_WorkGroupID.x / p.k_num;
+        split_k_index = gl_WorkGroupID.x % p.k_num;
+    } else if (p.gqa_ratio > 1) {
+        i = 0;
+        gqa_iq1 = gl_WorkGroupID.x;
+        split_k_index = 0;
+    } else {
+        i = gl_WorkGroupID.x;
+        gqa_iq1 = 0;
+        split_k_index = 0;
     }
 
     Tr = CEIL_DIV(N, Br);

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -90,7 +90,7 @@ void main() {
         barrier();
     }
 
-    uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
+    uint32_t q_offset = gqa_iq1*p.nb01 + (iq2*p.nb02+iq3*p.nb03) / 4;
 
     [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
         uint32_t d = (idx + tid) % (HSK / 4);
@@ -141,9 +141,9 @@ void main() {
     uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
-    uint32_t m_offset = 0;
+    uint32_t m_offset = gqa_iq1*KV;
     if (p.nem2 != 1 || p.nem3 != 1) {
-        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
+        m_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
     }
 
     [[dont_unroll]]
@@ -370,7 +370,8 @@ void main() {
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
+        // note: O and Q have swapped coord 1,2.
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
 
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
@@ -382,7 +383,7 @@ void main() {
             }
         }
 
-        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
+        o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
                 perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -428,7 +429,7 @@ void main() {
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
+    uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -111,7 +111,7 @@ void main() {
     coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseAccumulator> Q;
     coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA> Qf16;
 
-    uint32_t q_offset = iq2*p.nb02+iq3*p.nb03;
+    uint32_t q_offset = gqa_iq1*p.nb01*4/*sizeof(float)*/ + iq2*p.nb02+iq3*p.nb03;
     coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK_pad));
 
     Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA>(Q);
@@ -138,9 +138,9 @@ void main() {
         coopMatPerElementNV(slopeMat, slopeMat, perElemOpComputeSlope, iq2);
     }
 
-    uint32_t m_offset = 0;
+    uint32_t m_offset = gqa_iq1*KV * 2 /*sizeof(float16_t)*/;
     if (p.nem2 != 1 || p.nem3 != 1) {
-        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV * 2 /*sizeof(float16_t)*/;
+        m_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV * 2 /*sizeof(float16_t)*/;
     }
 
     [[dont_unroll]]
@@ -272,10 +272,11 @@ void main() {
     if (p.k_num > 1) {
         coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
 
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
+        // note: O and Q have swapped coord 1,2.
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
 
-        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
+        o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
         coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
         coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
         return;
@@ -325,7 +326,7 @@ void main() {
     [[unroll]] for (uint i = 0; i < O.length(); ++i) { O[i] = clamp(O[i], -ACC_TYPE_MAX, ACC_TYPE_MAX); }
 #endif
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
+    uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
 
     coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
     if (p.gqa_ratio > 1) {

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

@@ -12,7 +12,8 @@ layout (binding = 2) writeonly buffer D {float data_d[];};
 
 layout (push_constant) uniform parameter {
     uint D;
-    uint N;
+    uint ne1;
+    uint ne2;
     uint ne3;
     uint k_num;
     uint sinks;
@@ -24,15 +25,15 @@ void main() {
     // Each workgroup handles a row
     const uint n = gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
-    const uint iq3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.z % p.ne2;
+    const uint i3 = gl_WorkGroupID.z / p.ne2;
 
     uint D = p.D;
-    uint N = p.N;
     uint k_num = p.k_num;
 
-    uint l_offset = D * N * p.ne3 * k_num + N * iq3 * k_num * 2 + n;
-    uint m_offset = D * N * p.ne3 * k_num + N * iq3 * k_num * 2 + N + n;
-    uint lm_stride = N * 2;
+    uint l_offset = D * p.ne1 * p.ne2 * p.ne3 * k_num + p.ne1 * 2 * (0/*split_k_index*/ + p.k_num * (i2 + p.ne2 * i3)) + n;
+    uint m_offset = D * p.ne1 * p.ne2 * p.ne3 * k_num + p.ne1 * 2 * (0/*split_k_index*/ + p.k_num * (i2 + p.ne2 * i3)) + p.ne1 + n;
+    uint lm_stride = p.ne1 * 2;
 
     // Compute the max m value for the row
     float m_max = -1.0/0.0;
@@ -99,7 +100,7 @@ void main() {
     if (d < D) {
         float O = 0.0;
         [[unroll]] for (uint k = 0; k < k_num; ++k) {
-            uint o_offset = D * N * (k + iq3 * k_num) + D * n + d;
+            uint o_offset = D * p.ne1 * (k + p.k_num * (i2 + p.ne2 * i3)) + D * n + d;
             float m = data_a[m_offset + k * lm_stride];
             O += exp(m - m_max) * data_a[o_offset];
         }
@@ -115,6 +116,6 @@ void main() {
         const float FLT_MAX = uintBitsToFloat(0x7F7FFFFF);
         O = clamp(O, -FLT_MAX, FLT_MAX);
 
-        data_d[iq3 * D * N + D * n + d] = O;
+        data_d[(i3 * p.ne2 + i2) * p.ne1 * D + D * n + d] = O;
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl

@@ -29,6 +29,8 @@ layout (push_constant) uniform parameter
 #ifdef MUL_MAT_ID
     uint nei0;
     uint ne11;
+    uint expert_i1;
+    uint nbi1;
 #else
     uint ne02;
     uint ne12;
@@ -43,7 +45,7 @@ uint expert_id;
 
 void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.y;
+    const uint expert_i0 = gl_GlobalInvocationID.y;
 #else
     const uint batch_idx = gl_GlobalInvocationID.y;
 #endif
@@ -60,7 +62,7 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
         batch_idx_a = i03 * p.ne02 + i02;
     }
 #else
-    expert_id = data_ids[expert_idx];
+    expert_id = data_ids[expert_i0 + p.expert_i1 * p.nbi1];
 #endif
 
     a_offset =
@@ -71,13 +73,13 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #endif
     b_offset =
 #ifdef MUL_MAT_ID
-            (expert_idx % p.ne11) * p.stride_b;
+            (expert_i0 % p.ne11) * p.stride_b + p.expert_i1 * p.batch_stride_b;
 #else
             batch_idx * p.batch_stride_b;
 #endif
     d_offset =
 #ifdef MUL_MAT_ID
-            expert_idx * p.stride_d;
+            expert_i0 * p.stride_d + p.expert_i1 * p.batch_stride_d;
 #else
             batch_idx * p.batch_stride_d;
 #endif
@@ -103,12 +105,12 @@ void reduce_result(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t
                     temp[j][n] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_i0]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_i0]);
                 }
 #else
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
@@ -158,12 +160,12 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
                     temp[j][n] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_i0]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_i0]);
                 }
 #else
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
@@ -203,12 +205,12 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
                     tmpsh[j][n][0] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse0[expert_i0]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse1[expert_i0]);
                 }
 #else
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -1982,6 +1982,9 @@ static std::optional<webgpu_command> ggml_webgpu_encode_node(webgpu_context ctx,
     if (ggml_is_empty(node)) {
         return std::nullopt;
     }
+    if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+        return std::nullopt;
+    }
     WEBGPU_LOG_DEBUG("ggml_webgpu_encode_node(" << node << ", " << ggml_op_name(node->op) << ")");
 
     ggml_tensor * src0 = node->src[0];

ggml/src/ggml-zdnn/ggml-zdnn.cpp

@@ -58,6 +58,10 @@ static enum ggml_status ggml_zdnn_graph_compute(ggml_backend_t backend, ggml_cgr
             continue;
         }
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         bool ok = ggml_zdnn_compute_forward(ctx, node);
         if (!ok) {
             GGML_LOG_ERROR("%s: unsupported op %s (%s)\n",
@@ -368,7 +372,8 @@ static size_t ggml_backend_zdnn_buffer_type_get_alignment(ggml_backend_buffer_ty
 }
 
 static bool ggml_backend_zdnn_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    return true;
+    /* while it resides in host memory, additional transformation is needed */
+    return false;
 
     GGML_UNUSED(buft);
 }

ggml/src/ggml-zendnn/ggml-zendnn.cpp

@@ -211,6 +211,10 @@ static ggml_status ggml_backend_zendnn_graph_compute(ggml_backend_t backend, ggm
     for (int i = 0; i < cgraph->n_nodes; i++) {
         struct ggml_tensor * node = cgraph->nodes[i];
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         switch (node->op) {
             case GGML_OP_MUL_MAT:
                 ggml_zendnn_compute_forward_mul_mat(ctx, node);

ggml/src/ggml.c

@@ -3441,7 +3441,8 @@ struct ggml_tensor * ggml_cast(
 
     result->op     = GGML_OP_CPY;
     result->src[0] = a;
-    result->src[1] = result;
+    result->src[1] = result; // note: this self-reference might seem redundant, but it's actually needed by some
+                             //       backends for consistency with ggml_cpy_impl() above
 
     return result;
 }
@@ -6725,20 +6726,35 @@ static void ggml_compute_backward(
     GGML_ASSERT(!src2_needs_grads || ggml_are_same_shape(src2, cgraph->grads[isrc2]));
 }
 
-static size_t ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
-    // check if already visited
-    size_t node_hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
+static size_t ggml_visit_parents_graph(struct ggml_cgraph * cgraph, struct ggml_tensor * node, bool compute) {
+    if (node->op != GGML_OP_NONE && compute) {
+        node->flags |= GGML_TENSOR_FLAG_COMPUTE;
+    }
+
+    const size_t node_hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
     GGML_ASSERT(node_hash_pos != GGML_HASHSET_FULL);
-    if (!ggml_bitset_get(cgraph->visited_hash_set.used, node_hash_pos)) {
-        // This is the first time we see this node in the current graph.
-        cgraph->visited_hash_set.keys[node_hash_pos] = node;
-        ggml_bitset_set(cgraph->visited_hash_set.used, node_hash_pos);
-        cgraph->use_counts[node_hash_pos] = 0;
-    } else {
+
+    if (ggml_bitset_get(cgraph->visited_hash_set.used, node_hash_pos)) {
         // already visited
+
+        if (compute) {
+            // update the compute flag regardless
+            for (int i = 0; i < GGML_MAX_SRC; ++i) {
+                struct ggml_tensor * src = node->src[i];
+                if (src && ((src->flags & GGML_TENSOR_FLAG_COMPUTE) == 0)) {
+                    ggml_visit_parents_graph(cgraph, src, true);
+                }
+            }
+        }
+
         return node_hash_pos;
     }
 
+    // This is the first time we see this node in the current graph.
+    cgraph->visited_hash_set.keys[node_hash_pos] = node;
+    ggml_bitset_set(cgraph->visited_hash_set.used, node_hash_pos);
+    cgraph->use_counts[node_hash_pos] = 0;
+
     for (int i = 0; i < GGML_MAX_SRC; ++i) {
         const int k =
             (cgraph->order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? i :
@@ -6747,7 +6763,7 @@ static size_t ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor
 
         struct ggml_tensor * src = node->src[k];
         if (src) {
-            size_t src_hash_pos = ggml_visit_parents(cgraph, src);
+            const size_t src_hash_pos = ggml_visit_parents_graph(cgraph, src, compute);
 
             // Update the use count for this operand.
             cgraph->use_counts[src_hash_pos]++;
@@ -6778,17 +6794,17 @@ static size_t ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor
     return node_hash_pos;
 }
 
-static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand) {
+static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand, bool compute) {
     if (!expand) {
         // TODO: this branch isn't accessible anymore, maybe move this to ggml_build_forward_expand
         ggml_graph_clear(cgraph);
     }
 
-    const int n0 = cgraph->n_nodes;
+    const int n_old = cgraph->n_nodes;
 
-    ggml_visit_parents(cgraph, tensor);
+    ggml_visit_parents_graph(cgraph, tensor, compute);
 
-    const int n_new = cgraph->n_nodes - n0;
+    const int n_new = cgraph->n_nodes - n_old;
     GGML_PRINT_DEBUG("%s: visited %d new nodes\n", __func__, n_new);
 
     if (n_new > 0) {
@@ -6797,8 +6813,22 @@ static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_ten
     }
 }
 
+struct ggml_tensor * ggml_build_forward_select(
+        struct ggml_cgraph  * cgraph,
+        struct ggml_tensor ** tensors,
+        int                   n_tensors,
+        int                   idx) {
+    GGML_ASSERT(idx >= 0 && idx < n_tensors);
+
+    for (int i = 0; i < n_tensors; i++) {
+        ggml_build_forward_impl(cgraph, tensors[i], true, i == idx ? true : false);
+    }
+
+    return tensors[idx];
+}
+
 void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor) {
-    ggml_build_forward_impl(cgraph, tensor, true);
+    ggml_build_forward_impl(cgraph, tensor, true, true);
 }
 
 void ggml_build_backward_expand(
@@ -7229,6 +7259,10 @@ bool ggml_can_fuse_subgraph_ext(const struct ggml_cgraph * cgraph,
             return false;
         }
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            return false;
+        }
+
         if (ggml_node_list_find_tensor(cgraph, outputs, num_outputs, node) != -1) {
             continue;
         }
@@ -7310,7 +7344,7 @@ static void ggml_graph_dump_dot_leaf_edge(FILE * fp, struct ggml_tensor * node,
             label);
 }
 
-void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename) {
+void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * cgraph, const char * filename) {
     char color[16];
 
     FILE * fp = ggml_fopen(filename, "w");
@@ -7331,7 +7365,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
         if (node->flags & GGML_TENSOR_FLAG_PARAM) {
             snprintf(color, sizeof(color), "yellow");
         } else if (grad) {
-            if (ggml_graph_find(gf, node)) {
+            if (ggml_graph_find(cgraph, node)) {
                 snprintf(color, sizeof(color), "green");
             } else {
                 snprintf(color, sizeof(color), "lightblue");

ggml/src/gguf.cpp

@@ -734,7 +734,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
     FILE * file = ggml_fopen(fname, "rb");
 
     if (!file) {
-        GGML_LOG_ERROR("%s: failed to open GGUF file '%s'\n", __func__, fname);
+        GGML_LOG_ERROR("%s: failed to open GGUF file '%s' (%s)\n", __func__, fname, strerror(errno));
         return nullptr;
     }
 

requirements/requirements-tool_bench.txt

@@ -3,7 +3,7 @@ pytest~=8.3.3
 huggingface_hub>=0.34.0,<1.0
 matplotlib~=3.10.0
 numpy~=1.26.4
-openai~=1.55.3
+openai~=2.14.0
 pandas~=2.2.3
 prometheus-client~=0.20.0
 requests~=2.32.3

src/CMakeLists.txt

@@ -24,6 +24,7 @@ add_library(llama
             llama-kv-cache-iswa.cpp
             llama-memory.cpp
             llama-memory-hybrid.cpp
+            llama-memory-hybrid-iswa.cpp
             llama-memory-recurrent.cpp
             llama-mmap.cpp
             llama-model-loader.cpp

src/llama-graph.cpp

@@ -7,6 +7,7 @@
 #include "llama-kv-cache.h"
 #include "llama-kv-cache-iswa.h"
 #include "llama-memory-hybrid.h"
+#include "llama-memory-hybrid-iswa.h"
 #include "llama-memory-recurrent.h"
 
 #include <cassert>
@@ -510,6 +511,76 @@ bool llm_graph_input_mem_hybrid::can_reuse(const llm_graph_params & params) {
     return res;
 }
 
+void llm_graph_input_mem_hybrid_iswa::set_input(const llama_ubatch * ubatch) {
+    const auto * attn_ctx = mctx->get_attn();
+
+    // base tensors may not be allocated if there are no non-SWA attention layers
+    if (inp_attn->self_k_idxs && inp_attn->self_k_idxs->buffer) {
+        attn_ctx->get_base()->set_input_k_idxs(inp_attn->self_k_idxs, ubatch);
+        attn_ctx->get_base()->set_input_v_idxs(inp_attn->self_v_idxs, ubatch);
+
+        attn_ctx->get_base()->set_input_kq_mask(inp_attn->self_kq_mask, ubatch, cparams.causal_attn);
+    }
+
+    // swa tensors may not be allocated if there are no SWA attention layers
+    if (inp_attn->self_k_idxs_swa && inp_attn->self_k_idxs_swa->buffer) {
+        attn_ctx->get_swa()->set_input_k_idxs(inp_attn->self_k_idxs_swa, ubatch);
+        attn_ctx->get_swa()->set_input_v_idxs(inp_attn->self_v_idxs_swa, ubatch);
+
+        attn_ctx->get_swa()->set_input_kq_mask(inp_attn->self_kq_mask_swa, ubatch, cparams.causal_attn);
+    }
+
+    const int64_t n_rs = mctx->get_recr()->get_n_rs();
+
+    if (inp_rs->s_copy) {
+        GGML_ASSERT(ggml_backend_buffer_is_host(inp_rs->s_copy->buffer));
+        int32_t * data = (int32_t *) inp_rs->s_copy->data;
+
+        // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
+        for (uint32_t i = 0; i < n_rs; ++i) {
+            data[i] = mctx->get_recr()->s_copy(i);
+        }
+    }
+}
+
+bool llm_graph_input_mem_hybrid_iswa::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_memory_hybrid_iswa_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    const auto * attn_ctx = mctx->get_attn();
+
+    // base tensors may not be allocated if there are no non-SWA attention layers
+    if (inp_attn->self_k_idxs && inp_attn->self_k_idxs->buffer) {
+        res &= inp_attn->self_k_idxs->ne[0] == params.ubatch.n_tokens;
+      //res &= inp_attn->self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+        res &= inp_attn->self_kq_mask->ne[0] == attn_ctx->get_base()->get_n_kv();
+        res &= inp_attn->self_kq_mask->ne[1] == params.ubatch.n_tokens;
+    }
+
+    // swa tensors may not be allocated if there are no SWA attention layers
+    if (inp_attn->self_k_idxs_swa && inp_attn->self_k_idxs_swa->buffer) {
+        res &= inp_attn->self_k_idxs_swa->ne[0] == params.ubatch.n_tokens;
+      //res &= inp_attn->self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+        res &= inp_attn->self_kq_mask_swa->ne[0] == attn_ctx->get_swa()->get_n_kv();
+        res &= inp_attn->self_kq_mask_swa->ne[1] == params.ubatch.n_tokens;
+    }
+
+    res &= inp_rs->s_copy->ne[0] == mctx->get_recr()->get_n_rs();
+
+    res &= inp_rs->s_copy_main->ne[0]  == params.ubatch.n_seqs;
+    res &= inp_rs->s_copy_extra->ne[0] == mctx->get_recr()->get_n_rs() - params.ubatch.n_seqs;
+
+    res &= inp_rs->head == mctx->get_recr()->get_head();
+    res &= inp_rs->rs_z == mctx->get_recr()->get_rs_z();
+
+    return res;
+}
+
 void llm_graph_input_sampling::set_input(const llama_ubatch * ubatch) {
     // set the inputs only for the active samplers in the current ubatch
     std::unordered_set<llama_seq_id> active_samplers;
@@ -2056,6 +2127,47 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
     return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
 }
 
+llm_graph_input_mem_hybrid_iswa * llm_graph_context::build_inp_mem_hybrid_iswa() const {
+    const auto * mctx_cur = static_cast<const llama_memory_hybrid_iswa_context *>(mctx);
+
+    auto inp_rs = build_rs_inp_impl(ctx0, ubatch, mctx_cur->get_recr());
+
+    // build iswa attention input
+    const auto * attn_ctx = mctx_cur->get_attn();
+
+    auto inp_attn = std::make_unique<llm_graph_input_attn_kv_iswa>(hparams, cparams, attn_ctx);
+
+    const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
+
+    {
+        const auto n_kv = attn_ctx->get_base()->get_n_kv();
+
+        inp_attn->self_k_idxs = attn_ctx->get_base()->build_input_k_idxs(ctx0, ubatch);
+        inp_attn->self_v_idxs = attn_ctx->get_base()->build_input_v_idxs(ctx0, ubatch);
+
+        inp_attn->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+        ggml_set_input(inp_attn->self_kq_mask);
+
+        inp_attn->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp_attn->self_kq_mask, GGML_TYPE_F16) : inp_attn->self_kq_mask;
+    }
+
+    {
+        const auto n_kv = attn_ctx->get_swa()->get_n_kv();
+
+        inp_attn->self_k_idxs_swa = attn_ctx->get_swa()->build_input_k_idxs(ctx0, ubatch);
+        inp_attn->self_v_idxs_swa = attn_ctx->get_swa()->build_input_v_idxs(ctx0, ubatch);
+
+        inp_attn->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+        ggml_set_input(inp_attn->self_kq_mask_swa);
+
+        inp_attn->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp_attn->self_kq_mask_swa, GGML_TYPE_F16) : inp_attn->self_kq_mask_swa;
+    }
+
+    auto inp = std::make_unique<llm_graph_input_mem_hybrid_iswa>(cparams, std::move(inp_attn), std::move(inp_rs), mctx_cur);
+
+    return (llm_graph_input_mem_hybrid_iswa *) res->add_input(std::move(inp));
+}
+
 void llm_graph_context::build_dense_out(
     ggml_tensor * dense_2,
     ggml_tensor * dense_3) const {

src/llama-graph.h

@@ -24,6 +24,7 @@ class llama_kv_cache_context;
 class llama_kv_cache_iswa_context;
 class llama_memory_recurrent_context;
 class llama_memory_hybrid_context;
+class llama_memory_hybrid_iswa_context;
 
 // certain models (typically multi-modal) can produce different types of graphs
 enum llm_graph_type {
@@ -397,6 +398,34 @@ public:
     const llama_memory_hybrid_context * mctx;
 };
 
+class llm_graph_input_mem_hybrid_iswa : public llm_graph_input_i {
+public:
+    llm_graph_input_mem_hybrid_iswa(
+            const llama_cparams & cparams,
+            std::unique_ptr<llm_graph_input_attn_kv_iswa> inp_attn,
+            std::unique_ptr<llm_graph_input_rs>          inp_rs,
+            const llama_memory_hybrid_iswa_context *     mctx) :
+        inp_attn(std::move(inp_attn)),
+        inp_rs(std::move(inp_rs)),
+        cparams(cparams),
+        mctx(mctx) { }
+    virtual ~llm_graph_input_mem_hybrid_iswa() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    bool can_reuse(const llm_graph_params & params) override;
+
+    std::unique_ptr<llm_graph_input_attn_kv_iswa> inp_attn;
+    std::unique_ptr<llm_graph_input_rs>          inp_rs;
+
+    llm_graph_input_attn_kv_iswa * get_attn() const { return inp_attn.get(); }
+    llm_graph_input_rs           * get_recr() const { return inp_rs.get(); }
+
+    const llama_cparams cparams;
+
+    const llama_memory_hybrid_iswa_context * mctx;
+};
+
 class llm_graph_input_sampling : public llm_graph_input_i {
 public:
     llm_graph_input_sampling(std::map<llama_seq_id, llama_sampler *> samplers) :
@@ -881,6 +910,8 @@ struct llm_graph_context {
 
     llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const;
 
+    llm_graph_input_mem_hybrid_iswa * build_inp_mem_hybrid_iswa() const;
+
     //
     // pooling
     //

src/llama-memory-hybrid-iswa.cpp

@@ -0,0 +1,275 @@
+#include "llama-memory-hybrid-iswa.h"
+
+#include "llama-impl.h"
+#include "llama-model.h"
+#include "llama-context.h"
+
+//
+// llama_memory_hybrid_iswa
+//
+
+llama_memory_hybrid_iswa::llama_memory_hybrid_iswa(
+        const llama_model & model,
+                            /* attn */
+                ggml_type   type_k,
+                ggml_type   type_v,
+                     bool   v_trans,
+                     bool   swa_full,
+                 uint32_t   kv_size,
+                 uint32_t   n_ubatch,
+                 uint32_t   n_pad,
+                            /* recurrent */
+                ggml_type   type_r,
+                ggml_type   type_s,
+                 uint32_t   rs_size,
+                            /* common */
+                 uint32_t   n_seq_max,
+                     bool   offload,
+                     bool   unified,
+                            /* layer filters */
+    const layer_filter_cb & filter_attn,
+    const layer_filter_cb & filter_recr) :
+    hparams(model.hparams),
+    mem_attn(new llama_kv_cache_iswa(
+        model,
+        type_k,
+        type_v,
+        v_trans,
+        offload,
+        swa_full,
+        unified,
+        kv_size,
+        n_seq_max,
+        n_ubatch,
+        n_pad,
+        filter_attn == nullptr ?
+            [&](int32_t il) { return !hparams.is_recurrent(il); }
+            : filter_attn,
+        nullptr
+    )),
+    mem_recr(new llama_memory_recurrent(
+        model,
+        type_r,
+        type_s,
+        offload,
+        rs_size,
+        n_seq_max,
+        filter_recr == nullptr ?
+            [&](int32_t il) { return hparams.is_recurrent(il); }
+            : filter_recr
+    )) {}
+
+llama_memory_context_ptr llama_memory_hybrid_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+    do {
+        balloc.split_reset();
+
+        // follow the recurrent pattern for creating the ubatch splits
+        std::vector<llama_ubatch> ubatches;
+
+        while (true) {
+            llama_ubatch ubatch;
+
+            if (embd_all) {
+                // if all tokens are output, split by sequence
+                ubatch = balloc.split_seq(n_ubatch);
+            } else {
+                // TODO: non-sequential equal split can be done if using unified KV cache
+                //       for simplicity, we always use sequential equal split for now
+                ubatch = balloc.split_equal(n_ubatch, true);
+            }
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch)); // NOLINT
+        }
+
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
+        // prepare the recurrent batches first
+        if (!mem_recr->prepare(ubatches)) {
+            // TODO: will the recurrent cache be in an undefined context at this point?
+            LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__);
+            return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+        }
+
+        // prepare the attention cache (iswa version returns both base and swa slot infos)
+        auto sinfos_base = mem_attn->get_base()->prepare(ubatches);
+        if (sinfos_base.empty()) {
+            LLAMA_LOG_ERROR("%s: failed to prepare attention base ubatches\n", __func__);
+            return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+        }
+
+        auto sinfos_swa = mem_attn->get_swa()->prepare(ubatches);
+        if (sinfos_swa.empty()) {
+            LLAMA_LOG_ERROR("%s: failed to prepare attention swa ubatches\n", __func__);
+            return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+        }
+
+        return std::make_unique<llama_memory_hybrid_iswa_context>(
+                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
+    } while(false);
+
+    return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+}
+
+llama_memory_context_ptr llama_memory_hybrid_iswa::init_full() {
+    return std::make_unique<llama_memory_hybrid_iswa_context>(this);
+}
+
+llama_memory_context_ptr llama_memory_hybrid_iswa::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_memory_hybrid_iswa_context>(this, lctx, optimize);
+}
+
+bool llama_memory_hybrid_iswa::get_can_shift() const {
+    // Shifting is trivially supported for recurrent
+    return mem_attn->get_can_shift();
+}
+
+void llama_memory_hybrid_iswa::clear(bool data) {
+    mem_attn->clear(data);
+    mem_recr->clear(data);
+}
+
+bool llama_memory_hybrid_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    // Try removing from the recurrent cache first since it may fail. If it does
+    // fail, the cache will not have been mutated.
+    if (!mem_recr->seq_rm(seq_id, p0, p1)) {
+        return false;
+    }
+    return mem_attn->seq_rm(seq_id, p0, p1);
+}
+
+void llama_memory_hybrid_iswa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+    mem_attn->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    mem_recr->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_memory_hybrid_iswa::seq_keep(llama_seq_id seq_id) {
+    mem_attn->seq_keep(seq_id);
+    mem_recr->seq_keep(seq_id);
+}
+
+void llama_memory_hybrid_iswa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+    mem_attn->seq_add(seq_id, p0, p1, shift);
+    mem_recr->seq_add(seq_id, p0, p1, shift);
+}
+
+void llama_memory_hybrid_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+    mem_attn->seq_div(seq_id, p0, p1, d);
+    mem_recr->seq_div(seq_id, p0, p1, d);
+}
+
+llama_pos llama_memory_hybrid_iswa::seq_pos_min(llama_seq_id seq_id) const {
+    // the min of the total cache is the max of the two caches' min values
+    return std::max(mem_attn->seq_pos_min(seq_id), mem_recr->seq_pos_min(seq_id));
+}
+
+llama_pos llama_memory_hybrid_iswa::seq_pos_max(llama_seq_id seq_id) const {
+    // the max of the total cache is the min of the two caches' max values
+    return std::min(mem_attn->seq_pos_max(seq_id), mem_recr->seq_pos_max(seq_id));
+}
+
+std::map<ggml_backend_buffer_type_t, size_t> llama_memory_hybrid_iswa::memory_breakdown() const {
+    std::map<ggml_backend_buffer_type_t, size_t> mb = mem_attn->memory_breakdown();
+    for (const auto & buft_size : mem_recr->memory_breakdown()) {
+        mb[buft_size.first] += buft_size.second;
+    }
+    return mb;
+}
+
+void llama_memory_hybrid_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    mem_attn->state_write(io, seq_id, flags);
+    mem_recr->state_write(io, seq_id, flags);
+}
+
+void llama_memory_hybrid_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    mem_attn->state_read(io, seq_id, flags);
+    mem_recr->state_read(io, seq_id, flags);
+}
+
+llama_kv_cache_iswa * llama_memory_hybrid_iswa::get_mem_attn() const {
+    return mem_attn.get();
+}
+
+llama_memory_recurrent * llama_memory_hybrid_iswa::get_mem_recr() const {
+    return mem_recr.get();
+}
+
+//
+// llama_memory_hybrid_iswa_context
+//
+
+llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(llama_memory_status status) : status(status) {}
+
+llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(llama_memory_hybrid_iswa * mem) :
+    ctx_attn(mem->get_mem_attn()->init_full()),
+    ctx_recr(mem->get_mem_recr()->init_full()),
+    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
+}
+
+llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(
+        llama_memory_hybrid_iswa * mem,
+                   llama_context * lctx,
+                            bool   optimize) :
+    ctx_attn(mem->get_mem_attn()->init_update(lctx, optimize)),
+    ctx_recr(mem->get_mem_recr()->init_update(lctx, optimize)),
+    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
+}
+
+llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(
+           llama_memory_hybrid_iswa * mem,
+                    slot_info_vec_t   sinfos_base,
+                    slot_info_vec_t   sinfos_swa,
+          std::vector<llama_ubatch>   ubatches) :
+    ubatches(std::move(ubatches)),
+    // note: here we copy the ubatches. not sure if this is ideal
+    ctx_attn(new llama_kv_cache_iswa_context(mem->get_mem_attn(), std::move(sinfos_base), std::move(sinfos_swa), this->ubatches)),
+    ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(), this->ubatches)),
+    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
+}
+
+bool llama_memory_hybrid_iswa_context::next() {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    ctx_attn->next();
+    ctx_recr->next();
+
+    if (++i_next >= ubatches.size()) {
+        return false;
+    }
+
+    return true;
+}
+
+bool llama_memory_hybrid_iswa_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
+
+    bool res = true;
+
+    res = res & ctx_attn->apply();
+    res = res & ctx_recr->apply();
+
+    return res;
+}
+
+llama_memory_status llama_memory_hybrid_iswa_context::get_status() const {
+    return status;
+}
+
+const llama_ubatch & llama_memory_hybrid_iswa_context::get_ubatch() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+    return ubatches[i_next];
+}
+
+const llama_kv_cache_iswa_context * llama_memory_hybrid_iswa_context::get_attn() const {
+    return static_cast<const llama_kv_cache_iswa_context *>(ctx_attn.get());
+}
+
+const llama_memory_recurrent_context * llama_memory_hybrid_iswa_context::get_recr() const {
+    return static_cast<const llama_memory_recurrent_context *>(ctx_recr.get());
+}

src/llama-memory-hybrid-iswa.h

@@ -0,0 +1,140 @@
+#pragma once
+
+#include "llama-batch.h"
+#include "llama-graph.h"
+#include "llama-kv-cache-iswa.h"
+#include "llama-memory.h"
+#include "llama-memory-recurrent.h"
+
+#include <memory>
+#include <vector>
+
+//
+// llama_memory_hybrid_iswa
+//
+
+// utilizes instances of llama_memory_recurrent and llama_kv_cache_iswa to
+//   support models where each layer may be either attention-based (with SWA support) or recurrent
+
+class llama_memory_hybrid_iswa : public llama_memory_i {
+public:
+    llama_memory_hybrid_iswa(
+        const llama_model & model,
+                            /* attn */
+                ggml_type   type_k,
+                ggml_type   type_v,
+                     bool   v_trans,
+                     bool   swa_full,
+                 uint32_t   kv_size,
+                 uint32_t   n_ubatch,
+                 uint32_t   n_pad,
+                            /* recurrent */
+                ggml_type   type_r,
+                ggml_type   type_s,
+                 uint32_t   rs_size,
+                            /* common */
+                 uint32_t   n_seq_max,
+                     bool   offload,
+                     bool   unified,
+                            /* layer filters */
+    const layer_filter_cb & filter_attn = nullptr,
+    const layer_filter_cb & filter_recr = nullptr);
+
+    ~llama_memory_hybrid_iswa() = default;
+
+    //
+    // llama_memory_i
+    //
+
+    llama_memory_context_ptr init_batch(
+            llama_batch_allocr & balloc,
+            uint32_t n_ubatch,
+            bool embd_all) override;
+
+    llama_memory_context_ptr init_full() override;
+
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    bool get_can_shift() const override;
+
+    void clear(bool data) 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 shift) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_min(llama_seq_id seq_id) const override;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0)       override;
+
+    //
+    // llama_memory_hybrid_iswa specific API
+    //
+
+    llama_kv_cache_iswa * get_mem_attn() const;
+    llama_memory_recurrent * get_mem_recr() const;
+
+private:
+    const llama_hparams & hparams;
+
+    const std::unique_ptr<llama_kv_cache_iswa> mem_attn;
+    const std::unique_ptr<llama_memory_recurrent> mem_recr;
+};
+
+class llama_memory_hybrid_iswa_context : public llama_memory_context_i {
+public:
+    using slot_info_vec_t = llama_kv_cache::slot_info_vec_t;
+
+    // init failure
+    explicit llama_memory_hybrid_iswa_context(llama_memory_status status);
+
+    // init full
+    explicit llama_memory_hybrid_iswa_context(llama_memory_hybrid_iswa * mem);
+
+    // init update
+    explicit llama_memory_hybrid_iswa_context(
+        llama_memory_hybrid_iswa * mem,
+                   llama_context * lctx,
+                            bool   optimize);
+
+    // init success
+    llama_memory_hybrid_iswa_context(
+           llama_memory_hybrid_iswa * mem,
+                    slot_info_vec_t   sinfos_base,
+                    slot_info_vec_t   sinfos_swa,
+          std::vector<llama_ubatch>   ubatches);
+
+    ~llama_memory_hybrid_iswa_context() = default;
+
+    bool next()  override;
+    bool apply() override;
+
+    llama_memory_status  get_status() const override;
+    const llama_ubatch & get_ubatch() const override;
+
+    //
+    // llama_memory_hybrid_iswa_context
+    //
+
+    const llama_kv_cache_iswa_context * get_attn() const;
+    const llama_memory_recurrent_context * get_recr() const;
+
+private:
+    // the index of the next ubatch to process
+    size_t i_next = 0;
+
+    std::vector<llama_ubatch> ubatches;
+
+    const llama_memory_context_ptr ctx_attn;
+    const llama_memory_context_ptr ctx_recr;
+
+    const llama_memory_status status;
+};

src/llama-mmap.cpp

@@ -265,7 +265,8 @@ struct llama_file::impl {
                         continue;  // Interrupted by signal, retry
                     }
                     // Fallback to std::fread in case the DMA controller cannot access the buffer
-                    if (errno == EFAULT) {
+                    if (errno == EFAULT || errno == EINVAL) {
+                        LLAMA_LOG_WARN("%s: Falling back to buffered IO due to %s\n", __func__, strerror(errno));
                         auto curr_off = tell();
                         close(fd);
                         fd = -1;
@@ -384,6 +385,9 @@ int llama_file::file_id() const {
 #ifdef _WIN32
     return _fileno(pimpl->fp);
 #else
+    if (pimpl->fd != -1) {
+        return pimpl->fd;
+    }
 #if defined(fileno)
     return fileno(pimpl->fp);
 #else

src/llama-model-loader.cpp

@@ -539,12 +539,18 @@ llama_model_loader::llama_model_loader(
     files.emplace_back(new llama_file(fname.c_str(), "rb", use_direct_io));
     contexts.emplace_back(ctx);
 
-    use_direct_io = use_direct_io && files.back()->has_direct_io();
-
-    // Disable mmap in case Direct I/O is enabled and available
-    if (use_direct_io && use_mmap) {
-        use_mmap = false;
-        LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__);
+    if (use_mmap && use_direct_io) {
+        if (files.back()->has_direct_io()) {
+            // Disable mmap, as DirectIO is available
+            use_mmap = false;
+            LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__);
+        } else {
+            // Disable DirectIO and reopen file using std::fopen for mmap
+            use_direct_io = false;
+            files.pop_back();
+            files.emplace_back(new llama_file(fname.c_str(), "rb", false));
+            LLAMA_LOG_WARN("%s: direct I/O is not available, using mmap\n", __func__);
+        }
     }
 
     // Save tensors data offset of the main file.

src/llama-model.cpp

@@ -8,6 +8,7 @@
 #include "llama-kv-cache.h"
 #include "llama-kv-cache-iswa.h"
 #include "llama-memory-hybrid.h"
+#include "llama-memory-hybrid-iswa.h"
 #include "llama-memory-recurrent.h"
 
 #include "ggml-cpp.h"
@@ -1713,7 +1714,12 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
                     // for compatibility with existing DeepSeek V2 and V2.5 GGUFs
                     // that have no expert_gating_func model parameter set
-                    hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;
+                    if ((hparams.n_layer == 47 || hparams.n_layer == 48) && n_vocab == 154880) {
+                        // GLM 4.7 Lite
+                        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+                    } else {
+                        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;
+                    }
                 }
 
                 if (ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, 0.0f)) {
@@ -7523,23 +7529,44 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         };
                     }
 
-                    res = new llama_memory_hybrid(
-                        /* model             */ *this,
-                        /* attn_type_k       */ params.type_k,
-                        /* attn_type_v       */ params.type_v,
-                        /* attn_v_trans      */ !cparams.flash_attn,
-                        /* attn_kv_size      */ cparams.n_ctx,
-                        /* attn_n_pad        */ 1,
-                        /* attn_n_swa        */ hparams.n_swa,
-                        /* attn_swa_type     */ hparams.swa_type,
-                        /* recurrent_type_k  */ GGML_TYPE_F32,
-                        /* recurrent_type_v  */ GGML_TYPE_F32,
-                        /* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max),
-                        /* n_seq_max         */ cparams.n_seq_max,
-                        /* offload           */ cparams.offload_kqv,
-                        /* unified           */ cparams.kv_unified,
-                        /* filter_attn       */ std::move(filter_attn),
-                        /* filter_recr       */ std::move(filter_recr));
+                    if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
+                        // Use hybrid-iswa for hybrid models with SWA
+                        res = new llama_memory_hybrid_iswa(
+                            /* model             */ *this,
+                            /* attn_type_k       */ params.type_k,
+                            /* attn_type_v       */ params.type_v,
+                            /* attn_v_trans      */ !cparams.flash_attn,
+                            /* attn_swa_full     */ params.swa_full,
+                            /* attn_kv_size      */ cparams.n_ctx,
+                            /* attn_n_ubatch     */ cparams.n_ubatch,
+                            /* attn_n_pad        */ 1,
+                            /* recurrent_type_r  */ GGML_TYPE_F32,
+                            /* recurrent_type_s  */ GGML_TYPE_F32,
+                            /* recurrent_rs_size */ std::max((uint32_t) 1, cparams.n_seq_max),
+                            /* n_seq_max         */ cparams.n_seq_max,
+                            /* offload           */ cparams.offload_kqv,
+                            /* unified           */ cparams.kv_unified,
+                            /* filter_attn       */ std::move(filter_attn),
+                            /* filter_recr       */ std::move(filter_recr));
+                    } else {
+                        res = new llama_memory_hybrid(
+                            /* model             */ *this,
+                            /* attn_type_k       */ params.type_k,
+                            /* attn_type_v       */ params.type_v,
+                            /* attn_v_trans      */ !cparams.flash_attn,
+                            /* attn_kv_size      */ cparams.n_ctx,
+                            /* attn_n_pad        */ 1,
+                            /* attn_n_swa        */ hparams.n_swa,
+                            /* attn_swa_type     */ hparams.swa_type,
+                            /* recurrent_type_k  */ GGML_TYPE_F32,
+                            /* recurrent_type_v  */ GGML_TYPE_F32,
+                            /* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max),
+                            /* n_seq_max         */ cparams.n_seq_max,
+                            /* offload           */ cparams.offload_kqv,
+                            /* unified           */ cparams.kv_unified,
+                            /* filter_attn       */ std::move(filter_attn),
+                            /* filter_recr       */ std::move(filter_recr));
+                    }
                 } else {
                     llama_memory_i::layer_reuse_cb reuse = nullptr;
 

src/models/nemotron-h.cpp

@@ -67,7 +67,7 @@ ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *
                                                           const llama_model &       model,
                                                           const int64_t             n_embd_head,
                                                           const int                 il) {
-    // compute Q and K and (optionally) RoPE them
+    // compute Q and K
     ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
     cb(Qcur, "Qcur", il);
     if (model.layers[il].bq) {

tests/CMakeLists.txt

@@ -187,6 +187,7 @@ llama_build_and_test(test-chat-parser.cpp)
 llama_build_and_test(test-chat-peg-parser.cpp peg-parser/simple-tokenize.cpp)
 llama_build_and_test(test-chat-template.cpp)
 llama_build_and_test(test-jinja.cpp)
+llama_test(test-jinja NAME test-jinja-py ARGS -py LABEL python)
 llama_build_and_test(test-json-partial.cpp)
 llama_build_and_test(test-log.cpp)
 llama_build_and_test(

tests/test-backend-ops.cpp

@@ -8460,6 +8460,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     // Qwen3-VL-8B https://github.com/ggml-org/llama.cpp/issues/17012
     test_cases.emplace_back(new test_flash_attn_ext(72, 72, 16, {1, 1}, 5776, 5776, false, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
 
+    test_cases.emplace_back(new test_flash_attn_ext(64, 64, 8, {8, 1}, 7680, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
+    test_cases.emplace_back(new test_flash_attn_ext(64, 64, 8, {8, 1}, 7680, 4, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
+
     for (int kv : { 4096, 8192, 16384, }) {
         for (int hs : { 64, 128, }) {
             for (int nr : { 1, 4, }) {

tests/test-chat-parser.cpp

@@ -54,113 +54,109 @@ static void assert_throws(const std::function<void()> & fn, const std::string &
 static void test_reasoning() {
   //common_log_set_verbosity_thold(LOG_DEFAULT_DEBUG);
   {
-    common_chat_msg_parser builder("<tnk>Cogito</tnk>Ergo sum", /* is_partial= */ false, {
-        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_NONE,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ false,
-    });
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    params.reasoning_format = COMMON_REASONING_FORMAT_NONE;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = false;
+    common_chat_msg_parser builder("<tnk>Cogito</tnk>Ergo sum", /* is_partial= */ false, params);
     assert_equals(false, builder.try_parse_reasoning("<tnk>", "</tnk>"));
     assert_equals("<tnk>Cogito</tnk>Ergo sum", builder.consume_rest());
   }
   {
-    common_chat_msg_parser builder("<tnk>Cogito</tnk>Ergo sum", /* is_partial= */ false, {
-        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ false,
-    });
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = false;
+    common_chat_msg_parser builder("<tnk>Cogito</tnk>Ergo sum", /* is_partial= */ false, params);
     assert_equals(true, builder.try_parse_reasoning("<tnk>", "</tnk>"));
     assert_equals(std::string("Cogito"), builder.result().reasoning_content);
     assert_equals("Ergo sum", builder.consume_rest());
   }
   {
-    common_chat_msg_parser builder("Cogito</tnk>Ergo sum", /* is_partial= */ false, {
-        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_NONE,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ false,
-    });
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    params.reasoning_format = COMMON_REASONING_FORMAT_NONE;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = false;
+    common_chat_msg_parser builder("Cogito</tnk>Ergo sum", /* is_partial= */ false, params);
     assert_equals(false, builder.try_parse_reasoning("<tnk>", "</tnk>"));
     assert_equals("Cogito</tnk>Ergo sum", builder.consume_rest());
   }
   {
-    common_chat_msg_parser builder("Cogito</tnk>Ergo sum", /* is_partial= */ false, {
-        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ true,
-    });
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = true;
+    common_chat_msg_parser builder("Cogito</tnk>Ergo sum", /* is_partial= */ false, params);
     assert_equals(true, builder.try_parse_reasoning("<tnk>", "</tnk>"));
     assert_equals(std::string("Cogito"), builder.result().reasoning_content);
     assert_equals("Ergo sum", builder.consume_rest());
   }
   {
-    common_chat_msg_parser builder("Cogito</tnk>Ergo sum", /* is_partial= */ false, {
-        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ true,
-        /* .thinking_forced_open = */ true,
-    });
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = true;
+    params.thinking_forced_open = true;
+    common_chat_msg_parser builder("Cogito</tnk>Ergo sum", /* is_partial= */ false, params);
     assert_equals(true, builder.try_parse_reasoning("<tnk>", "</tnk>"));
     assert_equals("<think>Cogito</think>", builder.result().content);
     assert_equals("Ergo sum", builder.consume_rest());
   }
   {
     const std::string variant("content_only_inline_think");
-    common_chat_syntax syntax = {
-        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ false,
-        /* .parse_tool_calls = */ false,
-    };
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = false;
+    params.parse_tool_calls = false;
     const std::string input = "<think>Pense</think>Bonjour";
-    auto msg = common_chat_parse(input, false, syntax);
+    auto msg = common_chat_parse(input, false, params);
     assert_equals(variant, std::string("Pense"), msg.reasoning_content);
     assert_equals(variant, std::string("Bonjour"), msg.content);
   }
   {
     const std::string variant("llama_3_inline_think");
-    common_chat_syntax syntax = {
-        /* .format = */ COMMON_CHAT_FORMAT_LLAMA_3_X,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ false,
-        /* .parse_tool_calls = */ false,
-    };
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_LLAMA_3_X;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = false;
+    params.parse_tool_calls = false;
     const std::string input = "<think>Plan</think>Réponse";
-    auto msg = common_chat_parse(input, false, syntax);
+    auto msg = common_chat_parse(input, false, params);
     assert_equals(variant, std::string("Plan"), msg.reasoning_content);
     assert_equals(variant, std::string("Réponse"), msg.content);
   }
   // Test DeepSeek V3.1 parsing - reasoning content followed by "</think>" and then regular content
   {
-    common_chat_syntax syntax = {
-        /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ true,
-        /* .parse_tool_calls = */ true,
-    };
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = true;
+    params.parse_tool_calls = true;
     const std::string variant("deepseek_v3_1_reasoning_format_deepseek");
-    common_chat_msg_parser builder("REASONING</think>ok", /* is_partial= */ false, syntax);
+    common_chat_msg_parser builder("REASONING</think>ok", /* is_partial= */ false, params);
     assert_equals(variant, true, builder.try_parse_reasoning("<think>", "</think>"));
     assert_equals(variant, std::string("REASONING"), builder.result().reasoning_content);
     assert_equals(variant, std::string("ok"), builder.consume_rest());
   }
   // Test DeepSeek V3.1 parsing - reasoning_format none - reasoning content followed by "</think>" and then regular content
   {
-    common_chat_syntax syntax = {
-        /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_NONE,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ true,
-        /* .parse_tool_calls = */ true,
-    };
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+    params.reasoning_format = COMMON_REASONING_FORMAT_NONE;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = true;
+    params.parse_tool_calls = true;
     const std::string variant("deepseek_v3_1_reasoning_format_none");
     const std::string input = "REASONING</think>ok";
-    auto msg = common_chat_parse(input, false, syntax);
+    auto msg = common_chat_parse(input, false, params);
     assert_equals(variant, std::string("REASONING</think>ok"), msg.content);
     assert_equals(variant, std::string(""), msg.reasoning_content);
   }
@@ -256,15 +252,14 @@ static void test_deepseek_v3_1_tool_calls() {
     //common_log_set_verbosity_thold(LOG_DEFAULT_DEBUG);
     // variant: happy path for when it works as the model card says it should
     const std::string variant("simple");
-    common_chat_syntax syntax = {
-        /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ false,
-        /* .parse_tool_calls = */ true,
-    };
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = false;
+    params.parse_tool_calls = true;
     const std::string input = "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>";
-    auto msg = common_chat_parse(input, false, syntax);
+    auto msg = common_chat_parse(input, false, params);
     assert_equals<std::size_t>(variant, 1, msg.tool_calls.size());
     assert_equals(variant, std::string("get_time"), msg.tool_calls[0].name);
     // JSON arguments are dumped without spaces
@@ -274,16 +269,15 @@ static void test_deepseek_v3_1_tool_calls() {
 
     // variant: simple + thinking open
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ true,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = true;
+        params.parse_tool_calls = true;
         const std::string variant("simple_thinking");
         const std::string in = "REASONING</think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>";
-        auto m = common_chat_parse(in, false, syntax);
+        auto m = common_chat_parse(in, false, params);
         assert_equals<std::size_t>(variant, 1, m.tool_calls.size());
         assert_equals(variant, std::string("get_time"), m.tool_calls[0].name);
         assert_equals(variant, std::string("{\"city\":\"Tokyo\"}"), m.tool_calls[0].arguments);
@@ -292,16 +286,15 @@ static void test_deepseek_v3_1_tool_calls() {
     }
     // variant: simple + multiple tool calls
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ false,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = false;
+        params.parse_tool_calls = true;
         const std::string variant("simple_multiple_tool_calls");
         const std::string in = "CONTENT<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Paris\"}<｜tool▁call▁end｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{\"city\": \"Paris\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>";
-        auto m = common_chat_parse(in, false, syntax);
+        auto m = common_chat_parse(in, false, params);
         assert_equals<std::size_t>(variant, 2, m.tool_calls.size());
         assert_equals(variant, std::string("get_time"), m.tool_calls[0].name);
         assert_equals(variant, std::string("{\"city\":\"Paris\"}"), m.tool_calls[0].arguments);
@@ -314,16 +307,15 @@ static void test_deepseek_v3_1_tool_calls() {
 
     // variant: thinking forced open + tool call in reasoning content
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ true,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = true;
+        params.parse_tool_calls = true;
         const std::string variant("thinking_forced_open_tool_call_in_reasoning");
         const std::string in = "REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time2<｜tool▁sep｜>{\"city\": \"Tokyo2\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>REASONING</think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>";
-        auto m = common_chat_parse(in, false, syntax);
+        auto m = common_chat_parse(in, false, params);
         assert_equals<std::size_t>(variant, 1, m.tool_calls.size());
         assert_equals(variant, std::string("get_time"), m.tool_calls[0].name);
         assert_equals(variant, std::string("{\"city\":\"Tokyo\"}"), m.tool_calls[0].arguments);
@@ -336,16 +328,15 @@ static void test_deepseek_v3_1_tool_calls() {
     //          to make tool calls in reasoning content according to the model card, but it does sometimes, so
     //          add the reasoning content as regular content and parse the tool calls.
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ true,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = true;
+        params.parse_tool_calls = true;
         const std::string variant("thinking_forced_open_tool_call_in_reasoning_no_closing_think_not_partial");
         const std::string in = "REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>";
-        auto m = common_chat_parse(in, false, syntax);
+        auto m = common_chat_parse(in, false, params);
         assert_equals(variant, std::string("REASONING"), m.content);
         assert_equals(variant, std::string(""), m.reasoning_content);
         assert_equals<std::size_t>(variant, 1, m.tool_calls.size());
@@ -355,16 +346,15 @@ static void test_deepseek_v3_1_tool_calls() {
 
     // variant: thinking forced open + tool call in reasoning content + no closing think + partial
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ true,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = true;
+        params.parse_tool_calls = true;
         const std::string variant("thinking_forced_open_tool_call_in_reasoning_no_closing_think_partial");
         const std::string in = "REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>";
-        auto m = common_chat_parse(in, /* is_partial= */ true, syntax);
+        auto m = common_chat_parse(in, /* is_partial= */ true, params);
         assert_equals(variant, std::string("REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>"), m.reasoning_content);
         assert_equals(variant, std::string(""), m.content);
         assert_equals<std::size_t>(variant, 0, m.tool_calls.size());
@@ -372,32 +362,30 @@ static void test_deepseek_v3_1_tool_calls() {
 
     // variant: thinking not forced open + reasoning + regular content + no tool calls
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ true,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = true;
+        params.parse_tool_calls = true;
         const std::string variant("thinking_forced_open_reasoning_regular_content_no_tool_calls");
         const std::string in = "REASONING</think>CONTENT";
-        auto m = common_chat_parse(in, false, syntax);
+        auto m = common_chat_parse(in, false, params);
         assert_equals<std::size_t>(variant, 0, m.tool_calls.size());
         assert_equals(variant, std::string("CONTENT"), m.content);
         assert_equals(variant, std::string("REASONING"), m.reasoning_content);
     }
     // variant: thinking not forced open + missing reasoning + no tool calls
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ false,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = false;
+        params.parse_tool_calls = true;
         const std::string variant("thinking_not_forced_open_missing_reasoning_no_tool_calls");
         const std::string in = "CONTENT";
-        auto m = common_chat_parse(in, false, syntax);
+        auto m = common_chat_parse(in, false, params);
         assert_equals<std::size_t>(variant, 0, m.tool_calls.size());
         assert_equals(variant, std::string("CONTENT"), m.content);
         assert_equals(variant, std::string(""), m.reasoning_content);

tests/test-chat-peg-parser.cpp

@@ -616,15 +616,15 @@ void test_command7_parser_compare(testing & t) {
 
     auto test_legacy = [&](const std::string & input, bool need_more_input, bool print_results) {
         // Original common_chat_combinator_parser taken from chat.cpp
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_GENERIC;
+        params.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = false;
         common_chat_msg_parser builder(
             input,
             /* .is_partial = */ need_more_input,
-            {
-                /* .format = */ COMMON_CHAT_FORMAT_GENERIC,
-                /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
-                /* .reasoning_in_content = */ false,
-                /* .thinking_forced_open = */ false,
-            }
+            params
         );
 
         builder.try_parse_reasoning("<|START_THINKING|>", "<|END_THINKING|>");

tests/test-jinja.cpp

@@ -191,6 +191,84 @@ static void test_conditionals(testing & t) {
         json::object(),
         "yes"
     );
+
+    test_template(t, "is undefined falsy",
+        "{{ 'yes' if not y else 'no' }}",
+        json::object(),
+        "yes"
+    );
+
+    test_template(t, "is undefined attribute falsy",
+        "{{ 'yes' if not y.x else 'no' }}",
+        {{"y", true}},
+        "yes"
+    );
+
+    test_template(t, "is undefined key falsy",
+        "{{ 'yes' if not y['x'] else 'no' }}",
+        {{"y", {{}}}},
+        "yes"
+    );
+
+    test_template(t, "is empty array falsy",
+        "{{ 'yes' if not y else 'no' }}",
+        {{"y", json::array()}},
+        "yes"
+    );
+
+    test_template(t, "is empty object falsy",
+        "{{ 'yes' if not y else 'no' }}",
+        {{"y", json::object()}},
+        "yes"
+    );
+
+    test_template(t, "is empty string falsy",
+        "{{ 'yes' if not y else 'no' }}",
+        {{"y", ""}},
+        "yes"
+    );
+
+    test_template(t, "is 0 falsy",
+        "{{ 'yes' if not y else 'no' }}",
+        {{"y", 0}},
+        "yes"
+    );
+
+    test_template(t, "is 0.0 falsy",
+        "{{ 'yes' if not y else 'no' }}",
+        {{"y", 0.0}},
+        "yes"
+    );
+
+    test_template(t, "is non-empty array truthy",
+        "{{ 'yes' if y else 'no' }}",
+        {{"y", json::array({""})}},
+        "yes"
+    );
+
+    test_template(t, "is non-empty object truthy",
+        "{{ 'yes' if y else 'no' }}",
+        {{"y", {"x", false}}},
+        "yes"
+    );
+
+    test_template(t, "is non-empty string truthy",
+        "{{ 'yes' if y else 'no' }}",
+        {{"y", "0"}},
+        "yes"
+    );
+
+    test_template(t, "is 1 truthy",
+        "{{ 'yes' if y else 'no' }}",
+        {{"y", 1}},
+        "yes"
+    );
+
+    test_template(t, "is 1.0 truthy",
+        "{{ 'yes' if y else 'no' }}",
+        {{"y", 1.0}},
+        "yes"
+    );
 }
 
 static void test_loops(testing & t) {
@@ -531,6 +609,12 @@ static void test_filters(testing & t) {
         json::object(),
         "hello"
     );
+
+    test_template(t, "none to string",
+        "{{ x|string }}",
+        {{"x", nullptr}},
+        "None"
+    );
 }
 
 static void test_literals(testing & t) {

tools/cli/cli.cpp

@@ -66,19 +66,25 @@ struct cli_context {
         defaults.stream = true; // make sure we always use streaming mode
         defaults.timings_per_token = true; // in order to get timings even when we cancel mid-way
         // defaults.return_progress = true; // TODO: show progress
-        defaults.oaicompat_chat_syntax.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     }
 
     std::string generate_completion(result_timings & out_timings) {
         server_response_reader rd = ctx_server.get_response_reader();
+        auto chat_params = format_chat();
         {
             // TODO: reduce some copies here in the future
             server_task task = server_task(SERVER_TASK_TYPE_COMPLETION);
-            task.id        = rd.get_new_id();
-            task.index     = 0;
-            task.params    = defaults;    // copy
-            task.cli_input = messages;    // copy
-            task.cli_files = input_files; // copy
+            task.id         = rd.get_new_id();
+            task.index      = 0;
+            task.params     = defaults;           // copy
+            task.cli_prompt = chat_params.prompt; // copy
+            task.cli_files  = input_files;        // copy
+            task.cli        = true;
+
+            // chat template settings
+            task.params.chat_parser_params = common_chat_parser_params(chat_params);
+            task.params.chat_parser_params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+
             rd.post_task({std::move(task)});
         }
 
@@ -156,6 +162,25 @@ struct cli_context {
             return content;
         }
     }
+
+    common_chat_params format_chat() {
+        auto meta = ctx_server.get_meta();
+        auto & chat_params = meta.chat_params;
+
+        common_chat_templates_inputs inputs;
+        inputs.messages              = common_chat_msgs_parse_oaicompat(messages);
+        inputs.tools                 = {}; // TODO
+        inputs.tool_choice           = COMMON_CHAT_TOOL_CHOICE_NONE;
+        inputs.json_schema           = ""; // TODO
+        inputs.grammar               = ""; // TODO
+        inputs.use_jinja             = chat_params.use_jinja;
+        inputs.parallel_tool_calls   = false;
+        inputs.add_generation_prompt = true;
+        inputs.enable_thinking       = chat_params.enable_thinking;
+
+        // Apply chat template to the list of messages
+        return common_chat_templates_apply(chat_params.tmpls.get(), inputs);
+    }
 };
 
 int main(int argc, char ** argv) {

tools/server/README.md

@@ -6,7 +6,7 @@ Set of LLM REST APIs and a web UI to interact with llama.cpp.
 
 **Features:**
  * LLM inference of F16 and quantized models on GPU and CPU
- * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
+ * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions, responses, and embeddings routes
  * [Anthropic Messages API](https://docs.anthropic.com/en/api/messages) compatible chat completions
  * Reranking endpoint (https://github.com/ggml-org/llama.cpp/pull/9510)
  * Parallel decoding with multi-user support
@@ -1267,6 +1267,49 @@ This provides information on the performance of the server. It also allows calcu
 
 The total number of tokens in context is equal to `prompt_n + cache_n + predicted_n`
 
+### POST `/v1/responses`: OpenAI-compatible Responses API
+
+*Options:*
+
+See [OpenAI Responses API documentation](https://platform.openai.com/docs/api-reference/responses).
+
+*Examples:*
+
+You can use either Python `openai` library with appropriate checkpoints:
+
+```python
+import openai
+
+client = openai.OpenAI(
+    base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
+    api_key = "sk-no-key-required"
+)
+
+response = client.responses.create(
+  model="gpt-4.1",
+  instructions="You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests.",
+  input="Write a limerick about python exceptions"
+)
+
+print(response.output_text)
+```
+
+... or raw HTTP requests:
+
+```shell
+curl http://localhost:8080/v1/responses \
+-H "Content-Type: application/json" \
+-H "Authorization: Bearer no-key" \
+-d '{
+"model": "gpt-4.1",
+"instructions": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests.",
+"input": "Write a limerick about python exceptions"
+}'
+```
+
+This endpoint works by converting Responses request into Chat Completions request.
+
+
 ### POST `/v1/embeddings`: OpenAI-compatible embeddings API
 
 This endpoint requires that the model uses a pooling different than type `none`. The embeddings are normalized using the Eucledian norm.

tools/server/server-common.cpp

@@ -779,7 +779,6 @@ static void handle_media(
         // download remote image
         // TODO @ngxson : maybe make these params configurable
         common_remote_params params;
-        params.headers.push_back({"User-Agent", "llama.cpp/" + build_info});
         params.max_size = 1024 * 1024 * 10; // 10MB
         params.timeout  = 10; // seconds
         SRV_INF("downloading image from '%s'\n", url.c_str());
@@ -831,7 +830,7 @@ static void handle_media(
 // used by /chat/completions endpoint
 json oaicompat_chat_params_parse(
     json & body, /* openai api json semantics */
-    const oaicompat_parser_options & opt,
+    const server_chat_params & opt,
     std::vector<raw_buffer> & out_files)
 {
     json llama_params;
@@ -1012,7 +1011,7 @@ json oaicompat_chat_params_parse(
     }
 
     // Apply chat template to the list of messages
-    auto chat_params = common_chat_templates_apply(opt.tmpls, inputs);
+    auto chat_params = common_chat_templates_apply(opt.tmpls.get(), inputs);
 
     /* Append assistant prefilled message */
     if (prefill_assistant_message) {
@@ -1070,6 +1069,283 @@ json oaicompat_chat_params_parse(
     return llama_params;
 }
 
+json convert_responses_to_chatcmpl(const json & response_body) {
+    if (!response_body.contains("input")) {
+        throw std::invalid_argument("'input' is required");
+    }
+    if (!json_value(response_body, "previous_response_id", std::string{}).empty()) {
+        throw std::invalid_argument("llama.cpp does not support 'previous_response_id'.");
+    }
+
+    const json input_value = response_body.at("input");
+    json chatcmpl_body = response_body;
+    chatcmpl_body.erase("input");
+    std::vector<json> chatcmpl_messages;
+
+    if (response_body.contains("instructions")) {
+        chatcmpl_messages.push_back({
+            {"role",    "system"},
+            {"content", json_value(response_body, "instructions", std::string())},
+        });
+        chatcmpl_body.erase("instructions");
+    }
+
+    if (input_value.is_string()) {
+        // #responses_create-input-text_input
+        chatcmpl_messages.push_back({
+            {"role",    "user"},
+            {"content", input_value},
+        });
+    } else if (input_value.is_array()) {
+        // #responses_create-input-input_item_list
+
+        static auto exists_and_is_array = [](const json & j, const char * key) -> bool {
+            return j.contains(key) && j.at(key).is_array();
+        };
+        static auto exists_and_is_string = [](const json & j, const char * key) -> bool {
+            return j.contains(key) && j.at(key).is_string();
+        };
+
+        for (json item : input_value) {
+            if (exists_and_is_string(item, "content")) {
+                // #responses_create-input-input_item_list-input_message-content-text_input
+                // Only "Input message" contains item["content"]::string
+                // After converting item["content"]::string to item["content"]::array,
+                // we can treat "Input message" as sum of "Item-Input message" and "Item-Output message"
+                item["content"] = json::array({
+                    json {
+                        {"text", item.at("content")},
+                        {"type", "input_text"}
+                    }
+                });
+            }
+
+            if (exists_and_is_array(item, "content") &&
+                exists_and_is_string(item, "role") &&
+                (item.at("role") == "user" ||
+                    item.at("role") == "system" ||
+                    item.at("role") == "developer")
+            ) {
+                // #responses_create-input-input_item_list-item-input_message
+                std::vector<json> chatcmpl_content;
+
+                for (const json & input_item : item.at("content")) {
+                    const std::string type = json_value(input_item, "type", std::string());
+
+                    if (type == "input_text") {
+                        if (!input_item.contains("text")) {
+                            throw std::invalid_argument("'Input text' requires 'text'");
+                        }
+                        chatcmpl_content.push_back({
+                            {"text", input_item.at("text")},
+                            {"type", "text"},
+                        });
+                    } else if (type == "input_image") {
+                        // While `detail` is marked as required,
+                        // it has default value("auto") and can be omitted.
+
+                        if (!input_item.contains("image_url")) {
+                            throw std::invalid_argument("'image_url' is required");
+                        }
+                        chatcmpl_content.push_back({
+                            {"image_url", json {
+                                {"url", input_item.at("image_url")}
+                            }},
+                            {"type", "image_url"},
+                        });
+                    } else if (type == "input_file") {
+                        throw std::invalid_argument("'input_file' is not supported by llamacpp at this moment");
+                        // if (input_item.contains("file_url")) {
+                        //     // chat completion API does not support file_url
+                        //     throw std::invalid_argument("'file_url' is not supported");
+                        // }
+                        // if (!input_item.contains("file_data") || !input_item.contains("filename")) {
+                        //     throw std::invalid_argument("Both 'file_data' and 'filename' are required");
+                        // }
+                        // chatcmpl_content.push_back({
+                        //     {"file", json {
+                        //         {"file_data", input_item.at("file_data")},
+                        //         {"filename",  input_item.at("filename")},
+                        //     }},
+                        //     {"type", "file"},
+                        // });
+                    } else {
+                        throw std::invalid_argument("'type' must be one of 'input_text', 'input_image', or 'input_file'");
+                    }
+                }
+
+                if (item.contains("type")) {
+                    item.erase("type");
+                }
+                if (item.contains("status")) {
+                    item.erase("status");
+                }
+                item["content"] = chatcmpl_content;
+
+                chatcmpl_messages.push_back(item);
+            } else if (exists_and_is_array(item, "content") &&
+                exists_and_is_string(item, "role") &&
+                item.at("role") == "assistant" &&
+                // exists_and_is_string(item, "status") &&
+                // (item.at("status") == "in_progress" ||
+                //     item.at("status") == "completed" ||
+                //     item.at("status") == "incomplete") &&
+                // item["status"] not sent by codex-cli
+                exists_and_is_string(item, "type") &&
+                item.at("type") == "message"
+            ) {
+                // #responses_create-input-input_item_list-item-output_message
+                std::vector<json> chatcmpl_content;
+
+                for (const auto & output_text : item.at("content")) {
+                    const std::string type = json_value(output_text, "type", std::string());
+                    if (type != "output_text") {
+                        throw std::invalid_argument("'type' must be 'output_text'");
+                    }
+                    if (!exists_and_is_string(output_text, "text")) {
+                        throw std::invalid_argument("'Output text' requires 'text'");
+                    }
+                    // Ignore annotations and logprobs for now
+                    chatcmpl_content.push_back({
+                        {"text", output_text.at("text")},
+                        {"type", "text"},
+                    });
+                }
+
+                item.erase("status");
+                item.erase("type");
+                item["content"] = chatcmpl_content;
+                chatcmpl_messages.push_back(item);
+            } else if (exists_and_is_string(item, "arguments") &&
+                exists_and_is_string(item, "call_id") &&
+                exists_and_is_string(item, "name") &&
+                exists_and_is_string(item, "type") &&
+                item.at("type") == "function_call"
+            ) {
+                // #responses_create-input-input_item_list-item-function_tool_call
+                json msg = json {
+                    {"role", "assistant"},
+                    {"tool_calls", json::array({ json {
+                        {"function", json {
+                            {"arguments", item.at("arguments")},
+                            {"name",      item.at("name")},
+                        }},
+                        {"id",   item.at("call_id")},
+                        {"type", "function"},
+                    }})},
+                };
+
+                if (!chatcmpl_messages.empty() && chatcmpl_messages.back().contains("reasoning_content")) {
+                    // Move reasoning content from dummy message to tool call message
+                    msg["reasoning_content"] = chatcmpl_messages.back().at("reasoning_content");
+                    chatcmpl_messages.pop_back();
+                }
+                chatcmpl_messages.push_back(msg);
+            } else if (exists_and_is_string(item, "call_id") &&
+                (exists_and_is_string(item, "output") || exists_and_is_array(item, "output")) &&
+                exists_and_is_string(item, "type") &&
+                item.at("type") == "function_call_output"
+            ) {
+                // #responses_create-input-input_item_list-item-function_tool_call_output
+                if (item.at("output").is_string()) {
+                    chatcmpl_messages.push_back(json {
+                        {"content",      item.at("output")},
+                        {"role",         "tool"},
+                        {"tool_call_id", item.at("call_id")},
+                    });
+                } else {
+                    json chatcmpl_outputs = item.at("output");
+                    for (json & chatcmpl_output : chatcmpl_outputs) {
+                        if (!chatcmpl_output.contains("type") || chatcmpl_output.at("type") != "input_text") {
+                            throw std::invalid_argument("Output of tool call should be 'Input text'");
+                        }
+                        chatcmpl_output["type"] = "text";
+                    }
+                    chatcmpl_messages.push_back(json {
+                        {"content",      chatcmpl_outputs},
+                        {"role",         "tool"},
+                        {"tool_call_id", item.at("call_id")},
+                    });
+                }
+            } else if (// exists_and_is_string(item, "id") &&
+                // item["id"] not sent by codex-cli
+                exists_and_is_array(item, "summary") &&
+                exists_and_is_string(item, "type") &&
+                item.at("type") == "reasoning") {
+                // #responses_create-input-input_item_list-item-reasoning
+
+                if (!exists_and_is_array(item, "content")) {
+                    throw std::invalid_argument("item['content'] is not an array");
+                }
+                if (item.at("content").empty()) {
+                    throw std::invalid_argument("item['content'] is empty");
+                }
+                if (!exists_and_is_string(item.at("content")[0], "text")) {
+                    throw std::invalid_argument("item['content']['text'] is not a string");
+                }
+
+                // Pack reasoning content in dummy message
+                chatcmpl_messages.push_back(json {
+                    {"role", "assistant"},
+                    {"content", json::array()},
+                    {"reasoning_content", item.at("content")[0].at("text")},
+                });
+            } else {
+                throw std::invalid_argument("Cannot determine type of 'item'");
+            }
+        }
+    } else {
+        throw std::invalid_argument("'input' must be a string or array of objects");
+    }
+
+    // Remove unused dummy message which contains
+    // reasoning content not followed by tool call
+    chatcmpl_messages.erase(std::remove_if(
+        chatcmpl_messages.begin(),
+        chatcmpl_messages.end(),
+        [](const json & x){ return x.contains("role") &&
+            x.at("role") == "assistant" &&
+            x.contains("content") &&
+            x.at("content") == json::array() &&
+            x.contains("reasoning_content");
+        }),
+        chatcmpl_messages.end()
+    );
+
+    chatcmpl_body["messages"] = chatcmpl_messages;
+
+    if (response_body.contains("tools")) {
+        if (!response_body.at("tools").is_array()) {
+            throw std::invalid_argument("'tools' must be an array of objects");
+        }
+        std::vector<json> chatcmpl_tools;
+        for (json resp_tool : response_body.at("tools")) {
+            json chatcmpl_tool;
+
+            if (json_value(resp_tool, "type", std::string()) != "function") {
+                throw std::invalid_argument("'type' of tool must be 'function'");
+            }
+            resp_tool.erase("type");
+            chatcmpl_tool["type"] = "function";
+
+            if (!resp_tool.contains("strict")) {
+                resp_tool["strict"] = true;
+            }
+            chatcmpl_tool["function"] = resp_tool;
+            chatcmpl_tools.push_back(chatcmpl_tool);
+        }
+        chatcmpl_body.erase("tools");
+        chatcmpl_body["tools"] = chatcmpl_tools;
+    }
+
+    if (response_body.contains("max_output_tokens")) {
+        chatcmpl_body.erase("max_output_tokens");
+        chatcmpl_body["max_tokens"] = response_body["max_output_tokens"];
+    }
+
+    return chatcmpl_body;
+}
+
 json convert_anthropic_to_oai(const json & body) {
     json oai_body;
 
@@ -1483,6 +1759,24 @@ std::string format_oai_sse(const json & data) {
     return ss.str();
 }
 
+std::string format_oai_resp_sse(const json & data) {
+    std::ostringstream ss;
+    auto send_single = [&ss](const json & event_obj) {
+        ss << "event: " << event_obj.at("event").get<std::string>() << "\n";
+        ss << "data: " << safe_json_to_str(event_obj.at("data")) << "\n\n";
+    };
+
+    if (data.is_array()) {
+        for (const auto & item : data) {
+            send_single(item);
+        }
+    } else {
+        send_single(data);
+    }
+
+    return ss.str();
+}
+
 std::string format_anthropic_sse(const json & data) {
     std::ostringstream ss;
 

tools/server/server-common.h

@@ -13,8 +13,6 @@
 #include <vector>
 #include <cinttypes>
 
-const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
-
 using json = nlohmann::ordered_json;
 
 #define SLT_INF(slot, fmt, ...) LOG_INF("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, ((slot).task ? (slot).task->id : -1), __VA_ARGS__)
@@ -274,27 +272,31 @@ std::vector<server_tokens> tokenize_input_prompts(
 // OAI utils
 //
 
-// used by /completions endpoint
-json oaicompat_completion_params_parse(const json & body);
-
-struct oaicompat_parser_options {
+// global server parameters for chat formatting / parsing
+struct server_chat_params {
     bool use_jinja;
     bool prefill_assistant;
     common_reasoning_format reasoning_format;
-    std::map<std::string,std::string> chat_template_kwargs;
-    common_chat_templates * tmpls;
+    std::map<std::string, std::string> chat_template_kwargs; // mapping key --> json value
+    common_chat_templates_ptr tmpls;
     bool allow_image;
     bool allow_audio;
     bool enable_thinking = true;
     std::string media_path;
 };
 
+// used by /completions endpoint
+json oaicompat_completion_params_parse(const json & body);
+
 // used by /chat/completions endpoint
 json oaicompat_chat_params_parse(
     json & body, /* openai api json semantics */
-    const oaicompat_parser_options & opt,
+    const server_chat_params & opt,
     std::vector<raw_buffer> & out_files);
 
+// convert OpenAI Responses API format to OpenAI Chat Completions API format
+json convert_responses_to_chatcmpl(const json & body);
+
 // convert Anthropic Messages API format to OpenAI Chat Completions API format
 json convert_anthropic_to_oai(const json & body);
 
@@ -332,6 +334,8 @@ std::string tokens_to_output_formatted_string(const llama_context * ctx, const l
 // note: if data is a json array, it will be sent as multiple events, one per item
 std::string format_oai_sse(const json & data);
 
+std::string format_oai_resp_sse(const json & data);
+
 // format Anthropic-style SSE with event types
 std::string format_anthropic_sse(const json & data);
 

tools/server/server-context.cpp

@@ -534,8 +534,8 @@ public:
     server_queue    queue_tasks;
     server_response queue_results;
 
-    common_chat_templates_ptr chat_templates;
-    oaicompat_parser_options  oai_parser_opt;
+    // note: chat_params must not be refreshed upon existing sleeping state
+    server_chat_params chat_params;
 
     ~server_context_impl() {
         if (!sleeping) {
@@ -688,15 +688,6 @@ private:
             llama_init_dft->free_context();
         }
 
-        chat_templates = common_chat_templates_init(model, params_base.chat_template);
-        try {
-            common_chat_format_example(chat_templates.get(), params.use_jinja, params.default_template_kwargs);
-        } catch (const std::exception & e) {
-            SRV_WRN("%s: Chat template parsing error: %s\n", __func__, e.what());
-            SRV_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__);
-            chat_templates = common_chat_templates_init(model, "chatml");
-        }
-
         std::string & mmproj_path = params_base.mmproj.path;
         if (!mmproj_path.empty()) {
             if (!is_resume) {
@@ -845,30 +836,6 @@ private:
             model_name = model_path.filename().string();
         }
 
-        // thinking is enabled if:
-        // 1. It's not explicitly disabled (reasoning_budget == 0)
-        // 2. The chat template supports it
-        const bool enable_thinking = params_base.use_jinja && params_base.reasoning_budget != 0 && common_chat_templates_support_enable_thinking(chat_templates.get());
-        SRV_INF("thinking = %d\n", enable_thinking);
-
-        oai_parser_opt = {
-            /* use_jinja             */ params_base.use_jinja,
-            /* prefill_assistant     */ params_base.prefill_assistant,
-            /* reasoning_format      */ params_base.reasoning_format,
-            /* chat_template_kwargs  */ params_base.default_template_kwargs,
-            /* common_chat_templates */ chat_templates.get(),
-            /* allow_image           */ mctx ? mtmd_support_vision(mctx) : false,
-            /* allow_audio           */ mctx ? mtmd_support_audio (mctx) : false,
-            /* enable_thinking       */ enable_thinking,
-            /* media_path            */ params_base.media_path,
-        };
-
-        // print sample chat example to make it clear which template is used
-        // @ngxson modern templates are too long, spam the logs; printing the example is enough
-        LOG_INF("%s: chat template, example_format: '%s'\n", __func__,
-        //      common_chat_templates_source(chat_templates.get()),
-                common_chat_format_example(chat_templates.get(), params_base.use_jinja, params_base.default_template_kwargs).c_str());
-
         if (!is_resume) {
             return init();
         }
@@ -907,6 +874,42 @@ private:
             }
         }
 
+        // populate chat template params
+        {
+            common_chat_templates_ptr chat_templates;
+
+            try {
+                chat_templates = common_chat_templates_init(model, params_base.chat_template);
+
+                LOG_INF("%s: chat template, example_format: '%s'\n", __func__,
+                    common_chat_format_example(chat_templates.get(), params_base.use_jinja, params_base.default_template_kwargs).c_str());
+
+            } catch (const std::exception & e) {
+                SRV_ERR("%s: chat template parsing error: %s\n", __func__, e.what());
+                SRV_ERR("%s: please consider disabling jinja via --no-jinja, or use a custom chat template via --chat-template\n", __func__);
+                SRV_ERR("%s: for example: --no-jinja --chat-template chatml\n", __func__);
+                return false;
+            }
+
+            // thinking is enabled if:
+            // 1. It's not explicitly disabled (reasoning_budget == 0)
+            // 2. The chat template supports it
+            const bool enable_thinking = params_base.use_jinja && params_base.reasoning_budget != 0 && common_chat_templates_support_enable_thinking(chat_templates.get());
+            SRV_INF("%s: chat template, thinking = %d\n", __func__, enable_thinking);
+
+            chat_params = {
+                /* use_jinja             */ params_base.use_jinja,
+                /* prefill_assistant     */ params_base.prefill_assistant,
+                /* reasoning_format      */ params_base.reasoning_format,
+                /* chat_template_kwargs  */ params_base.default_template_kwargs,
+                /* tmpls                 */ std::move(chat_templates),
+                /* allow_image           */ mctx ? mtmd_support_vision(mctx) : false,
+                /* allow_audio           */ mctx ? mtmd_support_audio (mctx) : false,
+                /* enable_thinking       */ enable_thinking,
+                /* media_path            */ params_base.media_path,
+            };
+        }
+
         return true;
     }
 
@@ -1326,11 +1329,12 @@ private:
     }
 
     void populate_token_probs(const server_slot & slot, completion_token_output & result, bool post_sampling, bool special, int idx) const {
-        const size_t n_probs = slot.task->params.sampling.n_probs;
+        const size_t n_probs_request = slot.task->params.sampling.n_probs;
 
         if (post_sampling) {
             const auto * cur_p = common_sampler_get_candidates(slot.smpl.get(), true);
             const size_t max_probs = cur_p->size;
+            const size_t n_probs = std::min(max_probs, n_probs_request);
 
             // set probability for sampled token
             for (size_t i = 0; i < max_probs; i++) {
@@ -1341,8 +1345,8 @@ private:
             }
 
             // set probability for top n_probs tokens
-            result.probs.reserve(max_probs);
-            for (size_t i = 0; i < std::min(max_probs, n_probs); i++) {
+            result.probs.reserve(n_probs);
+            for (size_t i = 0; i < n_probs; i++) {
                 result.probs.push_back({
                     cur_p->data[i].id,
                     common_token_to_piece(ctx, cur_p->data[i].id, special),
@@ -1352,9 +1356,11 @@ private:
         } else {
             // TODO: optimize this with min-p optimization
             std::vector<llama_token_data> cur = get_token_probabilities(ctx, idx);
+            const size_t max_probs = cur.size();
+            const size_t n_probs = std::min(max_probs, n_probs_request);
 
             // set probability for sampled token
-            for (size_t i = 0; i < cur.size(); i++) {
+            for (size_t i = 0; i < max_probs; i++) {
                 // set probability for sampled token
                 if (cur[i].id == result.tok) {
                     result.prob = cur[i].p;
@@ -1364,7 +1370,7 @@ private:
 
             // set probability for top n_probs tokens
             result.probs.reserve(n_probs);
-            for (size_t i = 0; i < std::min(cur.size(), n_probs); i++) {
+            for (size_t i = 0; i < n_probs; i++) {
                 result.probs.push_back({
                     cur[i].id,
                     common_token_to_piece(ctx, cur[i].id, special),
@@ -1585,32 +1591,14 @@ private:
 
     // tokenize the input if it's set by CLI, return false on error
     bool tokenize_cli_input(server_task & task) {
-        GGML_ASSERT(task.cli_input != nullptr);
         try {
-            auto & opt = oai_parser_opt;
-            common_chat_templates_inputs inputs;
-            inputs.messages              = common_chat_msgs_parse_oaicompat(task.cli_input);
-            inputs.tools                 = {}; // TODO
-            inputs.tool_choice           = COMMON_CHAT_TOOL_CHOICE_NONE;
-            inputs.json_schema           = ""; // TODO
-            inputs.grammar               = ""; // TODO
-            inputs.use_jinja             = opt.use_jinja;
-            inputs.parallel_tool_calls   = false;
-            inputs.add_generation_prompt = true;
-            inputs.reasoning_format      = opt.reasoning_format;
-            inputs.enable_thinking       = opt.enable_thinking;
-
-            // Apply chat template to the list of messages
-            auto chat_params = common_chat_templates_apply(opt.tmpls, inputs);
-
-            // tokenize the resulting prompt
-            auto & prompt = chat_params.prompt;
+            auto & prompt = task.cli_prompt;
             if (mctx != nullptr) {
                 task.tokens = process_mtmd_prompt(mctx, prompt, task.cli_files);
             } else {
                 task.tokens = std::move(tokenize_input_prompts(vocab, mctx, prompt, true, true)[0]);
             }
-            task.cli_input.clear();
+            task.cli_prompt.clear();
             task.cli_files.clear();
         } catch (const std::exception & e) {
             send_error(task, std::string("Failed to format input: ") + e.what(), ERROR_TYPE_INVALID_REQUEST);
@@ -1686,7 +1674,7 @@ private:
                 {
                     // special case: if input is provided via CLI, tokenize it first
                     // otherwise, no need to tokenize as it's already done inside the HTTP thread
-                    if (task.cli_input != nullptr) {
+                    if (task.cli) {
                         if (!tokenize_cli_input(task)) {
                             break;
                         }
@@ -2898,8 +2886,6 @@ server_response_reader server_context::get_response_reader() {
 }
 
 server_context_meta server_context::get_meta() const {
-    auto tool_use_src = common_chat_templates_source(impl->chat_templates.get(), "tool_use");
-
     auto bos_id = llama_vocab_bos(impl->vocab);
     auto eos_id = llama_vocab_eos(impl->vocab);
     auto bos_token_str = bos_id != LLAMA_TOKEN_NULL ? common_token_to_piece(impl->ctx, bos_id, true) : "";
@@ -2910,14 +2896,13 @@ server_context_meta server_context::get_meta() const {
         /* model_name             */ impl->model_name,
         /* model_path             */ impl->params_base.model.path,
         /* has_mtmd               */ impl->mctx != nullptr,
-        /* has_inp_image          */ impl->oai_parser_opt.allow_image,
-        /* has_inp_audio          */ impl->oai_parser_opt.allow_audio,
+        /* has_inp_image          */ impl->chat_params.allow_image,
+        /* has_inp_audio          */ impl->chat_params.allow_audio,
         /* json_webui_settings    */ impl->json_webui_settings,
         /* slot_n_ctx             */ impl->get_slot_n_ctx(),
         /* pooling_type           */ llama_pooling_type(impl->ctx),
 
-        /* chat_template          */ common_chat_templates_source(impl->chat_templates.get()),
-        /* chat_template_tool_use */ tool_use_src ? tool_use_src : "",
+        /* chat_params            */ impl->chat_params,
 
         /* bos_token_str          */ bos_token_str,
         /* eos_token_str          */ eos_token_str,
@@ -3088,6 +3073,8 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
         json first_result_json = first_result->to_json();
         if (res_type == TASK_RESPONSE_TYPE_ANTHROPIC) {
             res->data = format_anthropic_sse(first_result_json);
+        } else if (res_type == TASK_RESPONSE_TYPE_OAI_RESP) {
+            res->data = format_oai_resp_sse(first_result_json);
         } else {
             res->data = format_oai_sse(first_result_json);
         }
@@ -3122,13 +3109,16 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
 
                 // check if there is more data
                 if (!rd.has_next()) {
-                    if (res_type == TASK_RESPONSE_TYPE_ANTHROPIC) {
-                        // Anthropic doesn't send [DONE], message_stop was already sent
-                        output = "";
-                    } else if (res_type != TASK_RESPONSE_TYPE_NONE) {
-                        output = "data: [DONE]\n\n";
-                    } else {
-                        output = "";
+                    switch (res_type) {
+                        case TASK_RESPONSE_TYPE_NONE:
+                        case TASK_RESPONSE_TYPE_OAI_RESP:
+                        case TASK_RESPONSE_TYPE_ANTHROPIC:
+                            output = "";
+                            break;
+
+                        default:
+                            output = "data: [DONE]\n\n";
+                            break;
                     }
                     SRV_DBG("%s", "all results received, terminating stream\n");
                     return false; // no more data, terminate
@@ -3156,6 +3146,8 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
                     json res_json = result->to_json();
                     if (res_type == TASK_RESPONSE_TYPE_ANTHROPIC) {
                         output = format_anthropic_sse(res_json);
+                    } else if (res_type == TASK_RESPONSE_TYPE_OAI_RESP) {
+                        output = format_oai_resp_sse(res_json);
                     } else {
                         output = format_oai_sse(res_json);
                     }
@@ -3199,8 +3191,8 @@ void server_routes::init_routes() {
 
         // this endpoint can be accessed during sleeping
         // the next LOC is to avoid someone accidentally use ctx_server
-        bool server_ctx; // do NOT delete this line
-        GGML_UNUSED(server_ctx);
+        bool ctx_server; // do NOT delete this line
+        GGML_UNUSED(ctx_server);
 
         res->ok({{"status", "ok"}});
         return res;
@@ -3390,8 +3382,8 @@ void server_routes::init_routes() {
 
         // this endpoint can be accessed during sleeping
         // the next LOC is to avoid someone accidentally use ctx_server
-        bool server_ctx; // do NOT delete this line
-        GGML_UNUSED(server_ctx);
+        bool ctx_server; // do NOT delete this line
+        GGML_UNUSED(ctx_server);
 
         task_params tparams;
         tparams.sampling = params.sampling;
@@ -3400,6 +3392,9 @@ void server_routes::init_routes() {
             { "n_ctx",  meta->slot_n_ctx },
         };
 
+        std::string tmpl_default = common_chat_templates_source(meta->chat_params.tmpls.get(), "");
+        std::string tmpl_tools   = common_chat_templates_source(meta->chat_params.tmpls.get(), "tool_use");
+
         json props = {
             { "default_generation_settings", default_generation_settings_for_props },
             { "total_slots",                 params.n_parallel },
@@ -3414,15 +3409,15 @@ void server_routes::init_routes() {
             { "endpoint_metrics",            params.endpoint_metrics },
             { "webui",                       params.webui },
             { "webui_settings",              meta->json_webui_settings },
-            { "chat_template",               meta->chat_template },
+            { "chat_template",               tmpl_default },
             { "bos_token",                   meta->bos_token_str },
             { "eos_token",                   meta->eos_token_str },
             { "build_info",                  meta->build_info },
             { "is_sleeping",                 queue_tasks.is_sleeping() },
         };
         if (params.use_jinja) {
-            if (!meta->chat_template_tool_use.empty()) {
-                props["chat_template_tool_use"] = meta->chat_template_tool_use;
+            if (!tmpl_tools.empty()) {
+                props["chat_template_tool_use"] = tmpl_tools;
             }
         }
         res->ok(props);
@@ -3443,6 +3438,7 @@ void server_routes::init_routes() {
 
     this->get_api_show = [this](const server_http_req &) {
         auto res = create_response();
+        std::string tmpl_default = common_chat_templates_source(meta->chat_params.tmpls.get(), "");
         json data = {
             {
                 "model_info", {
@@ -3451,7 +3447,7 @@ void server_routes::init_routes() {
             },
             {"modelfile", ""},
             {"parameters", ""},
-            {"template", meta->chat_template},
+            {"template", tmpl_default},
             {"details", {
                 {"parent_model", ""},
                 {"format", "gguf"},
@@ -3576,7 +3572,7 @@ void server_routes::init_routes() {
         json body = json::parse(req.body);
         json body_parsed = oaicompat_chat_params_parse(
             body,
-            ctx_server.oai_parser_opt,
+            meta->chat_params,
             files);
         return handle_completions_impl(
             req,
@@ -3586,13 +3582,29 @@ void server_routes::init_routes() {
             TASK_RESPONSE_TYPE_OAI_CHAT);
     };
 
+    this->post_responses_oai = [this](const server_http_req & req) {
+        auto res = create_response();
+        std::vector<raw_buffer> files;
+        json body = convert_responses_to_chatcmpl(json::parse(req.body));
+        json body_parsed = oaicompat_chat_params_parse(
+            body,
+            meta->chat_params,
+            files);
+        return handle_completions_impl(
+            req,
+            SERVER_TASK_TYPE_COMPLETION,
+            body_parsed,
+            files,
+            TASK_RESPONSE_TYPE_OAI_RESP);
+    };
+
     this->post_anthropic_messages = [this](const server_http_req & req) {
         auto res = create_response();
         std::vector<raw_buffer> files;
         json body = convert_anthropic_to_oai(json::parse(req.body));
         json body_parsed = oaicompat_chat_params_parse(
             body,
-            ctx_server.oai_parser_opt,
+            meta->chat_params,
             files);
         return handle_completions_impl(
             req,
@@ -3608,7 +3620,7 @@ void server_routes::init_routes() {
         json body = convert_anthropic_to_oai(json::parse(req.body));
         json body_parsed = oaicompat_chat_params_parse(
             body,
-            ctx_server.oai_parser_opt,
+            meta->chat_params,
             files);
 
         json prompt = body_parsed.at("prompt");
@@ -3624,7 +3636,7 @@ void server_routes::init_routes() {
         json body = json::parse(req.body);
         json data = oaicompat_chat_params_parse(
             body,
-            ctx_server.oai_parser_opt,
+            meta->chat_params,
             files);
         res->ok({{ "prompt", std::move(data.at("prompt")) }});
         return res;
@@ -3635,8 +3647,8 @@ void server_routes::init_routes() {
 
         // this endpoint can be accessed during sleeping
         // the next LOC is to avoid someone accidentally use ctx_server
-        bool server_ctx; // do NOT delete this line
-        GGML_UNUSED(server_ctx);
+        bool ctx_server; // do NOT delete this line
+        GGML_UNUSED(ctx_server);
 
         json models = {
             {"models", {

tools/server/server-context.h

@@ -20,9 +20,8 @@ struct server_context_meta {
     int slot_n_ctx;
     enum llama_pooling_type pooling_type;
 
-    // chat template
-    std::string chat_template;
-    std::string chat_template_tool_use;
+    // chat params
+    server_chat_params & chat_params;
 
     // tokens
     std::string bos_token_str;
@@ -95,6 +94,7 @@ struct server_routes {
     server_http_context::handler_t post_completions;
     server_http_context::handler_t post_completions_oai;
     server_http_context::handler_t post_chat_completions;
+    server_http_context::handler_t post_responses_oai;
     server_http_context::handler_t post_anthropic_messages;
     server_http_context::handler_t post_anthropic_count_tokens;
     server_http_context::handler_t post_apply_template;