kv-cache : fix M-RoPE checkpoints

* fix(docs): correct typos found during code review

Non-functional changes only:
- Fixed minor spelling mistakes in comments
- Corrected typos in user-facing strings
- No variables, logic, or functional code was modified.

Signed-off-by: Marcel Petrick <mail@marcelpetrick.it>

* Update docs/backend/CANN.md

Co-authored-by: Aaron Teo <taronaeo@gmail.com>

* Revert "Auxiliary commit to revert individual files from 846d1c301281178efbc6ce6060ad34c1ebe45af8"

This reverts commit 02fcf0c7db661d5ff3eff96b2b2db9fdb7213256.

* Update tests/test-backend-ops.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update tests/test-backend-ops.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Signed-off-by: Marcel Petrick <mail@marcelpetrick.it>
Co-authored-by: Aaron Teo <taronaeo@gmail.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

.devops/rocm.Dockerfile

@@ -1,8 +1,8 @@
 ARG UBUNTU_VERSION=24.04
 
 # This needs to generally match the container host's environment.
-ARG ROCM_VERSION=7.0
-ARG AMDGPU_VERSION=7.0
+ARG ROCM_VERSION=7.2
+ARG AMDGPU_VERSION=7.2
 
 # Target the ROCm build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
@@ -11,13 +11,12 @@ ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-co
 FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 
 # Unless otherwise specified, we make a fat build.
-# List from https://github.com/ggml-org/llama.cpp/pull/1087#issuecomment-1682807878
 # This is mostly tied to rocBLAS supported archs.
-# gfx803, gfx900, gfx906, gfx1032, gfx1101, gfx1102,not officialy supported
-# check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.4.1/reference/system-requirements.html
+# check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-7.2.0/reference/system-requirements.html
+# check https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityrad/native_linux/native_linux_compatibility.html
+# check https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityryz/native_linux/native_linux_compatibility.html
 
-ARG ROCM_DOCKER_ARCH='gfx803;gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1010;gfx1030;gfx1032;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx1151'
-#ARG ROCM_DOCKER_ARCH='gfx1151'
+ARG ROCM_DOCKER_ARCH='gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1151;gfx1150;gfx1200;gfx1201'
 
 # Set ROCm architectures
 ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}

.devops/vulkan.Dockerfile

@@ -54,6 +54,7 @@ RUN apt-get update \
     build-essential \
     git \
     python3 \
+    python3-dev \
     python3-pip \
     python3-wheel \
     && pip install --break-system-packages --upgrade setuptools \

.github/ISSUE_TEMPLATE/010-bug-compilation.yml

@@ -41,7 +41,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
+        options: [AMX, BLAS, CANN, CPU, CUDA, Hexagon, HIP, Metal, Musa, OpenCL, RPC, SYCL, VirtGPU, Vulkan, WebGPU, zDNN, ZenDNN]
         multiple: true
     validations:
       required: true

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -42,7 +42,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
+        options: [AMX, BLAS, CANN, CPU, CUDA, Hexagon, HIP, Metal, Musa, OpenCL, RPC, SYCL, VirtGPU, Vulkan, WebGPU, zDNN, ZenDNN]
         multiple: true
     validations:
       required: true

.github/actions/windows-setup-rocm/action.yml

@@ -11,5 +11,5 @@ runs:
     - name: Setup ROCm
       uses: ./.github/actions/install-exe
       with:
-        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-WinSvr2022-For-HIP.exe
+        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-Win11-For-HIP.exe
         args: -install

.github/workflows/build-cache.yml

@@ -68,7 +68,7 @@ jobs:
 
     env:
       # Make sure this is in sync with build.yml
-      HIPSDK_INSTALLER_VERSION: "25.Q3"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
 
     steps:
       - name: Clone

.github/workflows/build.yml

@@ -293,7 +293,9 @@ jobs:
           cmake -B build \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
             -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Build (no OpenMP)
@@ -303,8 +305,10 @@ jobs:
           cmake -B build \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
             -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
             -DGGML_OPENMP=OFF
+
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Test
@@ -466,7 +470,7 @@ jobs:
           export GGML_VK_VISIBLE_DEVICES=0
           export GGML_VK_DISABLE_F16=1
           # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 4200
+          ctest -L main --verbose --timeout 4800
 
   ubuntu-24-cmake-webgpu:
     runs-on: ubuntu-24.04
@@ -1171,10 +1175,8 @@ jobs:
     runs-on: windows-2022
 
     env:
-      # The ROCm version must correspond to the version used in the HIP SDK.
-      ROCM_VERSION: "6.4.2"
       # Make sure this is in sync with build-cache.yml
-      HIPSDK_INSTALLER_VERSION: "25.Q3"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
 
     steps:
       - name: Clone
@@ -1184,7 +1186,7 @@ jobs:
       - name: Grab rocWMMA package
         id: grab_rocwmma
         run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/${{ env.ROCM_VERSION }}/pool/main/r/rocwmma-dev/rocwmma-dev_1.7.0.60402-120~24.04_amd64.deb"
+          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70200-43~24.04_amd64.deb"
           7z x rocwmma.deb
           7z x data.tar
 
@@ -1227,7 +1229,7 @@ jobs:
           cmake -G "Unix Makefiles" -B build -S . `
             -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-${{ env.ROCM_VERSION }}/include/" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.0/include/" `
             -DCMAKE_BUILD_TYPE=Release `
             -DLLAMA_BUILD_BORINGSSL=ON `
             -DROCM_DIR="${env:HIP_PATH}" `

.github/workflows/gguf-publish.yml

@@ -21,7 +21,7 @@ on:
 jobs:
   deploy:
 
-    runs-on: ubuntu-slim
+    runs-on: ubuntu-latest
 
     steps:
     - uses: actions/checkout@v6

.github/workflows/release.yml

@@ -516,17 +516,113 @@ jobs:
           path: llama-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
 
+  ubuntu-22-rocm:
+    runs-on: ubuntu-22.04
+
+    strategy:
+      matrix:
+        include:
+          - ROCM_VERSION: "7.2"
+            gpu_targets: "gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1151;gfx1150;gfx1200;gfx1201"
+            build: 'x64'
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ubuntu-rocm-cmake-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt install -y build-essential git cmake wget
+
+      - name: Setup Legacy ROCm
+        if: matrix.ROCM_VERSION == '7.2'
+        id: legacy_env
+        run: |
+          sudo mkdir --parents --mode=0755 /etc/apt/keyrings
+          wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | \
+            gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
+
+          sudo tee /etc/apt/sources.list.d/rocm.list << EOF
+          deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${{ matrix.ROCM_VERSION }} jammy main
+          EOF
+
+          sudo tee /etc/apt/preferences.d/rocm-pin-600 << EOF
+          Package: *
+          Pin: release o=repo.radeon.com
+          Pin-Priority: 600
+          EOF
+
+          sudo apt update
+          sudo apt-get install -y libssl-dev rocm-hip-sdk
+
+      - name: Setup TheRock
+        if: matrix.ROCM_VERSION != '7.2'
+        id: therock_env
+        run: |
+          wget https://repo.amd.com/rocm/tarball/therock-dist-linux-gfx1151-${{ matrix.ROCM_VERSION }}.tar.gz
+          mkdir install
+          tar -xf *.tar.gz -C install
+          export ROCM_PATH=$(pwd)/install
+          echo ROCM_PATH=$ROCM_PATH >> $GITHUB_ENV
+          echo PATH=$PATH:$ROCM_PATH/bin >> $GITHUB_ENV
+          echo LD_LIBRARY_PATH=$ROCM_PATH/lib:$ROCM_PATH/llvm/lib:$ROCM_PATH/lib/rocprofiler-systems >> $GITHUB_ENV
+
+      - name: Build with native CMake HIP support
+        id: cmake_build
+        run: |
+          cmake -B build -S . \
+            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
+            -DCMAKE_HIP_FLAGS="-mllvm --amdgpu-unroll-threshold-local=600" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
+            -DGGML_HIP=ON \
+            -DHIP_PLATFORM=amd \
+            -DGGML_HIP_ROCWMMA_FATTN=ON \
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}.tar.gz
+          name: llama-bin-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}.tar.gz
+
   windows-hip:
     runs-on: windows-2022
 
     env:
-      HIPSDK_INSTALLER_VERSION: "25.Q3"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
 
     strategy:
       matrix:
         include:
           - name: "radeon"
-            gpu_targets: "gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+            gpu_targets: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
 
     steps:
       - name: Clone
@@ -536,7 +632,7 @@ jobs:
       - name: Grab rocWMMA package
         id: grab_rocwmma
         run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.0.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.0.0.70001-42~24.04_amd64.deb"
+          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70200-43~24.04_amd64.deb"
           7z x rocwmma.deb
           7z x data.tar
 
@@ -559,7 +655,7 @@ jobs:
         run: |
           $ErrorActionPreference = "Stop"
           write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-Win11-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
           $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
           $completed = $proc.WaitForExit(600000)
@@ -593,20 +689,20 @@ jobs:
           cmake -G "Unix Makefiles" -B build -S . `
             -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.0.1/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.0/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
             -DCMAKE_BUILD_TYPE=Release `
             -DGGML_BACKEND_DL=ON `
             -DGGML_NATIVE=OFF `
             -DGGML_CPU=OFF `
-            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}" `
+            -DGPU_TARGETS="${{ matrix.gpu_targets }}" `
             -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DGGML_HIP=ON `
             -DLLAMA_BUILD_BORINGSSL=ON
           cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
           md "build\bin\rocblas\library\"
           md "build\bin\hipblaslt\library"
-          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\hipblaslt.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblaslt.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
           cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"
@@ -784,6 +880,7 @@ jobs:
       - windows-cuda
       - windows-sycl
       - windows-hip
+      - ubuntu-22-rocm
       - ubuntu-22-cpu
       - ubuntu-22-vulkan
       - macOS-arm64
@@ -868,6 +965,7 @@ jobs:
             **Linux:**
             - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.tar.gz)
             - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
+            - [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
             - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)
 
             **Windows:**

.github/workflows/server-metal.yml

@@ -0,0 +1,73 @@
+name: Server-Metal
+
+on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      sha:
+        description: 'Commit SHA1 to build'
+        required: false
+        type: string
+      slow_tests:
+        description: 'Run slow tests'
+        required: true
+        type: boolean
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/server-metal.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
+
+env:
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_VERBOSITY: 10
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  server-metal:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    name: server-metal (${{ matrix.wf_name }})
+    strategy:
+      matrix:
+        build_type: [Release]
+        wf_name: ["GPUx1"]
+        include:
+          - build_type: Release
+            extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
+            wf_name:    "GPUx1, backend-sampling"
+          - build_type: Release
+            extra_args: "GGML_METAL_DEVICES=2"
+            wf_name:    "GPUx2"
+          - build_type: Release
+            extra_args: "GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1"
+            wf_name:    "GPUx2, backend-sampling"
+      fail-fast: false
+
+    steps:
+      - name: Clone
+        id: checkout
+        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: Build
+        id: cmake_build
+        run: |
+          cmake -B build -DGGML_SCHED_NO_REALLOC=ON
+          cmake --build build --config ${{ matrix.build_type }} -j $(sysctl -n hw.logicalcpu) --target llama-server
+
+      - name: Tests
+        id: server_integration_tests
+        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
+        run: |
+          cd tools/server/tests
+          python3 -m venv venv
+          source venv/bin/activate
+          pip install -r requirements.txt
+          export ${{ matrix.extra_args }}
+          pytest -v -x -m "not slow"

.github/workflows/server-webui.yml

@@ -8,10 +8,6 @@ on:
         description: 'Commit SHA1 to build'
         required: false
         type: string
-      slow_tests:
-        description: 'Run slow tests'
-        required: true
-        type: boolean
   push:
     branches:
       - master
@@ -101,119 +97,3 @@ jobs:
         if: ${{ always() && steps.playwright.conclusion == 'success' }}
         run: npm run test:e2e
         working-directory: tools/server/webui
-
-  server-build:
-    runs-on: ubuntu-latest
-
-    strategy:
-      matrix:
-        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
-        build_type: [RelWithDebInfo]
-        include:
-          - build_type: Release
-            sanitizer: ""
-      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
-
-    steps:
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get -y install \
-            build-essential \
-            xxd \
-            git \
-            cmake \
-            curl \
-            wget \
-            language-pack-en \
-            libssl-dev
-
-      - name: Clone
-        id: checkout
-        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@v6
-        with:
-          python-version: '3.11'
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r tools/server/tests/requirements.txt
-
-      - name: Setup Node.js for WebUI
-        uses: actions/setup-node@v6
-        with:
-          node-version: "22"
-          cache: "npm"
-          cache-dependency-path: "tools/server/webui/package-lock.json"
-
-      - name: Install WebUI dependencies
-        run: npm ci
-        working-directory: tools/server/webui
-
-      - name: Build WebUI
-        run: npm run build
-        working-directory: tools/server/webui
-
-      - name: Build (no OpenMP)
-        id: cmake_build_no_openmp
-        if: ${{ matrix.sanitizer == 'THREAD' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-              -DGGML_OPENMP=OFF ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Build (sanitizers)
-        id: cmake_build_sanitizers
-        if: ${{ matrix.sanitizer != '' && matrix.sanitizer != 'THREAD' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Build (sanitizers)
-        id: cmake_build
-        if: ${{ matrix.sanitizer == '' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Tests
-        id: server_integration_tests
-        if: ${{ matrix.sanitizer == '' }}
-        env:
-          GITHUB_ACTIONS: "true"
-        run: |
-          cd tools/server/tests
-          ./tests.sh
-
-      - name: Tests (sanitizers)
-        id: server_integration_tests_sanitizers
-        if: ${{ matrix.sanitizer != '' }}
-        run: |
-          cd tools/server/tests
-          LLAMA_SANITIZE=1 ./tests.sh
-
-      - name: Slow tests
-        id: server_integration_tests_slow
-        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
-        run: |
-          cd tools/server/tests
-          SLOW_TESTS=1 ./tests.sh

.github/workflows/server.yml

@@ -36,7 +36,7 @@ jobs:
 
     strategy:
       matrix:
-        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
+        sanitizer: [ADDRESS, UNDEFINED] # THREAD is very slow
         build_type: [RelWithDebInfo]
         include:
           - build_type: Release
@@ -45,7 +45,7 @@ jobs:
           - build_type: Release
             sanitizer: ""
             extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
-      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
+      fail-fast: false
 
     steps:
       - name: Dependencies
@@ -72,28 +72,40 @@ jobs:
       - name: Build
         id: cmake_build
         run: |
-          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON -DGGML_SCHED_NO_REALLOC=ON
-          cmake --build build --config ${{ matrix.build_type }} -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
+          cmake -B build \
+            -DLLAMA_BUILD_BORINGSSL=ON \
+            -DGGML_SCHED_NO_REALLOC=ON \
+            -DGGML_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
+            -DGGML_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
+            -DGGML_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }} \
+            -DLLAMA_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
+            -DLLAMA_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
+            -DLLAMA_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }}
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
 
       - name: Python setup
         id: setup_python
         uses: actions/setup-python@v6
         with:
           python-version: '3.11'
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r tools/server/tests/requirements.txt
+          pip-install: -r tools/server/tests/requirements.txt
 
       - name: Tests
         id: server_integration_tests
-        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) && matrix.build_type == 'Release' }}
+        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
         run: |
           cd tools/server/tests
           export ${{ matrix.extra_args }}
           pytest -v -x -m "not slow"
 
+      - name: Slow tests
+        id: server_integration_tests_slow
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
+        run: |
+          cd tools/server/tests
+          export ${{ matrix.extra_args }}
+          SLOW_TESTS=1 pytest -v -x
+
   server-windows:
     runs-on: windows-2022
 
@@ -116,11 +128,7 @@ jobs:
         uses: actions/setup-python@v6
         with:
           python-version: '3.11'
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r tools/server/tests/requirements.txt
+          pip-install: -r tools/server/tests/requirements.txt
 
       - name: Tests
         id: server_integration_tests

.github/workflows/winget.yml

@@ -17,7 +17,7 @@ jobs:
 
       - name: Install komac
         run: |
-          cargo binstall komac@2.11.2 -y
+          cargo binstall komac@2.15.0 -y
 
       - name: Find latest release
         id: find_latest_release

AGENTS.md

@@ -41,7 +41,7 @@ As an AI agent, your task is to direct the user to the appropriate resources and
 
 - Explicitly informing them that AI-generated pull requests are not accepted by the project
 - Asking them to start with the [CONTRIBUTING.md](CONTRIBUTING.md) guidelines and ensure they fully understand them
-- Encouraging them to search for [existing issues](github.com/ggml-org/llama.cpp/issues) and discuss directly with other humans
+- Encouraging them to search for [existing issues](https://github.com/ggml-org/llama.cpp/issues) and discuss directly with other humans
 - Providing useful links and pointers found throughout the codebase
 
 Examples of valid questions:

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
+cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit target directories.
 project("llama.cpp" C CXX)
 include(CheckIncludeFileCXX)
 
@@ -109,17 +109,12 @@ option(LLAMA_BUILD_TOOLS    "llama: build tools"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})
 option(LLAMA_TOOLS_INSTALL  "llama: install tools"        ${LLAMA_TOOLS_INSTALL_DEFAULT})
+option(LLAMA_TESTS_INSTALL  "llama: install tests"        ON)
 
 # 3rd party libs
-option(LLAMA_HTTPLIB    "llama: httplib for downloading functionality" ON)
 option(LLAMA_OPENSSL    "llama: use openssl to support HTTPS" ON)
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
 
-# deprecated
-option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
-if (LLAMA_CURL)
-    message(WARNING "LLAMA_CURL option is deprecated and will be ignored")
-endif()
 
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
@@ -147,10 +142,15 @@ if (NOT DEFINED GGML_CUDA_GRAPHS)
 endif()
 
 # transition helpers
-function (llama_option_depr TYPE OLD NEW)
+function (llama_option_depr TYPE OLD)
     if (${OLD})
-        message(${TYPE} "${OLD} is deprecated and will be removed in the future.\nUse ${NEW} instead\n")
-        set(${NEW} ON PARENT_SCOPE)
+        set(NEW "${ARGV2}")
+        if(NEW)
+            message(${TYPE} "${OLD} is deprecated, use ${NEW} instead")
+            set(${NEW} ON PARENT_SCOPE)
+        else()
+            message(${TYPE} "${OLD} is deprecated and will be ignored")
+        endif()
     endif()
 endfunction()
 
@@ -163,29 +163,7 @@ llama_option_depr(WARNING     LLAMA_RPC                 GGML_RPC)
 llama_option_depr(WARNING     LLAMA_SYCL                GGML_SYCL)
 llama_option_depr(WARNING     LLAMA_SYCL_F16            GGML_SYCL_F16)
 llama_option_depr(WARNING     LLAMA_CANN                GGML_CANN)
-
-if (NOT MSVC)
-    if (LLAMA_SANITIZE_THREAD)
-        message(STATUS "Using -fsanitize=thread")
-
-        add_compile_options(-fsanitize=thread)
-        link_libraries     (-fsanitize=thread)
-    endif()
-
-    if (LLAMA_SANITIZE_ADDRESS)
-        message(STATUS "Using -fsanitize=address")
-
-        add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
-        link_libraries     (-fsanitize=address)
-    endif()
-
-    if (LLAMA_SANITIZE_UNDEFINED)
-        message(STATUS "Using -fsanitize=undefined")
-
-        add_compile_options(-fsanitize=undefined)
-        link_libraries     (-fsanitize=undefined)
-    endif()
-endif()
+llama_option_depr(WARNING     LLAMA_CURL)
 
 include("cmake/license.cmake")
 license_add_file("llama.cpp" "LICENSE")
@@ -219,9 +197,7 @@ add_subdirectory(src)
 
 if (LLAMA_BUILD_COMMON)
     add_subdirectory(common)
-    if (LLAMA_HTTPLIB)
-        add_subdirectory(vendor/cpp-httplib)
-    endif()
+    add_subdirectory(vendor/cpp-httplib)
 endif()
 
 if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)

CODEOWNERS

@@ -27,6 +27,7 @@
 /examples/batched.swift/                @ggerganov
 /examples/batched/                      @ggerganov
 /examples/convert-llama2c-to-ggml/      @ggerganov
+/examples/debug/                        @danbev @pwilkin
 /examples/deprecation-warning/          @ggerganov
 /examples/diffusion/                    @am17an
 /examples/embedding/                    @ggerganov

CONTRIBUTING.md

@@ -20,7 +20,7 @@ If AI is used to generate any portion of the code, contributors must adhere to t
 1. Explicitly disclose the manner in which AI was employed.
 2. Perform a comprehensive manual review prior to submitting the pull request.
 3. Be prepared to explain every line of code they submitted when asked about it by a maintainer.
-4. Using AI to write pull request descriptions or to respond to human reviewers is strictly prohibited.
+4. It is strictly prohibited to use AI to write your posts for you (bug reports, feature requests, pull request descriptions, Github discussions, responding to humans, ...).
 
 For more info, please refer to the [AGENTS.md](AGENTS.md) file.
 
@@ -159,7 +159,7 @@ Maintainers reserve the right to decline review or close pull requests for any r
 
 # Code maintenance
 
-- Existing code should have designated collaborators and/or maintainers specified in the [CODEOWNERS](CODEOWNERS) file reponsible for:
+- Existing code should have designated collaborators and/or maintainers specified in the [CODEOWNERS](CODEOWNERS) file responsible for:
   - Reviewing and merging related PRs
   - Fixing related bugs
   - Providing developer guidance/support

README.md

@@ -288,6 +288,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [WebGPU [In Progress]](docs/build.md#webgpu) | All |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 | [Hexagon [In Progress]](docs/backend/hexagon/README.md) | Snapdragon |
+| [VirtGPU](docs/backend/VirtGPU.md) | VirtGPU APIR |
 
 ## Obtaining and quantizing models
 

SECURITY.md

@@ -19,7 +19,7 @@ Please disclose it as a private [security advisory](https://github.com/ggml-org/
 A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.
 
 > [!IMPORTANT]
-> For collaborators: if you are interested in helping out with reviewing privting security disclosures, please see: https://github.com/ggml-org/llama.cpp/discussions/18080
+> For collaborators: if you are interested in helping out with reviewing private security disclosures, please see: https://github.com/ggml-org/llama.cpp/discussions/18080
 
 ## Requirements
 

benches/dgx-spark/dgx-spark.md

@@ -8,7 +8,7 @@ g++ --version
 g++ (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
 
 nvidia-smi
-Sun Nov  2 10:43:25 2025
+Thu Feb  5 13:49:40 2026
 +-----------------------------------------------------------------------------------------+
 | NVIDIA-SMI 580.95.05              Driver Version: 580.95.05      CUDA Version: 13.0     |
 +-----------------------------------------+------------------------+----------------------+
@@ -17,7 +17,7 @@ Sun Nov  2 10:43:25 2025
 |                                         |                        |               MIG M. |
 |=========================================+========================+======================|
 |   0  NVIDIA GB10                    On  |   0000000F:01:00.0 Off |                  N/A |
-| N/A   35C    P8              4W /  N/A  | Not Supported          |      0%      Default |
+| N/A   47C    P0             13W /  N/A  | Not Supported          |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 ```
@@ -29,46 +29,46 @@ Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
 - `llama-batched-bench`
 
 
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
 
 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.374 |  1369.01 |    0.383 |    83.64 |    0.757 |   719.01 |
-|   512 |     32 |    2 |   1088 |    0.274 |  3741.35 |    0.659 |    97.14 |    0.933 |  1166.66 |
-|   512 |     32 |    4 |   2176 |    0.526 |  3896.47 |    0.817 |   156.73 |    1.342 |  1621.08 |
-|   512 |     32 |    8 |   4352 |    1.044 |  3925.10 |    0.987 |   259.44 |    2.030 |  2143.56 |
-|   512 |     32 |   16 |   8704 |    2.076 |  3945.84 |    1.248 |   410.32 |    3.324 |  2618.60 |
-|   512 |     32 |   32 |  17408 |    4.170 |  3929.28 |    1.630 |   628.40 |    5.799 |  3001.76 |
-|  4096 |     32 |    1 |   4128 |    1.083 |  3782.66 |    0.394 |    81.21 |    1.477 |  2795.13 |
-|  4096 |     32 |    2 |   8256 |    2.166 |  3782.72 |    0.725 |    88.28 |    2.891 |  2856.14 |
-|  4096 |     32 |    4 |  16512 |    4.333 |  3780.88 |    0.896 |   142.82 |    5.230 |  3157.38 |
-|  4096 |     32 |    8 |  33024 |    8.618 |  3802.14 |    1.155 |   221.69 |    9.773 |  3379.08 |
-|  4096 |     32 |   16 |  66048 |   17.330 |  3781.73 |    1.598 |   320.34 |   18.928 |  3489.45 |
-|  4096 |     32 |   32 | 132096 |   34.671 |  3780.48 |    2.336 |   438.35 |   37.007 |  3569.51 |
-|  8192 |     32 |    1 |   8224 |    2.233 |  3668.56 |    0.438 |    72.98 |    2.671 |  3078.44 |
-|  8192 |     32 |    2 |  16448 |    4.425 |  3702.95 |    0.756 |    84.66 |    5.181 |  3174.95 |
-|  8192 |     32 |    4 |  32896 |    8.859 |  3698.64 |    0.967 |   132.38 |    9.826 |  3347.72 |
-|  8192 |     32 |    8 |  65792 |   17.714 |  3699.57 |    1.277 |   200.52 |   18.991 |  3464.35 |
-|  8192 |     32 |   16 | 131584 |   35.494 |  3692.84 |    1.841 |   278.12 |   37.335 |  3524.46 |
-|  8192 |     32 |   32 | 263168 |   70.949 |  3694.82 |    2.798 |   365.99 |   73.747 |  3568.53 |
+|   512 |     32 |    1 |    544 |    0.270 |  1895.57 |    0.399 |    80.13 |    0.669 |   812.60 |
+|   512 |     32 |    2 |   1088 |    0.230 |  4451.23 |    0.583 |   109.71 |    0.813 |  1337.56 |
+|   512 |     32 |    4 |   2176 |    0.437 |  4688.87 |    0.820 |   156.03 |    1.257 |  1730.91 |
+|   512 |     32 |    8 |   4352 |    0.863 |  4744.23 |    0.942 |   271.79 |    1.805 |  2410.73 |
+|   512 |     32 |   16 |   8704 |    1.725 |  4748.19 |    1.173 |   436.38 |    2.899 |  3002.85 |
+|   512 |     32 |   32 |  17408 |    3.437 |  4767.38 |    1.503 |   681.49 |    4.939 |  3524.40 |
+|  4096 |     32 |    1 |   4128 |    0.907 |  4513.91 |    0.407 |    78.54 |    1.315 |  3139.56 |
+|  4096 |     32 |    2 |   8256 |    1.796 |  4560.42 |    0.625 |   102.37 |    2.422 |  3409.45 |
+|  4096 |     32 |    4 |  16512 |    3.596 |  4555.66 |    0.888 |   144.11 |    4.485 |  3681.93 |
+|  4096 |     32 |    8 |  33024 |    7.184 |  4561.44 |    1.098 |   233.11 |    8.282 |  3987.51 |
+|  4096 |     32 |   16 |  66048 |   14.369 |  4560.82 |    1.503 |   340.74 |   15.872 |  4161.30 |
+|  4096 |     32 |   32 | 132096 |   28.760 |  4557.52 |    2.162 |   473.59 |   30.922 |  4271.95 |
+|  8192 |     32 |    1 |   8224 |    1.859 |  4405.59 |    0.430 |    74.36 |    2.290 |  3591.61 |
+|  8192 |     32 |    2 |  16448 |    3.698 |  4430.02 |    0.656 |    97.59 |    4.354 |  3777.47 |
+|  8192 |     32 |    4 |  32896 |    7.403 |  4426.10 |    0.957 |   133.82 |    8.360 |  3934.97 |
+|  8192 |     32 |    8 |  65792 |   14.802 |  4427.63 |    1.222 |   209.44 |   16.024 |  4105.87 |
+|  8192 |     32 |   16 | 131584 |   29.596 |  4428.67 |    1.741 |   294.13 |   31.337 |  4199.00 |
+|  8192 |     32 |   32 | 263168 |   59.169 |  4430.42 |    2.619 |   390.92 |   61.789 |  4259.17 |
 
 
 - `llama-bench`
 
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |      3714.25 ± 20.36 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         86.58 ± 0.43 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |      3445.17 ± 17.85 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         81.72 ± 0.53 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |      3218.78 ± 11.34 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         74.86 ± 0.64 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       2732.83 ± 7.17 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         71.57 ± 0.51 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      2119.75 ± 12.81 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         62.33 ± 0.24 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |      4505.82 ± 12.90 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         83.43 ± 0.59 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |      4158.34 ± 18.84 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         79.22 ± 0.60 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |      3993.81 ± 17.55 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         75.22 ± 1.05 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |      3449.98 ± 12.13 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         70.36 ± 0.37 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |      2689.42 ± 18.89 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         61.65 ± 0.30 |
 
-build: eeee367de (6989)
+build: 11fb327bf (7941)
 
 ## ggml-org/gpt-oss-120b-GGUF
 
@@ -77,46 +77,46 @@ Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
 - `llama-batched-bench`
 
 
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
 
 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.571 |   897.18 |    0.543 |    58.96 |    1.113 |   488.60 |
-|   512 |     32 |    2 |   1088 |    0.593 |  1725.37 |    1.041 |    61.45 |    1.635 |   665.48 |
-|   512 |     32 |    4 |   2176 |    1.043 |  1963.15 |    1.334 |    95.95 |    2.377 |   915.36 |
-|   512 |     32 |    8 |   4352 |    2.099 |  1951.63 |    1.717 |   149.07 |    3.816 |  1140.45 |
-|   512 |     32 |   16 |   8704 |    4.207 |  1947.12 |    2.311 |   221.56 |    6.518 |  1335.35 |
-|   512 |     32 |   32 |  17408 |    8.422 |  1945.36 |    3.298 |   310.46 |   11.720 |  1485.27 |
-|  4096 |     32 |    1 |   4128 |    2.138 |  1915.88 |    0.571 |    56.09 |    2.708 |  1524.12 |
-|  4096 |     32 |    2 |   8256 |    4.266 |  1920.25 |    1.137 |    56.27 |    5.404 |  1527.90 |
-|  4096 |     32 |    4 |  16512 |    8.564 |  1913.02 |    1.471 |    86.99 |   10.036 |  1645.29 |
-|  4096 |     32 |    8 |  33024 |   17.092 |  1917.19 |    1.979 |   129.33 |   19.071 |  1731.63 |
-|  4096 |     32 |   16 |  66048 |   34.211 |  1915.65 |    2.850 |   179.66 |   37.061 |  1782.15 |
-|  4096 |     32 |   32 | 132096 |   68.394 |  1916.44 |    4.381 |   233.72 |   72.775 |  1815.13 |
-|  8192 |     32 |    1 |   8224 |    4.349 |  1883.45 |    0.620 |    51.65 |    4.969 |  1655.04 |
-|  8192 |     32 |    2 |  16448 |    8.674 |  1888.83 |    1.178 |    54.33 |    9.852 |  1669.48 |
-|  8192 |     32 |    4 |  32896 |   17.351 |  1888.55 |    1.580 |    81.01 |   18.931 |  1737.68 |
-|  8192 |     32 |    8 |  65792 |   34.743 |  1886.31 |    2.173 |   117.80 |   36.916 |  1782.20 |
-|  8192 |     32 |   16 | 131584 |   69.413 |  1888.29 |    3.297 |   155.28 |   72.710 |  1809.70 |
-|  8192 |     32 |   32 | 263168 |  138.903 |  1887.24 |    5.004 |   204.63 |  143.907 |  1828.73 |
+|   512 |     32 |    1 |    544 |    0.445 |  1151.80 |    0.560 |    57.14 |    1.005 |   541.53 |
+|   512 |     32 |    2 |   1088 |    0.472 |  2169.85 |    0.874 |    73.27 |    1.345 |   808.65 |
+|   512 |     32 |    4 |   2176 |    0.826 |  2480.33 |    1.299 |    98.51 |    2.125 |  1023.94 |
+|   512 |     32 |    8 |   4352 |    1.644 |  2491.67 |    1.608 |   159.18 |    3.252 |  1338.20 |
+|   512 |     32 |   16 |   8704 |    3.292 |  2488.35 |    2.117 |   241.85 |    5.409 |  1609.13 |
+|   512 |     32 |   32 |  17408 |    6.604 |  2481.07 |    2.898 |   353.31 |    9.502 |  1832.04 |
+|  4096 |     32 |    1 |   4128 |    1.698 |  2412.65 |    0.580 |    55.21 |    2.277 |  1812.66 |
+|  4096 |     32 |    2 |   8256 |    3.399 |  2409.88 |    0.934 |    68.53 |    4.333 |  1905.27 |
+|  4096 |     32 |    4 |  16512 |    6.823 |  2401.21 |    1.411 |    90.72 |    8.234 |  2005.30 |
+|  4096 |     32 |    8 |  33024 |   13.574 |  2413.97 |    1.841 |   139.07 |   15.415 |  2142.31 |
+|  4096 |     32 |   16 |  66048 |   27.176 |  2411.52 |    2.609 |   196.26 |   29.785 |  2217.49 |
+|  4096 |     32 |   32 | 132096 |   54.359 |  2411.23 |    3.905 |   262.20 |   58.264 |  2267.19 |
+|  8192 |     32 |    1 |   8224 |    3.491 |  2346.81 |    0.613 |    52.23 |    4.103 |  2004.21 |
+|  8192 |     32 |    2 |  16448 |    6.939 |  2361.03 |    0.981 |    65.21 |    7.921 |  2076.56 |
+|  8192 |     32 |    4 |  32896 |   13.888 |  2359.40 |    1.511 |    84.71 |   15.399 |  2136.21 |
+|  8192 |     32 |    8 |  65792 |   27.756 |  2361.18 |    2.034 |   125.86 |   29.790 |  2208.56 |
+|  8192 |     32 |   16 | 131584 |   55.554 |  2359.34 |    3.021 |   169.49 |   58.575 |  2246.41 |
+|  8192 |     32 |   32 | 263168 |  111.036 |  2360.89 |    4.537 |   225.72 |  115.573 |  2277.08 |
 
 
 - `llama-bench`
 
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       1919.36 ± 5.01 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         60.40 ± 0.30 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       1825.30 ± 6.37 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         56.94 ± 0.29 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       1739.19 ± 6.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         52.51 ± 0.42 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1536.75 ± 4.27 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         49.33 ± 0.27 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1255.85 ± 3.26 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         42.99 ± 0.18 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |       2443.91 ± 7.47 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         58.72 ± 0.20 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |       2309.84 ± 3.63 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         55.67 ± 0.35 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |      2216.68 ± 10.16 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         52.87 ± 0.43 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |       1956.31 ± 6.39 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         49.45 ± 0.20 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |      1567.08 ± 11.79 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         42.76 ± 0.14 |
 
-build: eeee367de (6989)
+build: 11fb327bf (7941)
 
 ## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
 
@@ -125,46 +125,46 @@ Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
 - `llama-batched-bench`
 
 
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
 
 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.398 |  1285.90 |    0.530 |    60.41 |    0.928 |   586.27 |
-|   512 |     32 |    2 |   1088 |    0.386 |  2651.65 |    0.948 |    67.50 |    1.334 |   815.38 |
-|   512 |     32 |    4 |   2176 |    0.666 |  3076.37 |    1.209 |   105.87 |    1.875 |  1160.71 |
-|   512 |     32 |    8 |   4352 |    1.325 |  3091.39 |    1.610 |   158.98 |    2.935 |  1482.65 |
-|   512 |     32 |   16 |   8704 |    2.664 |  3075.58 |    2.150 |   238.19 |    4.813 |  1808.39 |
-|   512 |     32 |   32 |  17408 |    5.336 |  3070.31 |    2.904 |   352.59 |    8.240 |  2112.50 |
-|  4096 |     32 |    1 |   4128 |    1.444 |  2836.81 |    0.581 |    55.09 |    2.025 |  2038.81 |
-|  4096 |     32 |    2 |   8256 |    2.872 |  2852.14 |    1.084 |    59.06 |    3.956 |  2086.99 |
-|  4096 |     32 |    4 |  16512 |    5.744 |  2852.32 |    1.440 |    88.90 |    7.184 |  2298.47 |
-|  4096 |     32 |    8 |  33024 |   11.463 |  2858.68 |    2.068 |   123.78 |   13.531 |  2440.65 |
-|  4096 |     32 |   16 |  66048 |   22.915 |  2859.95 |    3.018 |   169.67 |   25.933 |  2546.90 |
-|  4096 |     32 |   32 | 132096 |   45.956 |  2852.10 |    4.609 |   222.18 |   50.565 |  2612.39 |
-|  8192 |     32 |    1 |   8224 |    3.063 |  2674.72 |    0.693 |    46.20 |    3.755 |  2189.92 |
-|  8192 |     32 |    2 |  16448 |    6.109 |  2681.87 |    1.214 |    52.71 |    7.323 |  2245.98 |
-|  8192 |     32 |    4 |  32896 |   12.197 |  2686.63 |    1.682 |    76.11 |   13.878 |  2370.30 |
-|  8192 |     32 |    8 |  65792 |   24.409 |  2684.94 |    2.556 |   100.17 |   26.965 |  2439.95 |
-|  8192 |     32 |   16 | 131584 |   48.753 |  2688.50 |    3.994 |   128.20 |   52.747 |  2494.64 |
-|  8192 |     32 |   32 | 263168 |   97.508 |  2688.42 |    6.528 |   156.86 |  104.037 |  2529.57 |
+|   512 |     32 |    1 |    544 |    0.393 |  1303.73 |    0.548 |    58.36 |    0.941 |   578.10 |
+|   512 |     32 |    2 |   1088 |    0.387 |  2648.68 |    0.910 |    70.35 |    1.296 |   839.27 |
+|   512 |     32 |    4 |   2176 |    0.659 |  3107.63 |    1.302 |    98.33 |    1.961 |  1109.77 |
+|   512 |     32 |    8 |   4352 |    1.322 |  3099.35 |    1.669 |   153.42 |    2.990 |  1455.43 |
+|   512 |     32 |   16 |   8704 |    2.639 |  3104.63 |    2.212 |   231.44 |    4.851 |  1794.32 |
+|   512 |     32 |   32 |  17408 |    5.284 |  3100.80 |    2.955 |   346.53 |    8.239 |  2112.93 |
+|  4096 |     32 |    1 |   4128 |    1.417 |  2890.36 |    0.598 |    53.51 |    2.015 |  2048.45 |
+|  4096 |     32 |    2 |   8256 |    2.829 |  2895.62 |    1.019 |    62.82 |    3.848 |  2145.60 |
+|  4096 |     32 |    4 |  16512 |    5.656 |  2896.96 |    1.528 |    83.79 |    7.183 |  2298.71 |
+|  4096 |     32 |    8 |  33024 |   11.338 |  2890.02 |    2.127 |   120.36 |   13.465 |  2452.53 |
+|  4096 |     32 |   16 |  66048 |   22.709 |  2885.96 |    3.104 |   164.97 |   25.812 |  2558.79 |
+|  4096 |     32 |   32 | 132096 |   45.301 |  2893.35 |    4.723 |   216.80 |   50.024 |  2640.63 |
+|  8192 |     32 |    1 |   8224 |    3.022 |  2711.09 |    0.678 |    47.20 |    3.700 |  2222.89 |
+|  8192 |     32 |    2 |  16448 |    6.039 |  2713.01 |    1.149 |    55.70 |    7.188 |  2288.21 |
+|  8192 |     32 |    4 |  32896 |   12.050 |  2719.35 |    1.785 |    71.69 |   13.835 |  2377.67 |
+|  8192 |     32 |    8 |  65792 |   24.113 |  2717.90 |    2.629 |    97.39 |   26.741 |  2460.31 |
+|  8192 |     32 |   16 | 131584 |   48.178 |  2720.58 |    4.099 |   124.91 |   52.277 |  2517.06 |
+|  8192 |     32 |   32 | 263168 |   96.401 |  2719.31 |    6.696 |   152.93 |  103.097 |  2552.63 |
 
 
 - `llama-bench`
 
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       2925.55 ± 4.25 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         62.80 ± 0.27 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       2531.01 ± 6.79 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         55.86 ± 0.33 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       2244.39 ± 5.33 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         45.95 ± 0.33 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1783.17 ± 3.68 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.07 ± 0.10 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1241.90 ± 3.13 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         29.92 ± 0.06 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |      2986.97 ± 18.87 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         61.06 ± 0.23 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |       2633.45 ± 6.26 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         54.77 ± 0.28 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |       2354.14 ± 3.84 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         48.02 ± 0.40 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |       1908.86 ± 4.25 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         40.23 ± 0.10 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |       1348.17 ± 2.00 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         30.21 ± 0.04 |
 
-build: eeee367de (6989)
+build: 11fb327bf (7941)
 
 ## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
 
@@ -173,46 +173,46 @@ Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
 - `llama-batched-bench`
 
 
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
 
 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.211 |  2421.57 |    1.055 |    30.33 |    1.266 |   429.57 |
-|   512 |     32 |    2 |   1088 |    0.419 |  2441.34 |    1.130 |    56.65 |    1.549 |   702.32 |
-|   512 |     32 |    4 |   2176 |    0.873 |  2345.54 |    1.174 |   108.99 |    2.048 |  1062.74 |
-|   512 |     32 |    8 |   4352 |    1.727 |  2371.85 |    1.254 |   204.22 |    2.980 |  1460.19 |
-|   512 |     32 |   16 |   8704 |    3.452 |  2373.22 |    1.492 |   343.16 |    4.944 |  1760.56 |
-|   512 |     32 |   32 |  17408 |    6.916 |  2368.93 |    1.675 |   611.51 |    8.591 |  2026.36 |
-|  4096 |     32 |    1 |   4128 |    1.799 |  2277.26 |    1.084 |    29.51 |    2.883 |  1431.91 |
-|  4096 |     32 |    2 |   8256 |    3.577 |  2290.01 |    1.196 |    53.50 |    4.774 |  1729.51 |
-|  4096 |     32 |    4 |  16512 |    7.172 |  2284.36 |    1.313 |    97.50 |    8.485 |  1946.00 |
-|  4096 |     32 |    8 |  33024 |   14.341 |  2284.96 |    1.520 |   168.46 |   15.860 |  2082.18 |
-|  4096 |     32 |   16 |  66048 |   28.675 |  2285.44 |    1.983 |   258.21 |   30.658 |  2154.33 |
-|  4096 |     32 |   32 | 132096 |   57.354 |  2285.32 |    2.640 |   387.87 |   59.994 |  2201.82 |
-|  8192 |     32 |    1 |   8224 |    3.701 |  2213.75 |    1.119 |    28.59 |    4.820 |  1706.34 |
-|  8192 |     32 |    2 |  16448 |    7.410 |  2211.19 |    1.272 |    50.31 |    8.682 |  1894.56 |
-|  8192 |     32 |    4 |  32896 |   14.802 |  2213.83 |    1.460 |    87.68 |   16.261 |  2022.96 |
-|  8192 |     32 |    8 |  65792 |   29.609 |  2213.35 |    1.781 |   143.74 |   31.390 |  2095.93 |
-|  8192 |     32 |   16 | 131584 |   59.229 |  2212.96 |    2.495 |   205.17 |   61.725 |  2131.79 |
-|  8192 |     32 |   32 | 263168 |  118.449 |  2213.15 |    3.714 |   275.75 |  122.162 |  2154.25 |
+|   512 |     32 |    1 |    544 |    0.212 |  2420.12 |    1.100 |    29.10 |    1.311 |   414.85 |
+|   512 |     32 |    2 |   1088 |    0.428 |  2393.89 |    1.185 |    54.00 |    1.613 |   674.56 |
+|   512 |     32 |    4 |   2176 |    0.894 |  2290.41 |    1.229 |   104.17 |    2.123 |  1025.02 |
+|   512 |     32 |    8 |   4352 |    1.758 |  2330.36 |    1.319 |   194.15 |    3.076 |  1414.70 |
+|   512 |     32 |   16 |   8704 |    3.508 |  2335.21 |    1.543 |   331.90 |    5.051 |  1723.33 |
+|   512 |     32 |   32 |  17408 |    7.035 |  2328.93 |    1.738 |   589.21 |    8.773 |  1984.29 |
+|  4096 |     32 |    1 |   4128 |    1.831 |  2237.25 |    1.125 |    28.44 |    2.956 |  1396.42 |
+|  4096 |     32 |    2 |   8256 |    3.642 |  2249.48 |    1.253 |    51.07 |    4.895 |  1686.64 |
+|  4096 |     32 |    4 |  16512 |    7.274 |  2252.26 |    1.380 |    92.72 |    8.655 |  1907.81 |
+|  4096 |     32 |    8 |  33024 |   14.576 |  2248.09 |    1.617 |   158.29 |   16.193 |  2039.37 |
+|  4096 |     32 |   16 |  66048 |   29.138 |  2249.17 |    2.081 |   246.01 |   31.219 |  2115.63 |
+|  4096 |     32 |   32 | 132096 |   58.275 |  2249.19 |    2.814 |   363.87 |   61.089 |  2162.34 |
+|  8192 |     32 |    1 |   8224 |    3.757 |  2180.26 |    1.184 |    27.03 |    4.941 |  1664.37 |
+|  8192 |     32 |    2 |  16448 |    7.522 |  2178.05 |    1.341 |    47.73 |    8.863 |  1855.77 |
+|  8192 |     32 |    4 |  32896 |   15.043 |  2178.25 |    1.548 |    82.69 |   16.591 |  1982.74 |
+|  8192 |     32 |    8 |  65792 |   30.111 |  2176.49 |    1.937 |   132.13 |   32.048 |  2052.90 |
+|  8192 |     32 |   16 | 131584 |   60.405 |  2169.90 |    2.706 |   189.21 |   63.111 |  2084.97 |
+|  8192 |     32 |   32 | 263168 |  120.439 |  2176.58 |    3.993 |   256.46 |  124.432 |  2114.96 |
 
 
 - `llama-bench`
 
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       2272.74 ± 4.68 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         30.66 ± 0.02 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       2107.80 ± 9.55 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         29.71 ± 0.05 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       1937.80 ± 6.75 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         28.86 ± 0.04 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1641.12 ± 1.78 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         27.24 ± 0.04 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1296.02 ± 2.67 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         23.78 ± 0.03 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |       2250.28 ± 6.41 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         29.43 ± 0.02 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |       2100.19 ± 8.96 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         28.61 ± 0.02 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |       2007.56 ± 4.16 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         27.38 ± 0.09 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |       1779.11 ± 6.42 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         25.72 ± 0.03 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |       1471.23 ± 1.71 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         22.51 ± 0.02 |
 
-build: eeee367de (6989)
+build: 11fb327bf (7941)
 
 ## ggml-org/gemma-3-4b-it-qat-GGUF
 
@@ -221,44 +221,91 @@ Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
 - `llama-batched-bench`
 
 
-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
 
 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.094 |  5434.73 |    0.394 |    81.21 |    0.488 |  1114.15 |
-|   512 |     32 |    2 |   1088 |    0.168 |  6091.68 |    0.498 |   128.52 |    0.666 |  1633.41 |
-|   512 |     32 |    4 |   2176 |    0.341 |  6010.68 |    0.542 |   236.37 |    0.882 |  2466.43 |
-|   512 |     32 |    8 |   4352 |    0.665 |  6161.46 |    0.678 |   377.74 |    1.342 |  3241.72 |
-|   512 |     32 |   16 |   8704 |    1.323 |  6193.19 |    0.902 |   567.41 |    2.225 |  3911.74 |
-|   512 |     32 |   32 |  17408 |    2.642 |  6202.03 |    1.231 |   832.03 |    3.872 |  4495.36 |
-|  4096 |     32 |    1 |   4128 |    0.701 |  5840.49 |    0.439 |    72.95 |    1.140 |  3621.23 |
-|  4096 |     32 |    2 |   8256 |    1.387 |  5906.82 |    0.574 |   111.48 |    1.961 |  4210.12 |
-|  4096 |     32 |    4 |  16512 |    2.758 |  5940.33 |    0.651 |   196.58 |    3.409 |  4843.33 |
-|  4096 |     32 |    8 |  33024 |    5.491 |  5967.56 |    0.876 |   292.40 |    6.367 |  5187.12 |
-|  4096 |     32 |   16 |  66048 |   10.978 |  5969.58 |    1.275 |   401.69 |   12.253 |  5390.38 |
-|  4096 |     32 |   32 | 132096 |   21.944 |  5972.93 |    1.992 |   514.16 |   23.936 |  5518.73 |
-|  8192 |     32 |    1 |   8224 |    1.402 |  5841.91 |    0.452 |    70.73 |    1.855 |  4434.12 |
-|  8192 |     32 |    2 |  16448 |    2.793 |  5865.34 |    0.637 |   100.55 |    3.430 |  4795.51 |
-|  8192 |     32 |    4 |  32896 |    5.564 |  5889.64 |    0.770 |   166.26 |    6.334 |  5193.95 |
-|  8192 |     32 |    8 |  65792 |   11.114 |  5896.44 |    1.122 |   228.07 |   12.237 |  5376.51 |
-|  8192 |     32 |   16 | 131584 |   22.210 |  5901.38 |    1.789 |   286.15 |   24.000 |  5482.74 |
-|  8192 |     32 |   32 | 263168 |   44.382 |  5906.56 |    3.044 |   336.38 |   47.426 |  5549.02 |
+|   512 |     32 |    1 |    544 |    0.092 |  5566.97 |    0.412 |    77.63 |    0.504 |  1078.95 |
+|   512 |     32 |    2 |   1088 |    0.161 |  6345.67 |    0.522 |   122.70 |    0.683 |  1593.06 |
+|   512 |     32 |    4 |   2176 |    0.325 |  6309.87 |    0.562 |   227.68 |    0.887 |  2453.87 |
+|   512 |     32 |    8 |   4352 |    0.643 |  6374.42 |    0.685 |   373.67 |    1.328 |  3277.94 |
+|   512 |     32 |   16 |   8704 |    1.277 |  6413.64 |    0.915 |   559.47 |    2.192 |  3970.01 |
+|   512 |     32 |   32 |  17408 |    2.518 |  6506.57 |    1.249 |   819.61 |    3.767 |  4620.64 |
+|  4096 |     32 |    1 |   4128 |    0.674 |  6079.68 |    0.453 |    70.60 |    1.127 |  3662.88 |
+|  4096 |     32 |    2 |   8256 |    1.335 |  6137.82 |    0.627 |   102.03 |    1.962 |  4208.11 |
+|  4096 |     32 |    4 |  16512 |    2.657 |  6167.35 |    0.749 |   170.92 |    3.405 |  4848.71 |
+|  4096 |     32 |    8 |  33024 |    5.307 |  6173.91 |    0.974 |   262.89 |    6.281 |  5257.53 |
+|  4096 |     32 |   16 |  66048 |   10.610 |  6176.96 |    1.379 |   371.42 |   11.988 |  5509.40 |
+|  4096 |     32 |   32 | 132096 |   21.213 |  6178.89 |    2.122 |   482.50 |   23.335 |  5660.82 |
+|  8192 |     32 |    1 |   8224 |    1.359 |  6027.34 |    0.467 |    68.52 |    1.826 |  4503.48 |
+|  8192 |     32 |    2 |  16448 |    2.699 |  6069.68 |    0.653 |    98.03 |    3.352 |  4906.68 |
+|  8192 |     32 |    4 |  32896 |    5.366 |  6106.74 |    0.818 |   156.55 |    6.184 |  5319.96 |
+|  8192 |     32 |    8 |  65792 |   10.755 |  6093.50 |    1.174 |   218.04 |   11.929 |  5515.22 |
+|  8192 |     32 |   16 | 131584 |   21.484 |  6100.82 |    1.829 |   279.90 |   23.314 |  5644.11 |
+|  8192 |     32 |   32 | 263168 |   42.950 |  6103.40 |    3.058 |   334.91 |   46.008 |  5720.05 |
 
 
 - `llama-bench`
 
-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |      5810.04 ± 21.71 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         84.54 ± 0.18 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       5288.04 ± 3.54 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         78.82 ± 1.37 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |      4960.43 ± 16.64 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         74.13 ± 0.30 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |      4495.92 ± 31.11 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         72.37 ± 0.29 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      3746.90 ± 40.01 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         63.02 ± 0.20 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |      5948.74 ± 10.61 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         81.05 ± 0.20 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |      5652.69 ± 34.29 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         76.37 ± 0.58 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |      5509.57 ± 40.69 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         71.61 ± 0.80 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |      5340.86 ± 36.92 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         70.89 ± 0.34 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |      5023.30 ± 13.52 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         62.28 ± 0.30 |
 
-build: eeee367de (6989)
+build: 11fb327bf (7941)
 
+## ggml-org/GLM-4.7-Flash-GGUF
+
+Model: https://huggingface.co/ggml-org/GLM-4.7-Flash-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.433 |  1181.83 |    0.693 |    46.16 |    1.126 |   482.94 |
+|   512 |     32 |    2 |   1088 |    0.439 |  2334.46 |    1.034 |    61.89 |    1.473 |   738.75 |
+|   512 |     32 |    4 |   2176 |    0.772 |  2654.46 |    1.459 |    87.76 |    2.230 |   975.77 |
+|   512 |     32 |    8 |   4352 |    1.541 |  2658.78 |    2.043 |   125.31 |    3.583 |  1214.47 |
+|   512 |     32 |   16 |   8704 |    3.083 |  2656.91 |    2.675 |   191.42 |    5.758 |  1511.62 |
+|   512 |     32 |   32 |  17408 |    6.159 |  2660.12 |    3.615 |   283.24 |    9.774 |  1780.98 |
+|  4096 |     32 |    1 |   4128 |    1.915 |  2139.30 |    0.725 |    44.14 |    2.640 |  1563.83 |
+|  4096 |     32 |    2 |   8256 |    3.834 |  2136.40 |    1.119 |    57.21 |    4.953 |  1666.81 |
+|  4096 |     32 |    4 |  16512 |    7.636 |  2145.72 |    1.631 |    78.49 |    9.266 |  1781.93 |
+|  4096 |     32 |    8 |  33024 |   15.295 |  2142.40 |    2.344 |   109.21 |   17.639 |  1872.20 |
+|  4096 |     32 |   16 |  66048 |   30.573 |  2143.62 |    3.773 |   135.70 |   34.346 |  1923.04 |
+|  4096 |     32 |   32 | 132096 |   61.282 |  2138.82 |    5.795 |   176.71 |   67.077 |  1969.31 |
+|  8192 |     32 |    1 |   8224 |    4.510 |  1816.24 |    0.760 |    42.11 |    5.270 |  1560.44 |
+|  8192 |     32 |    2 |  16448 |    9.036 |  1813.19 |    1.206 |    53.06 |   10.242 |  1605.91 |
+|  8192 |     32 |    4 |  32896 |   18.070 |  1813.43 |    1.783 |    71.80 |   19.852 |  1657.03 |
+|  8192 |     32 |    8 |  65792 |   36.125 |  1814.15 |    2.635 |    97.14 |   38.760 |  1697.41 |
+|  8192 |     32 |   16 | 131584 |   72.367 |  1811.20 |    4.954 |   103.34 |   77.322 |  1701.77 |
+|  8192 |     32 |   32 | 263168 |  144.501 |  1814.13 |    8.103 |   126.37 |  152.604 |  1724.51 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --: | --------------: | -------------------: |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |          pp2048 |      2364.18 ± 11.43 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |            tg32 |         48.68 ± 0.12 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |  pp2048 @ d4096 |       1684.13 ± 1.24 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |    tg32 @ d4096 |         44.62 ± 0.22 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |  pp2048 @ d8192 |       1314.68 ± 1.41 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |    tg32 @ d8192 |         42.59 ± 0.11 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 | pp2048 @ d16384 |        914.05 ± 3.32 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |   tg32 @ d16384 |         38.72 ± 0.13 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 | pp2048 @ d32768 |        567.20 ± 0.90 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |   tg32 @ d32768 |         32.65 ± 0.09 |
+
+build: 11fb327bf (7941)

benches/mac-m2-ultra/mac-m2-ultra.md

@@ -0,0 +1,298 @@
+## System info
+
+```bash
+uname -a
+Darwin gg-studio 25.2.0 Darwin Kernel Version 25.2.0: Tue Nov 18 21:07:05 PST 2025; root:xnu-12377.61.12~1/RELEASE_ARM64_T6020 arm64
+
+g++ --version
+Apple clang version 17.0.0 (clang-1700.3.19.1)
+Target: arm64-apple-darwin25.2.0
+```
+
+## ggml-org/gpt-oss-20b-GGUF
+
+Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.215 |  2381.35 |    0.245 |   130.45 |    0.460 |  1181.81 |
+|   512 |     32 |    2 |   1088 |    0.379 |  2701.43 |    0.382 |   167.56 |    0.761 |  1429.67 |
+|   512 |     32 |    4 |   2176 |    0.721 |  2839.27 |    0.604 |   211.76 |    1.326 |  1641.32 |
+|   512 |     32 |    8 |   4352 |    1.433 |  2858.30 |    1.033 |   247.75 |    2.466 |  1764.57 |
+|   512 |     32 |   16 |   8704 |    2.853 |  2871.12 |    1.570 |   326.11 |    4.423 |  1967.77 |
+|   512 |     32 |   32 |  17408 |    5.699 |  2874.95 |    1.910 |   536.15 |    7.609 |  2287.88 |
+|  4096 |     32 |    1 |   4128 |    1.552 |  2638.56 |    0.334 |    95.72 |    1.887 |  2188.00 |
+|  4096 |     32 |    2 |   8256 |    3.084 |  2655.88 |    0.404 |   158.54 |    3.488 |  2366.86 |
+|  4096 |     32 |    4 |  16512 |    6.151 |  2663.78 |    0.652 |   196.39 |    6.802 |  2427.37 |
+|  4096 |     32 |    8 |  33024 |   12.288 |  2666.77 |    1.135 |   225.47 |   13.423 |  2460.27 |
+|  4096 |     32 |   16 |  66048 |   24.563 |  2668.12 |    1.762 |   290.55 |   26.325 |  2508.97 |
+|  4096 |     32 |   32 | 132096 |   49.114 |  2668.73 |    2.398 |   426.94 |   51.512 |  2564.35 |
+|  8192 |     32 |    1 |   8224 |    3.345 |  2448.78 |    0.275 |   116.46 |    3.620 |  2271.76 |
+|  8192 |     32 |    2 |  16448 |    6.665 |  2458.11 |    0.425 |   150.71 |    7.090 |  2319.91 |
+|  8192 |     32 |    4 |  32896 |   13.315 |  2460.92 |    0.691 |   185.21 |   14.006 |  2348.63 |
+|  8192 |     32 |    8 |  65792 |   26.611 |  2462.73 |    1.212 |   211.16 |   27.823 |  2364.62 |
+|  8192 |     32 |   16 | 131584 |   53.232 |  2462.27 |    1.919 |   266.83 |   55.151 |  2385.88 |
+|  8192 |     32 |   32 | 263168 |  110.455 |  2373.30 |    2.752 |   372.03 |  113.208 |  2324.64 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       2713.40 ± 3.56 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |        129.97 ± 3.90 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       2324.59 ± 3.01 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |        123.38 ± 0.17 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |      1989.82 ± 30.11 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |        117.39 ± 0.33 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |       1556.54 ± 6.22 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |        109.75 ± 0.42 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |       1122.63 ± 1.45 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         98.25 ± 0.08 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/gpt-oss-120b-GGUF
+
+Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.426 |  1200.92 |    0.361 |    88.56 |    0.788 |   690.64 |
+|   512 |     32 |    2 |   1088 |    0.683 |  1500.14 |    0.545 |   117.35 |    1.228 |   886.02 |
+|   512 |     32 |    4 |   2176 |    1.204 |  1701.56 |    0.847 |   151.19 |    2.050 |  1061.34 |
+|   512 |     32 |    8 |   4352 |    2.402 |  1705.20 |    1.455 |   176.00 |    3.857 |  1128.45 |
+|   512 |     32 |   16 |   8704 |    4.802 |  1705.90 |    2.349 |   217.93 |    7.152 |  1217.08 |
+|   512 |     32 |   32 |  17408 |    9.593 |  1707.85 |    3.665 |   279.42 |   13.258 |  1313.01 |
+|  4096 |     32 |    1 |   4128 |    2.581 |  1587.08 |    0.390 |    82.12 |    2.970 |  1389.67 |
+|  4096 |     32 |    2 |   8256 |    5.124 |  1598.79 |    0.589 |   108.62 |    5.713 |  1445.10 |
+|  4096 |     32 |    4 |  16512 |   10.231 |  1601.47 |    0.928 |   137.98 |   11.158 |  1479.80 |
+|  4096 |     32 |    8 |  33024 |   20.468 |  1600.94 |    1.606 |   159.38 |   22.074 |  1496.04 |
+|  4096 |     32 |   16 |  66048 |   40.924 |  1601.42 |    2.639 |   193.99 |   43.563 |  1516.15 |
+|  4096 |     32 |   32 | 132096 |   81.819 |  1601.98 |    4.466 |   229.29 |   86.284 |  1530.94 |
+|  8192 |     32 |    1 |   8224 |    5.517 |  1484.74 |    0.409 |    78.16 |    5.927 |  1387.58 |
+|  8192 |     32 |    2 |  16448 |   11.008 |  1488.43 |    0.622 |   102.92 |   11.629 |  1414.34 |
+|  8192 |     32 |    4 |  32896 |   22.002 |  1489.29 |    0.987 |   129.66 |   22.990 |  1430.90 |
+|  8192 |     32 |    8 |  65792 |   46.051 |  1423.11 |    1.858 |   137.79 |   47.909 |  1373.27 |
+|  8192 |     32 |   16 | 131584 |   97.680 |  1341.85 |    2.872 |   178.28 |  100.552 |  1308.62 |
+|  8192 |     32 |   32 | 263168 |  176.407 |  1486.02 |    5.048 |   202.85 |  181.455 |  1450.32 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       1648.69 ± 1.80 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         85.60 ± 0.52 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       1429.86 ± 1.01 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         82.03 ± 0.12 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       1257.90 ± 1.81 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         78.23 ± 0.33 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |       1013.49 ± 0.70 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         73.20 ± 0.28 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        721.11 ± 0.58 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         65.52 ± 0.10 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
+
+Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.243 |  2109.23 |    0.419 |    76.34 |    0.662 |   821.84 |
+|   512 |     32 |    2 |   1088 |    0.406 |  2521.40 |    0.575 |   111.36 |    0.981 |  1109.27 |
+|   512 |     32 |    4 |   2176 |    0.744 |  2751.65 |    0.841 |   152.22 |    1.585 |  1372.71 |
+|   512 |     32 |    8 |   4352 |    1.479 |  2770.20 |    1.330 |   192.48 |    2.809 |  1549.53 |
+|   512 |     32 |   16 |   8704 |    2.951 |  2776.20 |    2.572 |   199.05 |    5.523 |  1575.93 |
+|   512 |     32 |   32 |  17408 |    5.899 |  2777.64 |    2.603 |   393.34 |    8.502 |  2047.54 |
+|  4096 |     32 |    1 |   4128 |    1.901 |  2154.15 |    0.474 |    67.58 |    2.375 |  1738.14 |
+|  4096 |     32 |    2 |   8256 |    3.788 |  2162.89 |    0.652 |    98.17 |    4.439 |  1859.69 |
+|  4096 |     32 |    4 |  16512 |    7.564 |  2166.18 |    0.990 |   129.24 |    8.554 |  1930.34 |
+|  4096 |     32 |    8 |  33024 |   15.121 |  2166.98 |    1.632 |   156.82 |   16.754 |  1971.12 |
+|  4096 |     32 |   16 |  66048 |   30.241 |  2167.09 |    3.166 |   161.72 |   33.407 |  1977.04 |
+|  4096 |     32 |   32 | 132096 |   60.474 |  2167.42 |    3.780 |   270.93 |   64.254 |  2055.86 |
+|  8192 |     32 |    1 |   8224 |    4.733 |  1730.92 |    0.483 |    66.29 |    5.215 |  1576.85 |
+|  8192 |     32 |    2 |  16448 |    9.459 |  1732.09 |    0.722 |    88.58 |   10.182 |  1615.46 |
+|  8192 |     32 |    4 |  32896 |   18.912 |  1732.65 |    1.120 |   114.26 |   20.032 |  1642.14 |
+|  8192 |     32 |    8 |  65792 |   37.797 |  1733.91 |    1.873 |   136.67 |   39.670 |  1658.49 |
+|  8192 |     32 |   16 | 131584 |   84.133 |  1557.92 |    3.718 |   137.72 |   87.850 |  1497.82 |
+|  8192 |     32 |   32 | 263168 |  157.550 |  1663.88 |    4.854 |   210.98 |  162.403 |  1620.46 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       2453.11 ± 1.70 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         78.97 ± 0.46 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       1569.46 ± 1.97 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         71.18 ± 0.37 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       1145.51 ± 1.16 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         65.11 ± 0.36 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |        741.04 ± 0.74 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         56.87 ± 0.14 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        431.31 ± 0.31 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         45.26 ± 0.11 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
+
+Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.339 |  1509.22 |    0.409 |    78.17 |    0.749 |   726.67 |
+|   512 |     32 |    2 |   1088 |    0.646 |  1584.93 |    0.483 |   132.45 |    1.129 |   963.45 |
+|   512 |     32 |    4 |   2176 |    1.258 |  1627.50 |    0.585 |   218.67 |    1.844 |  1180.21 |
+|   512 |     32 |    8 |   4352 |    2.506 |  1634.41 |    1.005 |   254.83 |    3.511 |  1239.64 |
+|   512 |     32 |   16 |   8704 |    5.007 |  1635.99 |    1.595 |   321.07 |    6.602 |  1318.38 |
+|   512 |     32 |   32 |  17408 |   10.007 |  1637.19 |    1.676 |   611.12 |   11.683 |  1490.03 |
+|  4096 |     32 |    1 |   4128 |    2.730 |  1500.46 |    0.431 |    74.31 |    3.160 |  1306.12 |
+|  4096 |     32 |    2 |   8256 |    5.446 |  1504.33 |    0.524 |   122.04 |    5.970 |  1382.91 |
+|  4096 |     32 |    4 |  16512 |   10.875 |  1506.59 |    0.662 |   193.45 |   11.537 |  1431.28 |
+|  4096 |     32 |    8 |  33024 |   21.749 |  1506.61 |    1.158 |   221.11 |   22.907 |  1441.64 |
+|  4096 |     32 |   16 |  66048 |   43.477 |  1507.36 |    1.901 |   269.32 |   45.378 |  1455.49 |
+|  4096 |     32 |   32 | 132096 |   86.954 |  1507.37 |    2.325 |   440.42 |   89.279 |  1479.59 |
+|  8192 |     32 |    1 |   8224 |    5.940 |  1379.21 |    0.449 |    71.20 |    6.389 |  1287.20 |
+|  8192 |     32 |    2 |  16448 |   11.865 |  1380.84 |    0.559 |   114.59 |   12.424 |  1323.92 |
+|  8192 |     32 |    4 |  32896 |   23.723 |  1381.25 |    0.728 |   175.80 |   24.452 |  1345.35 |
+|  8192 |     32 |    8 |  65792 |   47.434 |  1381.63 |    1.279 |   200.09 |   48.713 |  1350.60 |
+|  8192 |     32 |   16 | 131584 |   94.864 |  1381.69 |    2.198 |   232.97 |   97.061 |  1355.68 |
+|  8192 |     32 |   32 | 263168 |  189.743 |  1381.57 |    3.052 |   335.50 |  192.795 |  1365.01 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       1565.91 ± 0.86 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         79.68 ± 0.39 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       1317.41 ± 1.02 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         74.70 ± 0.04 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       1134.65 ± 0.76 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         71.31 ± 0.12 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |        886.46 ± 0.78 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         65.93 ± 0.06 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        612.21 ± 0.30 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         56.83 ± 0.02 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/gemma-3-4b-it-qat-GGUF
+
+Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.186 |  2748.06 |    0.235 |   136.28 |    0.421 |  1291.78 |
+|   512 |     32 |    2 |   1088 |    0.342 |  2990.95 |    0.312 |   204.99 |    0.655 |  1662.15 |
+|   512 |     32 |    4 |   2176 |    0.662 |  3092.69 |    0.404 |   316.97 |    1.066 |  2041.21 |
+|   512 |     32 |    8 |   4352 |    1.317 |  3110.41 |    0.579 |   441.80 |    1.896 |  2294.97 |
+|   512 |     32 |   16 |   8704 |    2.625 |  3120.23 |    1.207 |   424.08 |    3.833 |  2270.93 |
+|   512 |     32 |   32 |  17408 |    5.242 |  3125.34 |    1.299 |   788.23 |    6.541 |  2661.19 |
+|  4096 |     32 |    1 |   4128 |    1.408 |  2909.90 |    0.296 |   108.07 |    1.704 |  2422.95 |
+|  4096 |     32 |    2 |   8256 |    2.793 |  2933.40 |    0.325 |   197.00 |    3.118 |  2648.25 |
+|  4096 |     32 |    4 |  16512 |    5.567 |  2943.22 |    0.440 |   291.07 |    6.006 |  2749.05 |
+|  4096 |     32 |    8 |  33024 |   11.114 |  2948.23 |    0.640 |   400.26 |   11.754 |  2809.59 |
+|  4096 |     32 |   16 |  66048 |   22.217 |  2949.76 |    1.327 |   385.83 |   23.544 |  2805.26 |
+|  4096 |     32 |   32 | 132096 |   44.420 |  2950.77 |    1.553 |   659.30 |   45.973 |  2873.36 |
+|  8192 |     32 |    1 |   8224 |    2.860 |  2864.58 |    0.250 |   127.90 |    3.110 |  2644.42 |
+|  8192 |     32 |    2 |  16448 |    5.702 |  2873.63 |    0.335 |   191.07 |    6.036 |  2724.77 |
+|  8192 |     32 |    4 |  32896 |   11.383 |  2878.69 |    0.456 |   280.72 |   11.839 |  2778.63 |
+|  8192 |     32 |    8 |  65792 |   22.750 |  2880.75 |    0.671 |   381.48 |   23.421 |  2809.14 |
+|  8192 |     32 |   16 | 131584 |   45.484 |  2881.74 |    1.406 |   364.04 |   46.890 |  2806.22 |
+|  8192 |     32 |   32 | 263168 |   90.956 |  2882.10 |    1.793 |   570.98 |   92.749 |  2837.41 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       2923.59 ± 3.10 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |        134.28 ± 1.29 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       2748.21 ± 3.05 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |        133.11 ± 0.08 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       2641.45 ± 2.31 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |        125.85 ± 0.35 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |       2446.20 ± 2.94 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |        125.00 ± 0.12 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |       2129.18 ± 7.43 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |        113.14 ± 0.10 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/GLM-4.7-Flash-GGUF
+
+Model: https://huggingface.co/ggml-org/GLM-4.7-Flash-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.326 |  1568.69 |    0.522 |    61.28 |    0.849 |   641.09 |
+|   512 |     32 |    2 |   1088 |    0.528 |  1939.42 |    0.744 |    86.07 |    1.272 |   855.63 |
+|   512 |     32 |    4 |   2176 |    0.968 |  2114.85 |    1.105 |   115.85 |    2.073 |  1049.56 |
+|   512 |     32 |    8 |   4352 |    1.928 |  2124.62 |    1.684 |   151.99 |    3.612 |  1204.82 |
+|   512 |     32 |   16 |   8704 |    3.844 |  2131.34 |    3.141 |   162.99 |    6.985 |  1246.11 |
+|   512 |     32 |   32 |  17408 |    7.683 |  2132.38 |    3.924 |   260.95 |   11.608 |  1499.71 |
+|  4096 |     32 |    1 |   4128 |    3.280 |  1248.75 |    0.723 |    44.29 |    4.003 |  1031.33 |
+|  4096 |     32 |    2 |   8256 |    6.545 |  1251.63 |    0.930 |    68.85 |    7.475 |  1104.53 |
+|  4096 |     32 |    4 |  16512 |   13.080 |  1252.64 |    1.454 |    88.03 |   14.534 |  1136.12 |
+|  4096 |     32 |    8 |  33024 |   26.154 |  1252.90 |    2.388 |   107.20 |   28.542 |  1157.04 |
+|  4096 |     32 |   16 |  66048 |   52.297 |  1253.14 |    4.724 |   108.37 |   57.022 |  1158.30 |
+|  4096 |     32 |   32 | 132096 |  104.578 |  1253.34 |    7.266 |   140.93 |  111.844 |  1181.08 |
+|  8192 |     32 |    1 |   8224 |    9.623 |   851.31 |    0.767 |    41.72 |   10.390 |   791.54 |
+|  8192 |     32 |    2 |  16448 |   20.916 |   783.32 |    1.148 |    55.74 |   22.064 |   745.45 |
+|  8192 |     32 |    4 |  32896 |   43.509 |   753.14 |    1.833 |    69.82 |   45.342 |   725.51 |
+|  8192 |     32 |    8 |  65792 |   79.621 |   823.10 |    3.180 |    80.50 |   82.801 |   794.58 |
+|  8192 |     32 |   16 | 131584 |  153.770 |   852.39 |    6.502 |    78.74 |  160.272 |   821.00 |
+|  8192 |     32 |   32 | 263168 |  307.539 |   852.39 |   10.839 |    94.48 |  318.378 |   826.59 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       1629.33 ± 0.27 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         59.58 ± 0.13 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |        732.67 ± 0.42 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         47.44 ± 0.15 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |        474.33 ± 0.33 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         40.20 ± 0.20 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |        277.46 ± 0.09 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         31.50 ± 0.93 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        151.44 ± 0.05 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         21.81 ± 0.01 |
+
+build: b828e18c7 (7948)

build-xcframework.sh

@@ -43,11 +43,6 @@ COMMON_CMAKE_ARGS=(
     -DGGML_OPENMP=${GGML_OPENMP}
 )
 
-XCODE_VERSION=$(xcodebuild -version 2>/dev/null | head -n1 | awk '{ print $2 }')
-MAJOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f1)
-MINOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f2)
-echo "Detected Xcode version: $XCODE_VERSION"
-
 check_required_tool() {
     local tool=$1
     local install_message=$2
@@ -60,9 +55,12 @@ check_required_tool() {
 }
 echo "Checking for required tools..."
 check_required_tool "cmake" "Please install CMake 3.28.0 or later (brew install cmake)"
-check_required_tool "xcodebuild" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
-check_required_tool "libtool" "Please install libtool which should be available with Xcode Command Line Tools (CLT). Make sure Xcode CLT is installed (xcode-select --install)"
-check_required_tool "dsymutil" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
+check_required_tool "xcrun" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)"
+
+XCODE_VERSION=$(xcrun xcodebuild -version 2>/dev/null | head -n1 | awk '{ print $2 }')
+MAJOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f1)
+MINOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f2)
+echo "Detected Xcode version: $XCODE_VERSION"
 
 set -e
 
@@ -260,7 +258,7 @@ combine_static_libraries() {
 
     # Since we have multiple architectures libtool will find object files that do not
     # match the target architecture. We suppress these warnings.
-    libtool -static -o "${temp_dir}/combined.a" "${libs[@]}" 2> /dev/null
+    xcrun libtool -static -o "${temp_dir}/combined.a" "${libs[@]}" 2> /dev/null
 
     # Determine SDK, architectures, and install_name based on platform and simulator flag.
     local sdk=""
@@ -333,7 +331,7 @@ combine_static_libraries() {
 
     # Platform-specific post-processing for device builds
     if [[ "$is_simulator" == "false" ]]; then
-        if command -v xcrun vtool &>/dev/null; then
+        if xcrun -f vtool &>/dev/null; then
             case "$platform" in
                 "ios")
                     echo "Marking binary as a framework binary for iOS..."
@@ -451,10 +449,9 @@ cmake -B build-visionos -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xros \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_HTTPLIB=OFF \
     -DLLAMA_BUILD_SERVER=OFF \
     -S .
 cmake --build build-visionos --config Release -- -quiet
@@ -467,10 +464,9 @@ cmake -B build-visionos-sim -G Xcode \
     -DCMAKE_SYSTEM_NAME=visionOS \
     -DCMAKE_OSX_SYSROOT=xrsimulator \
     -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
-    -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \
-    -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \
+    -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
+    -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
     -DLLAMA_OPENSSL=OFF \
-    -DLLAMA_HTTPLIB=OFF \
     -DLLAMA_BUILD_SERVER=OFF \
     -S .
 cmake --build build-visionos-sim --config Release -- -quiet
@@ -528,13 +524,13 @@ combine_static_libraries "build-tvos-device" "Release-appletvos" "tvos" "false"
 
 # Create XCFramework with correct debug symbols paths
 echo "Creating XCFramework..."
-xcodebuild -create-xcframework \
+xcrun xcodebuild -create-xcframework \
     -framework $(pwd)/build-ios-sim/framework/llama.framework \
     -debug-symbols $(pwd)/build-ios-sim/dSYMs/llama.dSYM \
     -framework $(pwd)/build-ios-device/framework/llama.framework \
     -debug-symbols $(pwd)/build-ios-device/dSYMs/llama.dSYM \
     -framework $(pwd)/build-macos/framework/llama.framework \
-    -debug-symbols $(pwd)/build-macos/dSYMS/llama.dSYM \
+    -debug-symbols $(pwd)/build-macos/dSYMs/llama.dSYM \
     -framework $(pwd)/build-visionos/framework/llama.framework \
     -debug-symbols $(pwd)/build-visionos/dSYMs/llama.dSYM \
     -framework $(pwd)/build-visionos-sim/framework/llama.framework \

cmake/common.cmake

@@ -32,4 +32,27 @@ function(llama_add_compile_flags)
             set(CXX_FLAGS "" PARENT_SCOPE)
         endif()
     endif()
+
+    if (NOT MSVC)
+        if (LLAMA_SANITIZE_THREAD)
+            message(STATUS "Using -fsanitize=thread")
+
+            add_compile_options(-fsanitize=thread)
+            link_libraries     (-fsanitize=thread)
+        endif()
+
+        if (LLAMA_SANITIZE_ADDRESS)
+            message(STATUS "Using -fsanitize=address")
+
+            add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
+            link_libraries     (-fsanitize=address)
+        endif()
+
+        if (LLAMA_SANITIZE_UNDEFINED)
+            message(STATUS "Using -fsanitize=undefined")
+
+            add_compile_options(-fsanitize=undefined)
+            link_libraries     (-fsanitize=undefined)
+        endif()
+    endif()
 endfunction()

common/CMakeLists.txt

@@ -5,7 +5,6 @@ find_package(Threads REQUIRED)
 llama_add_compile_flags()
 
 # Build info header
-#
 
 if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
     set(GIT_DIR "${PROJECT_SOURCE_DIR}/.git")
@@ -110,33 +109,16 @@ if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
-# TODO: use list(APPEND LLAMA_COMMON_EXTRA_LIBS ...)
-set(LLAMA_COMMON_EXTRA_LIBS build_info)
-
-if (LLAMA_HTTPLIB)
-    target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_HTTPLIB)
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} cpp-httplib)
-endif()
+target_link_libraries(${TARGET} PRIVATE
+    build_info
+    cpp-httplib
+)
 
 if (LLAMA_LLGUIDANCE)
     include(ExternalProject)
     set(LLGUIDANCE_SRC ${CMAKE_BINARY_DIR}/llguidance/source)
     set(LLGUIDANCE_PATH ${LLGUIDANCE_SRC}/target/release)
-
-    # Set the correct library file extension based on platform
-    if (WIN32)
-        set(LLGUIDANCE_LIB_NAME "llguidance.lib")
-        # Add Windows-specific libraries
-        set(LLGUIDANCE_PLATFORM_LIBS
-            ws2_32    # Windows Sockets API
-            userenv   # For GetUserProfileDirectoryW
-            ntdll     # For NT functions
-            bcrypt    # For BCryptGenRandom
-        )
-    else()
-        set(LLGUIDANCE_LIB_NAME "libllguidance.a")
-        set(LLGUIDANCE_PLATFORM_LIBS "")
-    endif()
+    set(LLGUIDANCE_LIB_NAME "${CMAKE_STATIC_LIBRARY_PREFIX}llguidance${CMAKE_STATIC_LIBRARY_SUFFIX}")
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
@@ -158,8 +140,10 @@ if (LLAMA_LLGUIDANCE)
     add_dependencies(llguidance llguidance_ext)
 
     target_include_directories(${TARGET} PRIVATE ${LLGUIDANCE_PATH})
-    # Add platform libraries to the main target
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance ${LLGUIDANCE_PLATFORM_LIBS})
-endif ()
+    target_link_libraries(${TARGET} PRIVATE llguidance)
+    if (WIN32)
+        target_link_libraries(${TARGET} PRIVATE ws2_32 userenv ntdll bcrypt)
+    endif()
+endif()
 
-target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
+target_link_libraries(${TARGET} PUBLIC llama Threads::Threads)

common/arg.cpp

@@ -1301,7 +1301,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, bool value) {
             params.kv_unified = value;
         }
-    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED, LLAMA_EXAMPLE_BENCH}));
+    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED, LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
     add_opt(common_arg(
         {"--context-shift"},
         {"--no-context-shift"},
@@ -1578,7 +1578,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_sparam());
     add_opt(common_arg(
-        {"--temp"}, "N",
+        {"--temp", "--temperature"}, "N",
         string_format("temperature (default: %.2f)", (double)params.sampling.temp),
         [](common_params & params, const std::string & value) {
             params.sampling.temp = std::stof(value);
@@ -1611,7 +1611,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_sparam());
     add_opt(common_arg(
-        {"--top-nsigma"}, "N",
+        {"--top-nsigma", "--top-n-sigma"}, "N",
         string_format("top-n-sigma sampling (default: %.2f, -1.0 = disabled)", params.sampling.top_n_sigma),
         [](common_params & params, const std::string & value) {
             params.sampling.top_n_sigma = std::stof(value);
@@ -1634,7 +1634,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_sparam());
     add_opt(common_arg(
-        {"--typical"}, "N",
+        {"--typical", "--typical-p"}, "N",
         string_format("locally typical sampling, parameter p (default: %.2f, 1.0 = disabled)", (double)params.sampling.typ_p),
         [](common_params & params, const std::string & value) {
             params.sampling.typ_p = std::stof(value);
@@ -2399,7 +2399,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 params.fit_params = false;
             } else {
                 throw std::runtime_error(
-                    string_format("error: unkown value for --fit: '%s'\n", value.c_str()));
+                    string_format("error: unknown value for --fit: '%s'\n", value.c_str()));
             }
         }
     ).set_env("LLAMA_ARG_FIT"));
@@ -2520,11 +2520,28 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ));
     add_opt(common_arg(
         {"-a", "--alias"}, "STRING",
-        "set alias for model name (to be used by REST API)",
+        "set model name aliases, comma-separated (to be used by API)",
         [](common_params & params, const std::string & value) {
-            params.model_alias = value;
+            for (auto & alias : string_split<std::string>(value, ',')) {
+                alias = string_strip(alias);
+                if (!alias.empty()) {
+                    params.model_alias.insert(alias);
+                }
+            }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ALIAS"));
+    add_opt(common_arg(
+        {"--tags"}, "STRING",
+        "set model tags, comma-separated (informational, not used for routing)",
+        [](common_params & params, const std::string & value) {
+            for (auto & tag : string_split<std::string>(value, ',')) {
+                tag = string_strip(tag);
+                if (!tag.empty()) {
+                    params.model_tags.insert(tag);
+                }
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TAGS"));
     add_opt(common_arg(
         {"-m", "--model"}, "FNAME",
         ex == LLAMA_EXAMPLE_EXPORT_LORA
@@ -3437,16 +3454,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.ngram_size_m = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
-    add_opt(common_arg(
-        {"--spec-ngram-check-rate"}, "N",
-        string_format("ngram check rate for ngram-simple/ngram-map speculative decoding (default: %d)", params.speculative.ngram_check_rate),
-        [](common_params & params, int value) {
-            if (value < 1) {
-                throw std::invalid_argument("ngram check rate must be at least 1");
-            }
-            params.speculative.ngram_check_rate = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--spec-ngram-min-hits"}, "N",
         string_format("minimum hits for ngram-map speculative decoding (default: %d)", params.speculative.ngram_min_hits),

common/chat-parser-xml-toolcall.cpp

@@ -803,7 +803,7 @@ inline void parse_msg_with_xml_tool_calls(common_chat_msg_parser & builder, cons
         }
 
         // remove potential partial suffix
-        if (builder.pos() == builder.input().size()) {
+        if (builder.pos() == builder.input().size() && builder.is_partial()) {
             if (unclosed_reasoning_content.empty()) {
                 rstrip(content);
                 trim_potential_partial_word(content);

common/chat-parser.cpp

@@ -893,23 +893,6 @@ static void common_chat_parse_minimax_m2(common_chat_msg_parser & builder) {
     builder.consume_reasoning_with_xml_tool_calls(form, "<think>", "</think>");
 }
 
-static void common_chat_parse_qwen3_coder_xml(common_chat_msg_parser & builder) {
-    static const xml_tool_call_format form = ([]() {
-        xml_tool_call_format form {};
-        form.scope_start = "<tool_call>";
-        form.tool_start  = "<function=";
-        form.tool_sep    = ">";
-        form.key_start   = "<parameter=";
-        form.key_val_sep = ">";
-        form.val_end     = "</parameter>";
-        form.tool_end    = "</function>";
-        form.scope_end   = "</tool_call>";
-        form.trim_raw_argval = true;
-        return form;
-    })();
-    builder.consume_reasoning_with_xml_tool_calls(form);
-}
-
 static void common_chat_parse_kimi_k2(common_chat_msg_parser & builder) {
     static const xml_tool_call_format form = ([]() {
         xml_tool_call_format form {};
@@ -1590,9 +1573,6 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
         case COMMON_CHAT_FORMAT_KIMI_K2:
             common_chat_parse_kimi_k2(builder);
             break;
-        case COMMON_CHAT_FORMAT_QWEN3_CODER_XML:
-            common_chat_parse_qwen3_coder_xml(builder);
-            break;
         case COMMON_CHAT_FORMAT_APRIEL_1_5:
             common_chat_parse_apriel_1_5(builder);
             break;

common/chat.cpp

@@ -65,14 +65,25 @@ json common_chat_msg::to_json_oaicompat(bool concat_typed_text) const {
     } else if (!content_parts.empty()) {
         if (concat_typed_text) {
             std::string text;
+            bool last_was_media_marker = false;
+            // join parts with newline, do not add newline before or after media markers
             for (const auto & part : content_parts) {
-                if (part.type != "text") {
+                bool add_new_line = true;
+                if (part.type == "text") {
+                    add_new_line = !last_was_media_marker && !text.empty();
+                    last_was_media_marker = false;
+                } else if (part.type == "media_marker") {
+                    add_new_line = false;
+                    last_was_media_marker = true;
+                } else {
                     LOG_WRN("Ignoring content part type: %s\n", part.type.c_str());
                     continue;
                 }
-                if (!text.empty()) {
+
+                if (add_new_line) {
                     text += '\n';
                 }
+
                 text += part.text;
             }
             jmsg["content"] = text;
@@ -319,7 +330,7 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
                             throw std::invalid_argument("Missing content part type: " + part.dump());
                         }
                         const auto & type = part.at("type");
-                        if (type != "text") {
+                        if (type != "text" && type != "media_marker") {
                             throw std::invalid_argument("Unsupported content part type: " + type.dump());
                         }
                         common_chat_msg_content_part msg_part;
@@ -380,15 +391,46 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
     return msgs;
 }
 
-json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
+static json render_message_to_json(const std::vector<common_chat_msg> & msgs, const jinja::caps & c) {
+    if (!c.supports_string_content && !c.supports_typed_content) {
+        LOG_WRN("%s: Neither string content nor typed content is supported by the template. This is unexpected and may lead to issues.\n", __func__);
+    }
+
+    bool only_string_accepted =  c.supports_string_content && !c.supports_typed_content;
+    bool only_typed_accepted  = !c.supports_string_content &&  c.supports_typed_content;
+
     json messages = json::array();
     for (const auto & msg : msgs) {
-        json jmsg = msg.to_json_oaicompat(concat_typed_text);
-        messages.push_back(jmsg);
+        if (only_string_accepted) {
+            json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ true);
+            messages.push_back(jmsg);
+        } else if (only_typed_accepted) {
+            json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
+            if (jmsg.at("content").is_string()) {
+                jmsg["content"] = json::array({
+                    json{
+                        {"type", "text"},
+                        {"text", jmsg.at("content").get<std::string>()},
+                    }
+                });
+            }
+            messages.push_back(jmsg);
+        } else {
+            json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
+            messages.push_back(jmsg);
+        }
     }
     return messages;
 }
 
+// DEPRECATED: only used in tests
+json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
+    jinja::caps c;
+    c.supports_string_content = true;
+    c.supports_typed_content = !concat_typed_text;
+    return render_message_to_json(msgs, c);
+}
+
 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & tools) {
     std::vector<common_chat_tool> result;
 
@@ -694,7 +736,6 @@ const char * common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_MINIMAX_M2: return "MiniMax-M2";
         case COMMON_CHAT_FORMAT_GLM_4_5: return "GLM 4.5";
         case COMMON_CHAT_FORMAT_KIMI_K2: return "Kimi K2";
-        case COMMON_CHAT_FORMAT_QWEN3_CODER_XML: return "Qwen3 Coder";
         case COMMON_CHAT_FORMAT_APRIEL_1_5: return "Apriel 1.5";
         case COMMON_CHAT_FORMAT_XIAOMI_MIMO: return "Xiaomi MiMo";
         case COMMON_CHAT_FORMAT_SOLAR_OPEN: return "Solar Open";
@@ -1480,14 +1521,17 @@ static common_chat_params common_chat_params_init_nemotron_v2(const common_chat_
     return data;
 }
 
-static common_chat_params common_chat_params_init_nemotron_v3(const common_chat_template & tmpl, const struct templates_params & inputs) {
+static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
 
     data.prompt = apply(tmpl, inputs);
     data.format = COMMON_CHAT_FORMAT_PEG_CONSTRUCTED;
 
+    // Nemotron Nano 3 and Step-3.5-Flash use the Qwen3 Coder tool calling with thinking
+    bool supports_reasoning = (tmpl.source().find("<think>") != std::string::npos);
+
     // Handle thinking tags appropriately based on inputs.enable_thinking
-    if (string_ends_with(data.prompt, "<think>\n")) {
+    if (supports_reasoning && string_ends_with(data.prompt, "<think>\n")) {
         if (!inputs.enable_thinking) {
             data.prompt += "</think>";
         } else {
@@ -1496,19 +1540,21 @@ static common_chat_params common_chat_params_init_nemotron_v3(const common_chat_
     }
 
     data.preserved_tokens = {
-        "<think>",
-        "</think>",
         "<tool_call>",
         "</tool_call>",
     };
 
+    if (supports_reasoning) {
+        data.preserved_tokens.insert(data.preserved_tokens.end(), {"<think>", "</think>"});
+    }
+
     auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
     auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
     auto include_grammar = true;
 
     auto parser = build_chat_peg_constructed_parser([&](auto & p) {
         auto reasoning = p.eps();
-        if (inputs.enable_thinking && extract_reasoning) {
+        if (supports_reasoning && inputs.enable_thinking && extract_reasoning) {
             auto reasoning_content = p.reasoning(p.until("</think>")) + ("</think>" | p.end());
             if (data.thinking_forced_open) {
                 reasoning = reasoning_content;
@@ -1846,38 +1892,6 @@ static common_chat_params common_chat_params_init_minimax_m2(const common_chat_t
     return data;
 }
 
-static common_chat_params common_chat_params_init_qwen3_coder_xml(const common_chat_template & tmpl, const struct templates_params & params) {
-    common_chat_params data;
-    data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-    data.prompt = apply(tmpl, params);
-    data.format = COMMON_CHAT_FORMAT_QWEN3_CODER_XML;
-
-    data.preserved_tokens = {
-        "<tool_call>",
-        "</tool_call>",
-        "<function=",
-        "</function>",
-        "<parameter=",
-        "</parameter>",
-    };
-
-    // build grammar for tool call
-    static const xml_tool_call_format form {
-        /* form.scope_start = */ "<tool_call>\n",
-        /* form.tool_start  = */ "<function=",
-        /* form.tool_sep    = */ ">\n",
-        /* form.key_start   = */ "<parameter=",
-        /* form.key_val_sep = */ ">\n",
-        /* form.val_end     = */ "\n</parameter>\n",
-        /* form.tool_end    = */ "</function>\n",
-        /* form.scope_end   = */ "</tool_call>",
-    };
-    build_grammar_xml_tool_call(data, params.tools, form);
-
-    return data;
-}
-
 static common_chat_params common_chat_params_init_kimi_k2(const common_chat_template & tmpl, const struct templates_params & params) {
     common_chat_params data;
     data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
@@ -2001,6 +2015,7 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
         if (has_reasoning_content && has_tool_calls) {
             auto adjusted_message = msg;
             adjusted_message["thinking"] = msg.at("reasoning_content");
+            adjusted_message.erase("content");
             adjusted_messages.push_back(adjusted_message);
         } else {
             adjusted_messages.push_back(msg);
@@ -3020,7 +3035,7 @@ static common_chat_params common_chat_templates_apply_jinja(
         : *tmpls->template_default;
     const auto & src = tmpl.source();
     const auto & caps = tmpl.original_caps();
-    params.messages = common_chat_msgs_to_json_oaicompat(inputs.messages, /* concat_text= */ !tmpl.original_caps().requires_typed_content);
+    params.messages = render_message_to_json(inputs.messages, tmpl.original_caps());
     params.add_generation_prompt = inputs.add_generation_prompt;
     params.tool_choice = inputs.tool_choice;
     params.reasoning_format = inputs.reasoning_format;
@@ -3098,19 +3113,13 @@ static common_chat_params common_chat_templates_apply_jinja(
     }
 
     // Qwen3-Coder XML format detection (must come before Hermes 2 Pro)
-    // Detect via explicit XML markers unique to Qwen3-Coder to avoid false positives in other templates.
-    // Require presence of <tool_call>, <function=...>, and <parameter=...> blocks.
+    // Detect via XML markers: <tool_call>, <function=...>, and <parameter=...> blocks.
+    // Also matches Step-3.5-Flash and Nemotron 3 Nano which use the same output format.
     if (src.find("<tool_call>") != std::string::npos &&
-        src.find("<function>") != std::string::npos &&
         src.find("<function=") != std::string::npos &&
-        src.find("<parameters>") != std::string::npos &&
         src.find("<parameter=") != std::string::npos) {
         workaround::func_args_not_string(params.messages);
-        // Nemotron 3 Nano 30B A3B
-        if (src.find("<think>") != std::string::npos) {
-            return common_chat_params_init_nemotron_v3(tmpl, params);
-        }
-        return common_chat_params_init_qwen3_coder_xml(tmpl, params);
+        return common_chat_params_init_qwen3_coder(tmpl, params);
     }
 
     // Xiaomi MiMo format detection (must come before Hermes 2 Pro)
@@ -3276,7 +3285,7 @@ static common_chat_params common_chat_templates_apply_legacy(
     for (const auto & msg : inputs.messages) {
         auto content = msg.content;
         for (const auto & part : msg.content_parts) {
-            if (part.type != "text") {
+            if (part.type != "text" && part.type != "media_marker") {
                 LOG_WRN("Ignoring non-text content part: %s\n", part.type.c_str());
                 continue;
             }

common/chat.h

@@ -128,7 +128,6 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_GLM_4_5,
     COMMON_CHAT_FORMAT_MINIMAX_M2,
     COMMON_CHAT_FORMAT_KIMI_K2,
-    COMMON_CHAT_FORMAT_QWEN3_CODER_XML,
     COMMON_CHAT_FORMAT_APRIEL_1_5,
     COMMON_CHAT_FORMAT_XIAOMI_MIMO,
     COMMON_CHAT_FORMAT_SOLAR_OPEN,
@@ -240,6 +239,8 @@ bool common_chat_templates_support_enable_thinking(const common_chat_templates *
 
 // Parses a JSON array of messages in OpenAI's chat completion API format.
 std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
+
+// DEPRECATED: only used in tests
 nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);

common/common.cpp

@@ -1,7 +1,3 @@
-#if defined(_MSC_VER)
-#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
-#endif
-
 #include "ggml.h"
 #include "gguf.h"
 
@@ -9,12 +5,12 @@
 #include "log.h"
 #include "llama.h"
 #include "sampling.h"
+#include "unicode.h"
 
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
 #include <cmath>
-#include <codecvt>
 #include <chrono>
 #include <cstdarg>
 #include <cstring>
@@ -456,34 +452,6 @@ void string_replace_all(std::string & s, const std::string & search, const std::
     s = std::move(builder);
 }
 
-bool string_ends_with(const std::string_view & str, const std::string_view & suffix) {
-    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
-}
-
-bool string_remove_suffix(std::string & str, const std::string_view & suffix) {
-    bool has_suffix = string_ends_with(str, suffix);
-    if (has_suffix) {
-        str = str.substr(0, str.size() - suffix.size());
-    }
-    return has_suffix;
-}
-
-size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop) {
-    if (!str.empty() && !stop.empty()) {
-        const char text_last_char = str.back();
-        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
-            if (stop[char_index] == text_last_char) {
-                const auto current_partial = stop.substr(0, char_index + 1);
-                if (string_ends_with(str, current_partial)) {
-                    return str.size() - char_index - 1;
-                }
-            }
-        }
-    }
-
-    return std::string::npos;
-}
-
 std::string regex_escape(const std::string & s) {
     static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
     return std::regex_replace(s, special_chars, "\\$&");
@@ -706,45 +674,28 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
         return false;
     }
 
-    std::u32string filename_utf32;
-    try {
-#if defined(__clang__)
-        // disable C++17 deprecation warning for std::codecvt_utf8
-#    pragma clang diagnostic push
-#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
-#endif
+    size_t offset = 0;
+    while (offset < filename.size()) {
+        utf8_parse_result result = parse_utf8_codepoint(filename, offset);
 
-        std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
-
-#if defined(__clang__)
-#    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
-#endif
-
-        filename_utf32 = converter.from_bytes(filename);
-
-        // If the reverse conversion mismatches, it means overlong UTF-8 sequences were used,
-        // or invalid encodings were encountered. Reject such attempts
-        std::string filename_reencoded = converter.to_bytes(filename_utf32);
-        if (filename_reencoded != filename) {
+        if (result.status != utf8_parse_result::SUCCESS) {
             return false;
         }
-    } catch (const std::exception &) {
-        return false;
-    }
+        uint32_t c = result.codepoint;
 
-    // Check for forbidden codepoints:
-    // - Control characters
-    // - Unicode equivalents of illegal characters
-    // - UTF-16 surrogate pairs
-    // - UTF-8 replacement character
-    // - Byte order mark (BOM)
-    // - Illegal characters: / \ : * ? " < > |
-    for (char32_t c : filename_utf32) {
+        if ((result.bytes_consumed == 2 && c < 0x80) ||
+            (result.bytes_consumed == 3 && c < 0x800) ||
+            (result.bytes_consumed == 4 && c < 0x10000)) {
+            return false;
+        }
+
+        // Check for forbidden codepoints:
+        // - Control characters
+        // - Unicode equivalents of illegal characters
+        // - UTF-16 surrogate pairs
+        // - UTF-8 replacement character
+        // - Byte order mark (BOM)
+        // - Illegal characters: / \ : * ? " < > |
         if (c <= 0x1F // Control characters (C0)
             || c == 0x7F // Control characters (DEL)
             || (c >= 0x80 && c <= 0x9F) // Control characters (C1)
@@ -752,6 +703,7 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
             || c == 0x2215 // Division Slash (forward slash equivalent)
             || c == 0x2216 // Set Minus (backslash equivalent)
             || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
+            || c > 0x10FFFF // Max Unicode limit
             || c == 0xFFFD // Replacement Character (UTF-8)
             || c == 0xFEFF // Byte Order Mark (BOM)
             || c == ':' || c == '*' // Illegal characters
@@ -762,6 +714,7 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
             // Subdirectories not allowed, reject path separators
             return false;
         }
+        offset += result.bytes_consumed;
     }
 
     // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
@@ -898,7 +851,8 @@ std::string fs_get_cache_directory() {
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
     } else {
-#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || defined(__OpenBSD__)
+#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) || \
+        defined(__OpenBSD__) || defined(__NetBSD__)
         if (std::getenv("XDG_CACHE_HOME")) {
             cache_directory = std::getenv("XDG_CACHE_HOME");
         } else if (std::getenv("HOME")) {
@@ -1242,7 +1196,7 @@ common_init_result_ptr common_init_from_params(common_params & params) {
             return res;
         }
 
-        int err = llama_apply_adapter_cvec(
+        int err = llama_set_adapter_cvec(
                 lctx,
                 cvec.data.data(),
                 cvec.data.size(),
@@ -1344,12 +1298,15 @@ std::string get_model_endpoint() {
 }
 
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
-    llama_clear_adapter_lora(ctx);
-    for (auto & la : lora) {
-        if (la.scale != 0.0f) {
-            llama_set_adapter_lora(ctx, la.ptr, la.scale);
-        }
+    std::vector<llama_adapter_lora *> loras;
+    std::vector<float> scales;
+
+    for (auto & la: lora) {
+        loras.push_back(la.ptr);
+        scales.push_back(la.scale);
     }
+
+    llama_set_adapters_lora(ctx, loras.data(), loras.size(), scales.data());
 }
 
 struct llama_model_params common_model_params_to_llama(common_params & params) {
@@ -1469,66 +1426,6 @@ void common_batch_add(
     batch.n_tokens++;
 }
 
-//
-// Token utils
-//
-
-size_t common_lcp(const llama_tokens & a, const llama_tokens & b) {
-    size_t i;
-    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
-
-    return i;
-}
-
-size_t common_lcs(const llama_tokens & a, const llama_tokens & b) {
-    // check for empty sequences
-    if (a.empty() || b.empty()) {
-        return 0;
-    }
-
-    // get the lengths of the input sequences
-    size_t a_len = a.size();
-    size_t b_len = b.size();
-
-    // initialize the maximum length of the longest common subsequence (LCS)
-    size_t max_length = 0;
-
-    // use two rows instead of a 2D matrix to optimize space
-    std::vector<size_t> prev_row(b_len + 1, 0);
-    std::vector<size_t> curr_row(b_len + 1, 0);
-
-    // iterate through the elements of a
-    for (size_t i = 1; i <= a_len; i++) {
-        // iterate through the elements of b
-        for (size_t j = 1; j <= b_len; j++) {
-            // if elements at the current positions match
-            if (a[i - 1] == b[j - 1]) {
-                // if it's the first element of either sequences, set LCS length to 1
-                if (i == 1 || j == 1) {
-                    curr_row[j] = 1;
-                } else {
-                    // increment LCS length by 1 compared to the previous element
-                    curr_row[j] = prev_row[j - 1] + 1;
-                }
-
-                // update max_length if necessary
-                if (curr_row[j] > max_length) {
-                    max_length = curr_row[j];
-                }
-            } else {
-                // reset LCS length if elements don't match
-                curr_row[j] = 0;
-            }
-        }
-
-        // update the previous row for the next iteration
-        prev_row = curr_row;
-    }
-
-    // return the maximum length of the LCS
-    return max_length;
-}
-
 //
 // Vocab utils
 //
@@ -1863,3 +1760,65 @@ float lr_opt::get_lr(float epoch) const {
     LOG_INF("epoch %.2g lr=%.2g\n", epoch, r);
     return r;
 }
+
+bool common_replay_last_token(struct llama_context * ctx, llama_token last_token, int32_t pos) {
+    llama_batch batch = llama_batch_get_one(&last_token, 1);
+    batch.pos = &pos;
+    if (llama_decode(ctx, batch)) {
+        LOG_ERR("%s: failed to replay last token\n", __func__);
+        return false;
+    }
+    return true;
+}
+
+bool common_prompt_batch_decode(
+              struct llama_context * ctx,
+    const std::vector<llama_token> & tokens,
+                               int & n_past,
+                               int   n_batch,
+                  std::string_view   state_path,
+                              bool   save_state) {
+    const int n_eval = tokens.size();
+    if (n_eval == 0) {
+        return true;
+    }
+
+    if (save_state && n_eval > 1) {
+        const int n_tokens_before_last = n_eval - 1;
+
+        GGML_ASSERT(n_eval <= n_batch);
+
+        // Decode all but the last token so we can save the memory state before decoding the last token.
+        // This is done so we can restore the session state later and replay the last token.
+        // Memory implementations in recurrent/hybrid models don't support removing tokens from their
+        // memory, so we can't just remove the last token from the memory and replay the last token which
+        // is the reason for this logic.
+        if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(tokens.data()), n_tokens_before_last))) {
+            LOG_ERR("%s : failed to eval\n", __func__);
+            return false;
+        }
+        n_past += n_tokens_before_last;
+
+        llama_state_save_file(ctx, state_path.data(), tokens.data(), n_tokens_before_last);
+        LOG_INF("saved session before last token to %s, n_tokens = %d\n", state_path.data(), n_tokens_before_last);
+
+        llama_token last_token = tokens.back();
+        llama_batch batch = llama_batch_get_one(&last_token, 1);
+        int32_t pos = n_past;
+        batch.pos = &pos;
+
+        if (llama_decode(ctx, batch)) {
+            LOG_ERR("%s : failed to eval last token\n", __func__);
+            return false;
+        }
+        n_past++;
+    } else {
+        if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(tokens.data()), n_eval))) {
+            LOG_ERR("%s : failed to eval\n", __func__);
+            return false;
+        }
+        n_past += n_eval;
+    }
+
+    return true;
+}

common/common.h

@@ -269,7 +269,6 @@ struct common_params_speculative {
 
     uint16_t ngram_size_n     = 12; // ngram size for lookup
     uint16_t ngram_size_m     = 48; // mgram size for speculative tokens
-    uint16_t ngram_check_rate =  1; // check rate for ngram lookup
     uint16_t ngram_min_hits   =  1; // minimum hits at ngram/mgram lookup for mgram to be proposed
 
     std::shared_ptr<common_ngram_mod> ngram_mod;
@@ -411,7 +410,8 @@ struct common_params {
 
     struct common_params_model model;
 
-    std::string model_alias          = ""; // model alias                                                   // NOLINT
+    std::set<std::string> model_alias;     // model aliases                                                 // NOLINT
+    std::set<std::string> model_tags;      // model tags (informational, not used for routing)              // NOLINT
     std::string hf_token             = ""; // HF token                                                      // NOLINT
     std::string prompt               = "";                                                                  // NOLINT
     std::string system_prompt        = "";                                                                  // NOLINT
@@ -671,30 +671,55 @@ static std::vector<T> string_split(const std::string & str, char delim) {
 }
 
 template<>
-std::vector<std::string> string_split<std::string>(const std::string & input, char separator)
+inline std::vector<std::string> string_split<std::string>(const std::string & str, char delim)
 {
     std::vector<std::string> parts;
     size_t begin_pos = 0;
-    size_t separator_pos = input.find(separator);
-    while (separator_pos != std::string::npos) {
-        std::string part = input.substr(begin_pos, separator_pos - begin_pos);
+    size_t delim_pos = str.find(delim);
+    while (delim_pos != std::string::npos) {
+        std::string part = str.substr(begin_pos, delim_pos - begin_pos);
         parts.emplace_back(part);
-        begin_pos = separator_pos + 1;
-        separator_pos = input.find(separator, begin_pos);
+        begin_pos = delim_pos + 1;
+        delim_pos = str.find(delim, begin_pos);
     }
-    parts.emplace_back(input.substr(begin_pos, separator_pos - begin_pos));
+    parts.emplace_back(str.substr(begin_pos));
     return parts;
 }
 
-static bool string_starts_with(const std::string & str,
-                               const std::string & prefix) {  // While we wait for C++20's std::string::starts_with...
-    return str.rfind(prefix, 0) == 0;
+// remove when moving to c++20
+inline bool string_starts_with(std::string_view str, std::string_view prefix) {
+    return str.size() >= prefix.size() &&
+           str.compare(0, prefix.size(), prefix) == 0;
 }
 
-// While we wait for C++20's std::string::ends_with...
-bool string_ends_with(const std::string_view & str, const std::string_view & suffix);
-bool string_remove_suffix(std::string & str, const std::string_view & suffix);
-size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop);
+// remove when moving to c++20
+inline bool string_ends_with(std::string_view str, std::string_view suffix) {
+    return str.size() >= suffix.size() &&
+           str.compare(str.size() - suffix.size(), suffix.size(), suffix) == 0;
+}
+
+inline bool string_remove_suffix(std::string & str, std::string_view suffix) {
+    if (string_ends_with(str, suffix)) {
+        str.resize(str.size() - suffix.size());
+        return true;
+    }
+    return false;
+}
+
+inline size_t string_find_partial_stop(std::string_view str, std::string_view stop) {
+    if (!str.empty() && !stop.empty()) {
+        const size_t max_len = std::min(str.size(), stop.size());
+        const char last_char = str.back();
+        for (size_t len = max_len; len > 0; --len) {
+            if (stop[len - 1] == last_char) {
+                if (string_ends_with(str, stop.substr(0, len))) {
+                    return str.size() - len;
+                }
+            }
+        }
+    }
+    return std::string::npos;
+}
 
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
@@ -780,15 +805,22 @@ void common_batch_add(
     const std::vector<llama_seq_id> & seq_ids,
                                bool   logits);
 
+// decodes a single batch of tokens for a prompt and manages session tokens
 //
-// Token utils
-//
+// Note: We save state before the last token so that we can replay it to ensure
+// compatibility with all memory types. Recurrent/hybrid models cannot remove
+// tokens from memory, so this approach works across all model architectures.
+bool common_prompt_batch_decode(
+              struct llama_context * ctx,
+    const std::vector<llama_token> & embd,
+                               int & n_past,
+                               int   n_batch,
+                  std::string_view   state_path,
+                              bool   save_state);
 
-// longest common prefix
-size_t common_lcp(const llama_tokens & a, const llama_tokens & b);
-
-// longet common subsequence
-size_t common_lcs(const llama_tokens & a, const llama_tokens & b);
+// replays the last token after loading state to regenerate logits
+// used after loading session state to ensure the sampling context has valid logits
+bool common_replay_last_token(struct llama_context * ctx, llama_token last_token, int32_t pos);
 
 //
 // Vocab utils
@@ -837,7 +869,7 @@ std::string common_detokenize(
 // Embedding utils
 //
 
-// TODO: repace embd_norm with an enum
+// TODO: replace embd_norm with an enum
 void common_embd_normalize(const float * inp, float * out, int n, int embd_norm);
 
 float common_embd_similarity_cos(const float * embd1, const float * embd2, int n);
@@ -881,11 +913,11 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 
 const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_(ch|)exps";
 
-static std::string llm_ffn_exps_block_regex(int idx) {
+inline std::string llm_ffn_exps_block_regex(int idx) {
     return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
 }
 
-static llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
+inline llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
     return { LLM_FFN_EXPS_REGEX, ggml_backend_cpu_buffer_type() };
 }
 

common/debug.cpp

@@ -45,6 +45,8 @@ static float common_ggml_get_float_value(const uint8_t * data,
     return v;
 }
 
+#define INDENT "    "
+
 template <bool abort>
 void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
     GGML_ASSERT(n > 0);
@@ -60,41 +62,41 @@ void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * n
         }
     }
     for (int64_t i3 = 0; i3 < ne[3]; i3++) {
-        LOG_ERR("                                     [\n");
+        LOG(INDENT "[\n");
         for (int64_t i2 = 0; i2 < ne[2]; i2++) {
             if (i2 == n && ne[2] > 2 * n) {
-                LOG_ERR("                                      ..., \n");
+                LOG(INDENT INDENT "..., \n");
                 i2 = ne[2] - n;
             }
-            LOG_ERR("                                      [\n");
+            LOG(INDENT INDENT "[\n");
             for (int64_t i1 = 0; i1 < ne[1]; i1++) {
                 if (i1 == n && ne[1] > 2 * n) {
-                    LOG_ERR("                                       ..., \n");
+                    LOG(INDENT INDENT INDENT "..., \n");
                     i1 = ne[1] - n;
                 }
-                LOG_ERR("                                       [");
+                LOG(INDENT INDENT INDENT "[");
                 for (int64_t i0 = 0; i0 < ne[0]; i0++) {
                     if (i0 == n && ne[0] > 2 * n) {
-                        LOG_ERR("..., ");
+                        LOG("   ..., ");
                         i0 = ne[0] - n;
                     }
                     const float v = common_ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
-                    LOG_ERR("%12.4f", v);
+                    LOG("%12.4f", v);
                     if (i0 < ne[0] - 1) {
-                        LOG_ERR(", ");
+                        LOG(", ");
                     }
                 }
-                LOG_ERR("],\n");
+                LOG("  ],\n");
             }
-            LOG_ERR("                                      ],\n");
+            LOG(INDENT INDENT "],\n");
         }
-        LOG_ERR("                                     ]\n");
-        LOG_ERR("                                     sum = %f\n", sum);
+        LOG(INDENT "]\n");
+        LOG(INDENT "sum = %f\n", sum);
     }
 
     if constexpr (abort) {
         if (std::isnan(sum)) {
-            LOG_ERR("encountered NaN - aborting\n");
+            LOG("encountered NaN - aborting\n");
             exit(0);
         }
     }
@@ -137,9 +139,9 @@ template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, b
     }
 
     if (matches_filter) {
-        LOG_ERR("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
-                ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
-                common_ggml_ne_string(t).c_str());
+        LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
+            ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
+            common_ggml_ne_string(t).c_str());
     }
 
     const bool is_host = ggml_backend_buffer_is_host(t->buffer);

common/debug.h

@@ -18,7 +18,7 @@ template <bool abort_on_nan> void common_debug_print_tensor(uint8_t * data, ggml
 // prints tensors that are processed in the computation graph
 // by default prints all tensors, but can be configured by creating a `base_callback_data` instance with
 // non-empty filter_patterns. See examples/debug.ccp for possible usage patterns
-// The template parameter determins whether an error should be thrown whenever a NaN is encountered
+// The template parameter determines whether an error should be thrown whenever a NaN is encountered
 // in a tensor (useful for stopping debug sessions on first erroneous tensor)
 // The callback data will be passed as the third parameter (user_data)
 template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data);

common/download.cpp

@@ -19,9 +19,7 @@
 #include <thread>
 #include <vector>
 
-#if defined(LLAMA_USE_HTTPLIB)
 #include "http.h"
-#endif
 
 #ifndef __EMSCRIPTEN__
 #ifdef __linux__
@@ -114,44 +112,18 @@ static void write_etag(const std::string & path, const std::string & etag) {
 }
 
 static std::string read_etag(const std::string & path) {
-    std::string none;
     const std::string etag_path = path + ".etag";
-
-    if (std::filesystem::exists(etag_path)) {
-        std::ifstream etag_in(etag_path);
-        if (!etag_in) {
-            LOG_ERR("%s: could not open .etag file for reading: %s\n", __func__, etag_path.c_str());
-            return none;
-        }
-        std::string etag;
-        std::getline(etag_in, etag);
-        return etag;
+    if (!std::filesystem::exists(etag_path)) {
+        return {};
     }
-
-    // no etag file, but maybe there is an old .json
-    // remove this code later
-    const std::string metadata_path = path + ".json";
-
-    if (std::filesystem::exists(metadata_path)) {
-        std::ifstream metadata_in(metadata_path);
-        try {
-            nlohmann::json metadata_json;
-            metadata_in >> metadata_json;
-            LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(),
-                    metadata_json.dump().c_str());
-            if (metadata_json.contains("etag") && metadata_json.at("etag").is_string()) {
-                std::string etag = metadata_json.at("etag");
-                write_etag(path, etag);
-                if (!std::filesystem::remove(metadata_path)) {
-                    LOG_WRN("%s: failed to delete old .json metadata file: %s\n", __func__, metadata_path.c_str());
-                }
-                return etag;
-            }
-        } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-        }
+    std::ifstream etag_in(etag_path);
+    if (!etag_in) {
+        LOG_ERR("%s: could not open .etag file for reading: %s\n", __func__, etag_path.c_str());
+        return {};
     }
-    return none;
+    std::string etag;
+    std::getline(etag_in, etag);
+    return etag;
 }
 
 static bool is_http_status_ok(int status) {
@@ -168,8 +140,6 @@ std::pair<std::string, std::string> common_download_split_repo_tag(const std::st
     return {hf_repo, tag};
 }
 
-#if defined(LLAMA_USE_HTTPLIB)
-
 class ProgressBar {
     static inline std::mutex mutex;
     static inline std::map<const ProgressBar *, int> lines;
@@ -305,7 +275,10 @@ static bool common_pull_file(httplib::Client & cli,
     );
 
     if (!res) {
-        LOG_ERR("%s: error during download. Status: %d\n", __func__, res ? res->status : -1);
+        LOG_ERR("%s: download failed: %s (status: %d)\n",
+                __func__,
+                httplib::to_string(res.error()).c_str(),
+                res ? res->status : -1);
         return false;
     }
 
@@ -344,62 +317,64 @@ static int common_download_file_single_online(const std::string        & url,
         LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
     }
 
-    for (int i = 0; i < max_attempts; ++i) {
-        auto head = cli.Head(parts.path);
-        bool head_ok = head && head->status >= 200 && head->status < 300;
-        if (!head_ok) {
-            LOG_WRN("%s: HEAD invalid http status code received: %d\n", __func__, head ? head->status : -1);
-            if (file_exists) {
-                LOG_INF("%s: Using cached file (HEAD failed): %s\n", __func__, path.c_str());
-                return 304; // 304 Not Modified - fake cached response
-            }
-            return head->status; // cannot use cached file, return raw status code
-            // TODO: maybe retry only on certain codes
-        }
-
-        std::string etag;
-        if (head_ok && head->has_header("ETag")) {
-            etag = head->get_header_value("ETag");
-        }
-
-        size_t total_size = 0;
-        if (head_ok && head->has_header("Content-Length")) {
-            try {
-                total_size = std::stoull(head->get_header_value("Content-Length"));
-            } catch (const std::exception& e) {
-                LOG_WRN("%s: Invalid Content-Length in HEAD response: %s\n", __func__, e.what());
-            }
-        }
-
-        bool supports_ranges = false;
-        if (head_ok && head->has_header("Accept-Ranges")) {
-            supports_ranges = head->get_header_value("Accept-Ranges") != "none";
-        }
-
-        bool should_download_from_scratch = false;
-        if (!last_etag.empty() && !etag.empty() && last_etag != etag) {
-            LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__,
-                    last_etag.c_str(), etag.c_str());
-            should_download_from_scratch = true;
-        }
-
+    auto head = cli.Head(parts.path);
+    if (!head || head->status < 200 || head->status >= 300) {
+        LOG_WRN("%s: HEAD failed, status: %d\n", __func__, head ? head->status : -1);
         if (file_exists) {
-            if (!should_download_from_scratch) {
-                LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
-                return 304; // 304 Not Modified - fake cached response
-            }
-            LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
-            if (remove(path.c_str()) != 0) {
-                LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
-                return -1;
-            }
+            LOG_INF("%s: using cached file (HEAD failed): %s\n", __func__, path.c_str());
+            return 304; // 304 Not Modified - fake cached response
+        }
+        return head ? head->status : -1;
+    }
+
+    std::string etag;
+    if (head->has_header("ETag")) {
+        etag = head->get_header_value("ETag");
+    }
+
+    size_t total_size = 0;
+    if (head->has_header("Content-Length")) {
+        try {
+            total_size = std::stoull(head->get_header_value("Content-Length"));
+        } catch (const std::exception& e) {
+            LOG_WRN("%s: invalid Content-Length in HEAD response: %s\n", __func__, e.what());
+        }
+    }
+
+    bool supports_ranges = false;
+    if (head->has_header("Accept-Ranges")) {
+        supports_ranges = head->get_header_value("Accept-Ranges") != "none";
+    }
+
+    if (file_exists) {
+        if (etag.empty()) {
+            LOG_INF("%s: using cached file (no server etag): %s\n", __func__, path.c_str());
+            return 304; // 304 Not Modified - fake cached response
+        }
+        if (!last_etag.empty() && last_etag == etag) {
+            LOG_INF("%s: using cached file (same etag): %s\n", __func__, path.c_str());
+            return 304; // 304 Not Modified - fake cached response
+        }
+        if (remove(path.c_str()) != 0) {
+            LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
+            return -1;
+        }
+    }
+
+    const std::string path_temporary = path + ".downloadInProgress";
+    int delay = retry_delay_seconds;
+
+    for (int i = 0; i < max_attempts; ++i) {
+        if (i) {
+            LOG_WRN("%s: retrying after %d seconds...\n", __func__, delay);
+            std::this_thread::sleep_for(std::chrono::seconds(delay));
+            delay *= retry_delay_seconds;
         }
 
-        const std::string path_temporary = path + ".downloadInProgress";
         size_t existing_size = 0;
 
         if (std::filesystem::exists(path_temporary)) {
-            if (supports_ranges && !should_download_from_scratch) {
+            if (supports_ranges) {
                 existing_size = std::filesystem::file_size(path_temporary);
             } else if (remove(path_temporary.c_str()) != 0) {
                 LOG_ERR("%s: unable to delete file: %s\n", __func__, path_temporary.c_str());
@@ -407,32 +382,23 @@ static int common_download_file_single_online(const std::string        & url,
             }
         }
 
-        // start the download
-        LOG_INF("%s: trying to download model from %s to %s (etag:%s)...\n",
-                __func__, common_http_show_masked_url(parts).c_str(), path_temporary.c_str(), etag.c_str());
-        const bool was_pull_successful = common_pull_file(cli, parts.path, path_temporary, supports_ranges, existing_size, total_size);
-        if (!was_pull_successful) {
-            if (i + 1 < max_attempts) {
-                const int exponential_backoff_delay = std::pow(retry_delay_seconds, i) * 1000;
-                LOG_WRN("%s: retrying after %d milliseconds...\n", __func__, exponential_backoff_delay);
-                std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
-            } else {
-                LOG_ERR("%s: download failed after %d attempts\n", __func__, max_attempts);
+        LOG_INF("%s: downloading from %s to %s (etag:%s)...\n",
+                __func__, common_http_show_masked_url(parts).c_str(),
+                path_temporary.c_str(), etag.c_str());
+
+        if (common_pull_file(cli, parts.path, path_temporary, supports_ranges, existing_size, total_size)) {
+            if (std::rename(path_temporary.c_str(), path.c_str()) != 0) {
+                LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
+                return -1;
             }
-            continue;
+            if (!etag.empty()) {
+                write_etag(path, etag);
+            }
+            return head->status;
         }
-
-        if (std::rename(path_temporary.c_str(), path.c_str()) != 0) {
-            LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
-            return -1;
-        }
-        if (!etag.empty()) {
-            write_etag(path, etag);
-        }
-
-        return head->status; // TODO: use actual GET status?
     }
 
+    LOG_ERR("%s: download failed after %d attempts\n", __func__, max_attempts);
     return -1; // max attempts reached
 }
 
@@ -798,30 +764,6 @@ std::string common_docker_resolve_model(const std::string & docker) {
     }
 }
 
-#else
-
-common_hf_file_res common_get_hf_file(const std::string &, const std::string &, bool, const common_header_list &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-bool common_download_model(const common_params_model &, const std::string &, bool, const common_header_list &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-std::string common_docker_resolve_model(const std::string &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-int common_download_file_single(const std::string &,
-                                const std::string &,
-                                const std::string &,
-                                bool,
-                                const common_header_list &) {
-    throw std::runtime_error("download functionality is not enabled in this build");
-}
-
-#endif // defined(LLAMA_USE_HTTPLIB)
-
 std::vector<common_cached_model_info> common_list_cached_models() {
     std::vector<common_cached_model_info> models;
     const std::string cache_dir = fs_get_cache_directory();

common/jinja/README.md

@@ -63,7 +63,7 @@ The llama.cpp Jinja engine introduces `jinja::string` (see `jinja/string.h`), wh
   - **One-to-many** (e.g., split): result is marked `is_input` **only if ALL** input parts are marked `is_input`
   - **Many-to-one** (e.g., join): same as one-to-many
 
-For string concatenation, string parts will be appended to the new string as-is, while perserving the `is_input` flag.
+For string concatenation, string parts will be appended to the new string as-is, while preserving the `is_input` flag.
 
 **Enabling Input Marking:**
 

common/jinja/caps.cpp

@@ -63,7 +63,8 @@ static void caps_print_stats(value & v, const std::string & path) {
 
 std::map<std::string, bool> caps::to_map() const {
     return {
-        {"requires_typed_content", requires_typed_content},
+        {"supports_string_content", supports_string_content},
+        {"supports_typed_content", supports_typed_content},
         {"supports_tools", supports_tools},
         {"supports_tool_calls", supports_tool_calls},
         {"supports_parallel_tool_calls", supports_parallel_tool_calls},
@@ -89,7 +90,7 @@ caps caps_get(jinja::program & prog) {
         return v->stats.ops.find(op_name) != v->stats.ops.end();
     };
 
-    // case: typed content requirement
+    // case: typed content support
     caps_try_execute(
         prog,
         [&]() {
@@ -105,12 +106,16 @@ caps caps_get(jinja::program & prog) {
             // tools
             return json{nullptr};
         },
-        [&](bool, value & messages, value &) {
+        [&](bool success, value & messages, value &) {
             auto & content = messages->at(0)->at("content");
             caps_print_stats(content, "messages[0].content");
             if (has_op(content, "selectattr") || has_op(content, "array_access")) {
                 // accessed as an array
-                result.requires_typed_content = true;
+                result.supports_typed_content = true;
+            }
+            if (!success) {
+                // failed to execute with content as string
+                result.supports_string_content = false;
             }
         }
     );

common/jinja/caps.h

@@ -14,7 +14,9 @@ struct caps {
     bool supports_parallel_tool_calls = true;
     bool supports_preserve_reasoning = false; // support assistant message with reasoning_content
 
-    bool requires_typed_content = false; // default: use string content
+    // one of the 2 content capabilities must be true
+    bool supports_string_content = true;
+    bool supports_typed_content = false;
 
     // for reporting on server
     std::map<std::string, bool> to_map() const;

common/jinja/runtime.cpp

@@ -85,7 +85,7 @@ value identifier::execute_impl(context & ctx) {
     auto builtins = global_builtins();
     if (!it->is_undefined()) {
         if (ctx.is_get_stats) {
-            it->stats.used = true;
+            value_t::stats_t::mark_used(it);
         }
         JJ_DEBUG("Identifier '%s' found, type = %s", val.c_str(), it->type().c_str());
         return it;
@@ -277,7 +277,7 @@ value binary_expression::execute_impl(context & ctx) {
 static value try_builtin_func(context & ctx, const std::string & name, value & input, bool undef_on_missing = false) {
     JJ_DEBUG("Trying built-in function '%s' for type %s", name.c_str(), input->type().c_str());
     if (ctx.is_get_stats) {
-        input->stats.used = true;
+        value_t::stats_t::mark_used(input);
         input->stats.ops.insert(name);
     }
     auto builtins = input->get_builtins();
@@ -446,6 +446,12 @@ value for_statement::execute_impl(context & ctx) {
 
     value iterable_val = iter_expr->execute(scope);
 
+    // mark the variable being iterated as used for stats
+    if (ctx.is_get_stats) {
+        value_t::stats_t::mark_used(iterable_val);
+        iterable_val->stats.ops.insert("array_access");
+    }
+
     if (iterable_val->is_undefined()) {
         JJ_DEBUG("%s", "For loop iterable is undefined, skipping loop");
         iterable_val = mk_val<value_array>();
@@ -464,7 +470,7 @@ value for_statement::execute_impl(context & ctx) {
             items.push_back(std::move(tuple));
         }
         if (ctx.is_get_stats) {
-            iterable_val->stats.used = true;
+            value_t::stats_t::mark_used(iterable_val);
             iterable_val->stats.ops.insert("object_access");
         }
     } else {
@@ -474,7 +480,7 @@ value for_statement::execute_impl(context & ctx) {
             items.push_back(item);
         }
         if (ctx.is_get_stats) {
-            iterable_val->stats.used = true;
+            value_t::stats_t::mark_used(iterable_val);
             iterable_val->stats.ops.insert("array_access");
         }
     }
@@ -715,6 +721,8 @@ value member_expression::execute_impl(context & ctx) {
         int64_t arr_size = 0;
         if (is_val<value_array>(object)) {
             arr_size = object->as_array().size();
+        } else if (is_val<value_string>(object)) {
+            arr_size = object->as_string().length();
         }
 
         if (is_stmt<slice_expression>(this->property)) {
@@ -811,8 +819,9 @@ value member_expression::execute_impl(context & ctx) {
     }
 
     if (ctx.is_get_stats && val && object && property) {
-        val->stats.used = true;
-        object->stats.used = true;
+        value_t::stats_t::mark_used(val);
+        value_t::stats_t::mark_used(object);
+        value_t::stats_t::mark_used(property);
         if (is_val<value_int>(property)) {
             object->stats.ops.insert("array_access");
         } else if (is_val<value_string>(property)) {

common/jinja/value.cpp

@@ -4,6 +4,7 @@
 // for converting from JSON to jinja values
 #include <nlohmann/json.hpp>
 
+#include <sstream>
 #include <string>
 #include <cctype>
 #include <vector>
@@ -160,6 +161,11 @@ static value tojson(const func_args & args) {
     value val_separators = args.get_kwarg_or_pos("separators",   3);
     value val_sort       = args.get_kwarg_or_pos("sort_keys",    4);
     int indent = -1;
+    if (args.ctx.is_get_stats) {
+        // mark as used (recursively) for stats
+        auto val_input = args.get_pos(0);
+        value_t::stats_t::mark_used(const_cast<value&>(val_input), true);
+    }
     if (is_val<value_int>(val_indent)) {
         indent = static_cast<int>(val_indent->as_int());
     }
@@ -715,8 +721,46 @@ const func_builtins & value_string_t::get_builtins() const {
             return args.get_pos(0);
         }},
         {"tojson", tojson},
-        {"indent", [](const func_args &) -> value {
-            throw not_implemented_exception("String indent builtin not implemented");
+        {"indent", [](const func_args &args) -> value {
+            args.ensure_count(1, 4);
+            value val_input  = args.get_pos(0);
+            value val_width  = args.get_kwarg_or_pos("width", 1);
+            const bool first = args.get_kwarg_or_pos("first", 2)->as_bool(); // undefined == false
+            const bool blank = args.get_kwarg_or_pos("blank", 3)->as_bool(); // undefined == false
+            if (!is_val<value_string>(val_input)) {
+                throw raised_exception("indent() first argument must be a string");
+            }
+            std::string indent;
+            if (is_val<value_int>(val_width)) {
+                indent.assign(val_width->as_int(), ' ');
+            } else if (is_val<value_string>(val_width)) {
+                indent = val_width->as_string().str();
+            } else {
+                indent = "    ";
+            }
+            std::string indented;
+            std::string input = val_input->as_string().str();
+            std::istringstream iss = std::istringstream(input);
+            std::string line;
+            while (std::getline(iss, line)) {
+                if (!indented.empty()) {
+                    indented.push_back('\n');
+                }
+                if ((indented.empty() ? first : (!line.empty() || blank))) {
+                    indented += indent;
+                }
+                indented += line;
+            }
+            if (!input.empty() && input.back() == '\n') {
+                indented.push_back('\n');
+                if (blank) {
+                    indented += indent;
+                }
+            }
+
+            auto res = mk_val<value_string>(indented);
+            res->val_str.mark_input_based_on(val_input->as_string());
+            return res;
         }},
         {"join", [](const func_args &) -> value {
             throw not_implemented_exception("String join builtin not implemented");
@@ -852,6 +896,11 @@ const func_builtins & value_array_t::get_builtins() const {
         }},
         {"string", [](const func_args & args) -> value {
             args.ensure_vals<value_array>();
+            if (args.ctx.is_get_stats) {
+                // mark as used (recursively) for stats
+                auto val_input = args.get_pos(0);
+                value_t::stats_t::mark_used(const_cast<value&>(val_input), true);
+            }
             return mk_val<value_string>(args.get_pos(0)->as_string());
         }},
         {"tojson", tojson},
@@ -1007,6 +1056,11 @@ const func_builtins & value_object_t::get_builtins() const {
         {"tojson", tojson},
         {"string", [](const func_args & args) -> value {
             args.ensure_vals<value_object>();
+            if (args.ctx.is_get_stats) {
+                // mark as used (recursively) for stats
+                auto val_input = args.get_pos(0);
+                value_t::stats_t::mark_used(const_cast<value&>(val_input), true);
+            }
             return mk_val<value_string>(args.get_pos(0)->as_string());
         }},
         {"length", [](const func_args & args) -> value {
@@ -1319,4 +1373,21 @@ std::string value_to_string_repr(const value & val) {
     }
 }
 
+// stats utility
+void value_t::stats_t::mark_used(value & val, bool deep) {
+    val->stats.used = true;
+    if (deep) {
+        if (is_val<value_array>(val)) {
+            for (auto & item : val->val_arr) {
+                mark_used(item, deep);
+            }
+        } else if (is_val<value_object>(val)) {
+            for (auto & pair : val->val_obj) {
+                mark_used(pair.first, deep);
+                mark_used(pair.second, deep);
+            }
+        }
+    }
+}
+
 } // namespace jinja

common/jinja/value.h

@@ -118,6 +118,8 @@ struct value_t {
         bool used = false;
         // ops can be builtin calls or operators: "array_access", "object_access"
         std::set<std::string> ops;
+        // utility to recursively mark value and its children as used
+        static void mark_used(value & val, bool deep = false);
     } stats;
 
     value_t() = default;

common/ngram-map.cpp

@@ -47,21 +47,15 @@ static std::string common_tokens_to_str(const llama_tokens & inp, size_t start,
  * @return Vector of draft tokens, empty if no matching pattern is found
  */
 llama_tokens common_ngram_simple_draft(
-        common_ngram_simple_state & state,
+        const common_ngram_simple_config & config,
         const llama_tokens & tokens, llama_token sampled) {
 
     // Simple implementation of self-speculative decoding without a draft model.
     //
     const size_t cur_len = tokens.size();
-    // Only check every check_rate tokens to save compute
-    // i.e., perform check if (cur_len - idx_last_check) >= check_rate
-    if (state.idx_last_check + state.config.check_rate > cur_len) {
-        llama_tokens draft_tokens;
-        return draft_tokens;
-    }
 
-    size_t n_draft_min = state.config.size_ngram; // size of n-gram to lookup in token history
-    size_t n_draft_max = state.config.size_mgram; // the m-gram following the found n-gram is used for draft
+    const size_t n_draft_min = config.size_ngram; // size of n-gram to lookup in token history
+    const size_t n_draft_max = config.size_mgram; // the m-gram following the found n-gram is used for draft
 
     // vector for tokens we want to verify.
     // return empty vector if there is no match.
@@ -80,9 +74,6 @@ llama_tokens common_ngram_simple_draft(
     }
     pattern.push_back(sampled); // add the last token to the pattern
 
-    // We do a search in the token history.
-    state.idx_last_check = cur_len;
-
     size_t match_pos = 0; // we ignore position 0, position 0 == no match
                           // search backwards, but skip the current match (we are currently there)
     for (size_t j = cur_len - n_draft_min - 1; j > 0; --j) {
@@ -240,10 +231,9 @@ void common_ngram_map_draft(common_ngram_map & map,
         GGML_ABORT("%s: cur_len exceeds UINT32_MAX: %zu", __func__, cur_len);
     }
 
-    // Only check every check_rate tokens to save compute
-    // i.e., perform check if (cur_len - idx_last_check) >= check_rate
-    if (map.idx_last_check + map.check_rate > cur_len) {
-        return;
+    if (map.idx_last_check  > cur_len) {
+        // Should not happen because of common_ngram_map_begin().
+        GGML_ABORT("%s: map.idx_last_check > cur_len: %zu > %zu", __func__, map.idx_last_check, cur_len);
     }
     map.idx_last_check = cur_len;
 
@@ -471,7 +461,7 @@ void common_ngram_map_draft(common_ngram_map & map,
             slot_max = v;
         }
     }
-    // What is sum of the other occurences?
+    // What is sum of the other occurrences?
     uint32_t sum_occur = 0;
     for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
         if (v == slot_max) {

common/ngram-map.h

@@ -24,26 +24,11 @@
 struct common_ngram_simple_config {
     uint16_t   size_ngram;      // size of n-grams to lookup in self-mode
     uint16_t   size_mgram;      // size of m-grams to draft in self-mode
-    uint16_t   check_rate;      // check for speculative decoding without draft model for each check_rate token
-};
-
-// current state (and config) of n-gram simple.
-struct common_ngram_simple_state {
-    common_ngram_simple_config config;
-
-    size_t idx_last_check = 0; // index of last check in context history (mutable)
-
-    common_ngram_simple_state(const common_ngram_simple_config & config)
-        : config(config) {}
 };
 
 // Searches for a n-gram in the history and checks whether a draft sequence should be generated.
-// state:              the ngram simple state to search in.
-// inp:                the tokens generated so far.
-// sampled:            the token that was just sampled.
-// draft:              vector to store the draft tokens, initially empty.
 llama_tokens common_ngram_simple_draft(
-        common_ngram_simple_state & state,
+        const common_ngram_simple_config & config,
         const llama_tokens & tokens, llama_token sampled);
 
 
@@ -59,7 +44,7 @@ llama_tokens common_ngram_simple_draft(
 // statistics of a m-gram after a known n-gram
 struct common_ngram_map_value {
     size_t   value_idx =  0;  // index of value m-gram in token-history (0 if unused)
-    uint16_t value_num =  0;  // number of occurences of this value m-gram after the key n-gram (0 in an unused values-slot)
+    uint16_t value_num =  0;  // number of occurrences of this value m-gram after the key n-gram (0 in an unused values-slot)
     int16_t n_accepted = -1;  // number of accepted tokens at last draft (-1 if unused)
 };
 
@@ -68,7 +53,7 @@ struct common_ngram_map_key {
     size_t   key_idx;   // index of key n-gram in token-history
     size_t   stat_idx;  // index of last token of stastistics computation (key_num, values)
 
-    uint16_t key_num;   // number of occurences of this key n-gram in token-history
+    uint16_t key_num;   // number of occurrences of this key n-gram in token-history
     common_ngram_map_value values[COMMON_NGRAM_MAX_VALUES]; // some known values after the key
 };
 
@@ -80,15 +65,14 @@ struct common_ngram_map {
     bool key_only;       // true if only key n-grams are used, no values.
 
     std::vector<common_ngram_map_key> keys; // key n-grams which occur several times in token-history
-    uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token
     uint16_t min_hits;   // minimum number of key hits to consider a draft
 
-    bool     show_key_map_stats = false; // true, if statitics of the key_map should be printed.
+    bool     show_key_map_stats = false; // true, if statistics of the key_map should be printed.
 
     common_ngram_map(uint16_t sz_key, uint16_t sz_value, bool only_keys,
-                     uint16_t check_rate, uint16_t min_hits)
+                     uint16_t min_hits)
         : size_key(sz_key), size_value(sz_value), key_only(only_keys),
-          check_rate(check_rate), min_hits(min_hits) {
+          min_hits(min_hits) {
         key_map.resize(COMMON_NGRAM_HASH_MAP_SIZE); // 2^18 hash entries, 0 entries if key_map shouldn't be used
     }
 

common/speculative.cpp

@@ -113,13 +113,14 @@ static bool common_speculative_are_compatible(
 struct common_speculative_state {
     const enum common_speculative_type type;
 
-    // TODO: rename to n_call_draft, n_gen_drafts, n_acc_drafts, n_gen_tokens, n_acc_tokens
-    // TODO: add n_call_begin, n_call_accept
-    size_t drafts_call_count       = 0; // number of times this implementation was called.
-    size_t drafts_generated_count  = 0; // number of times a draft or part was generated by this implementation.
-    size_t drafts_accepted_count   = 0; // number of times a draft or part was accepted by the target model.
-    size_t drafts_generated_tokens = 0; // number of tokens generated by this implementation.
-    size_t drafts_accepted_tokens  = 0; // number of tokens accepted by the target model.
+    size_t n_call_begin  = 0; // number of times this implementation was called for refresh.
+    size_t n_call_draft  = 0; // number of times this implementation was called for generation.
+    size_t n_call_accept = 0; // number of times this implementation was called for accumulation.
+
+    size_t n_gen_drafts = 0; // number of times a draft or part was generated by this implementation.
+    size_t n_acc_drafts = 0; // number of times a draft or part was accepted by the target model.
+    size_t n_gen_tokens = 0; // number of tokens generated by this implementation.
+    size_t n_acc_tokens = 0; // number of tokens accepted by the target model.
 
     // TODO: track performance of most recent calls
     const bool gen_perf = true; // whether to generate performance stats.
@@ -463,12 +464,12 @@ struct common_speculative_state_eagle3 : public common_speculative_state {
 
 // state of self-speculation (simple implementation, not ngram-map)
 struct common_speculative_state_ngram_simple : public common_speculative_state {
-    common_ngram_simple_state state;
+    common_ngram_simple_config config;
 
     common_speculative_state_ngram_simple(
             enum common_speculative_type type,
-            common_ngram_simple_state state)
-        : common_speculative_state(type), state(state) {}
+            common_ngram_simple_config config)
+        : common_speculative_state(type), config(config) {}
 
     void begin(const llama_tokens & prompt) override {
         GGML_UNUSED(prompt);
@@ -479,7 +480,8 @@ struct common_speculative_state_ngram_simple : public common_speculative_state {
             const llama_tokens & prompt_tgt,
             llama_token id_last,
             llama_tokens & result) override {
-        result = common_ngram_simple_draft(state, prompt_tgt, id_last);
+
+        result = common_ngram_simple_draft(config, prompt_tgt, id_last);
         GGML_UNUSED(params);
     }
 
@@ -744,10 +746,9 @@ static common_ngram_map get_common_ngram_map(const common_speculative_config & c
     uint16_t size_key   = config.params.ngram_size_n;
     uint16_t size_value = config.params.ngram_size_m;
     bool     key_only   = (config.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
-    uint16_t check_rate = config.params.ngram_check_rate;
     uint16_t min_hits   = config.params.ngram_min_hits;
 
-    return common_ngram_map(size_key, size_value, key_only, check_rate, min_hits);
+    return common_ngram_map(size_key, size_value, key_only, min_hits);
 }
 
 static common_speculative_state_ngram_cache create_state_ngram_cache(
@@ -797,6 +798,42 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
     return it->second;
 }
 
+bool common_speculative_is_compat(llama_context * ctx_tgt) {
+    auto * mem = llama_get_memory(ctx_tgt);
+    if (mem == nullptr) {
+        return false;
+    }
+
+    bool res = true;
+
+    llama_memory_clear(mem, true);
+
+    // eval 2 tokens to check if the context is compatible
+    std::vector<llama_token> tmp;
+    tmp.push_back(0);
+    tmp.push_back(0);
+
+    int ret = llama_decode(ctx_tgt, llama_batch_get_one(tmp.data(), tmp.size()));
+    if (ret != 0) {
+        LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
+        res = false;
+        goto done;
+    }
+
+    // try to remove the last tokens
+    if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
+        LOG_WRN("%s: the target context does not support partial sequence removal\n", __func__);
+        res = false;
+        goto done;
+    }
+
+done:
+    llama_memory_clear(mem, true);
+    llama_synchronize(ctx_tgt);
+
+    return res;
+}
+
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(
@@ -887,16 +924,14 @@ common_speculative * common_speculative_init(
 
                 uint16_t ngram_size_key   = ngram_map.size_key;
                 uint16_t mgram_size_value = ngram_map.size_value;
-                uint16_t check_rate       = ngram_map.check_rate;
 
-                auto config_simple = common_ngram_simple_config{
+                auto config_simple = common_ngram_simple_config {
                     /* .size_ngram      = */ ngram_size_key,
-                    /* .size_mgram      = */ mgram_size_value,
-                    /* .check_rate      = */ check_rate
+                    /* .size_mgram      = */ mgram_size_value
                 };
                 auto state = std::make_unique<common_speculative_state_ngram_simple>(
                     /* .type            = */ config.type,
-                    /* .state           = */ common_ngram_simple_state(config_simple)
+                    /* .state           = */ config_simple
                 );
                 impls.push_back(std::move(state));
                 break;
@@ -953,6 +988,7 @@ void common_speculative_begin(common_speculative * spec, const llama_tokens & pr
     for (auto & impl : spec->impls) {
         common_time_meas tm(impl->t_begin_us, !impl->gen_perf);
         impl->begin(prompt);
+        impl->n_call_begin++;
     }
 }
 
@@ -969,17 +1005,17 @@ llama_tokens common_speculative_draft(
         {
             common_time_meas tm(impl->t_draft_us, !impl->gen_perf);
             impl->draft(params, prompt_tgt, id_last, result);
-            impl->drafts_call_count++;
+            impl->n_call_draft++;
         }
 
         if (!result.empty()) {
             LOG_DBG("%s: called impl %s, hist size = %zu, call_count = %zu, gen = %zu\n", __func__,
                     common_speculative_type_to_str(impl.get()->type).c_str(), prompt_tgt.size(),
-                    impl.get()->drafts_call_count, result.size());
+                    impl.get()->n_call_draft, result.size());
 
             spec->curr_impl = impl.get(); // set current implementation for stats
-            impl->drafts_generated_count++;
-            impl->drafts_generated_tokens += result.size();
+            impl->n_gen_drafts++;
+            impl->n_gen_tokens += result.size();
 
             break; // We have a draft, so break out of the loop and return it.
         }
@@ -1000,11 +1036,12 @@ void common_speculative_accept(common_speculative * spec, uint16_t n_accepted) {
     {
         common_time_meas tm(impl->t_accept_us, !impl->gen_perf);
         if (n_accepted > 0) {
-            impl->drafts_accepted_count++;
-            impl->drafts_accepted_tokens += n_accepted;
+            impl->n_acc_drafts++;
+            impl->n_acc_tokens += n_accepted;
         }
 
         impl->accept(n_accepted);
+        impl->n_call_accept++;
     }
 }
 
@@ -1025,13 +1062,13 @@ void common_speculative_print_stats(const common_speculative * spec) {
             str_perf = "";
         }
 
-        LOG_INF("statistics %s: #calls = %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
+        LOG_INF("statistics %s: #calls(b,g,a) = %zu %zu %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
                 common_speculative_type_to_str(impl->type).c_str(),
-                impl->drafts_call_count,
-                impl->drafts_generated_count,
-                impl->drafts_accepted_count,
-                impl->drafts_generated_tokens,
-                impl->drafts_accepted_tokens,
+                impl->n_call_begin, impl->n_call_draft, impl->n_call_accept,
+                impl->n_gen_drafts,
+                impl->n_acc_drafts,
+                impl->n_gen_tokens,
+                impl->n_acc_tokens,
                 str_perf.c_str());
     }
 }

common/speculative.h

@@ -14,6 +14,10 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
 // convert type to string
 std::string common_speculative_type_to_str(enum common_speculative_type type);
 
+// check if the llama_context is compatible for speculative decoding
+// note: clears the memory of the context
+bool common_speculative_is_compat(llama_context * ctx_tgt);
+
 common_speculative * common_speculative_init(
         common_params_speculative & params,
         llama_context             * ctx_tgt);

convert_hf_to_gguf_update.py

@@ -99,6 +99,7 @@ models = [
     {"name": "stablelm2",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
     {"name": "refact",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
     {"name": "command-r",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
+    {"name": "tiny_aya",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereLabs/tiny-aya-base", },
     {"name": "qwen2",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
     {"name": "olmo",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
     {"name": "dbrx",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
@@ -106,6 +107,7 @@ models = [
     {"name": "jina-v2-en",       "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
     {"name": "jina-v2-es",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
     {"name": "jina-v2-de",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
+    {"name": "jina-v5-nano",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v5-text-nano", },
     {"name": "smaug-bpe",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
     {"name": "poro-chat",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", },
     {"name": "jina-v2-code",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", },
@@ -113,6 +115,7 @@ models = [
     {"name": "gemma",            "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", },
     {"name": "gemma-2",          "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", },
     {"name": "jais",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", },
+    {"name": "jais-2",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inceptionai/Jais-2-8B-Chat", },
     {"name": "t5",               "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", },
     {"name": "codeshell",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
     {"name": "tekken",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
@@ -148,6 +151,9 @@ models = [
     {"name": "youtu",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Youtu-LLM-2B", },
     {"name": "solar-open",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/upstage/Solar-Open-100B", },
     {"name": "exaone-moe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B", },
+    {"name": "qwen35",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3.5-9B-Instruct", },
+    {"name": "joyai-llm",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jdopensource/JoyAI-LLM-Flash", },
+    {"name": "kanana2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/kakaocorp/kanana-2-30b-a3b-instruct-2601", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -157,6 +163,7 @@ pre_computed_hashes = [
     {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516"},
     {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2"},
     {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.5-Air", "chkhsh": "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902"},
+    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.7-Flash", "chkhsh": "cdf5f35325780597efd76153d4d1c16778f766173908894c04afc20108536267"},
     {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
     {"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"},
     {"name": "hunyuan-dense", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-4B-Instruct", "chkhsh": "bba3b3366b646dbdded5dbc42d59598b849371afc42f7beafa914afaa5b70aa6"},
@@ -170,7 +177,6 @@ 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/CANN.md

@@ -20,7 +20,7 @@
 
 **Llama.cpp + CANN**
 
-The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the ability of AscendC and ACLNN which are intergrated to CANN Toolkit and kernels to using Ascend NPU directly.
+The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the ability of AscendC and ACLNN which are integrated to CANN Toolkit and kernels to using Ascend NPU directly.
 
 ## News
 
@@ -210,7 +210,7 @@ docker run --name llamacpp --device /dev/davinci0  --device /dev/davinci_manager
     # and install driver.
     sudo sh Ascend-hdk-910b-npu-firmware_x.x.x.x.X.run --full
     ```
-    If the following messaage appers, firmware is installed successfully.
+    If the following message appears, firmware is installed successfully.
     ```sh
     Firmware package installed successfully!
     ```
@@ -246,7 +246,7 @@ cmake --build build --config release
 
 1. **Retrieve and prepare model**
 
-    You can refer to the general [*Prepare and Quantize*](../../README.md#prepare-and-quantize) guide for model prepration.
+    You can refer to the general [*Obtaining and quantizing models*](../../README.md#obtaining-and-quantizing-models) guide for model prepration.
 
     **Notes**:
 

docs/backend/SYCL.md

@@ -281,7 +281,7 @@ as `-cl-fp32-correctly-rounded-divide-sqrt`
 
 #### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf?download=true) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
+You can refer to the general [*Obtaining and quantizing models*](../../README.md#obtaining-and-quantizing-models) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf?download=true) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
 
 ##### Check device
 
@@ -569,7 +569,7 @@ Once it is completed, final results will be in **build/Release/bin**
 
 #### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
+You can refer to the general [*Obtaining and quantizing models*](../../README.md#obtaining-and-quantizing-models) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
 
 ##### Check device
 
@@ -708,7 +708,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
   - Remove **build** folder or try a clean-build.
 
-- I can **not** see `[ext_oneapi_level_zero:gpu]` afer installing the GPU driver on Linux.
+- I can **not** see `[ext_oneapi_level_zero:gpu]` after installing the GPU driver on Linux.
 
   Please double-check with `sudo sycl-ls`.
 

docs/backend/VirtGPU.md

@@ -0,0 +1,182 @@
+# GGML-VirtGPU Backend
+
+The GGML-VirtGPU backend enables GGML applications to run machine
+learning computations on host hardware while the application itself
+runs inside a virtual machine.  It uses host-guest shared memory to
+efficiently share data buffers between the two sides.
+
+This backend relies on the virtio-gpu, and VirglRenderer API Remoting
+(APIR) component. The backend is split into two libraries:
+- a GGML implementation (the "remoting frontend"), running in the
+  guest and interacting with the virtgpu device
+- a VirglRenderer APIR compatible library (the "remoting backend"),
+  running in the host and interacting with Virglrenderer and an actual
+  GGML device backend.
+
+## OS support
+
+| OS       | Status            | Backend     | CI testing  | Notes
+| -------- | ----------------- | ----------- | ----------- | -----
+| MacOS 14 | Supported         | ggml-metal  | X           | Working when compiled on MacOS 14
+| MacOS 15 | Supported         | ggml-metal  | X           | Working when compiled on MacOS 14 or MacOS 15
+| MacOS 26 | Not tested        |             |             |
+| Linux    | Under development | ggml-vulkan | not working | Working locally, CI running into deadlocks
+
+
+## Architecture Overview
+
+The GGML-VirtGPU backend consists of three main components:
+
+```mermaid
+graph TD
+    %% Nodes
+
+ subgraph GuestVM ["Guest VM - Frontend"]
+        App([GGML Application<br/>llama.cpp, etc.])
+
+        direction TB
+        Interface[GGML Backend Interface]
+        Comm["GGML-VirtGPU<br/>(hypercalls + shared mem)"]
+
+        App --> Interface
+        Interface --> Comm
+    end
+
+    API[virtio-gpu / virglrenderer API]
+
+    subgraph HostSystem [Host System - Backend]
+        direction TB
+        Dispatcher[GGML-VirtGPU-Backend]
+        BackendLib[GGML Backend library<br/>Metal / Vulkan / CPU / ...]
+
+        Dispatcher --> BackendLib
+    end
+
+    %% Connections
+    Comm --> API
+    API --> HostSystem
+```
+
+### Key Components
+
+1. **Guest-side Frontend** (`ggml-virtgpu/`): Implements the GGML backend interface and forwards operations to the host
+2. **Host-side Backend** (`ggml-virtgpu/backend/`): Receives forwarded operations and executes them on actual hardware backends
+3. **Communication Layer**: Uses virtio-gpu hypercalls and shared memory for efficient data transfer
+
+## Features
+
+- **Dynamic backend loading** on the host side (CPU, CUDA, Metal, etc.)
+- **Zero-copy data transfer** via host-guest shared memory pages
+
+## Communication Protocol
+
+### Hypercalls and Shared Memory
+
+The backend uses two primary communication mechanisms:
+
+1. **Hypercalls (`DRM_IOCTL_VIRTGPU_EXECBUFFER`)**: Trigger remote execution from guest to host
+2. **Shared Memory Pages**: Zero-copy data transfer for tensors and parameters
+
+#### Shared Memory Layout
+
+Each connection uses two shared memory buffers:
+
+- **Data Buffer** (24 MiB): For command/response data and tensor transfers
+- **Reply Buffer** (16 KiB): For command replies and status information
+- **Data Buffers**: Dynamically allocated host-guest shared buffers
+  served as GGML buffers.
+
+### APIR Protocol
+
+The Virglrender API Remoting protocol defines three command types:
+
+- `HANDSHAKE`: Protocol version negotiation and capability discovery
+- `LOADLIBRARY`: Dynamic loading of backend libraries on the host
+- `FORWARD`: API function call forwarding
+
+### Binary Serialization
+
+Commands and data are serialized using a custom binary protocol with:
+
+- Fixed-size encoding for basic types
+- Variable-length arrays with size prefixes
+- Buffer bounds checking
+- Error recovery mechanisms
+
+## Supported Operations
+
+### Device Operations
+- Device enumeration and capability queries
+- Memory information (total/free)
+- Backend type detection
+
+### Buffer Operations
+- Buffer allocation and deallocation
+- Tensor data transfer (host ↔ guest)
+- Memory copying and clearing
+
+### Computation Operations
+- Graph execution forwarding
+
+## Build Requirements
+
+### Guest-side Dependencies
+- `libdrm` for DRM/virtio-gpu communication
+- C++20 compatible compiler
+- CMake 3.14+
+
+### Host-side Dependencies
+- virglrenderer with APIR support (pending upstream review)
+- Target backend libraries (libggml-metal, libggml-vulkan, etc.)
+
+## Configuration
+
+### Environment Variables
+
+- `GGML_VIRTGPU_BACKEND_LIBRARY`: Path to the host-side backend library
+- `GGML_VIRTGPU_DEBUG`: Enable debug logging
+
+### Build Options
+
+- `GGML_VIRTGPU`: Enable the VirtGPU backend (`ON` or `OFF`, default: `OFF`)
+- `GGML_VIRTGPU_BACKEND`: Build the host-side backend component (`ON`, `OFF` or `ONLY`, default: `OFF`)
+
+### System Requirements
+
+- VM with virtio-gpu support
+- VirglRenderer with APIR patches
+- Compatible backend libraries on host
+
+## Limitations
+
+- **VM-specific**: Only works in virtual machines with virtio-gpu support
+- **Host dependency**: Requires properly configured host-side backend
+- **Latency**: Small overhead from VM escaping for each operation
+- **Shared-memory size**: with the `libkrun` hypervisor, the RAM + VRAM
+  addressable memory is limited to 64 GB. So the maximum GPU memory
+  will be `64GB - RAM`, regardless of the hardware VRAM size.
+
+* This work is pending upstream changes in the VirglRenderer
+  project.
+  * The backend can be tested with Virglrenderer compiled from source
+  using this PR:
+  https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590
+* This work is pending changes in the VMM/hypervisor running the
+  virtual machine, which need to know how to route the newly
+  introduced APIR capset.
+  * The environment variable `VIRGL_ROUTE_VENUS_TO_APIR=1` allows
+    using the Venus capset, until the relevant hypervisors have been
+    patched. However, setting this flag breaks the Vulkan/Venus normal
+    behavior.
+  * The environment variable `GGML_REMOTING_USE_APIR_CAPSET` tells the
+    `ggml-virtgpu` backend to use the APIR capset. This will become
+    the default when the relevant hypervisors have been patched.
+
+* This work focused on improving the performance of llama.cpp running
+  on MacOS containers, and is mainly tested on this platform. The
+  linux support (via `krun`) is in progress.
+
+## See Also
+
+- [Development and Testing](VirtGPU/development.md)
+- [Backend configuration](VirtGPU/configuration.md)

docs/backend/VirtGPU/configuration.md

@@ -0,0 +1,174 @@
+# GGML-VirtGPU Backend Configuration
+
+This document describes the environment variables used by the ggml-virtgpu backend system, covering both the frontend (guest-side) and backend (host-side) components.
+
+## Environment Variables Overview
+
+The ggml-virtgpu backend uses environment variables for configuration across three main components:
+- **Frontend (Guest)**: GGML applications running in VMs
+- **Hypervisor**: Virglrenderer/APIR system
+- **Backend (Host)**: Host-side GGML backend integration
+
+## Frontend (Guest-side) Configuration
+
+### GGML_REMOTING_USE_APIR_CAPSET
+- **Location**: `ggml/src/ggml-virtgpu/virtgpu.cpp`
+- **Type**: Boolean flag (presence-based)
+- **Purpose**: Controls which virtio-gpu capability set to use for communication
+- **Values**:
+  - Set (any value): Use the APIR capset (long-term setup)
+  - Unset: Use the Venus capset (easier for testing with an unmodified hypervisor)
+- **Default**: Unset (Venus capset)
+- **Usage**:
+  ```bash
+  export GGML_REMOTING_USE_APIR_CAPSET=1  # Use APIR capset
+  # or leave unset for Venus capset
+  ```
+
+## Hypervisor (Virglrenderer/APIR) Configuration
+
+These environment variables are used during the transition phase for
+running with an unmodified hypervisor (not supporting the
+VirglRenderer APIR component). They will be removed in the future, and
+the hypervisor will instead configure VirglRenderer with the APIR
+_Configuration Key_.
+
+### VIRGL_APIR_BACKEND_LIBRARY
+- **Location**: `virglrenderer/src/apir/apir-context.c`
+- **Configuration Key**: `apir.load_library.path`
+- **Type**: File path string
+- **Purpose**: Path to the APIR backend library that virglrenderer should dynamically load
+- **Required**: Yes
+- **Example**:
+  ```bash
+  export VIRGL_APIR_BACKEND_LIBRARY="/path/to/libggml-remotingbackend.so"
+  ```
+
+### VIRGL_ROUTE_VENUS_TO_APIR
+- **Location**: `virglrenderer/src/apir/apir-renderer.h`
+- **Type**: Boolean flag (presence-based)
+- **Purpose**: Temporary workaround to route Venus capset calls to APIR during hypervisor transition period
+- **Status**: will be removed once hypervisors support APIR natively
+- **Warning**: Breaks normal Vulkan/Venus functionality
+- **Usage**:
+  ```bash
+  export VIRGL_ROUTE_VENUS_TO_APIR=1  # For testing with an unmodified hypervisor
+  ```
+
+### VIRGL_APIR_LOG_TO_FILE
+- **Location**: `virglrenderer/src/apir/apir-renderer.c`
+- **Environment Variable**: `VIRGL_APIR_LOG_TO_FILE`
+- **Type**: File path string
+- **Purpose**: Enable debug logging from the VirglRenderer APIR component to specified file
+- **Required**: No (optional debugging)
+- **Default**: Logging to `stderr`
+- **Usage**:
+  ```bash
+  export VIRGL_APIR_LOG_TO_FILE="/tmp/apir-debug.log"
+  ```
+
+## Backend (Host-side) Configuration
+
+These environment variables are used during the transition phase for
+running with an unmodified hypervisor (not supporting the
+VirglRenderer APIR component). They will be removed in the future, and
+the hypervisor will instead configure VirglRenderer with the APIR
+_Configuration Key_.
+
+### APIR_LLAMA_CPP_GGML_LIBRARY_PATH
+- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp`
+- **Environment Variable**: `APIR_LLAMA_CPP_GGML_LIBRARY_PATH`
+- **Configuration Key**: `ggml.library.path`
+- **Type**: File path string
+- **Purpose**: Path to the actual GGML backend library (Metal, CUDA, Vulkan, etc.)
+- **Required**: **Yes** - backend initialization fails without this
+- **Examples**:
+  ```bash
+  # macOS with Metal backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-metal.dylib"
+
+  # Linux with CUDA backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-cuda.so"
+
+  # macOS or Linux with Vulkan backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-vulkan.so"
+  ```
+
+### APIR_LLAMA_CPP_GGML_LIBRARY_REG
+- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp`
+- **Environment Variable**: `APIR_LLAMA_CPP_GGML_LIBRARY_REG`
+- **Configuration Key**: `ggml.library.reg`
+- **Type**: Function symbol name string
+- **Purpose**: Name of the backend registration function to call after loading the library
+- **Required**: No (defaults to `ggml_backend_init`)
+- **Default**: `ggml_backend_init`
+- **Examples**:
+  ```bash
+  # Metal backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_metal_reg"
+
+  # CUDA backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_cuda_reg"
+
+  # Vulkan backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_vulkan_reg"
+
+  # Generic fallback (default)
+  # export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_init"
+  ```
+
+### APIR_LLAMA_CPP_LOG_TO_FILE
+- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp:62`
+- **Environment Variable**: `APIR_LLAMA_CPP_LOG_TO_FILE`
+- **Type**: File path string
+- **Purpose**: Enable debug logging from the GGML backend to specified file
+- **Required**: No (optional debugging)
+- **Usage**:
+  ```bash
+  export APIR_LLAMA_CPP_LOG_TO_FILE="/tmp/ggml-backend-debug.log"
+  ```
+
+## Configuration Flow
+
+The configuration system works as follows:
+
+1. **Hypervisor Setup**: Virglrenderer loads the APIR backend library specified by `VIRGL_APIR_BACKEND_LIBRARY`
+
+2. **Context Creation**: When an APIR context is created, it populates a configuration table with environment variables:
+   - `apir.load_library.path` ← `VIRGL_APIR_BACKEND_LIBRARY`
+   - `ggml.library.path` ← `APIR_LLAMA_CPP_GGML_LIBRARY_PATH`
+   - `ggml.library.reg` ← `APIR_LLAMA_CPP_GGML_LIBRARY_REG`
+   - this step will eventually be performed by the hypervisor itself, with command-line arguments instead of environment variables.
+
+3. **Backend Initialization**: The backend queries the configuration via callbacks:
+   - `virgl_cbs->get_config(ctx_id, "ggml.library.path")` returns the library path
+   - `virgl_cbs->get_config(ctx_id, "ggml.library.reg")` returns the registration function
+
+4. **Library Loading**: The backend dynamically loads and initializes the specified GGML library
+
+## Error Messages
+
+Common error scenarios and their messages:
+
+- **Missing library path**: `"cannot open the GGML library: env var 'APIR_LLAMA_CPP_GGML_LIBRARY_PATH' not defined"`
+- **Missing registration function**: `"cannot register the GGML library: env var 'APIR_LLAMA_CPP_GGML_LIBRARY_REG' not defined"`
+
+## Example Complete Configuration
+
+Here's an example configuration for a macOS host with Metal backend:
+
+```bash
+# Hypervisor environment
+export VIRGL_APIR_BACKEND_LIBRARY="/opt/llama.cpp/lib/libggml-virtgpu-backend.dylib"
+
+# Backend configuration
+export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-metal.dylib"
+export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_metal_reg"
+
+# Optional logging
+export VIRGL_APIR_LOG_TO_FILE="/tmp/apir.log"
+export APIR_LLAMA_CPP_LOG_TO_FILE="/tmp/ggml.log"
+
+# Guest configuration
+export GGML_REMOTING_USE_APIR_CAPSET=1
+```

docs/backend/VirtGPU/development.md

@@ -0,0 +1,220 @@
+# Development and Testing
+
+## Development
+
+### Code Generation
+
+The backend uses code generation from YAML configuration:
+
+```bash
+# Regenerate protocol code
+cd ggml-virtgpu/
+python regenerate_remoting.py
+```
+
+### Adding New Operations
+
+1. Add function definition to `ggmlremoting_functions.yaml`
+2. Regenerate code with `regenerate_remoting.py`
+3. Implement guest-side forwarding in `virtgpu-forward-*.cpp`
+4. Implement host-side handling in `backend-dispatched-*.cpp`
+
+## Testing
+
+This document provides instructions for building and testing the GGML-VirtGPU backend on macOS with containers.
+
+### Prerequisites
+
+The testing setup requires:
+
+- macOS host system
+- Container runtime with `libkrun` provider (podman machine)
+- Access to development patchset for VirglRenderer
+
+### Required Patchsets
+
+The backend requires patches that are currently under review:
+
+- **Virglrenderer APIR upstream PR**: https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590 (for reference)
+- **MacOS Virglrenderer (for krunkit)**: https://gitlab.freedesktop.org/kpouget/virglrenderer/-/tree/main-macos
+- **Linux Virglrenderer (for krun)**: https://gitlab.freedesktop.org/kpouget/virglrenderer/-/tree/main-linux
+
+### Build Instructions
+
+#### 1. Build ggml-virtgpu-backend (Host-side, macOS)
+
+```bash
+# Build the backend that runs natively on macOS
+mkdir llama.cpp
+cd llama.cpp
+git clone https://github.com/ggml-org/llama.cpp.git src
+cd src
+
+LLAMA_MAC_BUILD=$PWD/build/ggml-virtgpu-backend
+
+cmake -S . -B $LLAMA_MAC_BUILD \
+      -DGGML_NATIVE=OFF \
+      -DLLAMA_CURL=ON \
+      -DGGML_REMOTINGBACKEND=ONLY \
+      -DGGML_METAL=ON
+
+TARGETS="ggml-metal"
+cmake --build $LLAMA_MAC_BUILD --parallel 8 --target $TARGETS
+
+# Build additional tools for native benchmarking
+EXTRA_TARGETS="llama-run llama-bench"
+cmake --build $LLAMA_MAC_BUILD --parallel 8 --target $EXTRA_TARGETS
+```
+
+#### 2. Build virglrenderer (Host-side, macOS)
+
+```bash
+# Build virglrenderer with APIR support
+mkdir virglrenderer
+git clone https://gitlab.freedesktop.org/kpouget/virglrenderer -b main-macos src
+cd src
+
+VIRGL_BUILD_DIR=$PWD/build
+
+# -Dvenus=true and VIRGL_ROUTE_VENUS_TO_APIR=1 route the APIR requests via the Venus backend, for easier testing without a patched hypervisor
+
+meson setup $VIRGL_BUILD_DIR \
+      -Dvenus=true \
+      -Dapir=true
+
+ninja -C $VIRGL_BUILD_DIR
+```
+
+#### 3. Build ggml-virtgpu (Guest-side, Linux)
+
+Option A: Build from a script:
+
+```bash
+# Inside a Linux container
+mkdir llama.cpp
+git clone https://github.com/ggml-org/llama.cpp.git src
+cd src
+
+LLAMA_LINUX_BUILD=$PWD//build-virtgpu
+
+cmake -S . -B $LLAMA_LINUX_BUILD \
+      -DGGML_VIRTGPU=ON
+
+ninja -C $LLAMA_LINUX_BUILD
+```
+
+Option B: Build container image with frontend:
+
+```bash
+cat << EOF > remoting.containerfile
+FROM quay.io/fedora/fedora:43
+USER 0
+
+WORKDIR /app/remoting
+
+ARG LLAMA_CPP_REPO="https://github.com/ggml-org/llama.cpp.git"
+ARG LLAMA_CPP_VERSION="master"
+ARG LLAMA_CPP_CMAKE_FLAGS="-DGGML_VIRTGPU=ON"
+ARG LLAMA_CPP_CMAKE_BUILD_FLAGS="--parallel 4"
+
+RUN dnf install -y git cmake gcc gcc-c++ libcurl-devel libdrm-devel
+
+RUN git clone "\${LLAMA_CPP_REPO}" src \\
+ && git -C src fetch origin \${LLAMA_CPP_VERSION} \\
+ && git -C src reset --hard FETCH_HEAD
+
+RUN mkdir -p build \\
+ && cd src \\
+ && set -o pipefail \\
+ && cmake -S . -B ../build \${LLAMA_CPP_CMAKE_FLAGS} \\
+ && cmake --build ../build/ \${LLAMA_CPP_CMAKE_BUILD_FLAGS}
+
+ENTRYPOINT ["/app/remoting/src/build/bin/llama-server"]
+EOF
+
+mkdir -p empty_dir
+podman build -f remoting.containerfile ./empty_dir -t localhost/llama-cpp.virtgpu
+```
+
+### Environment Setup
+
+#### Set krunkit Environment Variables
+
+```bash
+# Define the base directories (adapt these paths to your system)
+VIRGL_BUILD_DIR=$HOME/remoting/virglrenderer/build
+LLAMA_MAC_BUILD=$HOME/remoting/llama.cpp/build-backend
+
+# For krunkit to load the custom virglrenderer library
+export DYLD_LIBRARY_PATH=$VIRGL_BUILD_DIR/src
+
+# For Virglrenderer to load the ggml-remotingbackend library
+export VIRGL_APIR_BACKEND_LIBRARY="$LLAMA_MAC_BUILD/bin/libggml-virtgpu-backend.dylib"
+
+# For llama.cpp remotingbackend to load the ggml-metal backend
+export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="$LLAMA_MAC_BUILD/bin/libggml-metal.dylib"
+export APIR_LLAMA_CPP_GGML_LIBRARY_REG=ggml_backend_metal_reg
+```
+
+#### Launch Container Environment
+
+```bash
+# Set container provider to libkrun
+export CONTAINERS_MACHINE_PROVIDER=libkrun
+podman machine start
+```
+
+#### Verify Environment
+
+Confirm that krunkit is using the correct virglrenderer library:
+
+```bash
+lsof -c krunkit | grep virglrenderer
+# Expected output:
+# krunkit 50574 user  txt  REG  1,14  2273912  10849442 ($VIRGL_BUILD_DIR/src)/libvirglrenderer.1.dylib
+```
+
+### Running Tests
+
+#### Launch Test Container
+
+```bash
+# Optional model caching
+mkdir -p models
+PODMAN_CACHE_ARGS="-v models:/models --user root:root --cgroupns host --security-opt label=disable -w /models"
+
+podman run $PODMAN_CACHE_ARGS -it --rm --device /dev/dri localhost/llama-cpp.virtgpu
+```
+
+#### Test llama.cpp in Container
+
+```bash
+
+# Run performance benchmark
+/app/remoting/build/bin/llama-bench -m ./llama3.2
+```
+
+Expected output (performance may vary):
+```
+| model                          |       size |     params | backend    | ngl |          test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------------: | -------------------: |
+| llama 3B Q4_K - Medium         |   1.87 GiB |     3.21 B | ggml-virtgpu |  99 |         pp512 |        991.30 ± 0.66 |
+| llama 3B Q4_K - Medium         |   1.87 GiB |     3.21 B | ggml-virtgpu |  99 |         tg128 |         85.71 ± 0.11 |
+```
+
+### Troubleshooting
+
+#### SSH Environment Variable Issues
+
+⚠️ **Warning**: Setting `DYLD_LIBRARY_PATH` from SSH doesn't work on macOS. Here is a workaround:
+
+**Workaround 1: Replace system library**
+```bash
+VIRGL_BUILD_DIR=$HOME/remoting/virglrenderer/build  # ⚠️ adapt to your system
+BREW_VIRGL_DIR=/opt/homebrew/Cellar/virglrenderer/0.10.4d/lib
+VIRGL_LIB=libvirglrenderer.1.dylib
+
+cd $BREW_VIRGL_DIR
+mv $VIRGL_LIB ${VIRGL_LIB}.orig
+ln -s $VIRGL_BUILD_DIR/src/$VIRGL_LIB
+```

docs/backend/ZenDNN.md

@@ -22,7 +22,7 @@
 
 **Llama.cpp + ZenDNN**
 
-The llama.cpp ZenDNN backend leverages AMD's optimized matrix multiplication primitives to accelerate inference on AMD CPUs. It utilizes ZenDNN's **LowOHA (Low Overhead Hardware Accelerated)** MatMul operator for efficient GEMM operations with minimal execution overhead, built-in weight caching, and direct access to backend libraries (AOCL BLIS, LibXSMM, OneDNN).
+The llama.cpp ZenDNN backend leverages AMD's optimized matrix multiplication primitives to accelerate inference on AMD CPUs. It utilizes ZenDNN's **LowOHA (Low Overhead Hardware Accelerated)** MatMul operator for efficient GEMM operations with minimal execution overhead, built-in weight caching, and direct access to backend libraries (AOCL DLP, LibXSMM, OneDNN).
 
 For more information about ZenDNN, visit: https://www.amd.com/en/developer/zendnn.html
 
@@ -32,7 +32,7 @@ For more information about ZenDNN, visit: https://www.amd.com/en/developer/zendn
 |:-------:|:-------:|:----------------------------------------------:|
 | Linux   | Support | Ubuntu 20.04, 22.04, 24.04                     |
 
-For the latest list of supported operating systems, see the [ZenDNN Supported OS](https://github.com/amd/ZenDNN/blob/zendnnl/README.md#15-supported-os).
+For the latest list of supported operating systems, see the [ZenDNN Supported OS](https://github.com/amd/ZenDNN/blob/a18adf8c605fb5f5e52cefd7eda08a7b18febbaf/README.md#15-supported-os).
 
 ## Hardware
 
@@ -44,9 +44,9 @@ ZenDNN is optimized for AMD EPYC™ processors and AMD Ryzen™ processors based
 
 | CPU Family                    | Status  | Notes                              |
 |:-----------------------------:|:-------:|:----------------------------------:|
-| AMD EPYC™ 9005 Series (Turin)| Support | 5th Gen - Zen 5 architecture       |
-| AMD EPYC™ 9004 Series (Genoa)| Support | 4th Gen - Zen 4 architecture       |
-| AMD EPYC™ 7003 Series (Milan)| Support | 3rd Gen - Zen 3 architecture       |
+| AMD EPYC™ 9005 Series (Turin) | Support | 5th Gen - Zen 5 architecture       |
+| AMD EPYC™ 9004 Series (Genoa) | Support | 4th Gen - Zen 4 architecture       |
+| AMD EPYC™ 7003 Series (Milan) | Support | 3rd Gen - Zen 3 architecture       |
 | AMD Ryzen™ AI MAX (Strix Halo)| Support | High-performance mobile processors |
 
 *Notes:*
@@ -61,7 +61,7 @@ The ZenDNN backend currently accelerates **matrix multiplication (MUL_MAT)** ope
 
 | Operation    | Status  | Notes                                          |
 |:-------------|:-------:|:----------------------------------------------:|
-| MUL_MAT      |    ✓    | Accelerated via ZenDNN LowOHA MatMul           |
+| MUL_MAT      | Support | Accelerated via ZenDNN LowOHA MatMul           |
 
 *Note:* Since only MUL_MAT is accelerated, models will benefit most from ZenDNN when matrix multiplications dominate the computational workload (which is typical for transformer-based LLMs).
 
@@ -104,7 +104,6 @@ If you want to build ZenDNN yourself or use a specific version:
 # Clone ZenDNN repository
 git clone https://github.com/amd/ZenDNN.git
 cd ZenDNN
-git checkout zendnnl
 
 # Build and install (requires CMake >= 3.25)
 mkdir build && cd build
@@ -114,7 +113,7 @@ cmake --build . --target all
 
 Default installation path: `ZenDNN/build/install`
 
-**For detailed build instructions**, refer to the [ZenDNN README](https://github.com/amd/ZenDNN/blob/zendnnl/README.md).
+**For detailed build instructions**, refer to the [ZenDNN README](https://github.com/amd/ZenDNN/blob/a18adf8c605fb5f5e52cefd7eda08a7b18febbaf/README.md).
 
 **Step 2: Build llama.cpp with custom ZenDNN path**
 
@@ -146,8 +145,7 @@ Run llama.cpp server with ZenDNN acceleration:
 
 ```sh
 # Set optimal configuration
-export OMP_NUM_THREADS=64  # Adjust to your CPU core count
-export ZENDNNL_MATMUL_ALGO=2  # Blocked AOCL BLIS for best performance
+export ZENDNNL_MATMUL_ALGO=1    # Blocked AOCL DLP algo for best performance
 
 # Start server
 ./build/bin/llama-server \
@@ -160,62 +158,26 @@ export ZENDNNL_MATMUL_ALGO=2  # Blocked AOCL BLIS for best performance
 Access the server at `http://localhost:8080`.
 
 **Performance tips**:
-- Set `OMP_NUM_THREADS` to match your physical core count
-- Use `ZENDNNL_MATMUL_ALGO=2` for optimal performance
+- Use `ZENDNNL_MATMUL_ALGO=1` for optimal performance
 - For NUMA systems: `numactl --cpunodebind=0 --membind=0 ./build/bin/llama-server ...`
 
 ## Environment Variable
 
-### Build Time
+For environment variables related to ZenDNN, refer to the [ZenDNN Environment Variables Documentation](https://github.com/amd/ZenDNN/blob/a18adf8c605fb5f5e52cefd7eda08a7b18febbaf/docs/runtime_env.md).
 
-| Name               | Value                                 | Function                                    |
-|--------------------|---------------------------------------|---------------------------------------------|
-| GGML_ZENDNN        | ON/OFF                                | Enable ZenDNN backend support               |
-| ZENDNN_ROOT        | Path to ZenDNN installation           | Set ZenDNN installation directory           |
-| GGML_OPENMP        | ON/OFF (recommended: ON)              | Enable OpenMP for multi-threading           |
+### Performance Optimization
 
-### Runtime
-
-| Name                    | Value                    | Function                                                          |
-|-------------------------|--------------------------|-------------------------------------------------------------------|
-| OMP_NUM_THREADS         | Number (e.g., 64)        | Set number of OpenMP threads (recommended: physical core count)   |
-| ZENDNNL_MATMUL_ALGO     | 0-5                      | Select MatMul backend algorithm (see Performance Optimization)    |
-| ZENDNNL_PROFILE_LOG_LEVEL | 0-4                    | Profiling log level (0=disabled, 4=verbose)                       |
-| ZENDNNL_ENABLE_PROFILER | 0 or 1                   | Enable detailed profiling (1=enabled)                             |
-| ZENDNNL_API_LOG_LEVEL   | 0-4                      | API log level (0=disabled, 4=verbose)                             |
-
-**Example**:
+ZenDNN's LowOHA MatMul supports multiple backend algorithms. For **best performance**, use the **Blocked AOCL DLP** algorithm:
 
 ```sh
-export OMP_NUM_THREADS=64
-export ZENDNNL_MATMUL_ALGO=2  # Use Blocked AOCL BLIS for best performance
-./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Test" -n 100
+export ZENDNNL_MATMUL_ALGO=1    # Blocked AOCL DLP algo (recommended)
 ```
 
-## Performance Optimization
-
-### MatMul Algorithm Selection
-
-ZenDNN's LowOHA MatMul supports multiple backend algorithms. For **best performance**, use the **Blocked AOCL BLIS** algorithm:
-
-```sh
-export ZENDNNL_MATMUL_ALGO=2  # Blocked AOCL BLIS (recommended)
-```
-
-**Available algorithms**:
-
-| Value | Algorithm              | Description                                    |
-|:-----:|:-----------------------|:----------------------------------------------|
-| 0     | Dynamic Dispatch       | Automatic backend selection (default)         |
-| 1     | AOCL BLIS              | AOCL BLIS backend                             |
-| 2     | AOCL BLIS Blocked      | **Blocked AOCL BLIS (recommended)**           |
-| 3     | OneDNN                 | OneDNN backend                                |
-| 4     | OneDNN Blocked         | Blocked OneDNN                                |
-| 5     | LibXSMM                | LibXSMM backend                               |
+For more details on available algorithms, see the [ZenDNN MatMul Algorithm Documentation](https://github.com/amd/ZenDNN/blob/a18adf8c605fb5f5e52cefd7eda08a7b18febbaf/docs/runtime_env.md#algorithm-details).
 
 ### Profiling and Debugging
 
-For detailed profiling and logging options, refer to the [ZenDNN Logging Documentation](https://github.com/amd/ZenDNN/blob/zendnnl/docs/logging.md).
+For detailed profiling and logging options, refer to the [ZenDNN Logging Documentation](https://github.com/amd/ZenDNN/blob/a18adf8c605fb5f5e52cefd7eda08a7b18febbaf/docs/logging.md).
 
 ## Known Issues
 
@@ -245,10 +207,9 @@ A: Currently, ZenDNN primarily supports FP32 and BF16 data types. Quantized mode
 
 A: Ensure:
 1. You're using an AMD EPYC or Ryzen processor (Zen 2 or newer)
-2. `OMP_NUM_THREADS` is set appropriately (physical core count)
-3. `ZENDNNL_MATMUL_ALGO=2` is set for best performance (Blocked AOCL BLIS)
-4. You're using a sufficiently large model (small models may not benefit as much)
-5. Enable profiling to verify ZenDNN MatMul is being called
+2. `ZENDNNL_MATMUL_ALGO=1` is set for best performance (Blocked AOCL DLP)
+3. You're using a sufficiently large model (small models may not benefit as much)
+4. Enable profiling to verify ZenDNN MatMul is being called
 
 ### **GitHub Contribution**:
 Please add the **[ZenDNN]** prefix/tag in issues/PRs titles to help the ZenDNN-team check/address them without delay.

docs/backend/snapdragon/README.md

@@ -35,7 +35,7 @@ Adapt below build commands accordingly.
 Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
 
 ```
-[d]/workspace> cp docs/backend/hexagon/CMakeUserPresets.json .
+[d]/workspace> cp docs/backend/snapdragon/CMakeUserPresets.json .
 
 [d]/workspace> cmake --preset arm64-android-snapdragon-release -B build-snapdragon
 Preset CMake variables:
@@ -116,7 +116,7 @@ Llama-3.2-1B-Instruct-Q4_0.gguf: 1 file pushed, 0 skipped. 38.3 MB/s (773025920
 ### Windows
 
 All artifacts are already installed in the `pkg-snapdragon` folder.
-To run, adapt below instructions to use Powershell scrits in `scripts/snapdragon/windows`.
+To run, adapt below instructions to use Powershell scripts in `scripts/snapdragon/windows`.
 
 ## How to Run
 

docs/backend/snapdragon/windows.md

@@ -144,7 +144,7 @@ Once the build is complete HTP ops libraries will be installed like this
 -a----         1/22/2026   6:01 PM           4139 libggml-htp.cat
 ```
 
-The .cat file, the signature and proper certicate installation can be verified with
+The .cat file, the signature and proper certificate installation can be verified with
 
 ```
 > signtool.exe verify /v /pa .\pkg-snapdragon\lib\libggml-htp.cat

docs/build-s390x.md

@@ -242,10 +242,10 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 |------------|-------------|------|-------|
 | FP32       | ✅           | ✅    | ❓     |
 | FP16       | ✅           | ✅    | ❓     |
-| BF16       | 🚫           | ✅    | ❓     |
+| BF16       | ✅           | ✅    | ❓     |
 | Q4_0       | ✅           | ❓    | ❓     |
 | Q4_1       | ✅           | ❓    | ❓     |
-| MXFP4      | 🚫           | ❓    | ❓     |
+| MXFP4      | ✅           | ❓    | ❓     |
 | Q5_0       | ✅           | ❓    | ❓     |
 | Q5_1       | ✅           | ❓    | ❓     |
 | Q8_0       | ✅           | ❓    | ❓     |
@@ -272,4 +272,4 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 -   🚫 - acceleration unavailable, will still run using scalar implementation
 -   ❓ - acceleration unknown, please contribute if you can test it yourself
 
-Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Sep 7, 2025.
+Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on Feb 15, 2026.

docs/build.md

@@ -108,7 +108,7 @@ Building through oneAPI compilers will make avx_vnni instruction set available f
 - Using oneAPI docker image:
   If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above.
 
-Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
+Check [Optimizing and Running LLaMA2 on Intel® CPU](https://builders.intel.com/solutionslibrary/optimizing-and-running-llama2-on-intel-cpu) for more information.
 
 ### Other BLAS libraries
 
@@ -595,7 +595,7 @@ You can verify that KleidiAI is being used by running
 ```bash
 ./build/bin/llama-cli -m PATH_TO_MODEL -p "What is a car?"
 ```
-If KleidiAI is enabled, the ouput will contain a line similar to:
+If KleidiAI is enabled, the output will contain a line similar to:
 ```
 load_tensors: CPU_KLEIDIAI model buffer size =  3474.00 MiB
 ```
@@ -699,7 +699,7 @@ To read documentation for how to build on Android, [click here](./android.md)
 
 ## WebGPU [In Progress]
 
-The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The currrent implementation is up-to-date with Dawn commit `bed1a61`.
+The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The current implementation is up-to-date with Dawn commit `bed1a61`.
 
 In the llama.cpp directory, build with CMake:
 

docs/multimodal/MobileVLM.md

@@ -281,7 +281,7 @@ llama_print_timings:       total time =    5990.25 ms /   202 tokens
 
 Just the same as above.
 
-**ouput**
+**output**
 ```sh
 encode_image_with_clip: image embedding created: 144 tokens
 
@@ -305,7 +305,7 @@ llama_print_timings:       total time =   15513.95 ms /   412 tokens
 ## Run on Intel(R) Core(TM) Ultra7 115H
 ### operation system
 Windows11
-### comiple
+### compile
 ```sh
 make -j32
 ```

docs/ops.md

@@ -22,16 +22,16 @@ Legend:
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
 |                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
@@ -96,13 +96,13 @@ Legend:
 |                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |

docs/ops/SYCL.csv

@@ -77,8 +77,8 @@
 "SYCL0","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","FLOOR","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","FLOOR","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
-"SYCL0","CEIL","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
-"SYCL0","CEIL","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
+"SYCL0","CEIL","type=f16,ne_a=[128,2,2,2],v=1","support","1","yes","SYCL"
+"SYCL0","CEIL","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","ROUND","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","ROUND","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
 "SYCL0","TRUNC","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
@@ -161,8 +161,8 @@
 "SYCL0","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","FLOOR","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","FLOOR","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
-"SYCL0","CEIL","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
-"SYCL0","CEIL","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
+"SYCL0","CEIL","type=f32,ne_a=[128,2,2,2],v=1","support","1","yes","SYCL"
+"SYCL0","CEIL","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","ROUND","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","ROUND","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
 "SYCL0","TRUNC","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"

docs/ops/WebGPU.csv

@@ -8760,22 +8760,14 @@
 "WebGPU: WebGPU","ADD_ID","type_a=f32,type_b=f32,n_embd=129,n_experts=8,n_experts_used=4,n_token=1","support","0","no","WebGPU"
 "WebGPU: WebGPU","ADD_ID","type_a=f32,type_b=f32,n_embd=129,n_experts=8,n_experts_used=4,n_token=32","support","0","no","WebGPU"
 "WebGPU: WebGPU","ADD_ID","type_a=f32,type_b=f32,n_embd=129,n_experts=8,n_experts_used=4,n_token=129","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQR","type=f16,ne=[10,5,4,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQRT","type=f16,ne=[10,3,3,2]","support","0","no","WebGPU"
 "WebGPU: WebGPU","LOG","type=f16,ne=[10,5,4,3]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SIN","type=f16,ne=[10,2,2,2]","support","0","no","WebGPU"
-"WebGPU: WebGPU","COS","type=f16,ne=[10,2,2,2]","support","0","no","WebGPU"
 "WebGPU: WebGPU","CLAMP","type=f16,ne=[10,5,4,3],min=-0.500000,max=0.500000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f16,ne_a=[10,5,4,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","FLOOR","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","CEIL","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ROUND","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TRUNC","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SQR","type=f16,ne=[7,1,5,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQRT","type=f16,ne=[7,1,5,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","LOG","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SIN","type=f16,ne=[7,1,5,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","COS","type=f16,ne=[7,1,5,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","CLAMP","type=f16,ne=[7,1,5,3],min=-0.500000,max=0.500000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f16,ne_a=[7,1,5,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","FLOOR","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
@@ -8786,22 +8778,14 @@
 "WebGPU: WebGPU","ROUND","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TRUNC","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TRUNC","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SQR","type=f32,ne=[10,5,4,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQRT","type=f32,ne=[10,3,3,2]","support","0","no","WebGPU"
 "WebGPU: WebGPU","LOG","type=f32,ne=[10,5,4,3]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SIN","type=f32,ne=[10,2,2,2]","support","0","no","WebGPU"
-"WebGPU: WebGPU","COS","type=f32,ne=[10,2,2,2]","support","0","no","WebGPU"
 "WebGPU: WebGPU","CLAMP","type=f32,ne=[10,5,4,3],min=-0.500000,max=0.500000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f32,ne_a=[10,5,4,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","FLOOR","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","CEIL","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ROUND","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TRUNC","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SQR","type=f32,ne=[7,1,5,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQRT","type=f32,ne=[7,1,5,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","LOG","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SIN","type=f32,ne=[7,1,5,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","COS","type=f32,ne=[7,1,5,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","CLAMP","type=f32,ne=[7,1,5,3],min=-0.500000,max=0.500000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f32,ne_a=[7,1,5,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","FLOOR","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
@@ -9551,38 +9535,38 @@
 "WebGPU: WebGPU","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0,inplace=1","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=24,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=0,inplace=1","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=24,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0,inplace=1","support","1","yes","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","0","no","WebGPU"
-"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","0","no","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","1","yes","WebGPU"
+"WebGPU: WebGPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ARGSORT","type=f32,ne=[3,1,1,1],order=0","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ARGSORT","type=f32,ne=[4,1,1,1],order=0","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ARGSORT","type=f32,ne=[7,1,1,1],order=0","support","1","yes","WebGPU"
@@ -18901,3 +18885,27 @@
 "WebGPU: WebGPU","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","0","no","WebGPU"
 "WebGPU: WebGPU","OPT_STEP_ADAMW","type=f32,ne=[10,5,4,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","OPT_STEP_SGD","type=f32,ne=[10,5,4,3]","support","0","no","WebGPU"
+"WebGPU: WebGPU","SQR","type=f16,ne=[10,5,4,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f16,ne=[10,3,3,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f32,ne=[10,5,4,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f32,ne=[10,3,3,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f32,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f32,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f32,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f32,ne=[1024,1024,1,1]","support","1","yes","WebGPU"

docs/speculative.md

@@ -119,8 +119,6 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
                                         of lookup n-gram (default: 12)
 --spec-ngram-size-m N                   ngram size M for ngram-simple/ngram-map speculative decoding, length
                                         of draft m-gram (default: 48)
---spec-ngram-check-rate N               ngram check rate for ngram-simple/ngram-map speculative decoding
-                                        (default: 1)
 --spec-ngram-min-hits N                 minimum hits for ngram-map speculative decoding (default: 1)
 ```
 
@@ -153,10 +151,6 @@ Sets the size M of the draft m-gram for n-gram map based speculative decoding.
 The m-gram size determines how many tokens to draft when a match is found.
 Larger values can provide more speedup but may reduce acceptance rate.
 
-### `--spec-ngram-check-rate R`
-
-This option aims at performance if the n-gram lookup in history is to costly. A lookup will be executed at every R tokens (default is 1, every token).
-
 ### `--spec-ngram-min-hits H`
 
 This option defines how often a key has to appear in the token history to be used as a draft (default is 1).
@@ -175,7 +169,12 @@ draft acceptance rate = 0.70312 (   90 accepted /   128 generated)
 statistics ngram_mod: #calls = 810, #gen drafts = 15, #acc drafts = 15, #gen tokens = 960, #acc tokens = 730, dur(b,g,a) = 0.149, 0.347, 0.005 ms
 ```
 
-- `#calls`: number of calls of this implementations
+```
+statistics ngram_map_k: #calls(b,g,a) = 6 1690 26, #gen drafts = 26, #acc drafts = 26, #gen tokens = 1248, #acc tokens = 968, dur(b,g,a) = 2.234, 1.427, 0.016 ms
+```
+
+
+- `#calls(b,g,a)`: number of calls of begin (new prompt), generation and accumulation of this implementations
 - `#gen drafts`: number of drafts generated by this implementation
 - `#acc drafts`: number of drafts accepted (partially) by the main model
 - `#gen tokens`: number of tokens generated by this implementation (including rejected tokens)

examples/batched/batched.cpp

@@ -5,6 +5,7 @@
 #include "sampling.h"
 
 #include <algorithm>
+#include <clocale>
 #include <cstdio>
 #include <string>
 #include <vector>
@@ -16,6 +17,8 @@ static void print_usage(int, char ** argv) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     params.prompt = "Hello my name is";

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -5,14 +5,16 @@
 #include "common.h"
 #include "log.h"
 
+#include <algorithm>
+#include <cassert>
+#include <cinttypes>
+#include <climits>
+#include <clocale>
+#include <cstdarg>
+#include <cstring>
+#include <ctime>
 #include <unordered_map>
 #include <vector>
-#include <cassert>
-#include <climits>
-#include <cstring>
-#include <cstdarg>
-#include <cinttypes>
-#include <ctime>
 #include <random>
 #include <stdexcept>
 #include <sstream>
@@ -874,6 +876,8 @@ static std::string basename(const std::string &path) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_init();
 
     struct train_params params = get_default_train_params();

examples/debug/README.md

@@ -2,7 +2,7 @@
 
 This is a utility intended to help debug a model by registering a callback that
 logs GGML operations and tensor data. It can also store the generated logits or
-embeddings as well as the prompt and token ids for comparision with the original
+embeddings as well as the prompt and token ids for comparison with the original
 model.
 
 ### Usage

examples/deprecation-warning/deprecation-warning.cpp

@@ -1,11 +1,14 @@
 // Warns users that this filename was deprecated, and provides a link for more information.
 
+#include <clocale>
 #include <cstdio>
 #include <string>
 #include <unordered_map>
 
 // Main
 int main(int argc, char** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     std::string filename = "main";
     if (argc >= 1) {
         filename = argv[0];

examples/diffusion/README.md

@@ -43,12 +43,12 @@ Choose one of the following scheduling methods:
 - `-b`: Batch size
 
 ### Examples
-#### Dream architechture:
+#### Dream architecture:
 ```
 llama-diffusion-cli -m dream7b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-eps 0.001 --diffusion-algorithm 3 --diffusion-steps 256 --diffusion-visual
 ```
 
-#### LLaDA architechture:
+#### LLaDA architecture:
 ```
 llama-diffusion-cli -m llada-8b.gguf -p "write code to train MNIST in pytorch" -ub 512 --diffusion-block-length 32 --diffusion-steps 256 --diffusion-visual
 ```

examples/diffusion/diffusion-cli.cpp

@@ -7,6 +7,7 @@
 #include <limits.h>
 
 #include <algorithm>
+#include <clocale>
 #include <cmath>
 #include <cstring>
 #include <limits>
@@ -538,6 +539,8 @@ static std::string format_input_text(const std::string & prompt, const std::stri
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     ggml_time_init();
 
     common_params params;

examples/embedding/embedding.cpp

@@ -3,6 +3,7 @@
 #include "log.h"
 #include "llama.h"
 
+#include <clocale>
 #include <ctime>
 #include <algorithm>
 
@@ -94,6 +95,8 @@ static void print_raw_embeddings(const float * emb,
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) {

examples/eval-callback/eval-callback.cpp

@@ -4,6 +4,8 @@
 #include "log.h"
 #include "llama.h"
 #include "llama-cpp.h"
+
+#include <clocale>
 #include <string>
 #include <vector>
 
@@ -29,6 +31,8 @@ static bool run(llama_context * ctx, const common_params & params) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     base_callback_data cb_data;
 
     common_params params;

examples/gen-docs/gen-docs.cpp

@@ -1,6 +1,7 @@
 #include "arg.h"
 #include "common.h"
 
+#include <clocale>
 #include <fstream>
 #include <sstream>
 #include <string>
@@ -100,6 +101,8 @@ static void write_help(std::ostringstream & ss, const md_file & md) {
 }
 
 int main(int, char **) {
+    std::setlocale(LC_NUMERIC, "C");
+
     for (const auto & md : md_files) {
         std::ifstream infile(md.fname);
         if (!infile.is_open()) {

examples/gguf-hash/gguf-hash.cpp

@@ -1,13 +1,14 @@
 #include "ggml.h"
 #include "gguf.h"
 
-#include <cstdlib>   /* abort() */
+#include <algorithm>
+#include <clocale>
 #include <cstddef>
 #include <cstdio>
-#include <string>
-#include <stdexcept>
-#include <algorithm>
+#include <cstdlib>   /* abort() */
 #include <cstring>
+#include <stdexcept>
+#include <string>
 
 #include <sstream>
 #include <fstream>
@@ -626,6 +627,8 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) {
 }
 
 int main(int argc, const char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     hash_params params;
     manifest_check_params manifest_check;
     hash_params_parse(argc, argv, params);

examples/gguf/gguf.cpp

@@ -1,6 +1,7 @@
 #include "ggml.h"
 #include "gguf.h"
 
+#include <clocale>
 #include <cstdio>
 #include <string>
 #include <sstream>
@@ -240,6 +241,8 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     if (argc < 3) {
         printf("usage: %s data.gguf r|w [n]\n", argv[0]);
         printf("r: read data.gguf file\n");

examples/llama.vim

@@ -52,8 +52,8 @@ highlight llama_hl_info guifg=#77ff2f ctermfg=119
 "   n_prefix:         number of lines before the cursor location to include in the local prefix
 "   n_suffix:         number of lines after  the cursor location to include in the local suffix
 "   n_predict:        max number of tokens to predict
-"   t_max_prompt_ms:  max alloted time for the prompt processing (TODO: not yet supported)
-"   t_max_predict_ms: max alloted time for the prediction
+"   t_max_prompt_ms:  max allotted time for the prompt processing (TODO: not yet supported)
+"   t_max_predict_ms: max allotted time for the prediction
 "   show_info:        show extra info about the inference (0 - disabled, 1 - statusline, 2 - inline)
 "   auto_fim:         trigger FIM completion automatically on cursor movement
 "   max_line_suffix:  do not auto-trigger FIM completion if there are more than this number of characters to the right of the cursor

examples/lookahead/lookahead.cpp

@@ -4,10 +4,11 @@
 #include "log.h"
 #include "llama.h"
 
+#include <algorithm>
+#include <clocale>
 #include <cstdio>
 #include <string>
 #include <vector>
-#include <algorithm>
 
 struct ngram_data {
     bool active = false;
@@ -38,6 +39,8 @@ struct ngram_container {
 };
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {

examples/lookup/lookup-create.cpp

@@ -3,10 +3,13 @@
 #include "ngram-cache.h"
 #include "llama.h"
 
+#include <clocale>
 #include <string>
 #include <vector>
 
 int main(int argc, char ** argv){
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {

examples/lookup/lookup-merge.cpp

@@ -3,6 +3,7 @@
 #include "common.h"
 #include "ngram-cache.h"
 
+#include <clocale>
 #include <cstdint>
 #include <cstdio>
 #include <fstream>
@@ -17,6 +18,8 @@ static void print_usage(char* argv0) {
 }
 
 int main(int argc, char ** argv){
+    std::setlocale(LC_NUMERIC, "C");
+
     if (argc < 3) {
         print_usage(argv[0]);
         exit(1);

examples/lookup/lookup-stats.cpp

@@ -5,14 +5,17 @@
 #include "llama.h"
 #include "ggml.h"
 
+#include <cinttypes>
+#include <clocale>
 #include <cstdint>
 #include <cstdio>
-#include <cinttypes>
 #include <fstream>
 #include <string>
 #include <vector>
 
 int main(int argc, char ** argv){
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {

examples/lookup/lookup.cpp

@@ -6,6 +6,7 @@
 #include "log.h"
 #include "llama.h"
 
+#include <clocale>
 #include <cstdint>
 #include <cstdio>
 #include <fstream>
@@ -13,6 +14,8 @@
 #include <vector>
 
 int main(int argc, char ** argv){
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {

examples/model-conversion/Makefile

@@ -77,7 +77,10 @@ causal-verify-embeddings: causal-run-original-embeddings causal-run-converted-em
 	@./scripts/causal/compare-embeddings-logits.sh
 
 causal-inspect-original-model:
-	@./scripts/utils/inspect-org-model.py
+	@./scripts/utils/inspect-org-model.py --list-all -s
+
+causal-list-original-model-tensors:
+	@./scripts/utils/inspect-org-model.py --list-all-short -s
 
 causal-inspect-converted-model:
 	@./scripts/utils/inspect-converted-model.sh
@@ -153,7 +156,7 @@ embedding-verify-logits-st: embedding-run-original-model-st embedding-run-conver
 
 embedding-inspect-original-model:
 	$(call validate_embedding_model_path,embedding-inspect-original-model)
-	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/utils/inspect-org-model.py -m ${EMBEDDING_MODEL_PATH}
+	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/utils/inspect-org-model.py -m ${EMBEDDING_MODEL_PATH} --list-all -s
 
 embedding-inspect-converted-model:
 	@CONVERTED_EMBEDDING_MODEL="$(CONVERTED_EMBEDDING_MODEL)" ./scripts/utils/inspect-converted-model.sh ${CONVERTED_EMBEDDING_MODEL}

examples/model-conversion/README.md

@@ -69,7 +69,7 @@ Command line arguments take precedence over environment variables when both are
 
 In cases where the transformer implementation for the model has not been released
 yet it is possible to set the environment variable `UNRELEASED_MODEL_NAME` which
-will then cause the transformer implementation to be loaded explicitely and not
+will then cause the transformer implementation to be loaded explicitly and not
 use AutoModelForCausalLM:
 ```
 export UNRELEASED_MODEL_NAME=SomeNewModel
@@ -120,7 +120,7 @@ The converted model can be inspected using the following command:
 (venv) $ make causal-run-converted-model
 ```
 
-### Model logits verfication
+### Model logits verification
 The following target will run the original model and the converted model and
 compare the logits:
 ```console
@@ -235,7 +235,7 @@ new model the model can be converted to GGUF format using the following command:
 (venv) $ make embedding-run-converted-model
 ```
 
-### Model logits verfication
+### Model logits verification
 The following target will run the original model and the converted model (which
 was done manually in the previous steps) and compare the logits:
 ```console
@@ -335,7 +335,7 @@ $ make perplexity-run-full QUANTIZED_MODEL=~/path/to/quantized/model-Qxx.gguf LO
 
 ## HuggingFace utilities
 The following targets are useful for creating collections and model repositories
-on Hugging Face in the the ggml-org. These can be used when preparing a relase
+on Hugging Face in the the ggml-org. These can be used when preparing a release
 to script the process for new model releases.
 
 For the following targets a `HF_TOKEN` environment variable is required.
@@ -347,7 +347,7 @@ For the following targets a `HF_TOKEN` environment variable is required.
 > $ unset HF_TOKEN
 
 ### Create a new Hugging Face Model (model repository)
-This will create a new model repsository on Hugging Face with the specified
+This will create a new model repository on Hugging Face with the specified
 model name.
 ```console
 (venv) $ make hf-create-model MODEL_NAME='TestModel' NAMESPACE="danbev" ORIGINAL_BASE_MODEL="some-base-model"

examples/model-conversion/scripts/causal/run-org-model.py

@@ -42,11 +42,15 @@ def load_model_and_tokenizer(model_path, device="auto"):
         config = config.text_config
         multimodal = True
 
-    print("Vocab size:       ", config.vocab_size)
-    print("Hidden size:      ", config.hidden_size)
-    print("Number of layers: ", config.num_hidden_layers)
-    print("BOS token id:     ", config.bos_token_id)
-    print("EOS token id:     ", config.eos_token_id)
+    def print_if_exists(label, obj, attr, default="N/A"):
+        val = getattr(obj, attr) if hasattr(obj, attr) else default
+        print(f"{label}", val)
+
+    print_if_exists("Vocab size:       ", config, "vocab_size")
+    print_if_exists("Hidden size:      ", config, "hidden_size")
+    print_if_exists("Number of layers: ", config, "num_hidden_layers")
+    print_if_exists("BOS token id:     ", config, "bos_token_id")
+    print_if_exists("EOS token id:     ", config, "eos_token_id")
 
     unreleased_model_name = os.getenv("UNRELEASED_MODEL_NAME")
     if unreleased_model_name:

examples/model-conversion/scripts/utils/inspect-org-model.py

@@ -1,67 +1,290 @@
 #!/usr/bin/env python3
 
 import argparse
-import os
 import json
+import os
+import re
+import struct
+import sys
+from pathlib import Path
+from typing import Optional
 from safetensors import safe_open
-from collections import defaultdict
 
-parser = argparse.ArgumentParser(description='Process model with specified path')
-parser.add_argument('--model-path', '-m', help='Path to the model')
-args = parser.parse_args()
 
-model_path = os.environ.get('MODEL_PATH', args.model_path)
-if model_path is None:
-    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+MODEL_SAFETENSORS_FILE = "model.safetensors"
+MODEL_SAFETENSORS_INDEX = "model.safetensors.index.json"
 
-# Check if there's an index file (multi-file model)
-index_path = os.path.join(model_path, "model.safetensors.index.json")
-single_file_path = os.path.join(model_path, "model.safetensors")
+DTYPE_SIZES = {
+    "F64": 8, "I64": 8, "U64": 8,
+    "F32": 4, "I32": 4, "U32": 4,
+    "F16": 2, "BF16": 2, "I16": 2, "U16": 2,
+    "I8": 1, "U8": 1, "BOOL": 1,
+    "F8_E4M3": 1, "F8_E5M2": 1,
+}
 
-if os.path.exists(index_path):
-    # Multi-file model
-    print("Multi-file model detected")
+SIZE_UNITS = ['B', 'KB', 'MB', 'GB', 'TB']
 
-    with open(index_path, 'r') as f:
-        index_data = json.load(f)
 
-    # Get the weight map (tensor_name -> file_name)
-    weight_map = index_data.get("weight_map", {})
+def get_weight_map(model_path: Path) -> Optional[dict[str, str]]:
+    index_file = model_path / MODEL_SAFETENSORS_INDEX
 
-    # Group tensors by file for efficient processing
-    file_tensors = defaultdict(list)
-    for tensor_name, file_name in weight_map.items():
-        file_tensors[file_name].append(tensor_name)
+    if index_file.exists():
+        with open(index_file, 'r') as f:
+            index = json.load(f)
+            return index.get("weight_map", {})
 
-    print("Tensors in model:")
+    return None
 
-    # Process each shard file
-    for file_name, tensor_names in file_tensors.items():
-        file_path = os.path.join(model_path, file_name)
-        print(f"\n--- From {file_name} ---")
 
-        with safe_open(file_path, framework="pt") as f:
-            for tensor_name in sorted(tensor_names):
-                tensor = f.get_tensor(tensor_name)
-                print(f"- {tensor_name} : shape = {tensor.shape}, dtype = {tensor.dtype}")
+def get_all_tensor_names(model_path: Path) -> list[str]:
+    weight_map = get_weight_map(model_path)
 
-elif os.path.exists(single_file_path):
-    # Single file model (original behavior)
-    print("Single-file model detected")
+    if weight_map is not None:
+        return list(weight_map.keys())
 
-    with safe_open(single_file_path, framework="pt") as f:
-        keys = f.keys()
-        print("Tensors in model:")
-        for key in sorted(keys):
-            tensor = f.get_tensor(key)
-            print(f"- {key} : shape = {tensor.shape}, dtype = {tensor.dtype}")
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        try:
+            with safe_open(single_file, framework="pt", device="cpu") as f:
+                return list(f.keys())
+        except Exception as e:
+            print(f"Error reading {single_file}: {e}")
+            sys.exit(1)
 
-else:
-    print(f"Error: Neither 'model.safetensors.index.json' nor 'model.safetensors' found in {model_path}")
-    print("Available files:")
-    if os.path.exists(model_path):
-        for item in sorted(os.listdir(model_path)):
-            print(f"  {item}")
+    print(f"Error: No safetensors files found in {model_path}")
+    sys.exit(1)
+
+
+def find_tensor_file(model_path: Path, tensor_name: str) -> Optional[str]:
+    weight_map = get_weight_map(model_path)
+
+    if weight_map is not None:
+        return weight_map.get(tensor_name)
+
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        return single_file.name
+
+    return None
+
+
+def read_safetensors_header(file_path: Path) -> dict:
+    with open(file_path, 'rb') as f:
+        header_size = struct.unpack('<Q', f.read(8))[0]
+        return json.loads(f.read(header_size))
+
+
+def get_tensor_size_bytes(tensor_meta: dict) -> int:
+    offsets = tensor_meta.get("data_offsets")
+    if offsets and len(offsets) == 2:
+        return offsets[1] - offsets[0]
+    n_elements = 1
+    for d in tensor_meta.get("shape", []):
+        n_elements *= d
+    return n_elements * DTYPE_SIZES.get(tensor_meta.get("dtype", "F32"), 4)
+
+
+def format_size(size_bytes: int) -> str:
+    val = float(size_bytes)
+    for unit in SIZE_UNITS[:-1]:
+        if val < 1024.0:
+            return f"{val:.2f} {unit}"
+        val /= 1024.0
+    return f"{val:.2f} {SIZE_UNITS[-1]}"
+
+
+def get_all_tensor_metadata(model_path: Path) -> dict[str, dict]:
+    weight_map = get_weight_map(model_path)
+
+    if weight_map is not None:
+        file_to_tensors: dict[str, list[str]] = {}
+        for tensor_name, file_name in weight_map.items():
+            file_to_tensors.setdefault(file_name, []).append(tensor_name)
+
+        all_metadata: dict[str, dict] = {}
+        for file_name, tensor_names in file_to_tensors.items():
+            try:
+                header = read_safetensors_header(model_path / file_name)
+                for tensor_name in tensor_names:
+                    if tensor_name in header:
+                        all_metadata[tensor_name] = header[tensor_name]
+            except Exception as e:
+                print(f"Warning: Could not read header from {file_name}: {e}", file=sys.stderr)
+        return all_metadata
+
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        try:
+            header = read_safetensors_header(single_file)
+            return {k: v for k, v in header.items() if k != "__metadata__"}
+        except Exception as e:
+            print(f"Error reading {single_file}: {e}")
+            sys.exit(1)
+
+    print(f"Error: No safetensors files found in {model_path}")
+    sys.exit(1)
+
+
+def normalize_tensor_name(tensor_name: str) -> str:
+    normalized = re.sub(r'\.\d+\.', '.#.', tensor_name)
+    normalized = re.sub(r'\.\d+$', '.#', normalized)
+    return normalized
+
+
+def list_all_tensors(
+    model_path: Path,
+    short: bool = False,
+    show_sizes: bool = False,
+):
+    tensor_names = get_all_tensor_names(model_path)
+
+    metadata: Optional[dict[str, dict]] = None
+    if show_sizes:
+        metadata = get_all_tensor_metadata(model_path)
+
+    total_bytes = 0
+
+    if short:
+        seen: dict[str, str] = {}
+        for tensor_name in sorted(tensor_names):
+            normalized = normalize_tensor_name(tensor_name)
+            if normalized not in seen:
+                seen[normalized] = tensor_name
+        display_pairs = list(sorted(seen.items()))
+        name_width = max((len(n) for n, _ in display_pairs), default=0)
+        for normalized, first_name in display_pairs:
+            if metadata and first_name in metadata:
+                m = metadata[first_name]
+                size = get_tensor_size_bytes(m)
+                total_bytes += size
+                print(f"{normalized:{name_width}}  {m.get('dtype', '?'):6s}  {str(m.get('shape', '')):30s}  {format_size(size)}")
+            else:
+                print(normalized)
     else:
-        print(f"  Directory {model_path} does not exist")
-    exit(1)
+        name_width = max((len(n) for n in tensor_names), default=0)
+        for tensor_name in sorted(tensor_names):
+            if metadata and tensor_name in metadata:
+                m = metadata[tensor_name]
+                size = get_tensor_size_bytes(m)
+                total_bytes += size
+                print(f"{tensor_name:{name_width}}  {m.get('dtype', '?'):6s}  {str(m.get('shape', '')):30s}  {format_size(size)}")
+            else:
+                print(tensor_name)
+
+    if show_sizes:
+        print(f"\nTotal: {format_size(total_bytes)}")
+
+
+def print_tensor_info(model_path: Path, tensor_name: str, num_values: Optional[int] = None):
+    tensor_file = find_tensor_file(model_path, tensor_name)
+
+    if tensor_file is None:
+        print(f"Error: Could not find tensor '{tensor_name}' in model index")
+        print(f"Model path: {model_path}")
+        sys.exit(1)
+
+    file_path = model_path / tensor_file
+
+    try:
+        header = read_safetensors_header(file_path)
+        tensor_meta = header.get(tensor_name, {})
+        dtype_str = tensor_meta.get("dtype")
+
+        with safe_open(file_path, framework="pt", device="cpu") as f:
+            if tensor_name in f.keys():
+                tensor_slice = f.get_slice(tensor_name)
+                shape = tensor_slice.get_shape()
+                print(f"Tensor: {tensor_name}")
+                print(f"File:   {tensor_file}")
+                print(f"Shape:  {shape}")
+                if dtype_str:
+                    print(f"Dtype:  {dtype_str}")
+                if tensor_meta:
+                    print(f"Size:   {format_size(get_tensor_size_bytes(tensor_meta))}")
+                if num_values is not None:
+                    tensor = f.get_tensor(tensor_name)
+                    if not dtype_str:
+                        print(f"Dtype:  {tensor.dtype}")
+                    flat = tensor.flatten()
+                    n = min(num_values, flat.numel())
+                    print(f"Values: {flat[:n].tolist()}")
+            else:
+                print(f"Error: Tensor '{tensor_name}' not found in {tensor_file}")
+                sys.exit(1)
+
+    except FileNotFoundError:
+        print(f"Error: The file '{file_path}' was not found.")
+        sys.exit(1)
+    except Exception as e:
+        print(f"An error occurred: {e}")
+        sys.exit(1)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Print tensor information from a safetensors model"
+    )
+    parser.add_argument(
+        "tensor_name",
+        nargs="?",
+        help="Name of the tensor to inspect"
+    )
+    parser.add_argument(
+        "-m", "--model-path",
+        type=Path,
+        help="Path to the model directory (default: MODEL_PATH environment variable)"
+    )
+    parser.add_argument(
+        "-l", "--list-all-short",
+        action="store_true",
+        help="List unique tensor patterns (layer numbers replaced with #)"
+    )
+    parser.add_argument(
+        "-la", "--list-all",
+        action="store_true",
+        help="List all tensor names with actual layer numbers"
+    )
+    parser.add_argument(
+        "-n", "--num-values",
+        nargs="?",
+        const=10,
+        default=None,
+        type=int,
+        metavar="N",
+        help="Print the first N values of the tensor flattened (default: 10 if flag is given without a number)"
+    )
+    parser.add_argument(
+        "-s", "--sizes",
+        action="store_true",
+        help="Show dtype, shape, and size for each tensor when listing"
+    )
+
+    args = parser.parse_args()
+
+    model_path = args.model_path
+    if model_path is None:
+        model_path_str = os.environ.get("MODEL_PATH")
+        if model_path_str is None:
+            print("Error: --model-path not provided and MODEL_PATH environment variable not set")
+            sys.exit(1)
+        model_path = Path(model_path_str)
+
+    if not model_path.exists():
+        print(f"Error: Model path does not exist: {model_path}")
+        sys.exit(1)
+
+    if not model_path.is_dir():
+        print(f"Error: Model path is not a directory: {model_path}")
+        sys.exit(1)
+
+    if args.list_all_short or args.list_all:
+        list_all_tensors(model_path, short=args.list_all_short, show_sizes=args.sizes)
+    else:
+        if args.tensor_name is None:
+            print("Error: tensor_name is required when not using --list-all-short or --list-all")
+            sys.exit(1)
+        print_tensor_info(model_path, args.tensor_name, args.num_values)
+
+
+if __name__ == "__main__":
+    main()

examples/parallel/parallel.cpp

@@ -7,12 +7,13 @@
 #include "log.h"
 #include "llama.h"
 
+#include <algorithm>
+#include <clocale>
 #include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 #include <ctime>
-#include <algorithm>
 
 // trim whitespace from the beginning and end of a string
 static std::string trim(const std::string & str) {
@@ -153,6 +154,8 @@ static std::vector<std::string> split_string(const std::string& input, char deli
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     srand(1234);
 
     common_params params;

examples/passkey/passkey.cpp

@@ -3,6 +3,7 @@
 #include "log.h"
 #include "llama.h"
 
+#include <clocale>
 #include <cmath>
 #include <cstdio>
 #include <string>
@@ -16,6 +17,8 @@ static void print_usage(int, char ** argv) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     params.n_junk = 250;

examples/retrieval/retrieval.cpp

@@ -4,6 +4,7 @@
 #include "llama.h"
 
 #include <algorithm>
+#include <clocale>
 #include <fstream>
 #include <iostream> // TODO: remove me
 
@@ -112,6 +113,8 @@ static void batch_process(llama_context * ctx, llama_batch & batch, float * outp
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_RETRIEVAL, print_usage)) {

examples/save-load-state/save-load-state.cpp

@@ -2,15 +2,21 @@
 #include "common.h"
 #include "llama.h"
 
+#include <clocale>
 #include <vector>
 #include <cstdio>
 
+
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     params.prompt = "The quick brown fox";
     params.sampling.seed = 1234;
 
+    const std::string_view state_file = "dump_state.bin";
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
@@ -53,35 +59,16 @@ int main(int argc, char ** argv) {
     // tokenize prompt
     auto tokens = common_tokenize(ctx, params.prompt, true);
 
-    // prepare the batch
-    llama_batch batch = llama_batch_init(tokens.size(), 0, 1);
-    for (size_t i = 0; i < tokens.size(); i++) {
-        common_batch_add(batch, tokens[i], i, {0}, false);
+    const bool save_state = true;
+    if (!common_prompt_batch_decode(ctx, tokens, n_past, params.n_batch, state_file, save_state)) {
+        return 1;
     }
-    batch.logits[batch.n_tokens - 1] = true; // generate next token
-
-    // evaluate prompt
-    llama_decode(ctx, batch);
-    n_past += batch.n_tokens;
-
-    // save state (rng, logits, embedding and kv_cache) to file
-    {
-        std::vector<uint8_t> state_mem(llama_state_get_size(ctx));
-        const size_t written = llama_state_get_data(ctx, state_mem.data(), state_mem.size());
-
-        FILE *fp_write = fopen("dump_state.bin", "wb");
-        fwrite(state_mem.data(), 1, written, fp_write);
-        fclose(fp_write);
-
-        fprintf(stderr, "%s : serialized state into %zd out of a maximum of %zd bytes\n", __func__, written, state_mem.size());
-    }
-
-    // save state (last tokens)
-    const auto n_past_saved = n_past;
 
     // first run
     printf("\nfirst run: %s", params.prompt.c_str());
 
+    llama_batch batch = llama_batch_init(1, 0, 1);
+
     for (auto i = 0; i < params.n_predict; i++) {
         auto next_token     = llama_sampler_sample(smpl, ctx, -1);
         auto next_token_str = common_token_to_piece(ctx, next_token);
@@ -111,27 +98,23 @@ int main(int argc, char ** argv) {
 
     printf("\nsecond run: %s", params.prompt.c_str());
 
-    // load state (rng, logits, embedding and kv_cache) from file
-    {
-        std::vector<uint8_t> state_mem;
+    // load state from file
+    std::vector<llama_token> unused_sts(tokens.size()); // unused session tokens.
+    size_t n_token_count_out = 0;
 
-        FILE * fp_read = fopen("dump_state.bin", "rb");
-        fseek(fp_read, 0, SEEK_END);
-        state_mem.resize(ftell(fp_read));
-        fseek(fp_read, 0, SEEK_SET);
-        const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
-        fclose(fp_read);
-
-        if (read != llama_state_set_data(ctx2, state_mem.data(), state_mem.size())) {
-            fprintf(stderr, "\n%s : failed to read state\n", __func__);
-            return 1;
-        }
-
-        fprintf(stderr, "%s : deserialized state from %zd out of a maximum of %zd bytes\n", __func__, read, state_mem.size());
+    if (!llama_state_load_file(ctx2, state_file.data(), unused_sts.data(), unused_sts.size(), &n_token_count_out)) {
+        fprintf(stderr, "\n%s : failed to load state\n", __func__);
+        return 1;
     }
 
+    fprintf(stderr, "%s : loaded state with %zu tokens\n", __func__, n_token_count_out);
+
     // restore state (last tokens)
-    n_past = n_past_saved;
+    n_past = n_token_count_out;
+    if (!common_replay_last_token(ctx2, tokens.back(), n_past)) {
+        return 1;
+    }
+    ++n_past;
 
     // second run
     for (auto i = 0; i < params.n_predict; i++) {
@@ -160,7 +143,9 @@ int main(int argc, char ** argv) {
     }
 
     // make new context
-    llama_context * ctx3 = llama_init_from_model(model, common_context_params_to_llama(params));
+    auto params_ctx3 = common_context_params_to_llama(params);
+    params_ctx3.n_seq_max = 2;
+    llama_context * ctx3 = llama_init_from_model(model, params_ctx3);
 
     llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
 
@@ -169,26 +154,21 @@ int main(int argc, char ** argv) {
     printf("\nsingle seq run: %s", params.prompt.c_str());
 
     // load state (rng, logits, embedding and kv_cache) from file
-    {
-        std::vector<uint8_t> state_mem;
+    n_token_count_out = 0;
 
-        FILE * fp_read = fopen("dump_state.bin", "rb");
-        fseek(fp_read, 0, SEEK_END);
-        state_mem.resize(ftell(fp_read));
-        fseek(fp_read, 0, SEEK_SET);
-        const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
-        fclose(fp_read);
-
-        if (read != llama_state_set_data(ctx3, state_mem.data(), state_mem.size())) {
-            fprintf(stderr, "\n%s : failed to read state\n", __func__);
-            return 1;
-        }
-
-        fprintf(stderr, "%s : deserialized state from %zd out of a maximum of %zd bytes\n", __func__, read, state_mem.size());
+    if (!llama_state_load_file(ctx3, state_file.data(), unused_sts.data(), unused_sts.size(), &n_token_count_out)) {
+        fprintf(stderr, "\n%s : failed to load state\n", __func__);
+        return 1;
     }
 
+    fprintf(stderr, "%s : loaded state with %zu tokens\n", __func__, n_token_count_out);
+
     // restore state (last tokens)
-    n_past = n_past_saved;
+    n_past = n_token_count_out;
+    if (!common_replay_last_token(ctx3, tokens.back(), n_past)) {
+        return 1;
+    }
+    ++n_past;
 
     // save seq 0 and load into seq 1
     {

examples/simple-chat/simple-chat.cpp

@@ -1,4 +1,5 @@
 #include "llama.h"
+#include <clocale>
 #include <cstdio>
 #include <cstring>
 #include <iostream>
@@ -12,6 +13,8 @@ static void print_usage(int, char ** argv) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     std::string model_path;
     int ngl = 99;
     int n_ctx = 2048;

examples/simple/simple.cpp

@@ -1,4 +1,5 @@
 #include "llama.h"
+#include <clocale>
 #include <cstdio>
 #include <cstring>
 #include <string>
@@ -11,6 +12,8 @@ static void print_usage(int, char ** argv) {
 }
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     // path to the model gguf file
     std::string model_path;
     // prompt to generate text from

examples/speculative-simple/speculative-simple.cpp

@@ -5,12 +5,15 @@
 #include "log.h"
 #include "llama.h"
 
+#include <clocale>
 #include <cstdio>
 #include <cstring>
 #include <string>
 #include <vector>
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {

examples/speculative/speculative.cpp

@@ -5,6 +5,7 @@
 #include "llama.h"
 
 #include <algorithm>
+#include <clocale>
 #include <cstdio>
 #include <cstring>
 #include <random>
@@ -30,6 +31,8 @@ struct seq_draft {
 };
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
 
     // needed to get candidate probs even for temp <= 0.0

examples/sycl/README.md

@@ -6,11 +6,11 @@ This example program provides the tools for llama.cpp for SYCL on Intel GPU.
 
 |Tool Name| Function|Status|
 |-|-|-|
-|llama-ls-sycl-device| List all SYCL devices with ID, compute capability, max work group size, ect.|Support|
+|llama-ls-sycl-device| List all SYCL devices with ID, compute capability, max work group size, etc.|Support|
 
 ### llama-ls-sycl-device
 
-List all SYCL devices with ID, compute capability, max work group size, ect.
+List all SYCL devices with ID, compute capability, max work group size, etc.
 
 1. Build the llama.cpp for SYCL for the specified target *(using GGML_SYCL_TARGET)*.
 

examples/sycl/ls-sycl-device.cpp

@@ -6,8 +6,10 @@
 
 
 #include "ggml-sycl.h"
+#include <clocale>
 
 int main() {
+    std::setlocale(LC_NUMERIC, "C");
     ggml_backend_sycl_print_sycl_devices();
     return 0;
 }

examples/training/finetune.cpp

@@ -3,6 +3,7 @@
 #include "log.h"
 #include "llama.h"
 
+#include <clocale>
 #include <cmath>
 #include <cstdio>
 #include <cstring>
@@ -14,6 +15,8 @@
 #endif
 
 int main(int argc, char ** argv) {
+    std::setlocale(LC_NUMERIC, "C");
+
     common_params params;
     params.escape = false;
 

ggml/CMakeLists.txt

@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 9)
-set(GGML_VERSION_PATCH 5)
+set(GGML_VERSION_PATCH 7)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)

ggml/include/ggml-backend.h

@@ -259,7 +259,7 @@ extern "C" {
       Example usage:
 
         // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
-        // preferrably to run on the same backend as the buffer
+        // preferably to run on the same backend as the buffer
         ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
 
         sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false, true);

ggml/include/ggml-opt.h

@@ -138,7 +138,7 @@ extern "C" {
     GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
     GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
 
-    // set gradients to zero, initilize loss, and optionally reset the optimizer
+    // set gradients to zero, initialize loss, and optionally reset the optimizer
     GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
 
     GGML_API bool ggml_opt_static_graphs(ggml_opt_context_t opt_ctx); // whether the graphs are allocated_statically

ggml/include/ggml-virtgpu.h

@@ -7,8 +7,6 @@
 extern "C" {
 #endif
 
-#define GGML_REMOTING_FRONTEND_NAME "RemotingFrontend"
-
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_virtgpu_reg();
 
 #ifdef  __cplusplus

ggml/include/ggml.h

@@ -730,10 +730,6 @@ extern "C" {
     GGML_API size_t  ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
     GGML_API size_t  ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
 
-    GGML_DEPRECATED(
-    GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
-    "use ggml_row_size() instead");
-
     GGML_API const char * ggml_type_name(enum ggml_type type);
     GGML_API const char * ggml_op_name  (enum ggml_op   op);
     GGML_API const char * ggml_op_symbol(enum ggml_op   op);
@@ -752,6 +748,7 @@ extern "C" {
     GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_empty     (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_view      (const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_scalar    (const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_vector    (const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_matrix    (const struct ggml_tensor * tensor);
@@ -2578,7 +2575,7 @@ 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
+    // build forward multiple 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
     //

ggml/src/ggml-alloc.c

@@ -17,11 +17,6 @@
 //#define AT_PRINTF(...) GGML_LOG_DEBUG(__VA_ARGS__)
 #define AT_PRINTF(...)
 
-
-static bool ggml_is_view(const struct ggml_tensor * t) {
-    return t->view_src != NULL;
-}
-
 // ops that return true for this function must not use restrict pointers for their backend implementations
 bool ggml_op_can_inplace(enum ggml_op op) {
     switch (op) {
@@ -627,7 +622,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
     GGML_ASSERT(buffer_id >= 0);
     struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
 
-    if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
+    if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_impl_is_view(node)) {
         hn->allocated = true;
         assert(hn->addr.offset == 0);
 
@@ -658,7 +653,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
 
                 struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
                 if (p_hn->n_children == 1 && p_hn->n_views == 0) {
-                    if (ggml_is_view(parent)) {
+                    if (ggml_impl_is_view(parent)) {
                         struct ggml_tensor * view_src = parent->view_src;
                         struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
                         if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
@@ -739,7 +734,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
         // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
         // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
         // itself is never used and should not be considered a dependency
-        if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
+        if (ggml_impl_is_view(node) && node->op != GGML_OP_NONE) {
             struct ggml_tensor * view_src = node->view_src;
             ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
         }
@@ -806,7 +801,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
                 parent->name, p_hn->n_children, p_hn->n_views, p_hn->allocated);
 
             if (p_hn->n_children == 0 && p_hn->n_views == 0) {
-                if (ggml_is_view(parent)) {
+                if (ggml_impl_is_view(parent)) {
                     struct ggml_tensor * view_src = parent->view_src;
                     struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
                     view_src_hn->n_views -= 1;

ggml/src/ggml-backend-reg.cpp

@@ -471,9 +471,10 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
 
     int best_score = 0;
     fs::path best_path;
+    std::error_code ec;
 
     for (const auto & search_path : search_paths) {
-        if (std::error_code ec; !fs::exists(search_path, ec)) {
+        if (!fs::exists(search_path, ec)) {
             if (ec) {
                 GGML_LOG_DEBUG("%s: posix_stat(%s) failure, error-message: %s\n", __func__, path_str(search_path).c_str(), ec.message().c_str());
             } else {
@@ -483,7 +484,7 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
         }
         fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
         for (const auto & entry : dir_it) {
-            if (entry.is_regular_file()) {
+            if (entry.is_regular_file(ec)) {
                 auto filename = entry.path().filename();
                 auto ext = entry.path().extension();
                 if (filename.native().find(file_prefix) == 0 && ext == file_extension) {