Merge branch 'master' into xsn/common_json

.devops/nix/package.nix

@@ -31,7 +31,7 @@
     ]
     && blas.meta.available,
   useCuda ? config.cudaSupport,
-  useMetalKit ? stdenv.hostPlatform.isAarch64 && stdenv.hostPlatform.isDarwin,
+  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin,
   # Increases the runtime closure size by ~700M
   useMpi ? false,
   useRocm ? config.rocmSupport,
@@ -92,7 +92,7 @@ let
 
   cudaBuildInputs = with cudaPackages; [
     cuda_cudart
-    cccl # <nv/target>
+    cuda_cccl # <nv/target>
     libcublas
   ];
 
@@ -166,7 +166,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
   # `xcrun` is used find the path of the Metal compiler, which is varible
   # and not on $PATH
   # see https://github.com/ggml-org/llama.cpp/pull/6118 for discussion
-  __noChroot = effectiveStdenv.hostPlatform.isDarwin && useMetalKit && precompileMetalShaders;
+  __noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders;
 
   nativeBuildInputs =
     [
@@ -181,10 +181,10 @@ effectiveStdenv.mkDerivation (finalAttrs: {
       autoAddDriverRunpath
     ]
     ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [ glibc.static ]
-    ++ optionals (effectiveStdenv.hostPlatform.isDarwin && useMetalKit && precompileMetalShaders) [ xcrunHost ];
+    ++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [ xcrunHost ];
 
   buildInputs =
-    optionals effectiveStdenv.hostPlatform.isDarwin darwinBuildInputs
+    optionals effectiveStdenv.isDarwin darwinBuildInputs
     ++ optionals useCuda cudaBuildInputs
     ++ optionals useMpi [ mpi ]
     ++ optionals useRocm rocmBuildInputs
@@ -245,7 +245,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
 
     # Configurations that are known to result in build failures. Can be
     # overridden by importing Nixpkgs with `allowBroken = true`.
-    broken = (useMetalKit && !effectiveStdenv.hostPlatform.isDarwin);
+    broken = (useMetalKit && !effectiveStdenv.isDarwin);
 
     description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
     homepage = "https://github.com/ggml-org/llama.cpp/";

.devops/openvino.Dockerfile

@@ -1,12 +1,12 @@
-ARG OPENVINO_VERSION_MAJOR=2026.3.1
-ARG OPENVINO_VERSION_FULL=2026.3.1.22476.56d9685302d
+ARG OPENVINO_VERSION_MAJOR=2026.3
+ARG OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c
 ARG UBUNTU_VERSION=24.04
 
 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
-ARG IGC_VERSION=v2.40.13
-ARG IGC_VERSION_FULL=2_2.40.13+22418
-ARG COMPUTE_RUNTIME_VERSION=26.31.39395.13
-ARG COMPUTE_RUNTIME_VERSION_FULL=26.31.39395.13-0
+ARG IGC_VERSION=v2.38.2
+ARG IGC_VERSION_FULL=2_2.38.2+22051
+ARG COMPUTE_RUNTIME_VERSION=26.27.39122.11
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.27.39122.11-0
 ARG IGDGMM_VERSION=22.10.0
 
 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
@@ -90,9 +90,6 @@ RUN bash -c "source ${OpenVINO_DIR}/setupvars.sh && \
     cmake -B build/ReleaseOV -G Ninja \
         -DCMAKE_BUILD_TYPE=Release \
         -DLLAMA_BUILD_TESTS=OFF \
-        -DGGML_NATIVE=OFF \
-        -DGGML_BACKEND_DL=ON \
-        -DGGML_CPU_ALL_VARIANTS=ON \
         -DGGML_OPENVINO=ON && \
     cmake --build build/ReleaseOV --parallel "
 

.github/ISSUE_TEMPLATE/config.yml

@@ -1,4 +1,4 @@
-blank_issues_enabled: false
+blank_issues_enabled: true
 contact_links:
   - name: Got an idea?
     url: https://github.com/ggml-org/llama.cpp/discussions/categories/ideas

.github/actions/ccache-buckets/action.yml

@@ -1,95 +0,0 @@
-name: "ccache-buckets"
-description: "Save/restore latest GitHub Actions ccache matching a key prefix to/from HF buckets"
-inputs:
-  key:
-    description: "Cache key prefix to match and load"
-    required: true
-  folder:
-    description: "Bucket folder containing ccache files"
-    required: true
-  evict-old-files:
-    description: "Corresponds to the ccache --evict-older-than AGE option, where AGE is the number of seconds or days followed by the 's' or 'd' suffix respectively."
-    default: ''
-  save:
-    description: "Save ccache"
-    required: false
-    default: false
-    type: boolean
-  hf_bucket:
-    description: 'Hugging Face buckets path'
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Install Hugging Face Hub CLI
-      shell: bash
-      run: |
-        python3 -m venv .venv-hf
-        .venv-hf/bin/pip install -U huggingface_hub==1.28.0
-
-    - name: Restore ccache from buckets
-      if: ${{ inputs.save != 'true' }}
-      shell: bash
-      run: |
-        set +e -uo pipefail
-        source .venv-hf/bin/activate
-        CCACHE_DIR=$(ccache -k cache_dir)
-        if [[ -d "$CCACHE_DIR" ]]; then
-            CACHE_PATH=$(hf buckets list "hf://buckets/${{ inputs.hf_bucket }}/${{ inputs.folder }}" --json | jq -r '[.[] | select(.type == "file") | select(.path | startswith("${{ inputs.folder }}/${{ inputs.key }}") and endswith(".tar.gz"))] | sort_by(.path) | last | .path // ""')
-            if [[ -n "$CACHE_PATH" ]]; then
-                echo "Restoring ccache from '$CACHE_PATH'."
-                hf buckets cp "hf://buckets/${{ inputs.hf_bucket }}/$CACHE_PATH" ccache_bucket.tar.gz
-                mkdir -p ccache_bucket
-                if tar -xzf ccache_bucket.tar.gz -C ccache_bucket; then
-                    rm -rf "$CCACHE_DIR"
-                    mv ccache_bucket "$CCACHE_DIR"
-                    ccache -z
-                fi
-                rm ccache_bucket.tar.gz
-            else
-                echo "No ccache found."
-            fi
-        else
-            echo "'$CCACHE_DIR' not found."
-        fi
-
-    - name: Save ccache to buckets
-      if: ${{ inputs.save == 'true' }}
-      shell: bash
-      run: |
-        if [[ -n "$HF_TOKEN" ]]; then
-            set +e -uo pipefail
-            source .venv-hf/bin/activate
-            CCACHE_DIR=$(ccache -k cache_dir)
-            if [[ -d "$CCACHE_DIR" ]]; then
-                ccache -s
-                if [[ -n "${{ inputs.evict-old-files }}" ]]; then
-                    ccache --evict-older-than "${{ inputs.evict-old-files }}"
-                fi
-                DATESTAMP=$(date -u +'%Y-%m-%dT%H:%M:%SZ')
-                CACHEFILE="${{ inputs.key }}-$DATESTAMP.tar.gz"
-                if tar -czf ccache_bucket.tar.gz -C "$CCACHE_DIR" .; then
-                    hf buckets cp ccache_bucket.tar.gz "hf://buckets/${{ inputs.hf_bucket }}/${{ inputs.folder }}/$CACHEFILE"
-                fi
-                rm ccache_bucket.tar.gz
-            else
-                echo "'$CCACHE_DIR' not found."
-            fi
-        fi
-
-    - name: Remove old ccache files from buckets
-      if: ${{ inputs.save == 'true' }}
-      shell: bash
-      run: |
-        if [[ -n "$HF_TOKEN" ]]; then
-            set +e -uo pipefail
-            source .venv-hf/bin/activate
-            CACHE_FILES=$(hf buckets list "hf://buckets/${{ inputs.hf_bucket }}/${{ inputs.folder }}" --json | jq -r '[.[] | select(.type == "file") | select((.uploaded_at | .[:19]+"Z" | fromdateiso8601) < (now - 5 * 60)) | select(.path | startswith("${{ inputs.folder }}/${{ inputs.key }}") and endswith(".tar.gz"))] | sort_by(.path)[:-1] | .[] | [.path // ""] | @tsv')
-            if [[ -n "$CACHE_FILES" ]]; then
-                echo "Removing old ccache files..."
-                while IFS=$'\t' read -r CACHE_PATH; do
-                    hf buckets rm "hf://buckets/${{ inputs.hf_bucket }}/$CACHE_PATH" -y
-                done <<< "$CACHE_FILES"
-            fi
-        fi

.github/actions/ccache-clear/action.yml

@@ -21,30 +21,68 @@ inputs:
 runs:
   using: "composite"
   steps:
-    - name: Install GitHub CLI if missing
-      shell: bash
-      run: |
-        # e.g. in container jobs, where it is not preinstalled
-        if ! command -v gh >/dev/null 2>&1; then
-          echo "GitHub CLI not found, installing..."
-          if ! command -v curl >/dev/null 2>&1; then
-            apt-get update >/dev/null 2>&1 || true
-            apt-get install -y curl >/dev/null 2>&1 || true
-          fi
-          mkdir -p -m 755 /etc/apt/keyrings
-          curl -fsSL https://cli.github.com/packages/githubcli-archive-keyring.gpg | tee /etc/apt/keyrings/githubcli-archive-keyring.gpg >/dev/null
-          chmod go+r /etc/apt/keyrings/githubcli-archive-keyring.gpg
-          echo "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/githubcli-archive-keyring.gpg] https://cli.github.com/packages stable main" > /etc/apt/sources.list.d/github-cli.list
-          apt-get update >/dev/null 2>&1 || true
-          apt-get install -y gh || { echo "Failed to install GitHub CLI (gh)" >&2; exit 1; }
-        fi
-        command -v gh >/dev/null 2>&1 || { echo "GitHub CLI (gh) is required but could not be installed" >&2; exit 1; }
-
     - name: Clear caches
       shell: bash
+      env:
+        CLEAR_KEY: ${{ inputs.key }}
+        CLEAR_OLDER: ${{ inputs.older }}
+        CLEAR_MIN: ${{ inputs.min }}
+        CLEAR_DRY_RUN: ${{ inputs.dry-run }}
       run: |
-        bash scripts/ccache-clear.sh \
-          --key "${{ inputs.key }}" \
-          --older "${{ inputs.older }}" \
-          --min "${{ inputs.min }}" \
-          ${{ inputs.dry-run == 'true' && '--dry-run' || '' }}
+        # Convert a duration (e.g. 90m, 1h, 1d, plain seconds) to seconds
+        to_seconds() {
+          local val="$1"
+          [[ "$val" =~ ^[0-9]+$ ]] && { echo "$val"; return 0; }
+          local num="${val%?}" unit="${val: -1}" mult
+          [[ "$num" =~ ^[0-9]+$ ]] || return 1
+          case "$unit" in
+            s) mult=1 ;;
+            m) mult=60 ;;
+            h) mult=3600 ;;
+            d) mult=86400 ;;
+            *) return 1 ;;
+          esac
+          echo $((num * mult))
+        }
+
+        [[ "$CLEAR_MIN" =~ ^[0-9]+$ ]] || { echo "Invalid min value: $CLEAR_MIN" >&2; exit 1; }
+        [[ "$CLEAR_DRY_RUN" =~ ^(true|false)$ ]] || { echo "Invalid dry-run value: $CLEAR_DRY_RUN" >&2; exit 1; }
+
+        CACHES=$(gh cache list --key "ccache-$CLEAR_KEY" --json id,key,createdAt --jq '.[] | [.createdAt, .id, .key] | @tsv' 2>/dev/null | LC_ALL=C sort)
+        if [ -z "$CACHES" ]; then
+          echo "No caches found with key prefix: $CLEAR_KEY"
+          exit 0
+        fi
+
+        TOTAL=$(( $(wc -l <<< "$CACHES") ))
+
+        echo "Found $TOTAL cache(s) with key prefix: $CLEAR_KEY (oldest first):"
+        while IFS=$'\t' read -r CREATED ID KEY; do
+          printf '  %s  %s  %s\n' "$CREATED" "$ID" "$KEY"
+        done <<< "$CACHES"
+
+        CUTOFF=""
+        if [ -n "$CLEAR_OLDER" ]; then
+          OLDER_SECONDS=$(to_seconds "$CLEAR_OLDER") || { echo "Invalid older value: $CLEAR_OLDER (expected e.g. 90m, 1h, 1d)" >&2; exit 1; }
+          CUTOFF=$(( $(date +%s) - OLDER_SECONDS ))
+        fi
+
+        # Caches are sorted oldest first
+        DELETED=0
+        while IFS=$'\t' read -r CREATED ID KEY; do
+          if [ -n "$CUTOFF" ] && [ "$(date -d "$CREATED" +%s)" -ge "$CUTOFF" ]; then
+            echo "Rest are not older than $CLEAR_OLDER, stopping"
+            break
+          fi
+          if [ $((TOTAL - DELETED - 1)) -lt "$CLEAR_MIN" ]; then
+            echo "Keeping at least $CLEAR_MIN cache(s), stopping"
+            break
+          fi
+          if [ "$CLEAR_DRY_RUN" = "true" ]; then
+            echo "Would delete cache: $ID ($KEY)"
+          else
+            echo "Deleting cache: $ID ($KEY)"
+            gh cache delete "$ID"
+          fi
+          DELETED=$((DELETED + 1))
+        done <<< "$CACHES"

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

@@ -24,7 +24,7 @@ runs:
 
         write-host "Installing ROCm wheels for multi-arch support"
         # Install ROCm wheels for multi-arch support (this may take several minutes)
-        python -m pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "rocm[libraries,devel]==${{ inputs.version }}"
+        python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ inputs.version }}"
 
         # Pre-expand the devel tree so it is included in the cache
         write-host "Initializing ROCm devel tree"

.github/workflows/build-android.yml

@@ -110,7 +110,7 @@ jobs:
       #   cache on:  https://github.com/ggerganov/tmp2/actions/runs/26534713799/job/78224189394
       #
       #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.24
+      #  uses: ggml-org/ccache-action@v1.2.21
       #  with:
       #    key: android-ubuntu-arm64
       #    evict-old-files: 1d

.github/workflows/build-apple.yml

@@ -22,8 +22,7 @@ on:
     types: [opened, synchronize, reopened]
     paths: [
       '.github/workflows/build-apple.yml',
-      'ggml/src/ggml-metal/**',
-      'ggml/src/ggml-rpc/**'
+      'ggml/src/ggml-metal/**'
     ]
 
 concurrency:
@@ -47,19 +46,11 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: apple-arm64
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: apple-arm64
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build
         id: cmake_build
@@ -74,25 +65,7 @@ jobs:
             -DGGML_RPC=ON \
             -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
           time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
-
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: apple-arm64
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
-      - name: Check for leaks
-        run: |
-          cmd=(./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1)
-          leaks -atExit -- "${cmd[@]}"
-          # Graphics devices are leaked by Metal in Apple code sometimes, so we ignore those leaks
-          OBJC_DEBUG_MISSING_POOLS=YES "${cmd[@]}" 2>&1 | awk '{ print } index($0, "autoreleased with no pool in place") && !/class [a-zA-Z0-9]+Device autoreleased/ { found = 1 } END { exit found }'
+          leaks -atExit -- ./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1
 
       - name: Test
         id: cmake_test
@@ -109,19 +82,11 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: apple-x64
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: apple-x64
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build
         id: cmake_build
@@ -138,18 +103,6 @@ jobs:
             -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
           time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: apple-x64
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
       - name: Test
         id: cmake_test
         run: |
@@ -210,6 +163,14 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      # TODO: this likely does not do anything - if yes, remove it
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: apple-tvos
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
       - name: Build
         id: cmake_build
         run: |
@@ -235,6 +196,14 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      # TODO: this likely does not do anything - if yes, remove it
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: apple-visionos
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
       - name: Build
         id: cmake_build
         run: |
@@ -265,6 +234,14 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      # TODO: this likely does not do anything - if yes, remove it
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: apple-swift
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
       - name: Download xcframework artifact
         uses: actions/download-artifact@v7
         with:

.github/workflows/build-cache.yml

@@ -41,8 +41,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone
@@ -69,8 +69,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone

.github/workflows/build-cpu.yml

@@ -62,10 +62,11 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: cpu-${{ matrix.os }}
-          save: false
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build Dependencies
         id: build_depends
@@ -90,15 +91,6 @@ jobs:
           python3 -m pip install --upgrade pip setuptools
           pip3 install ./gguf-py
 
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: cpu-${{ matrix.os }}
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
-
       - name: Build
         id: cmake_build
         run: |
@@ -108,18 +100,6 @@ jobs:
             -DGGML_RPC=ON
           time cmake --build build --config Release -j $(nproc)
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: cpu-${{ matrix.os }}
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
       - name: Test
         id: cmake_test
         run: |
@@ -137,6 +117,18 @@ jobs:
           ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
           ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
+      # note: real deletion only on push to master (same condition as the ccache save),
+      #       dry-run otherwise (the token is read-only on PRs from forks)
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: cpu-${{ matrix.os }}
+          older: 1h
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+
   windows:
     name: windows / ${{ matrix.build }}
     runs-on: windows-2025
@@ -164,7 +156,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: cpu-windows-2025-${{ matrix.build }}
           variant: ccache
@@ -223,13 +215,3 @@ jobs:
       #     cd build
       #     $env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1
       #     & $sde -future -- ctest -L main -C Release --verbose --timeout 900
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: cpu-windows-2025-${{ matrix.build }}
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-cuda-ubuntu.yml

@@ -50,22 +50,14 @@ jobs:
           DEBIAN_FRONTEND: noninteractive
         run: |
           apt update
-          apt install -y cmake build-essential ninja-build libgomp1 git libssl-dev jq python3 python3-venv python3-pip
+          apt install -y cmake build-essential ninja-build libgomp1 git libssl-dev
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: cuda-ubuntu-24.04-cuda
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: cuda-ubuntu-24.04-cuda
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build with CMake
         # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
@@ -80,18 +72,6 @@ jobs:
             -DGGML_CUDA_CUB_3DOT2=ON
           cmake --build build
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: cuda-ubuntu-24.04-cuda
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
   hip:
     runs-on: ubuntu-22.04
     container: rocm/dev-ubuntu-22.04:6.1.2
@@ -105,22 +85,14 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev rocwmma-dev jq python3-venv
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev rocwmma-dev
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: cuda-ubuntu-22.04-hip
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: cuda-ubuntu-22.04-hip
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build with native CMake HIP support
         id: cmake_build
@@ -131,18 +103,6 @@ jobs:
             -DGGML_HIP=ON
           cmake --build build --config Release -j $(nproc)
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: cuda-ubuntu-22.04-hip
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
   musa:
     runs-on: ubuntu-22.04
     container: mthreads/musa:rc4.3.0-devel-ubuntu22.04-amd64
@@ -156,22 +116,14 @@ jobs:
         id: depends
         run: |
           apt-get update
-          apt-get install -y build-essential git cmake libssl-dev jq
+          apt-get install -y build-essential git cmake libssl-dev
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: cuda-ubuntu-22.04-musa
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: cuda-ubuntu-22.04-musa
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build with native CMake MUSA support
         id: cmake_build
@@ -179,15 +131,3 @@ jobs:
           cmake -B build -S . \
             -DGGML_MUSA=ON
           time cmake --build build --config Release -j $(nproc)
-
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: cuda-ubuntu-22.04-musa
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true

.github/workflows/build-cuda-windows.yml

@@ -47,7 +47,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
 
@@ -152,7 +152,7 @@ jobs:
           & "${env:HIP_PATH}\lib\llvm\bin\clang.exe" --version
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           # TODO: this build does not match the build in release.yml, so we use a different cache key
           #       ideally, the builds should match, similar to the CUDA build above so that we would be able

.github/workflows/build-msys.yml

@@ -35,7 +35,7 @@ jobs:
         uses: actions/checkout@v6
 
       #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.24
+      #  uses: ggml-org/ccache-action@v1.2.16
       #  with:
       #    key: msys-windows-2025-x64
       #    variant: ccache

.github/workflows/build-opencl.yml

@@ -44,7 +44,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: opencl-windows-2025-x64
           variant: ccache
@@ -80,13 +80,3 @@ jobs:
         run: |
           cmake -S . -B build -G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON -DLLAMA_BUILD_BORINGSSL=ON
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: opencl-windows-2025-x64
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-openvino.yml

@@ -32,8 +32,6 @@ env:
   LLAMA_ARG_LOG_COLORS: 1
   LLAMA_ARG_LOG_PREFIX: 1
   LLAMA_ARG_LOG_TIMESTAMPS: 1
-  # TODO: fix failing tests on OpenVINO backend
-  CTEST_EXCLUDE: "test-llama-archs|^test-recurrent-state-|test-backend-ops|test-save-load-state"
 
 jobs:
   ubuntu-24-openvino:
@@ -41,8 +39,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone
@@ -80,24 +78,26 @@ jobs:
 
       - name: Test (CPU)
         id: cmake_test_cpu
+        # TODO: fix and re-enable the `test-llama-archs` test below
         run: |
           cd ${{ github.workspace }}
-          ctest --test-dir build/ReleaseOV -L main -E "${{ env.CTEST_EXCLUDE }}" --verbose --timeout 3000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 2000
 
       - name: Test (GPU)
         id: cmake_test_gpu
+        # TODO: fix and re-enable the `test-llama-archs` test below
         run: |
           cd ${{ github.workspace }}
           export GGML_OPENVINO_DEVICE=GPU
-          ctest --test-dir build/ReleaseOV -L main -E "${{ env.CTEST_EXCLUDE }}" --verbose --timeout 3000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 3000
 
   openvino-windows-2022:
     runs-on: windows-2022
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone
@@ -105,7 +105,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: openvino-windows-2022
           variant: ccache
@@ -159,20 +159,11 @@ jobs:
       - name: Test (CPU)
         id: cmake_test_cpu
         shell: cmd
+        # TODO: fix and re-enable the `test-llama-archs` test below
         run: |
           REM Find extracted OpenVINO folder dynamically
           for /d %%i in (openvino_toolkit\*) do set OPENVINO_ROOT=%%i
           call "%OPENVINO_ROOT%\setupvars.bat"
 
           cd build
-          ctest --test-dir ReleaseOV -L main -E "${{ env.CTEST_EXCLUDE }}" -C Release --verbose --timeout 3000
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: openvino-windows-2022
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+          ctest --test-dir ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" -C Release --verbose --timeout 3000

.github/workflows/build-riscv.yml

@@ -67,7 +67,7 @@ jobs:
 
       # note: sparing some ccache since these jobs run on dedicated runners that are not part of the organitzation
       #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.24
+      #  uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
       #  with:
       #    key: riscv-ubuntu-native
       #    evict-old-files: 1d
@@ -137,7 +137,7 @@ jobs:
 
       # note: sparing some ccache since these jobs run on dedicated runners that are not part of the organitzation
       #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.24
+      #  uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
       #  with:
       #    key: riscv-ubuntu-native-sanitizer-${{ matrix.sanitizer }}-${{ matrix.build_type }}
       #    evict-old-files: 1d

.github/workflows/build-sanitize.yml

@@ -55,7 +55,7 @@ jobs:
         uses: actions/checkout@v6
 
       # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.24
+      #   uses: ggml-org/ccache-action@v1.2.21
       #   if: ${{ matrix.sanitizer != 'UNDEFINED' }}
       #   with:
       #     key: ctest-${{ matrix.sanitizer }}-ubuntu-24.04

.github/workflows/build-self-hosted.yml

@@ -288,8 +288,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone

.github/workflows/build-sycl.yml

@@ -75,19 +75,11 @@ jobs:
           sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: sycl-ubuntu-24-${{ matrix.build }}
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: sycl-ubuntu-24-${{ matrix.build }}
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build
         id: cmake_build
@@ -104,18 +96,6 @@ jobs:
             -DGGML_SYCL_F16=${{ matrix.fp16 }}
           time cmake --build build --config Release -j $(nproc)
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: sycl-ubuntu-24-${{ matrix.build }}
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
   windows-latest-sycl:
     runs-on: windows-2022
 
@@ -147,7 +127,7 @@ jobs:
           "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: sycl-windows-latest
           variant: ccache
@@ -159,13 +139,3 @@ jobs:
       - name: Build
         id: cmake_build
         run:  examples/sycl/win-build-sycl.bat
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: sycl-windows-latest
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-vulkan.yml

@@ -53,20 +53,12 @@ jobs:
           echo "CXX=g++-14" >> "$GITHUB_ENV"
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: vulkan-ubuntu-24.04-arm
+          key: vulkan-ubuntu-24.04-arm-new
           variant: ccache
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: vulkan-ubuntu-24.04-arm
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Configure
         id: cmake_configure
@@ -81,18 +73,6 @@ jobs:
         run: |
           time cmake --build build -j $(nproc)
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: vulkan-ubuntu-24.04-arm
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
   ubuntu-llvmpipe:
     runs-on: ubuntu-24.04
 
@@ -122,19 +102,11 @@ jobs:
           strip: 1
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: vulkan-ubuntu-24.04-llvmpipe
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: vulkan-ubuntu-24.04-llvmpipe
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build
         id: cmake_build
@@ -145,18 +117,6 @@ jobs:
             -DGGML_VULKAN=ON
           cmake --build build --config Release -j $(nproc)
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: vulkan-ubuntu-24.04-llvmpipe
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
       - name: Test
         id: cmake_test
         run: |
@@ -180,7 +140,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: cpu-windows-2025-x64-vulkan
           variant: ccache
@@ -220,13 +180,3 @@ jobs:
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: cpu-windows-2025-x64-vulkan
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-wasm.yml

@@ -54,10 +54,11 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: webgpu-ubuntu-24.04-arm-wasm
-          save: false
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Install Emscripten
         run: |
@@ -75,15 +76,6 @@ jobs:
             "https://github.com/google/dawn/releases/download/${DAWN_TAG}/${EMDAWN_PKG}"
           unzip emdawn.zip
 
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: webgpu-ubuntu-24.04-arm-wasm
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
-
       - name: Build WASM WebGPU
         run: |
           source emsdk/emsdk_env.sh
@@ -96,15 +88,3 @@ jobs:
             -DEMDAWNWEBGPU_DIR=emdawnwebgpu_pkg
 
           time cmake --build build-wasm --config Release --target test-backend-ops -j $(nproc)
-
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: webgpu-ubuntu-24.04-arm-wasm
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true

.github/workflows/build-webgpu.yml

@@ -69,10 +69,11 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: webgpu-macos-latest
-          save: false
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Dawn Dependency
         id: dawn-depends
@@ -87,15 +88,6 @@ jobs:
           mkdir dawn
           tar -xvf artifact.tar.gz -C dawn --strip-components=1
 
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: webgpu-macos-latest
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
-
       - name: Build
         id: cmake_build
         run: |
@@ -103,18 +95,6 @@ jobs:
           cmake -B build -G "Ninja" -DCMAKE_BUILD_TYPE=Release -DGGML_WEBGPU=ON -DGGML_METAL=OFF -DGGML_BLAS=OFF
           time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: webgpu-macos-latest
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
       - name: Test
         id: cmake_test
         run: |
@@ -130,10 +110,11 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: webgpu-ubuntu-24.04
-          save: false
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Dependencies
         id: depends
@@ -157,15 +138,6 @@ jobs:
           mkdir dawn
           tar -xvf artifact.tar.gz -C dawn --strip-components=1
 
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: webgpu-ubuntu-24.04
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
-
       - name: Build
         id: cmake_build
         run: |
@@ -174,18 +146,6 @@ jobs:
             -DGGML_WEBGPU=ON
           time cmake --build build --config Release -j $(nproc)
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: webgpu-ubuntu-24.04
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
       - name: Test
         id: cmake_test
         run: |

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

@@ -29,7 +29,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: copilot-setup-steps
           evict-old-files: 1d

.github/workflows/docker.yml

@@ -64,7 +64,7 @@ jobs:
     needs: create_tag
     uses: ./.github/workflows/ui-build.yml
     with:
-      ui_version: ${{ needs.create_tag.outputs.source_tag }}
+      hf_ui_version: ${{ needs.create_tag.outputs.source_tag }}
 
   prepare_matrices:
     name: Prepare Docker matrices
@@ -162,7 +162,7 @@ jobs:
         if: ${{ matrix.config.prebuilt_ui == true }}
         uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist
 
       - name: Set up QEMU

.github/workflows/hip-quality-check.yml

@@ -49,22 +49,14 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev python3 python3-venv python3-pip jq
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev python3
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: hip-quality-check-ubuntu-22.04
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: hip-quality-check-ubuntu-22.04
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build with Werror
         id: cmake_build
@@ -92,15 +84,3 @@ jobs:
           cd build
           make -j $(nproc) 2>&1 | tee metrics.log | grep -v 'Rpass-analysis=kernel-resource-usage\|remark:\|^$'
           python3 ../scripts/hip/gcn-cdna-vgpr-check.py metrics.log
-
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: hip-quality-check-ubuntu-22.04
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true

.github/workflows/make-release.yml

@@ -19,7 +19,6 @@ env:
 
 permissions:
   contents: write
-  packages: write
 
 jobs:
   make-release:
@@ -85,13 +84,11 @@ jobs:
 
             New version has been released.
 
-            ## Assets
-
             ${{ steps.desc.outputs.nightly }}
 
-            ## More info
+            **Web UI:** the `nightly-tag.txt` asset contains the tag of the corresponding nightly release
 
-            - [Releases and versioning of `ggml-org` projects](https://github.com/ggml-org/ggml/discussions/1579)
+            **More info:** [dist : releases and versioning of ggml-org projects](https://github.com/ggml-org/ggml/discussions/1579)
 
             ## ${{ steps.desc.outputs.changelog_title }}
 
@@ -114,29 +111,6 @@ jobs:
               data: await fs.readFileSync('./nightly-tag.txt')
             });
 
-      - name: Re-tag container images with release version
-        if: ${{ github.event.inputs.dry_run == 'false' && steps.desc.outputs.nightly_tag != '' }}
-        env:
-          GITHUB_REPOSITORY_OWNER: ${{ github.repository_owner }}
-        run: |
-          VERSION="${{ steps.checks.outputs.version }}"
-          NIGHTLY_TAG="${{ steps.desc.outputs.nightly_tag }}"
-          REPO_OWNER="${GITHUB_REPOSITORY_OWNER,,}"
-          IMAGE_REPO="ghcr.io/${REPO_OWNER}/${{ github.event.repository.name }}"
-
-          echo "${{ secrets.GITHUB_TOKEN }}" | docker login ghcr.io -u "${{ github.actor }}" --password-stdin
-
-          VARIANTS=("" "-cuda" "-cuda13" "-vulkan" "-rocm" "-intel" "-musa" "-openvino")
-          TYPES=("full" "light" "server")
-          for type in "${TYPES[@]}"; do
-            for variant in "${VARIANTS[@]}"; do
-              src="${IMAGE_REPO}:${type}${variant}-${NIGHTLY_TAG}"
-              dst="${IMAGE_REPO}:${type}${variant}-${VERSION}"
-              echo "Tagging ${src} -> ${dst}"
-              docker buildx imagetools create --tag "${dst}" "${src}"
-            done
-          done
-
       - name: Dry run summary
         if: ${{ github.event.inputs.dry_run == 'true' }}
         run: |

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

@@ -31,7 +31,7 @@ jobs:
         uses: actions/setup-python@v6
         with:
           python-version: "3.11"
-          pip-install: -r requirements/requirements-all.txt ty==0.0.78
+          pip-install: -r requirements/requirements-all.txt ty==0.0.35
       # - name: Type-check with Pyright
       #   uses: jakebailey/pyright-action@v2
       #   with:
@@ -40,4 +40,4 @@ jobs:
       #     warnings: true
       - name: Type-check with ty
         run: |
-            ty check --exit-zero-on-warning --output-format=github
+            ty check --output-format=github

.github/workflows/release.yml

@@ -61,8 +61,31 @@ jobs:
             echo "should_release=false" >> $GITHUB_OUTPUT
           fi
 
+  get-version:
+    runs-on: ubuntu-slim
+    outputs:
+      ui_version: ${{ steps.version.outputs.ui_version }}
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+      - id: version
+        run: |
+          # Resolve UI version: BUILD_NUMBER from cmake/build-info.cmake > git hash + epoch > fallback
+          version=""
+          if grep -q "BUILD_NUMBER" cmake/build-info.cmake; then
+            build_number=$(grep "set(BUILD_NUMBER" cmake/build-info.cmake | grep -oP '\d+')
+            if [ -n "$build_number" ] && [ "$build_number" -gt 0 ]; then
+              version="b${build_number}"
+            fi
+          fi
+          if [ -z "$version" ]; then
+            version=$(git rev-parse --short HEAD)-$(date +%s)
+          fi
+          echo "ui_version=${version}" >> $GITHUB_OUTPUT
+
   macos-cpu:
-    needs: [check-release, ui-build]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
     strategy:
       matrix:
@@ -96,14 +119,15 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-${{ matrix.os }}-${{ matrix.arch }}
 
@@ -117,6 +141,7 @@ jobs:
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_BUILD_BORINGSSL=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
@@ -142,7 +167,7 @@ jobs:
           key: release-${{ matrix.os }}-${{ matrix.arch }}
 
   ubuntu-cpu:
-    needs: [check-release, ui-build]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
     strategy:
       matrix:
@@ -166,11 +191,12 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Dependencies
         id: depends
@@ -187,7 +213,7 @@ jobs:
 
       - name: ccache
         if: ${{ matrix.build != 's390x' }}
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-${{ matrix.os }}-cpu
 
@@ -201,6 +227,7 @@ jobs:
             -DGGML_NATIVE=OFF \
             -DGGML_CPU_ALL_VARIANTS=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -227,7 +254,7 @@ jobs:
           key: release-${{ matrix.os }}-cpu
 
   ubuntu-vulkan:
-    needs: [check-release, ui-build]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     strategy:
@@ -250,11 +277,12 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Dependencies
         id: depends
@@ -272,7 +300,7 @@ jobs:
           fi
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-${{ matrix.os }}-vulkan
 
@@ -286,6 +314,7 @@ jobs:
             -DGGML_NATIVE=OFF \
             -DGGML_CPU_ALL_VARIANTS=ON \
             -DGGML_VULKAN=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -311,7 +340,7 @@ jobs:
           key: release-${{ matrix.os }}-vulkan
 
   android-arm64:
-    needs: [check-release, ui-build]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: ubuntu-latest
@@ -329,11 +358,12 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Set up JDK
         uses: actions/setup-java@v5
@@ -358,7 +388,7 @@ jobs:
       #   cache on:  https://github.com/ggerganov/tmp2/actions/runs/26534713799/job/78224189394
       #
       #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.24
+      #  uses: ggml-org/ccache-action@v1.2.21
       #  with:
       #    key: release-android-arm64
 
@@ -377,6 +407,7 @@ jobs:
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_OPENMP=OFF \
             -DLLAMA_BUILD_BORINGSSL=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -402,7 +433,7 @@ jobs:
           name: llama-bin-android-arm64.tar.gz
 
   ubuntu-24-openvino:
-    needs: [check-release, ui-build]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: ubuntu-24.04
@@ -415,8 +446,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Set OpenVINO version output
@@ -429,14 +460,15 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-ubuntu-24.04-openvino-release-no-preset-v1
 
@@ -476,6 +508,7 @@ jobs:
             -DGGML_OPENVINO=ON \
             -DCMAKE_INSTALL_RPATH='$ORIGIN' \
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build/ReleaseOV --config Release --parallel
 
@@ -519,7 +552,7 @@ jobs:
           key: release-ubuntu-24.04-openvino-release-no-preset-v1
 
   windows-openvino:
-    needs: [check-release, ui-build]
+    needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: windows-2022
@@ -529,8 +562,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Set OpenVINO version output
@@ -544,14 +577,15 @@ jobs:
         with:
             fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-windows-2022-openvino
           variant: ccache
@@ -648,7 +682,7 @@ jobs:
 
   windows-cpu:
     name: windows-cpu / ${{ matrix.arch }}
-    needs: [check-release, ui-build]
+    needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: windows-2025-vs2026
@@ -668,18 +702,19 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Install Ninja
         run: |
           choco install ninja
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
 
@@ -714,8 +749,6 @@ jobs:
         with:
           key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
 
-  # note: builds only the ggml-hip backend - llama-server is injected from the
-  #       windows-cpu zip during the release "Merge artifacts" step
   windows-rocm:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -725,7 +758,7 @@ jobs:
     strategy:
       matrix:
         include:
-          - ROCM_VERSION: "10.0.0"
+          - ROCM_VERSION: "7.14.0"
             gpu_targets: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
             build: x64
 
@@ -736,12 +769,8 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Install Ninja
-        run: |
-          choco install ninja
-
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
           evict-old-files: 1d
@@ -793,28 +822,33 @@ jobs:
 
       - name: Build
         run: |
-          cmake -S . -B build `
-            -G "Ninja Multi-Config" `
+          mkdir build
+          cd build
+          cmake .. `
+            -G "Unix Makefiles" `
             -DCMAKE_PREFIX_PATH="${env:HIP_PATH}" `
+            -DCMAKE_BUILD_TYPE=Release `
             -DGGML_BACKEND_DL=ON `
             -DGGML_NATIVE=OFF `
-            -DGGML_CPU=OFF `
+            -DGGML_CPU=ON `
+            -DGGML_CPU_ALL_VARIANTS=ON `
             -DGGML_HIP=ON `
             -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
             -DCMAKE_C_FLAGS="-Wno-error=incompatible-pointer-types" `
             -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
             -DHIP_PATH="${env:HIP_PATH}" `
+            -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}"
-          cmake --build build --config Release --parallel ${env:NUMBER_OF_PROCESSORS} --target ggml-hip
+          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}
 
       - name: Verify HIP backend was built
         run: |
-          $hipDll = Get-ChildItem -Path build\bin\Release -Filter "ggml-hip*.dll" -ErrorAction SilentlyContinue
+          $hipDll = Get-ChildItem -Path build\bin -Filter "ggml-hip*.dll" -ErrorAction SilentlyContinue
           if (-not $hipDll) {
             Write-Host "##[error]ggml-hip*.dll was NOT produced. The HIP backend silently failed to build."
-            Write-Host "Contents of build\bin\Release:"
-            Get-ChildItem build\bin\Release | Format-Table -AutoSize
+            Write-Host "Contents of build\bin:"
+            Get-ChildItem build\bin | Format-Table -AutoSize
             exit 1
           }
           Write-Host "HIP backend artifact found:"
@@ -829,40 +863,10 @@ jobs:
           $rocmVersionShort = ('${{ matrix.ROCM_VERSION }}'.Split('.')[0..1] -join '.')
           echo "ROCM_VERSION_SHORT=$rocmVersionShort" >> $env:GITHUB_ENV
 
-      - name: Bundle HIP runtime DLLs (amdhip64_7.dll, rocm_kpack.dll, amd_comgr.dll)
-        run: |
-          $ErrorActionPreference = "Stop"
-          # See issue https://github.com/ggml-org/llama.cpp/issues/26929.
-          # ggml-hip.dll loads amdhip64_7.dll at run time. The Adrenalin driver
-          # ships an amdhip64_7.dll in System32, which the loader searches before PATH,
-          # so a matching DLL from PATH cannot win. Copy amdhip64 next to the
-          # binaries (exe directory is searched before System32) so the correct
-          # runtime is used. rocm_kpack.dll is amdhip64_7's direct dependency, so
-          # copy the matching version too. amd_comgr is copied as well to keep it
-          # in sync with the bundled amdhip64, avoiding a version mismatch with a
-          # amd_comgr from System32.
-          # rocblas/hipblaslt kernels resolve fine via PATH and are not copied.
-          $binPath = (rocm-sdk path --bin).Trim()
-          if (-not $binPath) { throw "rocm-sdk path --bin returned empty" }
-          write-host "ROCm bin path: $binPath"
-
-          $patterns = @("amdhip64_7.dll", "rocm_kpack.dll", "amd_comgr.dll")
-          foreach ($pattern in $patterns) {
-            $files = Get-ChildItem -Path $binPath -Filter $pattern -ErrorAction SilentlyContinue
-            if (-not $files) { throw "no match for $pattern in $binPath" }
-            foreach ($f in $files) {
-              Copy-Item $f.FullName -Destination build\bin\Release -Force
-              write-host "  copied $($f.Name)"
-            }
-          }
-
       - name: Pack artifacts
         run: |
-          7z a -snl llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip `
-            .\build\bin\Release\ggml-hip.dll `
-            .\build\bin\Release\amdhip64_7.dll `
-            .\build\bin\Release\rocm_kpack.dll `
-            .\build\bin\Release\amd_comgr.dll
+          cp "LICENSE" "build\bin\"
+          7z a -snl llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip .\build\bin\*
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v6
@@ -875,8 +879,6 @@ jobs:
         with:
           key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
 
-  # note: builds only the backend library - llama-server (with the embedded UI)
-  #       is injected from the windows-cpu zip during the release "Merge artifacts" step
   windows:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -907,6 +909,13 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
       - name: Install Vulkan SDK
         id: get_vulkan
         if: ${{ matrix.backend == 'vulkan' }}
@@ -923,7 +932,7 @@ jobs:
 
       # TODO: these jobs need to use llvm toolchain in order to utilize the ccache
       #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.24
+      #  uses: ggml-org/ccache-action@v1.2.21
       #  with:
       #    key: release-windows-2025-${{ matrix.arch }}-${{ matrix.backend }}
 
@@ -969,8 +978,6 @@ jobs:
           path: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
           name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
 
-  # note: builds only the ggml-cuda backend - llama-server is injected from the
-  #       windows-cpu zip during the release "Merge artifacts" step
   windows-cuda:
     name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
     needs: [check-release]
@@ -999,6 +1006,13 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
       - name: Install Cuda Toolkit
         uses: ./.github/actions/windows-setup-cuda
         with:
@@ -1011,7 +1025,7 @@ jobs:
           choco install ninja
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
 
@@ -1070,8 +1084,6 @@ jobs:
         with:
           key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
 
-  # note: builds only the ggml-sycl backend - llama-server is injected from the
-  #       windows-cpu zip during the release "Merge artifacts" step
   windows-sycl:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1106,8 +1118,15 @@ jobs:
           Expand-Archive -Path "level-zero-win-sdk.zip" -DestinationPath "C:/level-zero-sdk" -Force
           "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append
 
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-windows-2022-x64-sycl
 
@@ -1176,7 +1195,7 @@ jobs:
           key: release-windows-2022-x64-sycl
 
   ubuntu-24-sycl:
-    needs: [check-release, ui-build]
+    needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     strategy:
@@ -1218,14 +1237,15 @@ jobs:
           wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
           sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-ubuntu-24.04-sycl-${{ matrix.build }}
 
@@ -1267,11 +1287,11 @@ jobs:
         with:
           key: release-ubuntu-24.04-sycl-${{ matrix.build }}
 
-  ubuntu-24-rocm:
-    needs: [check-release, ui-build]
+  ubuntu-22-rocm:
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-22.04
 
     permissions:
       actions: write
@@ -1279,7 +1299,7 @@ jobs:
     strategy:
       matrix:
         include:
-          - ROCM_VERSION: "10.0.0"
+          - ROCM_VERSION: "7.14.0"
             gpu_targets: "gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
             build: 'x64'
 
@@ -1290,11 +1310,12 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Free up disk space
         uses: ggml-org/free-disk-space@v1.3.1
@@ -1302,9 +1323,9 @@ jobs:
           tool-cache: true
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: release-ubuntu-24.04-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          key: release-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
           evict-old-files: 1d
           max-size: "1G"
 
@@ -1333,7 +1354,7 @@ jobs:
           # libraries = HIP runtime and CMake configs needed for linking
           # devel = compilers, headers, static libs
           python -m pip install --upgrade pip
-          python -m pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "rocm[libraries,devel]==${{ matrix.ROCM_VERSION }}"
+          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ matrix.ROCM_VERSION }}"
 
           # Get ROCm installation paths using the rocm-sdk CLI tool
           ROCM_PATH=$(rocm-sdk path --root)
@@ -1367,6 +1388,7 @@ jobs:
             -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
             -DGGML_HIP=ON \
             -DHIP_PLATFORM=amd \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -1392,10 +1414,10 @@ jobs:
       - name: ccache-clear
         uses: ./.github/actions/ccache-clear
         with:
-          key: release-ubuntu-24.04-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          key: release-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
 
   ios-xcode:
-    needs: [check-release]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
     runs-on: macos-26
 
@@ -1423,7 +1445,8 @@ jobs:
             -DLLAMA_BUILD_SERVER=OFF \
             -DCMAKE_SYSTEM_NAME=iOS \
             -DCMAKE_OSX_DEPLOYMENT_TARGET=16.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }}
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
 
       - name: xcodebuild for swift package
@@ -1546,9 +1569,11 @@ jobs:
 #          name: llama-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz
 
   ui-build:
-    needs: [check-release]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
     uses: ./.github/workflows/ui-build.yml
+    with:
+      hf_ui_version: ${{ needs.get-version.outputs.ui_version }}
 
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -1563,13 +1588,14 @@ jobs:
     runs-on: ubuntu-slim
 
     needs:
+      - get-version
       - windows
       - windows-cpu
       - windows-cuda
       - windows-sycl
       - windows-rocm
       - windows-openvino
-      - ubuntu-24-rocm
+      - ubuntu-22-rocm
       - ubuntu-cpu
       - ubuntu-vulkan
       - ubuntu-24-openvino
@@ -1602,27 +1628,24 @@ jobs:
           path: ./artifact
           merge-multiple: true
 
-      - name: Merge artifacts
+      - name: Move artifacts
         id: move_artifacts
         run: |
           mkdir -p release
 
-          # the windows-cpu zip contains the full toolset (llama-server with the embedded
-          # UI, ggml-cpu) - inject it into the other windows zips so that every archive
-          # ships the same binaries, only with a different backend library on top
-          echo "Injecting windows-cpu binaries (llama-server + CPU backend) into the backend zips..."
+          echo "Adding CPU backend files to existing zips..."
           for arch in x64 arm64; do
             cpu_zip="artifact/llama-bin-win-cpu-${arch}.zip"
             temp_dir=$(mktemp -d)
-            echo "Extracting windows-cpu-${arch} package..."
+            echo "Extracting CPU backend for $arch..."
             unzip "$cpu_zip" -d "$temp_dir"
 
-            echo "Merging into $arch zips..."
+            echo "Adding CPU files to $arch zips..."
             for target_zip in artifact/llama-bin-win-*-${arch}.zip; do
               if [[ "$target_zip" == "$cpu_zip" ]]; then
                 continue
               fi
-              echo "Injecting into $(basename "$target_zip")"
+              echo "Adding CPU backend to $(basename "$target_zip")"
               realpath_target_zip=$(realpath "$target_zip")
               (cd "$temp_dir" && zip -r "$realpath_target_zip" .)
             done
@@ -1646,7 +1669,7 @@ jobs:
         id: download_ui
         uses: actions/download-artifact@v7
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: ./ui-dist
 
       - name: Package UI
@@ -1703,7 +1726,7 @@ jobs:
             - [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)
             - [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 arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-arm64.tar.gz)
-            - [Ubuntu x64 (ROCm 10.0)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-10.0-x64.tar.gz)
+            - [Ubuntu x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.14-x64.tar.gz)
             - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
             - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp32-x64.tar.gz)
             - [Ubuntu x64 (SYCL FP16)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp16-x64.tar.gz)
@@ -1721,7 +1744,7 @@ jobs:
             - [Windows x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-vulkan-x64.zip)
             - [Windows x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ needs.windows-openvino.outputs.openvino_version }}-x64.zip)
             - [Windows x64 (SYCL)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip)
-            - [Windows x64 (ROCm 10.0)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-rocm-10.0-x64.zip)
+            - [Windows x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-rocm-7.14-x64.zip)
 
             **openEuler:**
             - [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23705)

.github/workflows/server-sanitize.yml

@@ -73,6 +73,13 @@ jobs:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
       - name: Build
         id: cmake_build
         run: |
@@ -103,7 +110,7 @@ jobs:
           source .venv/bin/activate
           cd tools/server/tests
           export ${{ matrix.extra_args }}
-          PYTEST_WORKERS=1 ./tests.sh
+          ./tests.sh
 
       - name: Slow tests
         id: server_integration_tests_slow
@@ -112,4 +119,4 @@ jobs:
           source .venv/bin/activate
           cd tools/server/tests
           export ${{ matrix.extra_args }}
-          PYTEST_WORKERS=1 SLOW_TESTS=1 ./tests.sh
+          SLOW_TESTS=1 ./tests.sh

.github/workflows/server-self-hosted.yml

@@ -72,7 +72,7 @@ jobs:
         run: |
           cd tools/server/tests
           source venv/bin/activate
-          PYTEST_WORKERS=1 ./tests.sh
+          ./tests.sh
 
       - name: Tests (GPUx1, backend-sampling)
         id: server_integration_tests_backend_sampling
@@ -81,7 +81,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          PYTEST_WORKERS=1 ./tests.sh
+          ./tests.sh
 
       - name: Tests (GPUx2)
         id: server_integration_tests_gpu2
@@ -90,7 +90,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_METAL_DEVICES=2
-          PYTEST_WORKERS=1 ./tests.sh
+          ./tests.sh
 
       - name: Tests (GPUx2, backend-sampling)
         id: server_integration_tests_gpu2_backend_sampling
@@ -99,10 +99,10 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          PYTEST_WORKERS=1 ./tests.sh
+          ./tests.sh
 
   server-cuda:
-    runs-on: "hf-jobs-t4-small:cuda13"
+    runs-on: [self-hosted, llama-server, Linux, NVIDIA]
 
     steps:
       - name: Clone
@@ -112,42 +112,12 @@ jobs:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
-      - name: Install dependencies
-        run: |
-          sudo apt update
-          sudo apt install -y cmake libssl-dev python3 python3-venv python3-pip
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
-        with:
-          restore: false
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        with:
-          key: self-hosted-server-cuda
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
-
       - name: Build
         id: cmake_build
         run: |
-          cmake -B build -DGGML_CUDA=ON -DGGML_SCHED_NO_REALLOC=ON -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc
+          cmake -B build -DGGML_CUDA=ON -DGGML_SCHED_NO_REALLOC=ON
           cmake --build build --config Release -j $(nproc) --target llama-server
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: self-hosted-server-cuda
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
       - name: Python setup
         id: setup_python
         run: |
@@ -162,7 +132,7 @@ jobs:
         run: |
           cd tools/server/tests
           source venv/bin/activate
-          PYTEST_WORKERS=1 ./tests.sh
+          ./tests.sh
 
       - name: Tests (GPUx1, backend-sampling)
         id: server_integration_tests_backend_sampling
@@ -171,7 +141,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export LLAMA_ARG_BACKEND_SAMPLING=1
-          PYTEST_WORKERS=1 ./tests.sh
+          ./tests.sh
 
       - name: Tests (GPUx2)
         id: server_integration_tests_gpu2
@@ -180,7 +150,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_CUDA_DEVICES=2
-          PYTEST_WORKERS=1 ./tests.sh
+          ./tests.sh
 
       - name: Tests (GPUx2, backend-sampling)
         id: server_integration_tests_gpu2_backend_sampling
@@ -189,7 +159,7 @@ jobs:
           cd tools/server/tests
           source venv/bin/activate
           export GGML_CUDA_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1
-          PYTEST_WORKERS=1 ./tests.sh
+          ./tests.sh
 
   server-kleidiai:
     runs-on: ah-ubuntu_22_04-c8g_8x

.github/workflows/server.yml

@@ -80,19 +80,11 @@ jobs:
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: server-ubuntu-24.04-arm
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CI }}
-        with:
-          key: server-ubuntu-24.04-arm
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build
         id: cmake_build
@@ -101,18 +93,6 @@ jobs:
             -DGGML_SCHED_NO_REALLOC=ON
           cmake --build build --config Release -j $(nproc) --target llama-server
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: server-ubuntu-24.04-arm
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
       - name: Python setup
         id: setup_python
         uses: actions/setup-python@v6
@@ -160,7 +140,7 @@ jobs:
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: server-windows-2025-x64
           evict-old-files: 1d
@@ -201,13 +181,3 @@ jobs:
           cd tools/server/tests
           export SLOW_TESTS="1"
           ./tests.sh
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: server-windows-2025-x64
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/ui-build-self-hosted.yml

@@ -17,9 +17,8 @@ jobs:
         uses: actions/setup-node@v6
         with:
           node-version: "24"
-          # cache: "npm"
-          # cache-dependency-path: "tools/ui/package-lock.json"
-          package-manager-cache: false
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Install dependencies
         run: npm ci
@@ -32,6 +31,6 @@ jobs:
       - name: Upload built UI
         uses: actions/upload-artifact@v6
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
           retention-days: 1

.github/workflows/ui-build.yml

@@ -3,8 +3,8 @@ name: UI Build
 on:
   workflow_call:
     inputs:
-      ui_version:
-        description: 'Version string embedded in build.json (e.g. b1234); defaults to b<commit-count>'
+      hf_ui_version:
+        description: 'Version string for version.json (e.g. 12345)'
         required: false
         type: string
 
@@ -17,25 +17,13 @@ jobs:
     steps:
       - name: Checkout code
         uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-
-      - name: Resolve UI version
-        id: version
-        run: |
-          version="${{ inputs.ui_version }}"
-          if [ -z "$version" ]; then
-            version="b$(git rev-list --count HEAD)"
-          fi
-          echo "ui_version=${version}" >> $GITHUB_OUTPUT
 
       - name: Setup Node.js
         uses: actions/setup-node@v6
         with:
           node-version: "24"
-          # cache: "npm"
-          # cache-dependency-path: "tools/ui/package-lock.json"
-          package-manager-cache: false
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Install dependencies
         run: npm ci
@@ -43,7 +31,8 @@ jobs:
 
       - name: Build application
         env:
-          LLAMA_BUILD_NUMBER: ${{ steps.version.outputs.ui_version }}
+          HF_UI_VERSION: ${{ inputs.hf_ui_version || '' }}
+          LLAMA_BUILD_NUMBER: ${{ inputs.hf_ui_version || 'b0000' }}
         run: npm run build
         working-directory: tools/ui
 
@@ -54,6 +43,6 @@ jobs:
       - name: Upload built UI
         uses: actions/upload-artifact@v6
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
           retention-days: 1

.github/workflows/ui-publish.yml

@@ -37,7 +37,7 @@ jobs:
       - name: Download UI build artifact
         uses: actions/download-artifact@v7
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
 
       - name: Create distribution archive

.github/workflows/ui-self-hosted.yml

@@ -64,7 +64,7 @@ jobs:
       - name: Download built UI artifacts
         uses: actions/download-artifact@v6
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
 
       - name: Run type checking
@@ -106,7 +106,7 @@ jobs:
       - name: Download built UI artifacts
         uses: actions/download-artifact@v6
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
 
       - name: Build Storybook

.github/workflows/ui.yml

@@ -57,14 +57,13 @@ jobs:
         uses: actions/setup-node@v6
         with:
           node-version: "24"
-          # cache: "npm"
-          # cache-dependency-path: "tools/ui/package-lock.json"
-          package-manager-cache: false
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Download built UI artifacts
         uses: actions/download-artifact@v6
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
 
       - name: Install dependencies
@@ -115,9 +114,8 @@ jobs:
         uses: actions/setup-node@v6
         with:
           node-version: "24"
-          # cache: "npm"
-          # cache-dependency-path: "tools/ui/package-lock.json"
-          package-manager-cache: false
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Install dependencies
         id: setup
@@ -128,7 +126,7 @@ jobs:
       - name: Download built UI artifacts (reuses ui-build)
         uses: actions/download-artifact@v6
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
 
       - name: Install Playwright browsers

AUTHORS

@@ -1,4 +1,4 @@
-# date: Fri Sep  4 10:06:46 EEST 2026
+# date: Tue Aug 18 14:32:43 EEST 2026
 # this file is auto-generated by scripts/gen-authors.sh
 
 Нияз Гарифзянов <112617865+garrnizon@users.noreply.github.com>
@@ -46,7 +46,6 @@ Abhijit Ramesh <abhijitramesh2k@gmail.com>
 abhijitb11 <113058133+abhijitb11@users.noreply.github.com>
 Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
 Abhinay Krishna <abhinaykrishna60@gmail.com>
-Abhiram <78226909+geckguy@users.noreply.github.com>
 Abhishek Gopinath K <31348521+overtunned@users.noreply.github.com>
 abotsis <github@bots.is>
 Abraham Gonzalez <theabecaster0@gmail.com>
@@ -88,7 +87,6 @@ akleine <alb.kleine@gmx.de>
 Al G <toasting@gmail.com>
 Al Mochkin <14274697+amochkin@users.noreply.github.com>
 Alan Gray <agray3@users.noreply.github.com>
-Alan Tseng <alanhc.tseng1999@gmail.com>
 Alawode Oluwandabira <dabiraalawode@yahoo.com>
 Albert Jin <albert.jin@gmail.com>
 Alberto <57916483+albbus-stack@users.noreply.github.com>
@@ -138,9 +136,7 @@ alonfaraj <alonfaraj@gmail.com>
 AlpinDale <52078762+AlpinDale@users.noreply.github.com>
 alwqx <kenan3015@gmail.com>
 Aman <amangupta052@gmail.com>
-Aman Chadha(IVIXMMI) <79802170+ac-mmi@users.noreply.github.com>
 Aman Gupta <amangupta052@gmail.com>
-Aman Karki <itsamankarki@gmail.com>
 amd-dwang <dong.wang@amd.com>
 amd-lalithnc <lalithnc@amd.com>
 Amir <amir_zia@outlook.com>
@@ -191,7 +187,6 @@ Anton Mitkov <anton.mitkov@codeplay.com>
 Antonis Makropoulos <benuix@gmail.com>
 Anudit Nagar <nagaranudit@gmail.com>
 Anuj Attri <anujattri01@gmail.com>
-anujj <ajalota@nvidia.com>
 anzz1 <anzz1@live.com>
 Aparna M P <aparmp@qti.qualcomm.com>
 Aparna M P <quic_aparmp@quicinc.com>
@@ -201,7 +196,6 @@ arch-btw <57669023+arch-btw@users.noreply.github.com>
 arcrank <arcrank@gmail.com>
 ardfork <134447697+ardfork@users.noreply.github.com>
 Arik Poznanski <arikpoz@users.noreply.github.com>
-Aritro Bandyopadhyay <71339004+AriBandyo@users.noreply.github.com>
 arlo-phoenix <140345165+arlo-phoenix@users.noreply.github.com>
 Armen Kaleshian <kriation@users.noreply.github.com>
 Arsen Arutunan <58118221+limloop@users.noreply.github.com>
@@ -236,7 +230,6 @@ bandoti <141645996+bandoti@users.noreply.github.com>
 Bar Haim <barvhaim@gmail.com>
 BarfingLemurs <128182951+BarfingLemurs@users.noreply.github.com>
 Bart Louwers <bart.louwers@gmail.com>
-Bartosz Taudul <wolf@nereid.pl>
 Bartowski <3266127+bartowski1182@users.noreply.github.com>
 Bartowski <ckealty1182@gmail.com>
 Bas Nijholt <basnijholt@gmail.com>
@@ -284,7 +277,6 @@ Bono Lv <lvscar@users.noreply.github.com>
 Borislav Stanimirov <b.stanimirov@abv.bg>
 Borislav Stanimirov <b@ibob.bg>
 Bowen Han <fancycode@gmail.com>
-Brad Smith <1472326+infinitewarp@users.noreply.github.com>
 Branden Butler <bwtbutler@hotmail.com>
 Brandon Squizzato <35474886+bsquizz@users.noreply.github.com>
 Brian <mofosyne@gmail.com>
@@ -295,7 +287,6 @@ Bryan Honof <bryanhonof@gmail.com>
 bryanSwk <93190252+bryanSwk@users.noreply.github.com>
 bsilvereagle <bsilvereagle@users.noreply.github.com>
 bssrdf <merlintiger@hotmail.com>
-Buğra Özgürsoy <13810383+ozgursoy@users.noreply.github.com>
 byte-6174 <88070277+byte-6174@users.noreply.github.com>
 Caleb DeLeeuw <143902425+SolshineCode@users.noreply.github.com>
 Calvin Laurenson <calvin@laurenson.dev>
@@ -335,7 +326,6 @@ Chenguang Li <757486878@qq.com>
 Chenguang Li <87689256+noemotiovon@users.noreply.github.com>
 Chipmunk <101038159+CHIPMUNK-T0T@users.noreply.github.com>
 chiranko <96988916+chiranko@users.noreply.github.com>
-Chris Danis <cdanis@gmail.com>
 Chris Elrod <elrodc@gmail.com>
 Chris Kuehl <ckuehl@ckuehl.me>
 Chris Lee <clee@mg8.org>
@@ -366,7 +356,6 @@ clyang <clyang@clyang.net>
 cmdr2 <secondary.cmdr2@gmail.com>
 cmdr2 <shashank.shekhar.global@gmail.com>
 cocktailpeanut <121128867+cocktailpeanut@users.noreply.github.com>
-codemonkey <441345965@qq.com>
 codezjx <code.zjx@gmail.com>
 coezbek <c.oezbek@gmail.com>
 comex <comexk@gmail.com>
@@ -378,8 +367,6 @@ Copilot <198982749+Copilot@users.noreply.github.com>
 Corentin REGAL <corentin.regal@gmail.com>
 cphlipot <9103367+cphlipot@users.noreply.github.com>
 cpumaxx <163466046+cpumaxx@users.noreply.github.com>
-cqderek <cqderek@gmail.com>
-cqderek <cqiang@qti.qualcomm.com>
 crasm <crasm@git.vczf.net>
 crasm <crasm@git.vczf.us>
 crat0z <11581854+crat0z@users.noreply.github.com>
@@ -440,7 +427,6 @@ DavidKorczynski <david@adalogics.com>
 davidrhodus <david@vacovideo.com>
 Dawid Potocki <github@dawidpotocki.com>
 Dawid Wysocki <62249621+TortillaZHawaii@users.noreply.github.com>
-Daya Adianto <addianto@users.noreply.github.com>
 ddh0 <chemist-mulches-39@icloud.com>
 ddh0 <dylanhalladay02@icloud.com>
 ddpasa <112642920+ddpasa@users.noreply.github.com>
@@ -477,7 +463,6 @@ Dmytro Romanov <casteldazur@gmail.com>
 Dobri Danchev <12420863+danchev@users.noreply.github.com>
 DocShotgun <126566557+DocShotgun@users.noreply.github.com>
 Doctor Shotgun <126566557+DocShotgun@users.noreply.github.com>
-Dominik Pantaleoni <95251853+dpantaleoni@users.noreply.github.com>
 Don Mahurin <dmahurin@users.noreply.github.com>
 Dong Won Kim <63934649+ddwkim@users.noreply.github.com>
 Donghyeon Jeong <54725479+djeong20@users.noreply.github.com>
@@ -519,7 +504,6 @@ Emmanuel Ferdman <emmanuelferdman@gmail.com>
 Emreerdog <34742675+Emreerdog@users.noreply.github.com>
 Engininja2 <139037756+Engininja2@users.noreply.github.com>
 Equim <sayaka@ekyu.moe>
-Eric A Stalee <87948564+Eric-A-Stalee@users.noreply.github.com>
 Eric Curtin <ecurtin@redhat.com>
 Eric Curtin <eric.curtin@docker.com>
 Eric Curtin <ericcurtin17@gmail.com>
@@ -535,7 +519,6 @@ Esko Toivonen <eskot98@gmail.com>
 Ethan Turner <eturner64@gmail.com>
 Ettore Di Giacinto <mudler@users.noreply.github.com>
 EugeoSynthesisThirtyTwo <gabriel.dhimoila@gmail.com>
-Eurekatic <eurekatic@eurekatic.eu>
 Evan Huus <eapache@gmail.com>
 Evan Jones <evan.q.jones@gmail.com>
 Evan Miller <emmiller@gmail.com>
@@ -694,7 +677,6 @@ HimariO <dsfhe49854@gmail.com>
 hipudding <huafengchun@gmail.com>
 Hitesh Chopra <34310832+hiteshchopra11@users.noreply.github.com>
 hksdpc255 <43977088+hksdpc255@users.noreply.github.com>
-hmirin <hmirin@users.noreply.github.com>
 hmscider <201289679+hmscider@users.noreply.github.com>
 Hoang Nguyen <hugo53@users.noreply.github.com>
 hoangmit <hoangmit@users.noreply.github.com>
@@ -719,7 +701,6 @@ Huawei Lin <huaweilin.cs@gmail.com>
 Hugo <hugo@whynothugo.nl>
 Hugo Roussel <hugo.rous@gmail.com>
 Huifeng Ou <79071290+ho2103@users.noreply.github.com>
-HumerousGorgon <31957201+HumerousGorgon@users.noreply.github.com>
 hutli <6594598+hutli@users.noreply.github.com>
 hutli <hutli@hutli.hu>
 hutli <jensstaermose@hotmail.com>
@@ -757,15 +738,12 @@ intelmatt <61025942+intelmatt@users.noreply.github.com>
 iohub <rickyang.pro@gmail.com>
 Ionoclast Laboratories <brigham@ionoclast.com>
 iron <lizhenneng@gmail.com>
-Isaac <34376531+init-22@users.noreply.github.com>
 Isaac McFadyen <isaac@imcf.me>
 IsaacDynamo <61521674+IsaacDynamo@users.noreply.github.com>
 Ishaan Gandhi <Ishaangandhi@gmail.com>
 iSma <ismail.senhaji@gmail.com>
 Ismail <115064057+AlrIsmail@users.noreply.github.com>
 issixx <46835150+issixx@users.noreply.github.com>
-itsnotoger <19309683+itsnotoger@users.noreply.github.com>
-itterative <190138728+itterative@users.noreply.github.com>
 Ivan <nekotekina@gmail.com>
 Ivan Chikish <nekotekina@gmail.com>
 Ivan Filipov <159561759+vanaka11@users.noreply.github.com>
@@ -790,7 +768,6 @@ Jakkala Mahesh <155058658+MaheshJakkala@users.noreply.github.com>
 Jakub N <jakubniemczyk97@gmail.com>
 JamePeng <jame_peng@sina.com>
 James A Capozzoli <157492257+jac-jim@users.noreply.github.com>
-James Francis <6763899+JamesFranc@users.noreply.github.com>
 James O'Leary <65884233+jpohhhh@users.noreply.github.com>
 James Reynolds <magnusviri@users.noreply.github.com>
 jameswu2014 <545426914@qq.com>
@@ -821,7 +798,6 @@ Jed Fox <git@jedfox.com>
 Jeff Bolz <jbolz@nvidia.com>
 Jeffrey Morgan <jmorganca@gmail.com>
 Jeffrey Quesnelle <emozilla@nousresearch.com>
-Jeremie Miller <jeremie.miller@gmail.com>
 Jeremy Demeule <jdemeule@users.noreply.github.com>
 Jeremy Rand <244188+JeremyRand@users.noreply.github.com>
 Jeroen Mostert <jeroen.mostert@cm.com>
@@ -833,7 +809,6 @@ Jesse Jojo Johnson <williamsaintgeorge@gmail.com>
 Jesse LaRose <jesse@taey.ai>
 Jesse Posner <jesse.posner@gmail.com>
 Jesus Talavera <145992175+jesus-talavera-ibm@users.noreply.github.com>
-Jetson Tan <tanzongyouyi@outlook.com>
 Jett Janiak <jettjaniak@gmail.com>
 Jeximo <jeximo@gmail.com>
 JFLFY2255 <JFLFY2255@163.com>
@@ -850,7 +825,6 @@ Jie Fu (傅杰) <jiefu@tencent.com>
 jiez <373447296@qq.com>
 Jillis ter Hove <j.terhove@gmail.com>
 Jim Wu <jimw567@users.noreply.github.com>
-Jingxin (Philip) Li <philipaslee@gmail.com>
 Jinwoo Jeong <33892306+williamjeong2@users.noreply.github.com>
 Jinyang He <hejinyang@loongson.cn>
 jinzihao <jinzihao1996@gmail.com>
@@ -876,13 +850,11 @@ John Balis <phobossystems@gmail.com>
 John Bean <113509988+johnbean393@users.noreply.github.com>
 John Eismeier <42679190+jeis4wpi@users.noreply.github.com>
 John Smith <67539080+kingsidelee@users.noreply.github.com>
-John-Henry Lim <42513874+Interpause@users.noreply.github.com>
 Johnathan Craig Maudlin <13183098+jcmdln@users.noreply.github.com>
 JohnnyB <jboero@users.noreply.github.com>
 johnson442 <56517414+johnson442@users.noreply.github.com>
 jojorne <jojorne@users.noreply.github.com>
 jon-chuang <9093549+jon-chuang@users.noreply.github.com>
-Jonas J <111707981+John-194@users.noreply.github.com>
 Jonas Jankaitis <111707981+John-194@users.noreply.github.com>
 Jonas Wunderlich <32615971+jonas-w@users.noreply.github.com>
 Jonathan <47618606+jbuchananr@users.noreply.github.com>
@@ -952,7 +924,6 @@ Karsten Weiss <knweiss@gmail.com>
 Karthick <j.karthic2004@gmail.com>
 Karthik Kumar Viswanathan <195178+guilt@users.noreply.github.com>
 Karthik Sethuraman <k.seth1993@gmail.com>
-Kartik Gulia <kgulia@nvidia.com>
 Kartik Sirohi <99896785+sirohikartik@users.noreply.github.com>
 Kashif Rasul <kashif.rasul@gmail.com>
 KASR <karim.asrih@gmail.com>
@@ -960,7 +931,6 @@ Kasumi <90275229+kasumi-1@users.noreply.github.com>
 Katostrofik <georgiopapairo@gmail.com>
 katsu560 <118887472+katsu560@users.noreply.github.com>
 Kawrakow <48489457+ikawrakow@users.noreply.github.com>
-kbenkhaled <khalilbenkhaled01@gmail.com>
 kchro3 <62481661+kchro3@users.noreply.github.com>
 kdkd <2569413+kdkd@users.noreply.github.com>
 Keiichi Tabata <keiichi.tabata@outlook.com>
@@ -969,7 +939,6 @@ Kenvix ⭐ <kenvixzure@live.com>
 Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
 Kevin Gibbons <bakkot@gmail.com>
 Kevin Hannon <kehannon@redhat.com>
-Kevin Hopper <93635715+kh0pper@users.noreply.github.com>
 Kevin Ji <1146876+kevinji@users.noreply.github.com>
 Kevin Kwok <antimatter15@gmail.com>
 Kevin Liu <4396kevinliu@gmail.com>
@@ -995,14 +964,12 @@ Konstantin Herud <konstantin.herud@denkbares.com>
 Konstantin Zhuravlyov <konstantin.zhuravlyov@amd.com>
 Krishna Sridhar <99914379+srikris-sridhar@users.noreply.github.com>
 krystiancha <krystian@krystianch.com>
-krzsztf <krzysztof@witkowscy.org>
 kubawoo <k-wach@o2.pl>
 kumaal <44551860+kumaal@users.noreply.github.com>
 kunal-vaishnavi <115581922+kunal-vaishnavi@users.noreply.github.com>
 kunnis <kunnis@users.noreply.github.com>
 Kunshang Ji <kunshang.ji@intel.com>
 kuronekosaiko <EvanChanJ@163.com>
-kurquhar <kurquhar@qti.qualcomm.com>
 Kusha Gharahi <3326002+kushagharahi@users.noreply.github.com>
 kustaaya <58045274+kustaaya@users.noreply.github.com>
 kuvaus <22169537+kuvaus@users.noreply.github.com>
@@ -1014,7 +981,6 @@ Kyle Liang <liangmanlai@gmail.com>
 Kyle Mistele <kyle@mistele.com>
 KyleHagy <59183061+KyleHagy@users.noreply.github.com>
 Kylin <56434533+KyL0N@users.noreply.github.com>
-Kyozzz <1147385157@qq.com>
 l-austenfeld <53152202+l-austenfeld@users.noreply.github.com>
 l3utterfly <gc.pthzfoldr@gmail.com>
 l8bloom <l8bloomapi@gmail.com>
@@ -1026,7 +992,6 @@ Lars Sonchocky-Helldorf <lars.sonchocky-helldorf@hamburg.de>
 las7 <98077186+las7@users.noreply.github.com>
 Lasse Lauwerys <65569591+Iemand005@users.noreply.github.com>
 Laura <Tijntje_7@msn.com>
-Laurent Zuijdwijk <laurent.zuijdwijk@gmail.com>
 Law Po Ying <30721578+yingying0906@users.noreply.github.com>
 lcy <lcy0321@users.noreply.github.com>
 ldwang <ftgreat@163.com>
@@ -1074,8 +1039,6 @@ Ludovic Henry <git@ludovic.dev>
 Ludovic Henry <ludovic@rivosinc.com>
 Lukas Straub <lukasstraub2@web.de>
 Łukasz Ślusarczyk <112692748+lslusarczyk@users.noreply.github.com>
-Lukasz Stolcman <4583553+lstolcman@users.noreply.github.com>
-LunalFresh <165352784+LunalFresh@users.noreply.github.com>
 Luo Tian <lt@basecity.com>
 luoyu-intel <yu.luo@intel.com>
 luyhcsu <110711054+luyhcsu@users.noreply.github.com>
@@ -1091,7 +1054,6 @@ Maarten ter Huurne <maarten@treewalker.org>
 Maciej Lisowski <39798354+MaciejDromin@users.noreply.github.com>
 Mack Straight <eiz@users.noreply.github.com>
 maddes8cht <55592906+maddes8cht@users.noreply.github.com>
-Mads Marquart <mads@marquart.dk>
 Maël Kerbiriou <m431.kerbiriou@gmail.com>
 MaggotHATE <clay1326@gmail.com>
 MagicExists <106458387+gugugiyu@users.noreply.github.com>
@@ -1253,8 +1215,6 @@ Naco Siren <naco-siren@users.noreply.github.com>
 Nam D. Tran <42194884+namtranase@users.noreply.github.com>
 nanahi <130121847+na-na-hi@users.noreply.github.com>
 Nathan Epstein <nate2@umbc.edu>
-Nathan Wilson <67372905+Nathanw1014@users.noreply.github.com>
-Nathanw1014 <67372905+Nathanw1014@users.noreply.github.com>
 Natsu <chino@hotococoa.moe>
 Nauful Shaikh <nauful@gmail.com>
 NawafAlansari <72708095+NawafAlansari@users.noreply.github.com>
@@ -1277,7 +1237,6 @@ niansa/tuxifan <tuxifan@posteo.de>
 Nicholai Tukanov <nicholaitukanov@gmail.com>
 Nicholas Sparks <157740354+nisparks@users.noreply.github.com>
 Nick <0x0b4ac@gmail.com>
-Nick Farrell <nick.farrell@aiven.io>
 nick huang <nickhuang99@hotmail.com>
 Nick Lafleur <55208706+nicklafleur@users.noreply.github.com>
 Nick Towle <ntowle@gmail.com>
@@ -1300,7 +1259,6 @@ NikolaiLyssogor <59844691+NikolaiLyssogor@users.noreply.github.com>
 Nikolaos Pothitos <pothitos@di.uoa.gr>
 Nikolas <127742645+nneubacher@users.noreply.github.com>
 Nikolay Popov <131475237+npopov-vst@users.noreply.github.com>
-Nils Gladitz <nilsgladitz@gmail.com>
 Nindaleth <Nindaleth@users.noreply.github.com>
 ningshanwutuobang <ningshanwutuobang@gmail.com>
 Noah <99681487+NoahOksuz@users.noreply.github.com>
@@ -1397,7 +1355,6 @@ Pop Flamingo <trevor.annedenise@icloud.com>
 postmasters <namnguyen@google.com>
 Pouya <PooyaGhahramanian@Gmail.com>
 pqnet <119850+pqnet@users.noreply.github.com>
-Prabhsimran Singh <pskrunner14@gmail.com>
 Prabod <prabod@maincode.com>
 Prajwal B Mehendarkar <prajwal.b.mehendarkar@ibm.com>
 Pranav Dhinakar <pdhinaka@qti.qualcomm.com>
@@ -1421,7 +1378,6 @@ qouoq <qouoq@fastmail.com>
 Qu Zongfu <43257352+yancaoweidaode@users.noreply.github.com>
 quei <56998528+quei4r@users.noreply.github.com>
 Quentin Bramas <quentin.bramas@gmail.com>
-QuintinShaw <github@xyt.email>
 QuintinShaw <yx6f20@soton.ac.uk>
 qunash <anzoria@gmail.com>
 quyentonndbs <raynaedgar8677@outlook.com>
@@ -1506,7 +1462,6 @@ robertomeroni <150194833+robertomeroni@users.noreply.github.com>
 Robey Holderith <robey@flaminglunchbox.net>
 Robin Davidsson <40024429+R-Dson@users.noreply.github.com>
 Robyn <robyngraf@users.noreply.github.com>
-Rock Chen <rockchen.tw@gmail.com>
 Rőczey Barnabás <31726601+An0nie@users.noreply.github.com>
 RodriMora <bullerwins@gmail.com>
 Roger Chen <chenrui@gmail.com>
@@ -1544,21 +1499,17 @@ runfuture <runfuture@users.noreply.github.com>
 RunningLeon <maningsheng@sensetime.com>
 RunningLeon <mnsheng@yeah.net>
 Russyyds <161207317+Russyyds@users.noreply.github.com>
-Ryan C <ryan5rdx@users.noreply.github.com>
 Ryan Goulden <percontation@gmail.com>
 Ryan Landay <rlanday@gmail.com>
 Ryan Mangeno <160974989+ryan-mangeno@users.noreply.github.com>
 Ryder Wishart <ryderwishart@gmail.com>
 Ryuei <louixs@users.noreply.github.com>
 s-goto-11 <206795233+s-goto-11@users.noreply.github.com>
-s0mecode <213953308+s0mecode@users.noreply.github.com>
 s8322 <s0527684199@gmail.com>
-Saad Ali <NIXKnight@users.noreply.github.com>
 Saba Fallah <10401143+sfallah@users.noreply.github.com>
 Saba Fallah <sabafallah@gmail.com>
 Sachin Desai <smdesai@gmail.com>
 Sachin Sharma <sachin@zettabolt.com>
-Safi Ullah <safiullah.3915@gmail.com>
 safranowith <bsh155762@gmail.com>
 SakuraUmi <yukinon244@gmail.com>
 Salvador E. Tropea <stropea@inti.gob.ar>
@@ -1601,7 +1552,6 @@ Sergey Alirzaev <l29ah@riseup.net>
 Sergey Alirzaev <zl29ah@gmail.com>
 Sergey Fedorov <vital.had@gmail.com>
 Sergey Malinin <sergmalinin@gmail.com>
-Sergey Sklyarov <sergey.sklyarov@gmail.com>
 Sergio López <slp@redhat.com>
 Sergio López <slp@sinrega.org>
 Sergiu <8598216+mzsergiu@users.noreply.github.com>
@@ -1632,13 +1582,11 @@ Shawn Gu <shawngu@qti.qualcomm.com>
 Shawn yang <137684499+Yangxiaoz@users.noreply.github.com>
 Shelby Jenkins <47464908+ShelbyJenkins@users.noreply.github.com>
 Sheldon Robinson <sheldon.robinson@live.com>
-Shenghan Yang <ysharke@sjtu.edu.cn>
 shibe2 <shibe@tuta.io>
 Shijie <821898965@qq.com>
 Shin-myoung-serp <relent95@naver.com>
 Shintarou Okada <kokuzen@gmail.com>
 shivamkumard-ctrl <shivamkumard@nvidia.com>
-Shobhit <sobhit.me@gmail.com>
 Shouyu <65317431+joeldushouyu@users.noreply.github.com>
 Shouzheng Liu <61452103+lshzh-ww@users.noreply.github.com>
 Shouzheng Liu <lshzh.hi@gmail.com>
@@ -1659,7 +1607,6 @@ Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
 Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
 simevo <github@simevo.com>
 Simon Redman <simon@ergotech.com>
-Simon Teixidor <simon@flaskpost.me>
 Simon Willison <swillison@gmail.com>
 simon886212 <37953122+simon886212@users.noreply.github.com>
 Simranjeet Singh <105192966+simrnsingh@users.noreply.github.com>
@@ -1716,7 +1663,6 @@ stevenkuang <stevenkuang@tencent.com>
 Steward Garcia <57494570+FSSRepo@users.noreply.github.com>
 StrangeBytesDev <141275258+StrangeBytesDev@users.noreply.github.com>
 strawberrymelonpanda <152940198+strawberrymelonpanda@users.noreply.github.com>
-Strongtut <Strongtut@users.noreply.github.com>
 Suaj Carrot <72162667+SuajCarrot@users.noreply.github.com>
 sudhiarm <sudhi.sathyavathy@arm.com>
 Sukriti Sharma <Ssukriti@users.noreply.github.com>
@@ -1741,7 +1687,6 @@ Tamar <Tamar0812@outlook.co.il>
 tamarPal <tamarp3385@gmail.com>
 Tameem <113388789+AhmadTameem@users.noreply.github.com>
 Tamotsu Takahashi <ttakah+github@gmail.com>
-Tanner Bruhn <66120666+tannerbruhn@users.noreply.github.com>
 tarcey <cey.tarik@gmail.com>
 Tarek Dakhran <t.dakhran@gmail.com>
 Tarek Dakhran <tarek@liquid.ai>
@@ -1751,7 +1696,6 @@ Taylor <quantumtraveling@gmail.com>
 tc-mb <157115220+tc-mb@users.noreply.github.com>
 TecJesh <qdvm5gl@163.com>
 Tei Home <taiteitonghome@proton.me>
-Tekin Ertekin <tekin.ertekin@gmail.com>
 Tekin Ertekin <tekinertekin@gmail.com>
 tempstudio <49735574+tempstudio@users.noreply.github.com>
 teo <TeoZosa@users.noreply.github.com>
@@ -1793,7 +1737,6 @@ Ting Lou <louting@189.cn>
 Ting Lou <ting.lou@gmail.com>
 Ting Sun <suntcrick@gmail.com>
 Titaniumtown <titaniumtown@proton.me>
-Tiwei Bie <tiwei.btw@antgroup.com>
 tjohnman <tjohnman@users.noreply.github.com>
 Tobias Lütke <tobi@shopify.com>
 Toby <25832191+aetherbird@users.noreply.github.com>
@@ -1870,7 +1813,6 @@ Vishal Agarwal <vishalagarwal.jss@gmail.com>
 Vishal Singh <vishal@zettabolt.com>
 Vitali Lovich <vlovich+github@gmail.com>
 Vivian <vynride@gmail.com>
-vk <89937361+itsvedantkumar@users.noreply.github.com>
 Vlad <spitfireage@gmail.com>
 Vladimir <bogdad@gmail.com>
 Vladimir Malyutin <first-leon@yandex.ru>
@@ -1955,7 +1897,6 @@ Yaiko <elyaiko@hotmail.com>
 Yakine Tahtah <96926916+ReinforcedKnowledge@users.noreply.github.com>
 YangLe <smilingpoplar@gmail.com>
 yangli2 <yangli2@gmail.com>
-Yaniss Amazouz <yaniss91600@gmail.com>
 Yann Follet <131855179+YannFollet@users.noreply.github.com>
 Yanzhao Wang <yanzhaow@qti.qualcomm.com>
 Yarden Tal <yardent@qti.qualcomm.com>

CMakeLists.txt

@@ -4,7 +4,7 @@ include(CheckIncludeFileCXX)
 
 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 4)
+set(LLAMA_VERSION_MINOR 2)
 set(LLAMA_VERSION_PATCH 0)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")
 
@@ -134,8 +134,8 @@ option(LLAMA_BUILD_TOOLS     "llama: build tools"
 option(LLAMA_BUILD_EXAMPLES  "llama: build examples"                                                             ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER    "llama: build server example"                                                       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_APP       "llama: build the unified binary"                                                   ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_UI        "llama: build the embedded Web UI for server"                                       OFF)
-option(LLAMA_USE_PREBUILT_UI "llama: use prebuilt UI from HF Bucket when available" ON)
+option(LLAMA_BUILD_UI        "llama: build the embedded Web UI for server"                                       ON)
+option(LLAMA_USE_PREBUILT_UI "llama: use prebuilt UI from HF Bucket when available (requires LLAMA_BUILD_UI=ON)" ON)
 
 option(LLAMA_TOOLS_INSTALL "llama: install tools" ${LLAMA_TOOLS_INSTALL_DEFAULT})
 option(LLAMA_TESTS_INSTALL "llama: install tests" ON)

CODEOWNERS

@@ -57,7 +57,6 @@
 /ggml/src/ggml-cann/                    @ggml-org/ggml-cann
 /ggml/src/ggml-common.h                 @ggerganov
 /ggml/src/ggml-cpu/                     @ggerganov
-/ggml/src/ggml-cpu/iqp.*                @bartowski1182
 /ggml/src/ggml-cpu/spacemit/            @alex-spacemit
 /ggml/src/ggml-cuda/                    @ggml-org/ggml-cuda
 /ggml/src/ggml-cuda/vendors/hip.h       @IMbackK

CONTRIBUTING.md

@@ -74,7 +74,6 @@ For more info, please refer to the [AGENTS.md](AGENTS.md) file.
 - If a PR does not warrant a new release, add `[no release]` in the squashed commit to spare CI resources
 - Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)
 - Add the ["merge ready"](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+is%3Aopen+draft%3Ano+sort%3Aupdated-desc+label%3A%22merge+ready%22+) label to a PR to indicate when a PR can be fast-merged without waiting for 2 independent reviews. [(more info)](https://github.com/ggml-org/llama.cpp/pull/26178)
-- Wait for CI results before merging
 
 Maintainers reserve the right to decline review or close pull requests for any reason, without any questions, particularly under any of the following conditions:
 - The proposed change is already mentioned in the roadmap or an existing issue, and it has been assigned to someone.

README.md

@@ -13,7 +13,7 @@
 [![Docker](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/docker.yml?label=Docker)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
 [![Winget](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/winget.yml?label=Winget)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)
 
-[ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Ajhen0409%20OR%20author%3Abartowski1182%20OR%20author%3Anikwen%20OR%20author%3Ahipudding%20OR%20author%3Aravi9%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3Amarty1885%20OR%20author%3A0cc4m%20OR%20author%3ATitaniumtown%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Awine99%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [dev stats](https://github.com/ggml-org/llama.cpp-dev) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
+[manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
 
 </div>
 
@@ -74,7 +74,7 @@ The `llama.cpp` project is build on top of the [ggml](https://github.com/ggml-or
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [CUDA](docs/build.md#cuda) | Nvidia GPU |
 | [HIP](docs/build.md#hip) | AMD GPU |
-| [Hexagon](docs/backend/snapdragon/README.md) | Snapdragon |
+| [Hexagon [In Progress]](docs/backend/snapdragon/README.md) | Snapdragon |
 | [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
 | [MUSA](docs/build.md#musa) | Moore Threads GPU |
 | [Metal](docs/build.md#metal-build) | Apple Silicon |

app/llama.cpp

@@ -80,7 +80,7 @@ static const command cmds[] = {
 #undef UPDATE_HIDDEN
 
 static int version(int /*argc*/, char ** /*argv*/) {
-    llama_print_build_info(llama_version(), stdout);
+    llama_print_build_info(llama_version());
     return 0;
 }
 

build-xcframework.sh

@@ -18,7 +18,7 @@ LLAMA_BUILD_TESTS=OFF
 LLAMA_BUILD_SERVER=OFF
 LLAMA_BUILD_MTMD=ON
 GGML_METAL=ON
-GGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY:-ON}
+GGML_METAL_EMBED_LIBRARY=ON
 GGML_BLAS_DEFAULT=ON
 GGML_OPENMP=OFF
 
@@ -169,14 +169,6 @@ setup_framework_structure() {
     cp tools/mtmd/mtmd.h           ${header_path}
     cp tools/mtmd/mtmd-helper.h    ${header_path}
 
-    if [[ "$GGML_METAL_EMBED_LIBRARY" == "OFF" ]]; then
-        if [[ "$platform" == "macos" ]]; then
-            cp ${build_dir}/bin/*.metallib ${build_dir}/framework/${framework_name}.framework/Versions/A/Resources/
-        else
-            cp ${build_dir}/bin/*.metallib ${build_dir}/framework/${framework_name}.framework/
-        fi
-    fi
-
     # Create module map (common for all platforms)
     cat > ${module_path}module.modulemap << EOF
 framework module llama {
@@ -458,7 +450,6 @@ build_ios_sim() {
         -DIOS=ON \
         -DCMAKE_SYSTEM_NAME=iOS \
         -DCMAKE_OSX_SYSROOT=iphonesimulator \
-        -DGGML_METAL_TARGET_OS=ios \
         -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
         -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
         -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
@@ -476,7 +467,6 @@ build_ios_device() {
         -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
         -DCMAKE_SYSTEM_NAME=iOS \
         -DCMAKE_OSX_SYSROOT=iphoneos \
-        -DGGML_METAL_TARGET_OS=ios \
         -DCMAKE_OSX_ARCHITECTURES="arm64" \
         -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
         -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
@@ -508,7 +498,6 @@ build_visionos() {
         -DCMAKE_OSX_ARCHITECTURES="arm64" \
         -DCMAKE_SYSTEM_NAME=visionOS \
         -DCMAKE_OSX_SYSROOT=xros \
-        -DGGML_METAL_TARGET_OS=xros \
         -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
         -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
         -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
@@ -527,7 +516,6 @@ build_visionos_sim() {
         -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
         -DCMAKE_SYSTEM_NAME=visionOS \
         -DCMAKE_OSX_SYSROOT=xrsimulator \
-        -DGGML_METAL_TARGET_OS=xros \
         -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
         -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
         -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
@@ -546,7 +534,6 @@ build_tvos_sim() {
         -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
         -DCMAKE_SYSTEM_NAME=tvOS \
         -DCMAKE_OSX_SYSROOT=appletvsimulator \
-        -DGGML_METAL_TARGET_OS=tvos \
         -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
         -DGGML_METAL=ON \
         -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
@@ -565,7 +552,6 @@ build_tvos_device() {
         -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
         -DCMAKE_SYSTEM_NAME=tvOS \
         -DCMAKE_OSX_SYSROOT=appletvos \
-        -DGGML_METAL_TARGET_OS=tvos \
         -DCMAKE_OSX_ARCHITECTURES="arm64" \
         -DGGML_METAL=ON \
         -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \

ci/run.sh

@@ -189,8 +189,8 @@ if [ ! -z ${GG_BUILD_OPENVINO} ]; then
     fi
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"
 
-    # TODO: fix failing tests on OpenVINO backend
-    CTEST_EXTRA="-E test-llama-archs|^test-recurrent-state-|test-backend-ops|test-save-load-state"
+    # TODO: fix and re-enable the `test-llama-archs` test below
+    CTEST_EXTRA="-E test-llama-archs|test-recurrent-state-rollback-nemotron-h"
 fi
 
 ## helpers
@@ -300,40 +300,6 @@ function gg_sum_ctest_release {
     gg_printf '```\n'
 }
 
-# test_llama_archs_tensor_split
-
-function gg_run_test_llama_archs_tensor_split {
-    cd ${SRC}
-
-    set -e
-
-    if [ ! -z ${GG_BUILD_CUDA} ]; then
-        GGML_CUDA_DEVICES=1 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_CUDA_DEVICES=2 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_CUDA_DEVICES=3 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_CUDA_DEVICES=4 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-    fi
-
-    if [ ! -z ${GG_BUILD_METAL} ]; then
-        GGML_METAL_DEVICES=1 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_METAL_DEVICES=2 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_METAL_DEVICES=3 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_METAL_DEVICES=4 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-    fi
-
-    set +e
-}
-
-function gg_sum_test_llama_archs_tensor_split {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'Runs test-llama-archs with 1 to 4 devices\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '```\n'
-    gg_printf '%s\n' "$(cat $OUT/${ci}.log)"
-    gg_printf '```\n'
-}
-
 # test_scripts
 
 function gg_run_test_scripts {
@@ -732,11 +698,6 @@ function gg_check_build_requirements {
         gg_printf 'ctest not found, please install\n'
         exit 1
     fi
-
-    if ! command -v unzip &> /dev/null; then
-        gg_printf 'unzip not found, please install\n'
-        exit 1
-    fi
 }
 
 function gg_run_test_backend_ops_cpu {
@@ -790,8 +751,6 @@ ret=0
 test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
-test $ret -eq 0 && gg_run test_llama_archs_tensor_split
-
 if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
     test $ret -eq 0 && gg_run test_backend_ops_cpu
 fi

common/CMakeLists.txt

@@ -53,10 +53,7 @@ endif()
 
 set(TARGET llama-common)
 
-include(parsers/sources.cmake)
-
 add_library(${TARGET}
-    ${LLAMA_CHAT_PARSERS_SOURCES}
     arg.cpp
     arg.h
     base64.hpp

common/arg.cpp

@@ -894,12 +894,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
     postprocess_cpu_params(params.speculative.draft.cpuparams,       &params.cpuparams);
     postprocess_cpu_params(params.speculative.draft.cpuparams_batch, &params.cpuparams_batch);
 
-    // default the mmproj device to the global device selection if not set explicitly with -mmdev
-    if (params.mmproj_use_gpu && params.mmproj_device == nullptr && !params.devices.empty()) {
-        params.mmproj_device = params.devices.front();
-        params.mmproj_use_gpu = params.mmproj_device != nullptr;
-    }
-
     if (params.prompt_cache_all && (params.interactive || params.interactive_first)) {
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
@@ -966,11 +960,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         ));
     }
 
-    // if the preserve_reasoning kwarg was not specified explicitly, enable it by default
-    if (!params.default_template_kwargs.count("preserve_reasoning")) {
-        params.default_template_kwargs["preserve_reasoning"] = "true";
-    }
-
     return true;
 }
 
@@ -1654,14 +1643,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_env("LLAMA_ARG_CTX_SIZE"));
-    add_opt(common_arg(
-        { "--kv-unified-per-slot" }, "N",
-        "context limit per parallel slot (default: unset, behavior unchanged).\n"
-        "when set without -c/--ctx-size, the shared KV pool is sized to n_parallel*N",
-        [](common_params & params, int value) {
-            params.kv_unified_per_slot = value;
-        }
-    ).set_env("LLAMA_ARG_KV_UNIFIED_PER_SLOT").set_examples({ LLAMA_EXAMPLE_SERVER }));
     add_opt(common_arg(
         {"-n", "--predict", "--n-predict"}, "N",
         string_format(
@@ -2616,7 +2597,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         // note: "-mmdev" must sort after "--rpc" in the preset map, else RPC devices are not registered yet
         {"-mmdev", "--mmproj-device"}, "DEVICE",
-        "device to use for multimodal projector (none = don't offload, default: follows --device)\n"
+        "device to use for multimodal projector (none = don't offload, default: auto)\n"
         "use --list-devices to see a list of available devices",
         [](common_params & params, const std::string & value) {
             if (value == "none") {
@@ -2663,27 +2644,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.mtmd_batch_max_tokens = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MTMD_BATCH_MAX_TOKENS"));
-    add_opt(common_arg(
-        {"--video-fps"}, "N",
-        string_format("target video frame rate (default: %.1f)", params.video_fps),
-        [](common_params & params, const std::string & value) {
-            params.video_fps = std::stof(value);
-        }
-    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_VIDEO_FPS"));
-    add_opt(common_arg(
-        {"--video-timestamp-interval"}, "N",
-        string_format("interval in milliseconds between text timestamps (default: %" PRId64 ")", params.video_timestamp_interval_ms),
-        [](common_params & params, int value) {
-            params.video_timestamp_interval_ms = value;
-        }
-    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_VIDEO_TIMESTAMP_INTERVAL"));
-    add_opt(common_arg(
-        {"--video-ffmpeg-dir"}, "DIR",
-        "path to the directory containing ffmpeg and ffprobe (default: search in PATH)",
-        [](common_params & params, const std::string & value) {
-            params.video_ffmpeg_bin_dir = value;
-        }
-    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_VIDEO_FFMPEG_DIR"));
     if (params.is_gen_docs || llama_supports_rpc()) {
         add_opt(common_arg(
             {"--rpc"}, "SERVERS",
@@ -2739,19 +2699,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             else { throw std::invalid_argument("invalid value"); }
         }
     ).set_env("LLAMA_ARG_LOAD_MODE"));
-    add_opt(common_arg(
-        {"-lzm", "--lazy-mode"}, "MODE",
-        "on-demand reading of certain tensors, for example per-layer embeddings (default: auto)\n"
-        "- on: read the rows of such tensors from disk on demand instead of keeping them resident (requires mmap)\n"
-        "- auto: on, but only for tensors larger than 4 GiB\n"
-        "- off: always keep them resident",
-        [](common_params & params, const std::string & value) {
-            /**/ if (value == "on")   { params.lazy_mode = LLAMA_LAZY_MODE_ON;   }
-            else if (value == "auto") { params.lazy_mode = LLAMA_LAZY_MODE_AUTO; }
-            else if (value == "off")  { params.lazy_mode = LLAMA_LAZY_MODE_OFF;  }
-            else { throw std::invalid_argument("invalid value"); }
-        }
-    ).set_env("LLAMA_ARG_LAZY_MODE"));
     add_opt(common_arg(
         {"--numa"}, "TYPE",
         "attempt optimizations that help on some NUMA systems\n"
@@ -2803,20 +2750,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             if (value < 0) {
                 throw std::invalid_argument("invalid value");
             }
-            llm_add_n_cpu_ffn_overrides(value, LLM_FFN_EXPS_REGEX, params.tensor_buft_overrides);
+            for (int i = 0; i < value; ++i) {
+                // keep strings alive and avoid leaking memory by storing them in a static vector
+                static std::list<std::string> buft_overrides;
+                buft_overrides.push_back(llm_ffn_exps_block_regex(i));
+                params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), ggml_backend_cpu_buffer_type()});
+            }
         }
     ).set_env("LLAMA_ARG_N_CPU_MOE"));
-    add_opt(common_arg(
-        {"-ncffn", "--n-cpu-ffn"}, "N",
-        "keep the dense FFN weights of the first N layers in the CPU\n"
-        "(dense models; for MoE expert weights use --n-cpu-moe)",
-        [](common_params & params, int value) {
-            if (value < 0) {
-                throw std::invalid_argument("invalid value");
-            }
-            llm_add_n_cpu_ffn_overrides(value, LLM_FFN_DENSE_REGEX, params.tensor_buft_overrides);
-        }
-    ).set_env("LLAMA_ARG_N_CPU_FFN"));
     GGML_ASSERT(params.n_gpu_layers < 0); // string_format would need to be extended for a default >= 0
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
@@ -3564,10 +3505,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                     LOG_WRN("Setting 'enable_thinking' via --chat-template-kwargs is deprecated. "
                             "Use --reasoning on / --reasoning off instead.\n");
                 }
-                if (item.key() == "preserve_reasoning") {
-                    LOG_WRN("Setting 'preserve_reasoning' via --chat-template-kwargs is deprecated. "
-                            "Use --reasoning-preserve / --no-reasoning-preserve instead.\n");
-                }
                 params.default_template_kwargs[item.key()] = item.value().dump();
             }
         }
@@ -3758,7 +3695,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"--reasoning-preserve"},
         {"--no-reasoning-preserve"},
-        "preserve reasoning trace in the full history, not just the last assistant message (default: enabled)\n"
+        "preserve reasoning trace in the full history, not just the last assistant message (default: template default)\n"
         "compatible with certain templates having 'supports_preserve_reasoning' capability\n"
         "example: https://docs.z.ai/guides/capabilities/thinking-mode#preserved-thinking",
         [](common_params & params, bool value) {
@@ -3767,7 +3704,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             } else {
                 params.default_template_kwargs["preserve_reasoning"] = "false";
             }
-            params.preserve_reasoning_specified = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_REASONING_PRESERVE"));
     add_opt(common_arg(
@@ -3907,14 +3843,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             common_log_set_file(common_log_main(), value.c_str());
         }
     ).set_env("LLAMA_ARG_LOG_FILE"));
-    add_opt(common_arg(
-        {"--log-jsonl"},
-        {"--no-log-jsonl"},
-        "Log as JSONL (one JSON object per line) to stdout, this also disables colored logging (default: disabled)",
-        [](common_params &, bool value) {
-            common_log_set_jsonl(common_log_main(), value);
-        }
-    ).set_env("LLAMA_ARG_LOG_JSONL"));
     add_opt(common_arg(
         {"--log-prompts-dir"}, "PATH",
         "Log prompts to directory (auto-created if not present; only used for debugging, default: disabled)",
@@ -4156,7 +4084,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             if (value < 0) {
                 throw std::invalid_argument("invalid value");
             }
-            llm_add_n_cpu_ffn_overrides(value, LLM_FFN_EXPS_REGEX, params.speculative.draft.tensor_buft_overrides);
+            for (int i = 0; i < value; ++i) {
+                static std::list<std::string> buft_overrides_draft;
+                buft_overrides_draft.push_back(llm_ffn_exps_block_regex(i));
+                params.speculative.draft.tensor_buft_overrides.push_back({buft_overrides_draft.back().c_str(), ggml_backend_cpu_buffer_type()});
+            }
         }
     ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_N_CPU_MOE"));
 
@@ -4177,38 +4109,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.draft.n_min = value;
         }
     ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_N_MIN"));
-    add_opt(common_arg(
-        {"--spec-synth-len"}, "L",
-        "target mean synthetic acceptance length, including the target token (benchmarking only)",
-        [](common_params & params, const std::string & value) {
-            const std::string text = string_strip(value);
-            size_t pos = 0;
-            const double length = std::stod(text, &pos);
-            if (pos != text.size() || length == -1.0) {
-                throw std::invalid_argument("invalid value");
-            }
-            params.speculative.synth_len = length;
-        }
-    ).set_spec().set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_SYNTH_LEN"));
-    add_opt(common_arg(
-        {"--spec-synth-rates"}, "P0,P1,...",
-        "comma-separated unconditional per-position synthetic acceptance probabilities (benchmarking only)",
-        [](common_params & params, const std::string & value) {
-            const auto values = string_split<std::string>(value, ',');
-            std::vector<double> rates;
-            rates.reserve(values.size());
-            for (const auto & raw : values) {
-                const std::string text = string_strip(raw);
-                size_t pos = 0;
-                const double rate = std::stod(text, &pos);
-                if (pos != text.size()) {
-                    throw std::invalid_argument("invalid value");
-                }
-                rates.push_back(rate);
-            }
-            params.speculative.synth_rates = std::move(rates);
-        }
-    ).set_spec().set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_SYNTH_RATES"));
 
     add_opt(common_arg(
         {"--spec-draft-p-split", "--draft-p-split"}, "P",
@@ -4235,7 +4135,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_BACKEND_SAMPLING"));
     add_opt(common_arg(
         {"--spec-draft-device", "-devd", "--device-draft"}, "<dev1,dev2,..>",
-        "comma-separated list of devices to use for offloading the draft model (none = don't offload, default: follows --device)\n"
+        "comma-separated list of devices to use for offloading the draft model (none = don't offload)\n"
         "use --list-devices to see a list of available devices",
         [](common_params & params, const std::string & value) {
             params.speculative.draft.devices = parse_device_list(value);

common/build-info.cpp.in

@@ -29,7 +29,7 @@ const char * llama_build_info(void) {
     return s.c_str();
 }
 
-void llama_print_build_info(const char * llama_version, FILE * stream) {
-    fprintf(stream, "version: %s (build %d, commit %s)\n", llama_version, llama_build_number(), llama_commit());
-    fprintf(stream, "built with %s for %s\n", llama_compiler(), llama_build_target());
+void llama_print_build_info(const char * llama_version) {
+    fprintf(stderr, "version: %s (build %d, commit %s)\n", llama_version, llama_build_number(), llama_commit());
+    fprintf(stderr, "built with %s for %s\n", llama_compiler(), llama_build_target());
 }

common/build-info.h

@@ -1,7 +1,5 @@
 #pragma once
 
-#include <cstdio>
-
 int llama_build_number(void);
 
 const char * llama_commit(void);
@@ -10,4 +8,4 @@ const char * llama_compiler(void);
 const char * llama_build_target(void);
 const char * llama_build_info(void);
 
-void llama_print_build_info(const char *, FILE * = stderr);
+void llama_print_build_info(const char *);

common/common.cpp

@@ -402,11 +402,10 @@ void common_params_print_info(const common_params & params, bool print_devices)
 #endif
     COM_TRC("%s: build %d (%s) with %s for %s%s\n", __func__, llama_build_number(), llama_commit(), llama_compiler(), llama_build_target(), build_type);
 
-    const int verbosity = common_log_get_verbosity_thold();
-    COM_INF("%s: verbosity = %d (adjust with the `-lv N` CLI arg)\n", __func__, verbosity);
+    COM_INF("%s: verbosity = %d (adjust with the `-lv N` CLI arg)\n", __func__, common_log_get_verbosity_thold());
 
     // device enumeration creates a primary context on CUDA backends, skip it when the caller does not own any device
-    if (print_devices && verbosity >= LOG_LEVEL_TRACE) {
+    if (print_devices) {
         COM_TRC("%s", "device_info:\n");
         for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
             auto * dev = ggml_backend_dev_get(i);
@@ -1295,34 +1294,11 @@ common_init_result::common_init_result(common_params & params, bool model_only)
     if (params.fit_params) {
         COM_TRC("%s", "fitting params to device memory ...\n");
         COM_TRC("%s", "(for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)\n");
-
-        // the draft context is created from the same base params and follows the main context, fit both together
-        const bool has_draft = params.speculative.has_dft();
-        const bool spec_mtp  = std::find(params.speculative.types.begin(), params.speculative.types.end(),
-            COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
-
-        common_params params_dft = common_base_params_to_speculative(params);
-
-        auto mparams_dft = common_model_params_to_llama(params_dft);
-        auto cparams_dft = common_context_params_to_llama(params_dft);
-        if (spec_mtp) {
-            cparams_dft.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
-        }
-        cparams_dft.n_rs_seq = 0;
-
-        const common_fit_extra_model extra = {
-            /*.path_model   =*/ params_dft.model.path.c_str(),
-            /*.mparams      =*/ &mparams_dft,
-            /*.cparams      =*/ &cparams_dft,
-            /*.shares_model =*/ !has_draft, // an MTP context runs on the weights of the main model
-        };
-
         common_fit_params(params.model.path.c_str(), &mparams, &cparams,
             params.tensor_split,
             params.tensor_buft_overrides.data(),
             params.fit_params_target.data(),
             params.fit_params_min_ctx,
-            has_draft || spec_mtp ? &extra : nullptr,
             params.verbosity >= LOG_LEVEL_DEBUG ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     }
 
@@ -1688,7 +1664,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.load_mode       = params.load_mode;
-    mparams.lazy_mode = params.lazy_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.check_tensors   = params.check_tensors;
     mparams.use_extra_bufts = !params.no_extra_bufts;

common/common.h

@@ -8,7 +8,6 @@
 #include "ggml.h"
 #include "llama.h"
 
-#include <list>
 #include <set>
 #include <sstream>
 #include <string>
@@ -270,7 +269,7 @@ struct common_params_sampling {
         COMMON_SAMPLER_TYPE_TEMPERATURE,
     };
 
-    common_grammar                      grammar;          // optional grammar constraint (user / output-format / tool-calls)
+    common_grammar              grammar;      // optional grammar constraint (user / output-format / tool-calls)
     bool                                grammar_lazy = false;
     std::vector<common_grammar_trigger> grammar_triggers; // optional triggers (for lazy grammars)
     std::set<llama_token>               preserved_tokens;
@@ -370,9 +369,6 @@ struct common_params_speculative_ngram_cache {
 struct common_params_speculative {
     std::vector<enum common_speculative_type> types = { COMMON_SPECULATIVE_TYPE_NONE };
 
-    double synth_len = -1.0;
-    std::vector<double> synth_rates;
-
     // used by Simple, MTP, Eagle3, etc. - all methods that require some kind of draft model
     common_params_speculative_draft draft;
 
@@ -387,10 +383,6 @@ struct common_params_speculative {
         return !draft.mparams.empty();
     }
 
-    bool has_synth() const {
-        return synth_len != -1.0 || !synth_rates.empty();
-    }
-
     uint32_t need_n_rs_seq() const {
         bool needs_rs_seq = std::any_of(types.begin(), types.end(), [&](auto t) {
             return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP || t == COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3 || t == COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH || t == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK;
@@ -483,8 +475,6 @@ struct common_params {
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
     enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_AUTO; // how to load the model
 
-    enum llama_lazy_mode lazy_mode = LLAMA_LAZY_MODE_AUTO; // on-demand reading of tensors marked by the arch
-
     common_cpu_params cpuparams;
     common_cpu_params cpuparams_batch;
 
@@ -599,11 +589,6 @@ struct common_params {
     int image_max_tokens = -1;
     int mtmd_batch_max_tokens = 1024;
 
-    // for video input
-    float       video_fps                   = 4.0f;
-    int64_t     video_timestamp_interval_ms = 5000;
-    std::string video_ffmpeg_bin_dir        = "";
-
     // finetune
     struct lr_opt lr;
     enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
@@ -627,7 +612,6 @@ struct common_params {
     bool    cache_prompt        = true;  // whether to enable prompt caching
     bool    cache_idle_slots    = true;  // save and clear idle slots upon starting a new task
     int32_t n_ctx_checkpoints   = 32;    // max number of context checkpoints per slot
-    int32_t kv_unified_per_slot = 0;     // max context per parallel slot; 0 = unset
     int32_t checkpoint_min_step = 8192;  // minimum spacing between context checkpoints
     int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
 
@@ -657,7 +641,6 @@ struct common_params {
     std::string ssl_file_cert = "";                                                                         // NOLINT
 
     std::map<std::string, std::string> default_template_kwargs;
-    bool preserve_reasoning_specified = false;
 
     // CLI params
     std::string server_base; // if set, connect to this server instead of starting a new one
@@ -1125,30 +1108,19 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 }
 
 //
-// FFN offload utils
+// MoE utils
 //
 
 const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate|gate_up)_(ch|)exps";
 
-const char * const LLM_FFN_DENSE_REGEX = "\\.ffn_(up|down|gate)\\.";
-
-inline std::string llm_ffn_block_regex(int idx, const char * ffn_regex) {
-    return string_format("blk\\.%d%s", idx, ffn_regex);
+inline std::string llm_ffn_exps_block_regex(int idx) {
+    return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
 }
 
 inline llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
     return { LLM_FFN_EXPS_REGEX, ggml_backend_cpu_buffer_type() };
 }
 
-inline void llm_add_n_cpu_ffn_overrides(int n, const char * ffn_regex, std::vector<llama_model_tensor_buft_override> & overrides) {
-    // keep strings alive and avoid leaking memory by storing them in a static list
-    static std::list<std::string> buft_override_strings;
-    for (int i = 0; i < n; ++i) {
-        buft_override_strings.push_back(llm_ffn_block_regex(i, ffn_regex));
-        overrides.push_back({buft_override_strings.back().c_str(), ggml_backend_cpu_buffer_type()});
-    }
-}
-
 //
 // training utils
 //

common/fit.cpp

@@ -178,7 +178,7 @@ common_device_memory_data_vec common_get_device_memory_data(
 static void common_params_fit_impl(
         const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
         float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
-        size_t * margins_s, uint32_t n_ctx_min, const common_fit_extra_model * extra, enum ggml_log_level log_level) {
+        size_t * margins_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
     if (mparams->split_mode == LLAMA_SPLIT_MODE_TENSOR) {
         throw common_params_fit_exception("llama_params_fit is not implemented for SPLIT_MODE_TENSOR, abort");
     }
@@ -191,92 +191,10 @@ static void common_params_fit_impl(
     uint32_t hp_nct = 0; // hparams.n_ctx_train
     uint32_t hp_nex = 0; // hparams.n_expert
 
-    // with non-unified kv, we need to take into account n_streams
-    // for example, if memory can hold more than model's trained context size, we must extend the n_ctx to hold enough n_streams
-    const uint32_t n_streams  = cparams->kv_unified ? 1 : std::max<uint32_t>(1, cparams->n_seq_max);
-    const bool     n_ctx_auto = cparams->n_ctx == 0;
-
-    dmds_t   dmds_extra;       // memory of the extra model, laid out on the devices of the main model
-    uint32_t n_ctx_extra = 0;  // context that memory was measured at
-
-    // the extra model competes for the same memory as the main model, add it to every measurement
-    // its memory is measured again whenever the context it follows changes
-    auto add_extra_memory = [&](dmds_t & dmds) {
-        if (extra == nullptr) {
-            return;
-        }
-
-        if (dmds_extra.empty() || n_ctx_extra != cparams->n_ctx) {
-            std::vector<ggml_backend_dev_t> devs_extra;
-            uint32_t ngl_extra = 0;
-            uint32_t nct_extra = 0;
-            uint32_t nex_extra = 0;
-
-            extra->cparams->n_ctx = cparams->n_ctx;
-
-            LOG_TRC("%s: getting device memory data for the extra model at a context size of %" PRIu32 ":\n",
-                __func__, cparams->n_ctx);
-
-            dmds_t measured;
-            try {
-                measured = common_get_device_memory_data_impl(
-                    extra->path_model, extra->mparams, extra->cparams, devs_extra, ngl_extra, nct_extra, nex_extra, log_level);
-            } catch (const std::runtime_error & e) {
-                // the extra model is optional, fit the main model alone rather than giving up
-                LOG_WRN("%s: failed to measure the memory of the extra model, fitting without it: %s\n", __func__, e.what());
-                dmds_extra = dmds_t(devs.size() + 1);
-                n_ctx_extra = cparams->n_ctx;
-                return;
-            }
-
-            dmds_extra = dmds_t(devs.size() + 1);
-            dmds_extra.back().mb = measured.back().mb;
-            for (size_t je = 0; je < devs_extra.size(); je++) {
-                for (size_t id = 0; id < devs.size(); id++) {
-                    if (devs_extra[je] == devs[id]) {
-                        dmds_extra[id].mb.model   += measured[je].mb.model;
-                        dmds_extra[id].mb.context += measured[je].mb.context;
-                        dmds_extra[id].mb.compute += measured[je].mb.compute;
-                        break;
-                    }
-                }
-            }
-            if (extra->shares_model) {
-                for (llama_device_memory_data & dmd : dmds_extra) {
-                    dmd.mb.model = 0;
-                }
-            }
-
-            n_ctx_extra = cparams->n_ctx;
-        }
-
-        for (size_t id = 0; id < dmds.size(); id++) {
-            dmds[id].mb.model   += dmds_extra[id].mb.model;
-            dmds[id].mb.context += dmds_extra[id].mb.context;
-            dmds[id].mb.compute += dmds_extra[id].mb.compute;
-        }
-    };
-
     // step 1: get data for default parameters and check whether any changes are necessary in the first place
 
     LOG_TRC("%s: getting device memory data for initial parameters:\n", __func__);
-    dmds_t dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-
-    // saturate instead of overflowing, this also preserves the UINT32_MAX sentinel of n_ctx_min:
-    const uint32_t n_ctx_max       = (uint32_t) std::min<uint64_t>(uint64_t(hp_nct)    * n_streams, UINT32_MAX);
-    const uint32_t n_ctx_min_total = (uint32_t) std::min<uint64_t>(uint64_t(n_ctx_min) * n_streams, UINT32_MAX);
-
-    // llama_context would use only hp_nct in total for n_ctx == 0, resolve the context before measuring anything else:
-    if (n_ctx_auto) {
-        cparams->n_ctx = n_ctx_max;
-        if (n_streams > 1) {
-            LOG_TRC("%s: context size unset and KV cache not unified -> using %" PRIu32 " for %" PRIu32 " sequences:\n",
-                __func__, n_ctx_max, n_streams);
-            dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-        }
-    }
-    add_extra_memory(dmds_full);
-
+    const dmds_t dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
     const size_t nd = devs.size(); // number of devices
 
     std::vector<int64_t> margins; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
@@ -389,8 +307,8 @@ static void common_params_fit_impl(
                     "%s: cannot meet free memory targets on all devices, need to use %" PRId64 " MiB less in total\n",
                     __func__, -global_surplus/MiB);
             }
-            if (n_ctx_auto) {
-                if (n_ctx_max > n_ctx_min_total) {
+            if (cparams->n_ctx == 0) {
+                if (hp_nct > n_ctx_min) {
                     int64_t sum_used_target = sum_free;
                     if (nd == 0) {
                         sum_used_target -= margins[0];
@@ -410,9 +328,8 @@ static void common_params_fit_impl(
                     }
 
                     int64_t sum_projected_used_min_ctx = 0;
-                    cparams->n_ctx = n_ctx_min_total;
-                    dmds_t dmds_min_ctx = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-                    add_extra_memory(dmds_min_ctx);
+                    cparams->n_ctx = n_ctx_min;
+                    const dmds_t dmds_min_ctx = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
                     if (nd == 0) {
                         sum_projected_used_min_ctx = dmds_min_ctx.back().mb.total();
                     } else {
@@ -422,16 +339,14 @@ static void common_params_fit_impl(
                     }
                     if (sum_used_target > sum_projected_used_min_ctx) {
                         // linear interpolation between minimum and maximum context size:
-                        cparams->n_ctx += (n_ctx_max - n_ctx_min_total) * (sum_used_target - sum_projected_used_min_ctx)
+                        cparams->n_ctx += (hp_nct - n_ctx_min) * (sum_used_target - sum_projected_used_min_ctx)
                             / (sum_projected_used - sum_projected_used_min_ctx);
-                        // round down context for CUDA backend, keep it divisible by the number of streams:
-                        const uint32_t align = 256 * n_streams;
-                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % align, n_ctx_min_total);
+                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
 
-                        const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (n_ctx_max - n_ctx_min_total);
-                        const int64_t memory_reduction = (n_ctx_max - cparams->n_ctx) * bytes_per_ctx;
+                        const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (hp_nct - n_ctx_min);
+                        const int64_t memory_reduction = (hp_nct - cparams->n_ctx) * bytes_per_ctx;
                         LOG_TRC("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
-                            __func__, n_ctx_max, cparams->n_ctx, memory_reduction/MiB);
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                         if (nd <= 1) {
                             LOG_TRC("%s: entire model can be fit by reducing context\n", __func__);
                             return;
@@ -440,14 +355,14 @@ static void common_params_fit_impl(
                     } else {
                         const int64_t memory_reduction = sum_projected_used - sum_projected_used_min_ctx;
                         LOG_TRC("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
-                            __func__, n_ctx_max, cparams->n_ctx, memory_reduction/MiB);
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                     }
                 } else {
                     if (n_ctx_min == UINT32_MAX) {
-                        LOG_TRC("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, n_ctx_max);
+                        LOG_TRC("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct);
                     } else {
                         LOG_TRC("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
-                            __func__, n_ctx_max, n_ctx_min_total);
+                            __func__, hp_nct, n_ctx_min);
                     }
                 }
             } else {
@@ -592,9 +507,8 @@ static void common_params_fit_impl(
         llama_model_params mparams_copy = *mparams;
         set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy);
 
-        dmds_t dmd_nl = common_get_device_memory_data_impl(
+        const dmds_t dmd_nl = common_get_device_memory_data_impl(
             path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-        add_extra_memory(dmd_nl);
 
         LOG_TRC("%s: memory for test allocation by device:\n", func_name);
         for (size_t id = 0; id < nd; id++) {
@@ -621,9 +535,8 @@ static void common_params_fit_impl(
         mparams->tensor_buft_overrides = tensor_buft_overrides;
 
         LOG_TRC("%s: getting device memory data with all MoE tensors moved to system memory:\n", __func__);
-        dmds_t dmds_cpu_moe = common_get_device_memory_data_impl(
+        const dmds_t dmds_cpu_moe = common_get_device_memory_data_impl(
             path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-        add_extra_memory(dmds_cpu_moe);
 
         for (size_t id = 0; id < nd; id++) {
             global_surplus_cpu_moe += dmds_cpu_moe[id].free;
@@ -883,12 +796,11 @@ enum common_params_fit_status common_fit_params(
         llama_model_tensor_buft_override * tensor_buft_overrides,
         size_t * margins,
         uint32_t n_ctx_min,
-        const common_fit_extra_model * extra,
         ggml_log_level log_level) {
     const int64_t t0_us = llama_time_us();
     common_params_fit_status status = COMMON_PARAMS_FIT_STATUS_SUCCESS;
     try {
-        common_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, extra, log_level);
+        common_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, log_level);
         LOG_TRC("%s: successfully fit params to free device memory\n", __func__);
     } catch (const common_params_fit_exception & e) {
         LOG_WRN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());

common/fit.h

@@ -11,16 +11,6 @@ enum common_params_fit_status {
     COMMON_PARAMS_FIT_STATUS_ERROR   = 2, // a hard error occurred, e.g. because no model could be found at the specified path
 };
 
-// a second model that shares the devices of the main model, e.g. a draft model
-//   - its context follows the context of the main model, so its memory is measured again whenever that context changes
-//   - shares_model tells the fit that the weights are already counted in the main model, as for an MTP context
-struct common_fit_extra_model {
-    const char * path_model;
-    llama_model_params * mparams;
-    llama_context_params * cparams;
-    bool shares_model;
-};
-
 // fits mparams and cparams to free device memory (assumes system memory is unlimited)
 //   - returns true if the parameters could be successfully modified to fit device memory
 //   - this function is NOT thread safe because it modifies the global llama logger state
@@ -34,7 +24,6 @@ common_params_fit_status common_fit_params(
    llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
                              size_t * margins,               // margins of memory to leave per device in bytes
                            uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
-      const common_fit_extra_model * extra,                  // model to fit alongside the main one, nullptr if there is none
                      ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
 
 // print estimated memory to stdout

common/jinja/caps.cpp

@@ -117,7 +117,6 @@ caps caps_get(jinja::program & prog) {
 
     JJ_DEBUG("%s\n", ">>> Running capability check: typed content");
 
-    bool checks_for_string = false;
     static const std::string content_marker = "STRING_MARKER";
 
     // case: typed content support
@@ -137,10 +136,6 @@ caps caps_get(jinja::program & prog) {
         [&](context &, bool success, value & messages, value &, const std::string & rendered) {
             auto & content = messages->at(0)->at("content");
             caps_print_stats(content, "messages[0].content");
-            if (has_op(content, "test_is_string")) {
-                // checked if content is string
-                checks_for_string = true;
-            }
             bool used_as_array = has_op(content, "selectattr") || has_op(content, "array_access");
             if (used_as_array) {
                 // accessed as an array
@@ -156,33 +151,6 @@ caps caps_get(jinja::program & prog) {
         }
     );
 
-    if (checks_for_string) {
-        caps_try_execute(
-            prog,
-            [&]() {
-                // messages
-                return json::array({
-                    {
-                        {"role", "user"},
-                        {"content", json::array({
-                        })}
-                    }
-                });
-            },
-            nullptr, // ctx_fn
-            nullptr, // tools_fn
-            [&](context &, bool success, value & messages, value &, const std::string &) {
-                auto & content = messages->at(0)->at("content");
-                caps_print_stats(content, "messages[0].content");
-                bool used_as_array = has_op(content, "selectattr") || has_op(content, "array_access");
-                if (used_as_array && success) {
-                    // accessed as an array
-                    result.supports_typed_content = true;
-                }
-            }
-        );
-    }
-
     JJ_DEBUG("%s\n", ">>> Running capability check: system prompt");
 
     // case: system prompt support

common/jinja/runtime.cpp

@@ -412,18 +412,12 @@ value test_expression::execute_impl(context & ctx) {
         throw std::runtime_error("Invalid test expression");
     }
 
-    const std::string test_name = "test_is_" + test_id;
-    auto it = builtins.find(test_name);
-    JJ_DEBUG("Test expression %s '%s' %s (using function '%s')", operand->type().c_str(), test_id.c_str(), negate ? "(negate)" : "", test_name.c_str());
+    auto it = builtins.find("test_is_" + test_id);
+    JJ_DEBUG("Test expression %s '%s' %s (using function 'test_is_%s')", operand->type().c_str(), test_id.c_str(), negate ? "(negate)" : "", test_id.c_str());
     if (it == builtins.end()) {
         throw std::runtime_error("Unknown test '" + test_id + "'");
     }
 
-    if (ctx.is_get_stats) {
-        value_t::stats_t::mark_used(input);
-        input->stats.ops.insert(test_name);
-    }
-
     auto res = it->second(args);
 
     if (negate) {

common/json-schema-to-grammar.cpp

@@ -748,10 +748,6 @@ private:
             optional_props.push_back("*");
         }
 
-        if (required_props.empty() && optional_props.empty()) {
-            return "\"{\" space \"}\"";
-        }
-
         std::string rule = "\"{\" space ";
         for (size_t i = 0; i < required_props.size(); i++) {
             if (i > 0) {

common/json.cpp

@@ -78,21 +78,19 @@ common_json_value::common_json_value(const common_json & val) :
 common_json_value::common_json_value(common_json && val) :
     type(VAL_JSON), val_json(std::make_shared<common_json>(std::move(val))) {}
 
-// the ctors and get<T>() below are explicit specializations, giving strong symbols
-// an explicit instantiation is a weak symbol, dropped by some LTO builds (clang-cl)
 template <typename T>
-static std::shared_ptr<common_json> set_json(const std::set<T> & vals) {
+common_json_value::common_json_value(const std::set<T> & vals) : type(VAL_JSON) {
     common_json out = common_json::array();
 
     for (const auto & val : vals) {
         out.push_back(val);
     }
 
-    return std::make_shared<common_json>(std::move(out));
+    val_json = std::make_shared<common_json>(std::move(out));
 }
 
 // a set value is usable only for the types below
-#define COMMON_JSON_SET(...) template <> common_json_value::common_json_value(const std::set<__VA_ARGS__> & vals) : type(VAL_JSON), val_json(set_json(vals)) {}
+#define COMMON_JSON_SET(...) template common_json_value::common_json_value(const std::set<__VA_ARGS__> &);
 
 COMMON_JSON_SET(int)
 COMMON_JSON_SET(std::string)
@@ -100,45 +98,56 @@ COMMON_JSON_SET(std::string)
 #undef COMMON_JSON_SET
 
 template <typename T>
-static std::shared_ptr<common_json> map_json(const T & vals) {
+common_json_value::common_json_value(const std::map<std::string, T> & vals) : type(VAL_JSON) {
     common_json out = common_json::object();
 
     for (const auto & val : vals) {
         out.set({ val.first, val.second });
     }
 
-    return std::make_shared<common_json>(std::move(out));
+    val_json = std::make_shared<common_json>(std::move(out));
 }
 
 // a map value is usable only for the types below
-#define COMMON_JSON_MAP(...) template <> common_json_value::common_json_value(const std::map<std::string, __VA_ARGS__> & vals) : type(VAL_JSON), val_json(map_json(vals)) {}
+#define COMMON_JSON_MAP(...) template common_json_value::common_json_value(const std::map<std::string, __VA_ARGS__> &);
 
 COMMON_JSON_MAP(bool)
 COMMON_JSON_MAP(std::string)
 
 #undef COMMON_JSON_MAP
 
+template <typename T>
+common_json_value::common_json_value(const std::unordered_map<std::string, T> & vals) : type(VAL_JSON) {
+    common_json out = common_json::object();
+
+    for (const auto & val : vals) {
+        out.set({ val.first, val.second });
+    }
+
+    val_json = std::make_shared<common_json>(std::move(out));
+}
+
 // an unordered map value is usable only for the types below
-#define COMMON_JSON_UMAP(...) template <> common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> & vals) : type(VAL_JSON), val_json(map_json(vals)) {}
+#define COMMON_JSON_UMAP(...) template common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> &);
 
 COMMON_JSON_UMAP(size_t)
 
 #undef COMMON_JSON_UMAP
 
 template <typename T>
-static std::shared_ptr<common_json> vec_json(const std::vector<T> & vals) {
+common_json_value::common_json_value(const std::vector<T> & vals) : type(VAL_JSON) {
     common_json out = common_json::array();
 
     for (const auto & val : vals) {
         out.push_back(val);
     }
 
-    return std::make_shared<common_json>(std::move(out));
+    val_json = std::make_shared<common_json>(std::move(out));
 }
 
 // a vector value is usable only for the types below
 // note: std::vector<bool> is not here, its proxy reference does not convert
-#define COMMON_JSON_VEC(...) template <> common_json_value::common_json_value(const std::vector<__VA_ARGS__> & vals) : type(VAL_JSON), val_json(vec_json(vals)) {}
+#define COMMON_JSON_VEC(...) template common_json_value::common_json_value(const std::vector<__VA_ARGS__> &);
 
 COMMON_JSON_VEC(int)
 COMMON_JSON_VEC(unsigned char)
@@ -395,6 +404,10 @@ common_json::items_view common_json::items() const {
     return items_view(const_cast<common_json *>(this), size());
 }
 
+template <typename T> T common_json::get() const {
+    return guard([&] { return as_json(this).get<T>(); });
+}
+
 // the backing library cannot build a common_json, so this one is just a copy
 template <> common_json common_json::get<common_json>() const {
     return *this;
@@ -402,7 +415,7 @@ template <> common_json common_json::get<common_json>() const {
 
 // get<T>() is usable only for the types below
 
-#define COMMON_JSON_GET(...) template <> __VA_ARGS__ common_json::get<__VA_ARGS__>() const { return guard([&] { return as_json(this).get<__VA_ARGS__>(); }); }
+#define COMMON_JSON_GET(...) template __VA_ARGS__ common_json::get<__VA_ARGS__>() const;
 
 COMMON_JSON_GET(bool)
 COMMON_JSON_GET(int)
@@ -422,12 +435,3 @@ COMMON_JSON_GET(std::vector<size_t>)
 COMMON_JSON_GET(std::unordered_map<std::string, size_t>)
 
 #undef COMMON_JSON_GET
-
-// must stay below the get<std::string> specialization
-common_json::operator std::string() const {
-    return get<std::string>();
-}
-
-std::string common_json::value(const std::string & key, const char * def) const {
-    return contains(key) ? at(key).get<std::string>() : std::string(def);
-}

common/json.h

@@ -221,14 +221,16 @@ class common_json {
     // implicit get<T>() for plain values, so they can be assigned to their C++ type directly
     // note: kept to this short list on purpose, a wider one makes j["key"] ambiguous
     // note: a numeric one would make "str = json;" ambiguous, a number converts to char too
-    operator std::string() const;
+    operator std::string() const { return get<std::string>(); }
 
     template <typename T>
     T value(const std::string & key, T def) const {
         return contains(key) ? at(key).get<T>() : def;
     }
 
-    std::string value(const std::string & key, const char * def) const;
+    std::string value(const std::string & key, const char * def) const {
+        return contains(key) ? at(key).get<std::string>() : std::string(def);
+    }
 
     // a JSON default needs no get<T>(), it is already the right type
     common_json value(const std::string & key, const common_json & def) const {

common/log.cpp

@@ -1,6 +1,5 @@
 #include "common.h"
 #include "log.h"
-#include "json.h"
 
 #include <chrono>
 #include <condition_variable>
@@ -67,17 +66,6 @@ static const char* g_col[] = {
     "",
 };
 
-static const char * level_str(enum ggml_log_level level) {
-    switch (level) {
-        case GGML_LOG_LEVEL_DEBUG: return "debug";
-        case GGML_LOG_LEVEL_INFO:  return "info";
-        case GGML_LOG_LEVEL_WARN:  return "warn";
-        case GGML_LOG_LEVEL_ERROR: return "error";
-        case GGML_LOG_LEVEL_CONT:  return "cont";
-        default:                   return "none";
-    }
-}
-
 struct common_log_entry {
     enum ggml_log_level level {GGML_LOG_LEVEL_INFO};
 
@@ -86,7 +74,6 @@ struct common_log_entry {
     int64_t timestamp { 0 };
     bool is_end       { false }; // signals the worker thread to stop
     bool prefix       { false };
-    bool jsonl        { false };
 
     common_log_entry(size_t size = 256) : msg(size) { }
 
@@ -101,23 +88,11 @@ struct common_log_entry {
 
             fcur = stdout;
 
-            if (level != GGML_LOG_LEVEL_NONE && !jsonl) {
+            if (level != GGML_LOG_LEVEL_NONE) {
                 fcur = stderr;
             }
         }
 
-        if (jsonl) {
-            common_json obj = {
-                {"type",  "log"},
-                {"time",  timestamp},
-                {"level", level_str(level)},
-                {"msg",   msg.data()},
-            };
-            fprintf(fcur, "%s\n", obj.dump_safe().c_str());
-            fflush(fcur);
-            return;
-        }
-
         if (level != GGML_LOG_LEVEL_NONE && level != GGML_LOG_LEVEL_CONT && prefix) {
             if (timestamp) {
                 // [M.s.ms.us]
@@ -156,7 +131,6 @@ struct common_log {
         file       = nullptr;
         prefix     = false;
         timestamps = false;
-        jsonl      = false;
         running    = false;
         t_start    = t_us();
 
@@ -184,7 +158,6 @@ private:
 
     bool prefix;
     bool timestamps;
-    bool jsonl;
     bool running;
 
     int64_t t_start;
@@ -273,7 +246,6 @@ public:
         entry.is_end    = false;
         entry.level     = level;
         entry.prefix    = prefix;
-        entry.jsonl     = jsonl;
         entry.timestamp = 0;
         if (timestamps) {
             entry.timestamp = t_us() - t_start;
@@ -388,12 +360,6 @@ public:
 
         this->timestamps = timestamps;
     }
-
-    void set_jsonl(bool jsonl) {
-        std::lock_guard<std::mutex> lock(mtx);
-
-        this->jsonl = jsonl;
-    }
 };
 
 //
@@ -467,16 +433,12 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }
 
-void common_log_set_jsonl(struct common_log * log, bool jsonl) {
-    log->set_jsonl(jsonl);
-}
-
 void common_log_flush(struct common_log * log) {
     log->pause();
     log->resume();
 }
 
-int common_log_get_verbosity(enum ggml_log_level level) {
+static int common_get_verbosity(enum ggml_log_level level) {
     switch (level) {
         case GGML_LOG_LEVEL_DEBUG: return LOG_LEVEL_DEBUG;
         case GGML_LOG_LEVEL_INFO:  return LOG_LEVEL_TRACE;
@@ -490,7 +452,7 @@ int common_log_get_verbosity(enum ggml_log_level level) {
 }
 
 void common_log_default_callback(enum ggml_log_level level, const char * text, void * /*user_data*/) {
-    auto verbosity = common_log_get_verbosity(level);
+    auto verbosity = common_get_verbosity(level);
     if (verbosity <= common_log_verbosity_thold) {
         common_log_add(common_log_main(), level, "%s", text);
     }

common/log.h

@@ -43,8 +43,6 @@ int  common_log_get_verbosity_thold(void);
 
 void common_log_set_verbosity_thold(int verbosity); // not thread-safe
 
-int common_log_get_verbosity(enum ggml_log_level level);
-
 void common_log_default_callback(enum ggml_log_level level, const char * text, void * user_data);
 
 // the common_log uses an internal worker thread to print/write log messages
@@ -91,7 +89,6 @@ void common_log_set_file      (struct common_log * log, const char * file); // n
 void common_log_set_colors    (struct common_log * log, log_colors colors); // not thread-safe
 void common_log_set_prefix    (struct common_log * log, bool prefix);       // whether to output prefix to each log
 void common_log_set_timestamps(struct common_log * log, bool timestamps);   // whether to output timestamps in the prefix
-void common_log_set_jsonl     (struct common_log * log, bool jsonl);        // print each log as a JSON object on one line, not thread-safe
 void common_log_flush         (struct common_log * log);                    // flush all pending log messages
 
 // helper macros for logging

common/parsers/cohere2moe.cpp

@@ -1,150 +0,0 @@
-#include "parsers.h"
-
-// Cohere2 MoE (a.k.a. "North Code") parser.
-//
-// The assistant turn is fully marker-wrapped:
-//   <|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>
-//     <|START_THINKING|>{reasoning}<|END_THINKING|>
-//     then EITHER content:    <|START_TEXT|>{content}<|END_TEXT|>
-//          OR     tool calls: <|START_ACTION|>[
-//                                 {"tool_call_id": "0", "tool_name": "f", "parameters": {...}}, ...
-//                             ]<|END_ACTION|>
-//   <|END_OF_TURN_TOKEN|>
-//
-// The generation prompt forces a leading <|START_THINKING|> (when reasoning is enabled, which is
-// the template default), so the model's output continues from *inside* the thinking block. The
-// parser literal therefore only covers the stable <|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|> prefix
-// and the reasoning rule consumes the <|START_THINKING|> ... <|END_THINKING|> markers itself,
-// regardless of whether they came from the generation prompt or the generated text.
-common_chat_params common_chat_params_init_cohere2moe(const common_chat_template &          tmpl,
-                                                              const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    const std::string TURN_START    = "<|START_OF_TURN_TOKEN|>";
-    const std::string TURN_END      = "<|END_OF_TURN_TOKEN|>";
-    const std::string CHATBOT       = "<|CHATBOT_TOKEN|>";
-    const std::string USER          = "<|USER_TOKEN|>";
-    const std::string SYSTEM        = "<|SYSTEM_TOKEN|>";
-    const std::string THINK_START   = "<|START_THINKING|>";
-    const std::string THINK_END     = "<|END_THINKING|>";
-    const std::string TEXT_START    = "<|START_TEXT|>";
-    const std::string TEXT_END      = "<|END_TEXT|>";
-    const std::string ACTION_START  = "<|START_ACTION|>";
-    const std::string ACTION_END    = "<|END_ACTION|>";
-    const std::string RESULT_START  = "<|START_TOOL_RESULT|>";
-    const std::string RESULT_END    = "<|END_TOOL_RESULT|>";
-
-    // Stable prefix of the generation prompt that precedes the (forced) <|START_THINKING|> marker.
-    const std::string GEN_PREFIX = TURN_START + CHATBOT;
-
-    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking  = true;
-    data.thinking_start_tag = THINK_START;
-    data.thinking_end_tags  = {THINK_END};
-    data.preserved_tokens   = {
-        TURN_START, TURN_END, CHATBOT, USER, SYSTEM,
-        THINK_START, THINK_END,
-        TEXT_START, TEXT_END,
-        ACTION_START, ACTION_END,
-        RESULT_START, RESULT_END,
-    };
-
-    // Declare per-role message delimiters. Tool results are rendered with the
-    // system token followed by <|START_TOOL_RESULT|>, so the "tool" delimiter must be listed before
-    // the plain "system" one (it is a strict superset, and the role split tries delimiters in order).
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, GEN_PREFIX },
-        { COMMON_CHAT_ROLE_USER,      TURN_START + USER },
-        { COMMON_CHAT_ROLE_TOOL,      TURN_START + SYSTEM + RESULT_START },
-        { COMMON_CHAT_ROLE_SYSTEM,    TURN_START + SYSTEM },
-    };
-
-    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
-    auto has_response_format = inputs.json_schema.is_object() && !inputs.json_schema.empty();
-    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = GEN_PREFIX + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + TEXT_START + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.literal(GEN_PREFIX);
-        auto end               = p.end();
-
-        // The thinking block is always present (the generation prompt forces <|START_THINKING|>).
-        // When extracting reasoning, capture its body; otherwise keep the whole block (markers
-        // included) inline as content, matching reasoning_format=NONE conventions.
-        common_peg_parser reasoning = p.eps();
-        if (extract_reasoning) {
-            reasoning = p.optional(p.literal(THINK_START) +
-                                   p.reasoning(p.until_one_of({ THINK_END, TEXT_START, ACTION_START })) +
-                                   p.optional(p.literal(THINK_END)));
-        } else {
-            reasoning = p.optional(p.content(p.literal(THINK_START) +
-                                             p.until_one_of({ THINK_END, TEXT_START, ACTION_START }) +
-                                             p.optional(p.literal(THINK_END))));
-        }
-
-        auto text_content = has_response_format
-            ? p.literal(TEXT_START) +
-                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)) +
-                p.optional(p.literal(TEXT_END))
-            : p.literal(TEXT_START) + p.content(p.until(TEXT_END)) + p.optional(p.literal(TEXT_END));
-
-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-            return generation_prompt + reasoning + text_content + p.optional(p.literal(TURN_END)) + end;
-        }
-
-        auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-        // <|START_ACTION|>[ {"tool_call_id": "0", "tool_name": "f", "parameters": {...}}, ... ]<|END_ACTION|>
-        auto tool_calls = p.standard_json_tools(ACTION_START, ACTION_END, inputs.tools, inputs.parallel_tool_calls,
-                                                /* force_tool_calls = */ true,
-                                                /* name_key         = */ "tool_name",
-                                                /* args_key         = */ "parameters",
-                                                /* array_wrapped    = */ true,
-                                                /* function_is_key  = */ false,
-                                                /* call_id_key      = */ "",
-                                                /* gen_call_id_key  = */ "tool_call_id",
-                                                /* parameters_order = */ { "tool_call_id", "tool_name", "parameters" });
-
-        // Content and tool calls are mutually exclusive in this format.
-        common_peg_parser body = require_tools ? tool_calls : p.choice({ tool_calls, text_content });
-
-        return generation_prompt + reasoning + body + p.optional(p.literal(TURN_END)) + end;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = !has_response_format && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, ACTION_START }
-        };
-    }
-
-    return data;
-}

common/parsers/deepseek.cpp

@@ -1,287 +0,0 @@
-#include "parsers.h"
-
-// The DeepSeek V4 reference implementation renders consecutive tool results into a single
-// user block, ordered by the tool call order of the preceding assistant message (matched
-// by tool call id) rather than by the order they appear in the conversation.
-static json deepseek_v4_sort_tool_results(const json & messages) {
-    json adjusted = messages;
-    std::map<std::string, size_t> call_order;
-
-    for (size_t i = 0; i < adjusted.size();) {
-        const auto & msg  = adjusted[i];
-        const auto   role = msg.value("role", "");
-
-        if (role == "assistant" && msg.contains("tool_calls") &&
-                msg.at("tool_calls").is_array() && !msg.at("tool_calls").empty()) {
-            call_order.clear();
-            const auto & tool_calls = msg.at("tool_calls");
-            for (size_t idx = 0; idx < tool_calls.size(); idx++) {
-                auto id = tool_calls[idx].value("id", "");
-                if (!id.empty()) {
-                    call_order[id] = idx;
-                }
-            }
-            i++;
-            continue;
-        }
-
-        if (role != "user" && role != "tool") {
-            i++;
-            continue;
-        }
-
-        // collect a maximal run of user/tool messages - they render into one user block
-        std::vector<size_t> tool_positions;
-        size_t run_end = i;
-        for (; run_end < adjusted.size(); run_end++) {
-            const auto r = adjusted[run_end].value("role", "");
-            if (r == "tool") {
-                tool_positions.push_back(run_end);
-            } else if (r != "user") {
-                break;
-            }
-        }
-
-        if (tool_positions.size() > 1 && !call_order.empty()) {
-            std::vector<json> results;
-            results.reserve(tool_positions.size());
-            for (auto pos : tool_positions) {
-                results.push_back(adjusted[pos]);
-            }
-            std::stable_sort(results.begin(), results.end(), [&](const json & a, const json & b) {
-                const auto order = [&](const json & m) {
-                    auto it = call_order.find(m.value("tool_call_id", ""));
-                    return it == call_order.end() ? (size_t) 0 : it->second;
-                };
-                return order(a) < order(b);
-            });
-            for (size_t k = 0; k < tool_positions.size(); k++) {
-                adjusted[tool_positions[k]] = std::move(results[k]);
-            }
-        }
-
-        i = run_end;
-    }
-
-    return adjusted;
-}
-
-common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_template &    tmpl,
-                                                                 const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    // V4 uses the same DSML markup as V3.2, but names the tool call block "tool_calls"
-    // instead of "function_calls", renders tool results in tool call order and its
-    // non-thinking generation prompt ends with a bare </think> instead of an empty
-    // <think></think> pair.
-    const bool is_v4 = tmpl.source().find("function_calls") == std::string::npos;
-
-    std::optional<json> adjusted_messages;
-    if (is_v4) {
-        adjusted_messages = deepseek_v4_sort_tool_results(inputs.messages);
-    }
-
-    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
-    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
-    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
-
-    std::optional<json> additional_context;
-    if (is_v4 && has_response_format) {
-        additional_context = json{ { "response_format", inputs.json_schema } };
-    }
-
-    const std::string DSML         = "｜DSML｜";
-    const std::string THINK_START  = "<think>";
-    const std::string THINK_END    = "</think>";
-    const std::string TC_BLOCK     = is_v4 ? "tool_calls" : "function_calls";
-    const std::string FC_START     = "<" + DSML + TC_BLOCK + ">";
-    const std::string FC_END       = "</" + DSML + TC_BLOCK + ">";
-    const std::string INVOKE_START = "<" + DSML + "invoke";
-    const std::string INVOKE_END   = "</" + DSML + "invoke>";
-    const std::string PARAM_START  = "<" + DSML + "parameter";
-    const std::string PARAM_END    = "</" + DSML + "parameter>";
-    const std::string GEN_PROMPT   = "<｜Assistant｜>";
-    const std::string TC_SEPARATOR = "\n\n";
-
-    data.prompt = common_chat_template_direct_apply_impl(
-        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(
-        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
-    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking  = true;
-    data.thinking_start_tag = THINK_START;
-    data.thinking_end_tags  = {THINK_END, FC_START};
-    data.preserved_tokens   = {
-        DSML,
-        THINK_START,
-        THINK_END,
-    };
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        if (is_v4 && msg.reasoning_content.empty()) {
-            data.generation_prompt = GEN_PROMPT + THINK_END;
-            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-                data.generation_prompt += msg.render_content();
-            }
-        } else {
-            data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-                data.generation_prompt += THINK_END + msg.render_content();
-            }
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    bool require_tools   = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-    bool has_tool_calls = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.literal(GEN_PROMPT);
-        auto end               = p.end();
-
-        // build tool call section first since we might need it in reasoning
-        auto tool_choice = p.choice();
-        if (has_tool_calls) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                std::string  name     = function.at("name");
-                auto         params   = function.contains("parameters") ? function.at("parameters") : json::object();
-                const auto & props    = params.contains("properties") ? params.at("properties") : json::object();
-
-                std::set<std::string> required;
-                if (params.contains("required")) {
-                    required = params.at("required").get<std::set<std::string>>();
-                }
-
-                auto schema_info = common_schema_info();
-                schema_info.resolve_refs(params);
-
-                std::vector<common_peg_parser> required_parsers;
-                std::vector<common_peg_parser> optional_parsers;
-                for (const auto & [param_name, param_schema] : props.items()) {
-                    bool is_required = required.find(param_name) != required.end();
-                    bool is_string   = schema_info.resolves_to_string(param_schema);
-
-                    auto arg = p.tool_arg(
-                        p.tool_arg_open(p.literal(PARAM_START + " name=\"") + p.tool_arg_name(p.literal(param_name)) +
-                                        p.literal("\" string=\"" + std::string(is_string ? "true" : "false") + "\">")) +
-                        (is_string ?
-                             p.tool_arg_string_value(p.until(PARAM_END)) :
-                             p.tool_arg_json_value(p.schema(p.json(), "tool-" + name + "-arg-" + param_name + "-schema",
-                                                            param_schema, false))) +
-                        p.tool_arg_close(p.literal(PARAM_END)));
-
-                    auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
-                    if (is_required) {
-                        required_parsers.push_back(named_arg);
-                    } else {
-                        optional_parsers.push_back(named_arg);
-                    }
-                }
-
-                common_peg_parser args_seq = p.eps();
-                for (size_t i = 0; i < required_parsers.size(); i++) {
-                    if (i > 0) {
-                        args_seq = args_seq + p.space();
-                    }
-                    args_seq = args_seq + required_parsers[i];
-                }
-
-                if (!optional_parsers.empty()) {
-                    common_peg_parser any_opt = p.choice();
-                    for (const auto & opt : optional_parsers) {
-                        any_opt |= opt;
-                    }
-                    args_seq = args_seq + p.repeat(p.space() + any_opt, 0, -1);
-                }
-
-                common_peg_parser invoke_body = args_seq;
-                auto              func_parser = p.tool(p.tool_open(p.literal(INVOKE_START + " name=\"") +
-                                                                   p.tool_name(p.literal(name)) + p.literal("\">\n")) +
-                                                       invoke_body + p.space() + p.tool_close(p.literal(INVOKE_END)));
-
-                tool_choice |= p.rule("tool-" + name, func_parser);
-            });
-        }
-
-        common_peg_parser tool_calls = p.eps();
-        if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call",
-                p.literal(FC_START) + p.space() + tool_choice +
-                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
-        } else {
-            tool_calls = p.trigger_rule("tool-call",
-                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
-        }
-
-        auto reasoning = p.eps();
-        auto reasoning_with_tc = p.eps();
-        auto obligatory_tool_calls = tool_calls;
-        bool allow_reasoning_with_tc = false;
-
-        if (!require_tools) {
-            tool_calls = p.optional(tool_calls);
-        }
-
-        if (extract_reasoning && inputs.enable_thinking) {
-            reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
-            reasoning_with_tc = THINK_START +
-                p.reasoning(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START, THINK_END })) +
-                p.space() + obligatory_tool_calls;
-            allow_reasoning_with_tc = true;
-        } else if (extract_reasoning) {
-            // Thinking disabled but reasoning extraction requested: the generation prompt
-            // contains an empty <think></think> pair (V3.2) or a bare </think> (V4) that
-            // must still be consumed.
-            reasoning = is_v4
-                ? p.optional(p.literal(THINK_END))
-                : p.optional(p.literal(THINK_START) + p.until(THINK_END) + p.literal(THINK_END));
-        }
-
-        if (has_response_format) {
-            auto response_format = p.rule("response-format",
-                p.literal("```json") + p.space() +
-                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)) +
-                p.space() + p.literal("```"));
-            return generation_prompt + reasoning + response_format + end;
-        }
-
-        if (!has_tool_calls) {
-            return generation_prompt + reasoning + p.content(p.rest()) + end;
-        }
-
-        auto content_before_tools = p.negate(p.literal(THINK_START)) +
-            p.content(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START })) +
-            p.space();
-        return allow_reasoning_with_tc ? generation_prompt + (reasoning_with_tc | (reasoning + content_before_tools + tool_calls)) + end :
-            generation_prompt + reasoning + content_before_tools + tool_calls + end;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = has_tools && !require_tools;
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, FC_START },
-        };
-    }
-
-    return data;
-}

common/parsers/functionary-v3-2.cpp

@@ -1,101 +0,0 @@
-#include "parsers.h"
-
-// Functionary v3.2 - uses recipient-based format: >>>recipient\n{content}
-common_chat_params common_chat_params_init_functionary_v3_2(const common_chat_template &    tmpl,
-                                                                   const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.preserved_tokens  = {
-        ">>>all",
-    };
-
-    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
-    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-        data.generation_prompt = "<|start_header_id|>assistant<|end_header_id|>\n\n>>>all\n" + msg.render_content();
-        data.prompt += data.generation_prompt;
-    }
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        // Functionary v3.2 format:
-        // - Normal content: >>>all\n{content}
-        // - Tool calls: >>>function_name\n{json_args}
-        // Generation prompt ends with ">>>" so model outputs recipient immediately
-
-        // Build content parser for >>>all\n{content}
-        // When tools are present, content stops before the next ">>>" (tool call)
-        // When no tools, content goes until end
-        auto content_until_tool = p.literal("all\n") + p.content(p.until(">>>"));
-        auto content_until_end  = p.literal("all\n") + p.content(p.rest());
-        auto generation_prompt  = p.literal("<|start_header_id|>assistant<|end_header_id|>\n\n>>>");
-
-        // If no tools or tool_choice is NONE, just parse content
-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-            // When no tools, just match the prefix and capture everything after
-            return generation_prompt + content_until_end + p.end();
-        }
-
-        // Build tool call parsers for each available function
-        auto tool_choice = p.choice();
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            std::string  name     = function.at("name");
-            const auto & schema   = function.at("parameters");
-
-            // Tool format: >>>function_name\n{json_args}
-            auto tool_parser = p.tool(
-                p.tool_open(p.tool_name(p.literal(name)) + p.literal("\n")) +
-                p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema))
-            );
-
-            tool_choice |= p.rule("tool-" + name, tool_parser);
-        });
-
-        auto content_only = content_until_end;
-        auto tools_only = p.trigger_rule("tools", p.one_or_more(tool_choice));
-        auto content_and_tools = content_until_tool + tools_only;
-
-        auto ret = p.eps();
-        if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
-            if (inputs.parallel_tool_calls) {
-                ret = p.choice({ content_and_tools, tools_only }) + p.end();
-            } else {
-                ret = p.choice({ content_until_tool + tool_choice, tools_only }) + p.end();
-            }
-        } else if (inputs.parallel_tool_calls) {
-            ret = p.choice({ content_and_tools, content_only, tools_only }) + p.end();
-        } else {
-            auto content_and_tool = content_until_tool + tool_choice;
-            ret = p.choice({ content_and_tool, content_only, tool_choice }) + p.end();
-        }
-        return generation_prompt + ret;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
-
-        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        // Grammar trigger for when the model starts outputting a tool call
-        // (after the initial ">>>" in the generation prompt but recipient other than "all")
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, ">>>(?!all)" }
-        };
-    }
-
-    return data;
-}

common/parsers/gemma4.cpp

@@ -1,312 +0,0 @@
-#include "parsers.h"
-
-namespace workaround {
-
-// Gemma4 uses a custom tool_responses field instead of role:tool messages.
-//
-// This will transform a sequence of messages:
-//   assistant(tool_call+) -> tool+ -> assistant(content)
-//
-// Into a single assistant message containing a tool_responses field:
-//   assistant(content + tool_call + tool_responses)
-//
-// This is necessary for the Gemma4 chat template to properly format the prompt.
-// See https://ai.google.dev/gemma/docs/core/prompt-formatting-gemma4
-struct gemma4_model_turn_builder {
-    json & messages;
-    size_t pos;
-    json tool_calls = json::array();
-    json tool_responses = json::array();
-    json content;
-    json reasoning_content;
-
-    gemma4_model_turn_builder(json & msgs, size_t pos) : messages(msgs), pos(pos) {}
-
-    void collect() {
-        // Collect the first assistant message
-        auto & msg = messages[pos];
-        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
-            // According to the prompt formatting guide, we need to preserve reasoning_content
-            // between function calls. The current chat templates do not support this, but we will do it anyway.
-            reasoning_content = msg.at("reasoning_content");
-        }
-        for (auto & tc : msg.at("tool_calls")) {
-            tool_calls.push_back(tc);
-        }
-        pos++;
-
-        // Collect tool call results
-        while (pos < messages.size() && messages[pos].value("role", "") == "tool") {
-            collect_result(messages[pos]);
-            pos++;
-        }
-
-        // Check if the next assistant message is the final message
-        if (pos < messages.size() && messages[pos].value("role", "") == "assistant") {
-            auto & next = messages[pos];
-            if (!has_tool_calls(next) && has_content(next)) {
-                content = next.at("content");
-                pos++;
-            }
-        }
-    }
-
-    void collect_result(const json & curr) {
-        json response;
-        if (curr.contains("content")) {
-            const auto & content = curr.at("content");
-            if (content.is_string()) {
-                // Try to parse the content as JSON; fall back to raw string
-                try {
-                    response = json::parse(content.get<std::string>());
-                } catch (...) {
-                    response = content;
-                }
-            } else {
-                response = content;
-            }
-        }
-
-        std::string name;
-
-        // Match name with corresponding tool call
-        size_t idx = tool_responses.size();
-        if (idx < tool_calls.size()) {
-            auto & tc = tool_calls[idx];
-            if (tc.contains("function")) {
-                name = tc.at("function").value("name", "");
-            }
-        }
-
-        // Fallback to the tool call id
-        if (name.empty()) {
-            name = curr.value("tool_call_id", "");
-        }
-
-        tool_responses.push_back({{"name", name}, {"response", response}});
-    }
-
-    json build() {
-        collect();
-
-        json msg = {
-            {"role", "assistant"},
-            {"tool_calls", tool_calls},
-        };
-        if (!tool_responses.empty()) {
-            msg["tool_responses"] = tool_responses;
-        }
-        if (!content.is_null()) {
-            msg["content"] = content;
-        }
-        if (!reasoning_content.is_null()) {
-            msg["reasoning_content"] = reasoning_content;
-        }
-        return msg;
-    }
-
-    static bool has_content(const json & msg) {
-        if (!msg.contains("content") || msg.at("content").is_null()) {
-            return false;
-        }
-        const auto & content = msg.at("content");
-        if (content.is_string() && !content.get<std::string>().empty()) {
-            return true;
-        }
-        if (content.is_array() && !content.empty()) {
-            return true;
-        }
-        return false;
-    }
-
-    static bool has_tool_calls(const json & msg) {
-        return msg.contains("tool_calls") && msg.at("tool_calls").is_array() && !msg.at("tool_calls").empty();
-    }
-};
-
-void convert_tool_responses_gemma4(json & messages) {
-    json result = json::array();
-    size_t i = 0;
-
-    while (i < messages.size()) {
-        auto & msg = messages[i];
-
-        if (msg.value("role", "") != "assistant" || !msg.contains("tool_calls") ||
-            !msg.at("tool_calls").is_array() || msg.at("tool_calls").empty()) {
-            result.push_back(msg);
-            i++;
-            continue;
-        }
-
-        gemma4_model_turn_builder builder(messages, i);
-        result.push_back(builder.build());
-        i = builder.pos;
-    }
-
-    messages = result;
-}
-
-}
-
-common_chat_params common_chat_params_init_gemma4(const common_chat_template &    tmpl,
-                                                         const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
-
-    if (inputs.add_generation_prompt && string_ends_with(data.prompt, "<turn|>\n")) {
-        // This may happen if the model generates content + tool_call, the
-        // template does not add the model's next turn and confuses the model
-        // from emitting its proper reasoning token sequence.
-        data.generation_prompt = "<|turn>model\n";
-        data.prompt += data.generation_prompt;
-    }
-
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_USER,      "<|turn>user"  },
-        { COMMON_CHAT_ROLE_ASSISTANT, "<|turn>model" },
-    };
-
-    data.format            = COMMON_CHAT_FORMAT_PEG_GEMMA4;
-    data.supports_thinking  = true;
-    data.thinking_start_tag = "<|channel>thought";
-    data.thinking_end_tags  = {"<channel|>"};
-
-    data.preserved_tokens = {
-        "<|channel>",
-        "<channel|>",
-        "<|tool_call>",
-        "<tool_call|>",
-        "<|turn>",
-    };
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = string_ends_with(data.prompt, "<turn|>\n") ? "<|turn>model\n" : "";
-        data.generation_prompt += "<|channel>thought\n" + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += "<channel|>" + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
-    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
-    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
-    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto start = p.rule("start", p.optional(p.literal("<|turn>model\n")));
-
-        if (extract_reasoning) {
-            p.rule("thought", p.literal("<|channel>thought") + p.space() + p.reasoning(p.until("<channel|>")) + p.literal("<channel|>"));
-        } else {
-            p.rule("thought", p.content(p.literal("<|channel>thought") + p.space() + p.until("<channel|>") + p.literal("<channel|>")));
-        }
-
-        auto consume_empty_channels = p.gbnf(p.zero_or_more(p.literal("<|channel>") + p.negate(p.literal("thought"))), "");
-        auto thought = (p.peek(p.literal("<|channel>")) + consume_empty_channels + p.ref("thought")) | p.negate(p.literal("<|channel>"));
-
-        if (has_response_format) {
-            auto response_format = p.literal("```json") <<
-                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)) <<
-                p.literal("```");
-            return start + p.optional(thought) + response_format;
-        }
-
-        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
-            // Gemma4 tool calling syntax
-            // Rules should match traversal logic in gemma4_to_json()
-            p.rule("gemma4-string-content", p.until("<|\"|>"));
-            p.rule("gemma4-string", p.literal("<|\"|>") + p.ref("gemma4-string-content") + p.literal("<|\"|>"));
-            p.rule("gemma4-bool", p.json_bool());
-            p.rule("gemma4-null", p.json_null());
-            p.rule("gemma4-number", p.json_number());
-            p.rule("gemma4-dict-key", p.rule("gemma4-dict-key-name", p.chars("[^:}]", 1, -1)) + p.literal(":"));
-            p.rule("gemma4-dict-kv", p.ref("gemma4-dict-key") + p.space() + p.ref("gemma4-value"));
-            p.rule("gemma4-dict", [&]() {
-                auto ws = p.space();
-                auto member = p.ref("gemma4-dict-kv");
-                auto members = p.sequence({member, p.zero_or_more(p.sequence({p.literal(","), ws, member}))});
-                return p.sequence({
-                    p.literal("{"), ws,
-                    p.choice({p.literal("}"), p.sequence({members, ws, p.literal("}")})})
-                });
-            });
-            p.rule("gemma4-array", [&]() {
-                auto ws = p.space();
-                auto value = p.ref("gemma4-value");
-                auto elements = p.sequence({value, p.zero_or_more(p.sequence({p.literal(","), ws, value}))});
-                return p.sequence({
-                    p.literal("["), ws,
-                    p.choice({p.literal("]"), p.sequence({elements, ws, p.literal("]")})})
-                });
-            });
-            p.rule("gemma4-value", [&]() {
-                return p.choice({
-                    p.ref("gemma4-string"), p.ref("gemma4-dict"), p.ref("gemma4-array"),
-                    p.ref("gemma4-number"), p.ref("gemma4-bool"), p.ref("gemma4-null")
-                });
-            });
-
-            auto tool_choice = p.choice();
-
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                std::string  name     = function.at("name");
-                // TODO @aldehir : need to extend json-schema-to-grammar to produce more than JSON rules
-                // const auto & params   = function.at("parameters");
-
-                tool_choice |= p.rule("tool-" + name, p.tool(p.sequence({
-                    p.tool_open(p.tool_name(p.literal(name)) + p.peek(p.literal("{"))),
-                    p.tool_args(p.ref("gemma4-dict")),
-                })));
-            });
-
-            auto tool_call = p.trigger_rule("tool-call", p.repeat(
-                "<|tool_call>call:" + tool_choice + "<tool_call|>",
-                /* min = */ inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0,
-                /* max = */ inputs.parallel_tool_calls ? -1 : 1
-            ));
-
-            auto scan_to_toolcall = p.rule("scan-to-toolcall", p.until("<|tool_call>"));
-            auto content = p.rule("content", p.content(p.until_one_of({"<|channel>", "<channel|>", "<|tool_call>"})));
-            auto message = p.rule("message", thought + content);
-            return start + p.zero_or_more(message) + scan_to_toolcall + tool_call;
-        }
-
-        // Gemma 4 may emit an extra <|channel>thought\n<channel|> at the end of the content. It may
-        // also emit a single trailing <channel|> token. Consume all complete reasoning blocks and
-        // then stop at the first unmatched <channel|> token.
-        auto content = p.rule("content", p.content(p.until_one_of({"<|channel>", "<channel|>"})));
-        auto message = p.rule("message", thought + content);
-        return start + p.one_or_more(message);
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|tool_call>" },
-        };
-    }
-
-    return data;
-}

common/parsers/gigachat-v3.cpp

@@ -1,81 +0,0 @@
-#include "parsers.h"
-
-common_chat_params common_chat_params_init_gigachat_v3(
-        const common_chat_template & tmpl,
-        const autoparser::generation_params & inputs) {
-
-    common_chat_params data;
-
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = false;
-    data.preserved_tokens  = {
-        "<|message_sep|>\n\n",
-        "<|role_sep|>\n",
-    };
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-        data.generation_prompt = "assistant<|role_sep|>\n" + msg.render_content();
-        data.prompt += data.generation_prompt;
-    }
-
-    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
-    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
-    const auto *tool_call_start_prefix = "<|message_sep|>\n\nfunction call<|role_sep|>\n";
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto ret = p.eps();
-        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
-            // Build a choice of all available tools
-            auto tool_choice = p.choice();
-            for (const auto & tool : inputs.tools) {
-                const auto & function = tool.at("function");
-                std::string name = function.at("name");
-                const auto & schema = function.at("parameters");
-
-                auto tool_name = p.json_member("name", "\"" + p.tool_name(p.literal(name)) + "\"");
-                auto tool_args = p.json_member("arguments", p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema)));
-
-                auto tool_open = p.tool_open(p.literal("{") << tool_name);
-
-                tool_choice |= p.rule("tool-" + name, tool_open << "," << tool_args << "}");
-            }
-
-            // Define the tool call structure
-            auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
-            auto max_calls = 1; // parallel toolcalls are not supported
-            auto tool_call = p.rule("tool-call", p.literal(tool_call_start_prefix) + tool_choice);
-            auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(tool_call, /* min = */ min_calls, /* max = */ max_calls));
-
-            ret = p.content(p.until("<|message_sep|>\n\n")) << tool_calls;
-        } else {
-            // Content only parser
-            include_grammar = false;
-            ret = p.content(p.rest());
-        }
-
-        return p.literal("assistant<|role_sep|>\n") + ret;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
-
-        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto schema = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            {COMMON_GRAMMAR_TRIGGER_TYPE_WORD, tool_call_start_prefix}
-        };
-    }
-    return data;
-}

common/parsers/gpt-oss.cpp

@@ -1,167 +0,0 @@
-#include "parsers.h"
-
-common_chat_params common_chat_params_init_gpt_oss(const common_chat_template &    tmpl,
-                                                          const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    // Copy reasoning to the "thinking" field as expected by the gpt-oss template
-    auto adjusted_messages = json::array();
-    for (auto msg : inputs.messages) {
-        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
-            msg["thinking"] = msg.at("reasoning_content");
-            if (msg.contains("tool_calls") && msg.at("tool_calls").is_array() && !msg.at("tool_calls").empty()) {
-                msg.erase("content");
-            }
-        }
-        adjusted_messages.push_back(msg);
-    }
-
-    auto prompt = common_chat_template_direct_apply_impl(tmpl, inputs, /* messages_override= */ adjusted_messages);
-
-    // Check if we need to replace the return token with end token during
-    // inference and without generation prompt. For more details see:
-    // https://github.com/ggml-org/llama.cpp/issues/15417
-    if (inputs.is_inference && !inputs.add_generation_prompt) {
-        static constexpr std::string_view return_token = "<|return|>";
-        static constexpr std::string_view end_token    = "<|end|>";
-        if (size_t pos = prompt.rfind(return_token); pos != std::string::npos) {
-            prompt.replace(pos, return_token.length(), end_token);
-        }
-    }
-
-    data.prompt            = prompt;
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, /* messages_override= */ adjusted_messages);
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, "<|start|>assistant" },
-        { COMMON_CHAT_ROLE_USER,      "<|start|>user"      },
-        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>developer" },
-        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>system"    },
-        { COMMON_CHAT_ROLE_TOOL,      "<|start|>functions" },
-    };
-
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
-
-    data.thinking_start_tag = "<|channel|>analysis<|message|>";
-    data.thinking_end_tags  = {"<|end|>"};
-
-    // These special tokens are required to parse properly, so we include them
-    // even if parse_tool_calls is false.
-    data.preserved_tokens = {
-        "<|channel|>", "<|constrain|>", "<|message|>", "<|start|>", "<|end|>",
-    };
-
-    // Adjust prompt for continuation
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = "<|start|>assistant<|channel|>analysis<|message|>" + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += "<|end|><|start|>assistant<|channel|>final<|message|>" + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
-    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
-    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
-    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto start           = p.rule("start", p.literal("<|start|>assistant"));
-        auto end             = p.rule("end", p.literal("<|end|>"));
-        auto content         = p.rule("message-content", p.until("<|end|>"));
-        auto channel         = p.literal("<|channel|>") + (p.literal("commentary") | p.literal("analysis"));
-        auto constrain_type  = p.chars("[A-Za-z0-9_-]", 1, -1);
-
-        // Occasionally, gpt-oss-20b will prefix channels with this commentary
-        auto stray_commentary = p.optional(p.literal("<|channel|>commentary") + p.optional(p.literal(" to=assistant")));
-        auto start_analysis = stray_commentary + p.literal("<|channel|>analysis<|message|>");
-
-        if (extract_reasoning) {
-            p.rule("analysis", start_analysis + p.reasoning(content) + end);
-        } else {
-            p.rule("analysis", p.content(start_analysis + content + end));
-        }
-
-        auto analysis = p.ref("analysis");
-        auto preamble = p.rule("preamble", p.literal("<|channel|>commentary<|message|>") + p.content(content) + end);
-        auto final_msg = p.rule("final", stray_commentary + p.literal("<|channel|>final<|message|>") + p.content(content));
-
-        // Consume any unsolicited tool calls, e.g. builtin functions
-        auto unsolicited = p.rule("unsolicited", p.atomic(p.optional(channel) + p.literal(" to=") + content + end));
-
-        auto any = p.rule("any", preamble | analysis);
-
-        if (has_response_format) {
-            auto constraint = p.optional(p.space() + p.optional(p.literal("<|constrain|>")) + constrain_type);
-            auto response_format = p.rule("response-format",
-                p.literal("<|channel|>final") + constraint + p.literal("<|message|>") +
-                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)));
-
-            return p.zero_or_more(start + analysis) + start + response_format;
-        }
-
-        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
-            auto tool_choice = p.choice();
-
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                std::string  name     = function.at("name");
-                const auto & params   = function.at("parameters");
-
-                auto func_name  = p.literal(" to=functions.") + p.tool_name(p.literal(name));
-                auto constraint = p.optional(p.space() + p.optional(p.literal("<|constrain|>")) + constrain_type);
-                auto args       = p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", params));
-
-                // recipient in role header
-                //   <|start|>assistant to=functions.NAME<|channel|>(commentary|analysis)[constraint]<|message|>ARGS
-                auto tool_in_role = p.tool(p.tool_open(func_name + channel + constraint + p.literal("<|message|>")) + args);
-
-                // recipient in channel header
-                //   <|channel|>(commentary|analysis) to=functions.NAME[constraint]<|message|>ARGS
-                auto tool_in_channel = p.tool(p.tool_open(channel + func_name + constraint + p.literal("<|message|>")) + args);
-
-                tool_choice |= p.rule("tool-" + name, tool_in_role | tool_in_channel);
-            });
-
-            auto tool_call  = p.trigger_rule("tool-call", tool_choice);
-
-            if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
-                return p.zero_or_more(start + any) + start + tool_call;
-            }
-
-            return p.zero_or_more(start + any) + start + (tool_call | final_msg);
-        }
-
-        return p.zero_or_more(start + any) + start + (final_msg | unsolicited);
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "^\\s+to$" },
-            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "^<\\|channel\\|>(?:commentary|analysis)\\s+to=functions$" },
-            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "<\\|start\\|>assistant(\\s+to)" },
-            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, "<\\|start\\|>assistant(<\\|channel\\|>(?:commentary|analysis)\\s+to)" }
-        };
-    }
-
-    return data;
-}

common/parsers/kimi-k2.cpp

@@ -1,133 +0,0 @@
-#include "parsers.h"
-
-// Kimi K2 Thinking - uses unique tool call ID format: functions.<name>:<index>
-// The ID contains both the function name and an incrementing counter
-common_chat_params common_chat_params_init_kimi_k2(const common_chat_template &    tmpl,
-                                                          const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
-    data.preserved_tokens  = {
-        "<|tool_calls_section_begin|>",
-        "<|tool_calls_section_end|>",
-        "<|tool_call_begin|>",
-        "<|tool_call_argument_begin|>",
-        "<|tool_call_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   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
-
-    const std::string SECTION_BEGIN = "<|tool_calls_section_begin|>";
-    const std::string SECTION_END   = "<|tool_calls_section_end|>";
-    const std::string CALL_BEGIN    = "<|tool_call_begin|>";
-    const std::string ARGS_BEGIN    = "<|tool_call_argument_begin|>";
-    const std::string CALL_END      = "<|tool_call_end|>";
-
-    const std::string THINK_START = "<think>";
-    const std::string THINK_END   = "</think>";
-    const std::string GEN_PROMPT  = "<|im_assistant|>assistant<|im_middle|>";
-
-    data.thinking_start_tag = THINK_START;
-    data.thinking_end_tags  = {THINK_END};
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        // Kimi K2 Thinking format:
-        // - Reasoning: <think>{reasoning}</think>
-        // - Content: text after reasoning
-        // - Tool calls section:
-        //   <|tool_calls_section_begin|>
-        //   <|tool_call_begin|>functions.<name>:<index><|tool_call_argument_begin|>{json_args}<|tool_call_end|>
-        //   ...
-        //   <|tool_calls_section_end|>
-        // The ID format is: functions.<function_name>:<counter> where counter is 0, 1, 2, ...
-
-        // Tool call markers
-        auto end = p.end();
-
-        // Note: this model is CRAZY. It can diverge from its supposed tool calling pattern in so many ways it's not funny.
-        // For example, it can call tools at the end of reasoning without closing reasoning...
-        auto reasoning = extract_reasoning ? p.optional(THINK_START + p.reasoning(
-            p.until_one_of({ THINK_END, "<|tool_calls_section_begin|>", "<|tool_call_begin|>" })) +
-            p.optional(p.literal(THINK_END))) : p.eps();
-        auto generation_prompt = p.literal(GEN_PROMPT);
-
-
-        // Content only parser (no tools)
-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-            return generation_prompt + reasoning + p.content(p.rest()) + end;
-        }
-
-        // Build tool call parsers for each available function
-        // The ID format is: functions.<name>:<index>
-        // We need to match: functions.<name>:<digits>
-        auto tool_choice = p.choice();
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            std::string  name     = function.at("name");
-            const auto & schema   = function.at("parameters");
-
-            // Match: functions.<name>:<digits>
-            // Capture the full call id (functions.<name>:<digits>) using tool_id tag
-            auto tool_id = p.tool_id(p.literal("functions.") + p.tool_name(p.literal(name)) + p.literal(":") + p.chars("[0-9]", 1, -1));
-            auto tool_parser = p.tool(
-                p.tool_open(tool_id + p.literal(ARGS_BEGIN)) +
-                p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema)) +
-                p.tool_close(p.optional((p.literal(CALL_END))))
-            );
-
-            tool_choice |= p.rule("tool-" + name, tool_parser);
-        });
-
-        // Tool calls section: <|tool_calls_section_begin|> tool_calls <|tool_calls_section_end|>
-        auto min_calls  = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
-        auto max_calls  = inputs.parallel_tool_calls ? -1 : 1;
-        // Use trigger_rule so grammar generator knows where to start generating rules
-        auto tool_calls = p.rule("tool-calls",
-            p.optional(p.literal(SECTION_BEGIN)) +
-            p.trigger_rule("tool-call", p.repeat(CALL_BEGIN + tool_choice, min_calls, max_calls) +
-                p.optional(p.literal(SECTION_END)))
-        );
-
-        auto content_before_tools = p.content(p.until_one_of({ SECTION_BEGIN, CALL_BEGIN }));
-
-        return generation_prompt + reasoning + content_before_tools + tool_calls + end;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|tool_call_begin|>" }
-        };
-    }
-
-    return data;
-}

common/parsers/kimi-k3.cpp

@@ -1,174 +0,0 @@
-#include "parsers.h"
-
-// Kimi K3 - XTML tagged format, built by open_tag/close_tag macros:
-//   open_tag(t, attrs) = <|open|>t k="v"...<|sep|>   close_tag(t) = <|close|>t<|sep|>
-//   assistant := [think] [response] [tools] close_tag(message) <|end_of_msg|>
-// the generation prompt already opens the think (or response) section, so the
-// section opener is optional here - same as Kimi K2 Thinking
-common_chat_params common_chat_params_init_kimi_k3(const common_chat_template &          tmpl,
-                                                          const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
-
-    const std::string SEP         = "<|sep|>";
-    const std::string MSG_START   = "<|open|>message role=\"assistant\"<|sep|>";
-    const std::string THINK_START = "<|open|>think<|sep|>";
-    const std::string THINK_END   = "<|close|>think<|sep|>";
-    const std::string RESP_START  = "<|open|>response<|sep|>";
-    const std::string RESP_END    = "<|close|>response<|sep|>";
-    const std::string TOOLS_START = "<|open|>tools<|sep|>";
-    const std::string TOOLS_END   = "<|close|>tools<|sep|>";
-    const std::string CALL_START  = "<|open|>call tool=\"";
-    const std::string CALL_END    = "<|close|>call<|sep|>";
-    const std::string ARG_START   = "<|open|>argument key=\"";
-    const std::string ARG_END     = "<|close|>argument<|sep|>";
-    const std::string MSG_END     = "<|close|>message<|sep|>";
-    const std::string EOM_TOKEN   = "<|end_of_msg|>";
-
-    // only the markers are special tokens. tag names ("think", "response", ...) are
-    // normal tokens and must not be preserved, or prose with those words is broken
-    data.preserved_tokens = {
-        "<|open|>",
-        "<|close|>",
-        "<|sep|>",
-        "<|end_of_msg|>",
-    };
-
-    data.thinking_start_tag = THINK_START;
-    data.thinking_end_tags  = { THINK_END };
-
-    // per-role message-start delimiters. user/assistant messages only have the role
-    // attribute, so the full opener is used. system and tool messages have more
-    // attributes, so those delimiters stop after the closing quote of the role
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, "<|open|>message role=\"assistant\"<|sep|>" },
-        { COMMON_CHAT_ROLE_USER,      "<|open|>message role=\"user\"<|sep|>"      },
-        { COMMON_CHAT_ROLE_TOOL,      "<|open|>message role=\"tool\""             },
-        { COMMON_CHAT_ROLE_SYSTEM,    "<|open|>message role=\"system\""           },
-    };
-
-    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
-    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = MSG_START + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + RESP_START + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto end = p.end();
-
-        auto start = p.optional(p.literal(MSG_START));
-
-        // the think section is always consumed, even with reasoning extraction off:
-        // the generation prompt ends with open_tag('think'), so it is always present.
-        // reasoning stops at its own closer, or at the response opener if the model
-        // skips the closer
-        auto think_body = extract_reasoning ? p.reasoning(p.until_one_of({ THINK_END, RESP_START })) :
-                                              p.content(p.until_one_of({ THINK_END, RESP_START }));
-
-        auto reasoning = p.optional(p.optional(p.literal(THINK_START)) + think_body +
-                                    p.optional(p.literal(THINK_END)));
-
-        // content runs to the response closer, or to the next section if truncated
-        auto response = p.optional(p.literal(RESP_START)) +
-                        p.content(p.until_one_of({ RESP_END, TOOLS_START, MSG_END })) +
-                        p.optional(p.literal(RESP_END));
-
-        // the EOG token after the message closer reaches the parser as text,
-        // so it must be consumed or the parse stays incomplete
-        auto trailer = p.optional(p.literal(MSG_END)) + p.optional(p.literal(EOM_TOKEN));
-
-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-            return start + reasoning + response + trailer + end;
-        }
-
-        auto tool_choices = p.choice();
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            std::string  name     = function.at("name");
-            const json   schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-
-            // arguments come one tag per key, with the JSON type in a type="..."
-            // attribute. the type is taken from the tool schema instead, as it tells
-            // us if the value is JSON or a literal string
-            auto args = p.eps();
-            if (schema.contains("properties") && !schema.at("properties").empty()) {
-                auto arg_choices = p.choice();
-                for (const auto & prop : schema.at("properties").items()) {
-                    const std::string & key = prop.key();
-
-                    std::string type = "string";
-                    if (prop.value().is_object() && prop.value().contains("type") &&
-                        prop.value().at("type").is_string()) {
-                        type = prop.value().at("type").get<std::string>();
-                    }
-
-                    auto value = type == "string" ? p.tool_arg_string_value(p.until(ARG_END)) :
-                                                    p.tool_arg_value(p.until(ARG_END));
-
-                    // skip the trailing type="..." attribute: anything up to <|sep|>
-                    arg_choices |= p.rule("kimi-k3-arg-" + name + "-" + key,
-                                          p.tool_arg(p.tool_arg_open(p.literal(ARG_START)) +
-                                                     p.tool_arg_name(p.literal(key)) + p.literal("\"") +
-                                                     p.until(SEP) + p.literal(SEP) + value +
-                                                     p.tool_arg_close(p.literal(ARG_END))));
-                }
-                args = p.zero_or_more(arg_choices);
-            }
-
-            // skip the trailing index="N" attribute the same way
-            auto call = p.tool(p.tool_open(p.literal(CALL_START) + p.tool_name(p.literal(name)) + p.literal("\"") +
-                                           p.until(SEP) + p.literal(SEP)) +
-                               p.tool_args(args) + p.tool_close(p.literal(CALL_END)));
-
-            tool_choices |= p.rule("kimi-k3-tool-" + name, call);
-        });
-
-        // all calls go inside one tools section, then the message is closed. the
-        // message closer is part of the trigger rule, or else the lazy grammar
-        // rejects it once tool calls have started
-        auto tools_section =
-            p.trigger_rule("kimi-k3-tool-call", p.literal(TOOLS_START) + p.one_or_more(tool_choices) +
-                                                    p.literal(TOOLS_END) + p.optional(p.literal(MSG_END)) +
-                                                    p.optional(p.literal(EOM_TOKEN)));
-
-        auto tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? tools_section :
-                                                                              p.optional(tools_section);
-
-        return start + reasoning + response + tools + trailer + end;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                if (function.contains("parameters")) {
-                    auto schema = function.at("parameters");
-                    builder.resolve_refs(schema);
-                }
-            });
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, TOOLS_START },
-        };
-    }
-
-    return data;
-}

common/parsers/lfm2.cpp

@@ -1,119 +0,0 @@
-#include "parsers.h"
-
-// LFM2 format detection: template uses <|tool_list_start|>[...]<|tool_list_end|> around the tool list
-// and <|tool_call_start|>[...]<|tool_call_end|> around each tool call
-bool is_lfm2_template(const std::string & src) {
-    return src.find("<|tool_list_start|>") != std::string::npos &&
-           src.find("<|tool_list_end|>")   != std::string::npos;
-}
-
-// LFM2/LFM2.5 parser. Tool calls are almost Python-style and parallel-capable
-// (except dotted names and JSON literals true/false/null).
-// Always wrapped in <|tool_call_start|>[name(args)]<|tool_call_end|> with optional <think> reasoning.
-// tool_list_tokens preserves LFM2 system tool-list markers.
-common_chat_params common_chat_params_init_lfm2(const common_chat_template &          tmpl,
-                                                       const autoparser::generation_params & inputs,
-                                                       bool tool_list_tokens) {
-    common_chat_params data;
-
-    const std::string TOOL_CALL_START = "<|tool_call_start|>";
-    const std::string TOOL_CALL_END   = "<|tool_call_end|>";
-    const std::string TOOL_LIST_START = "<|tool_list_start|>";
-    const std::string TOOL_LIST_END   = "<|tool_list_end|>";
-    const std::string THINK_START     = "<think>";
-    const std::string THINK_END       = "</think>";
-    const std::string GEN_PROMPT      = "<|im_start|>assistant\n";
-
-    // Copy reasoning to the "thinking" field the template expects
-    auto adjusted_messages = json::array();
-    for (auto msg : inputs.messages) {
-        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
-            msg["thinking"] = msg.at("reasoning_content");
-        }
-        adjusted_messages.push_back(msg);
-    }
-
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs, adjusted_messages);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, adjusted_messages);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
-    data.preserved_tokens  = { TOOL_CALL_START, TOOL_CALL_END, THINK_START, THINK_END };
-    if (tool_list_tokens) {
-        data.preserved_tokens.push_back(TOOL_LIST_START);
-        data.preserved_tokens.push_back(TOOL_LIST_END);
-    }
-
-    data.thinking_start_tag = THINK_START;
-    data.thinking_end_tags  = {THINK_END};
-
-    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
-    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
-    // Gate by reasoning format and whether the template supports <think>
-    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE &&
-                             tmpl.source().find(THINK_START) != std::string::npos;
-    auto include_grammar   = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.literal(GEN_PROMPT);
-        auto end = p.end();
-
-        auto reasoning = p.eps();
-        if (extract_reasoning) {
-            reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
-        }
-
-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-            if (has_response_format) {
-                auto response_format = p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema));
-                return generation_prompt + reasoning + response_format + end;
-            }
-            return generation_prompt + reasoning + p.content(p.rest()) + end;
-        }
-        auto tool_calls = p.rule("tool-calls",
-            p.trigger_rule("tool-call",
-                p.literal(TOOL_CALL_START) +
-                p.python_style_tool_calls(inputs.tools, inputs.parallel_tool_calls, /* allow_json_literals = */ true) +
-                p.literal(TOOL_CALL_END)
-            )
-        );
-
-        auto content = p.content(p.until(TOOL_CALL_START));
-
-        return generation_prompt + reasoning + content + tool_calls + end;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, TOOL_CALL_START }
-        };
-    }
-
-    return data;
-}

common/parsers/minicpm5.cpp

@@ -1,144 +0,0 @@
-#include "parsers.h"
-
-// MiniCPM5 format:
-// - Reasoning: <think>{reasoning}</think> (optional)
-// - Tool calls: <function name="foo"><param name="bar">value</param></function>
-common_chat_params common_chat_params_init_minicpm5(const common_chat_template &          tmpl,
-                                                           const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
-    data.preserved_tokens  = {
-        "<function",
-        "<param",
-        "</function>",
-        "</param>",
-        "<think>",
-        "</think>",
-    };
-
-    data.thinking_start_tag = "<think>";
-    data.thinking_end_tags  = {"</think>"};
-
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, "<|im_start|>assistant"             },
-        { COMMON_CHAT_ROLE_TOOL,      "<|im_start|>user\n<tool_response>" },
-        { COMMON_CHAT_ROLE_USER,      "<|im_start|>user"                  },
-        { COMMON_CHAT_ROLE_SYSTEM,    "<|im_start|>system"                },
-    };
-
-    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
-    auto has_response_format = inputs.json_schema.is_object() && !inputs.json_schema.empty();
-    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = "<|im_start|>assistant\n<think>\n" + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += "\n</think>\n\n" + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.literal("<|im_start|>assistant\n");
-
-        auto reasoning = p.eps();
-        if (extract_reasoning) {
-            reasoning = ("<think>" << p.reasoning(p.until("</think>")) << "</think>") + p.space();
-        }
-
-        // Response format parser
-        if (has_response_format) {
-            return generation_prompt + reasoning + p.content(p.schema(p.json(), "response-format", inputs.json_schema));
-        }
-
-        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
-            // CDATA lets a value carry characters that would otherwise close the tag (e.g.
-            // </param>); capture the inner text only, excluding the CDATA markers.
-            auto string_value = p.choice({
-                p.literal("<![CDATA[") + p.ac(p.tool_arg_string_value(p.until("]]>")) + p.literal("]]>"), "]]>") + p.tool_arg_close(p.literal("</param>")),
-                p.negate(p.literal("<![CDATA[")) + p.ac(p.tool_arg_string_value(p.until("</param>")) + p.tool_arg_close(p.literal("</param>")), "</param>")
-            });
-
-            auto tool_choice = p.choice();
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto &      function = tool.at("function");
-                const std::string name     = function.at("name");
-                auto              params   = function.contains("parameters") ? function.at("parameters") : json::object();
-
-                auto args = p.eps();
-                if (params.contains("properties") && params.at("properties").is_object() && !params.at("properties").empty()) {
-                    auto schema_info = common_schema_info();
-                    schema_info.resolve_refs(params);
-
-                    auto arg_choice = p.choice();
-                    for (const auto & [prop_name, prop_schema] : params.at("properties").items()) {
-                        auto value_parser = p.eps();
-                        if (schema_info.resolves_to_string(prop_schema)) {
-                            value_parser = string_value;
-                        } else {
-                            value_parser = p.tool_arg_json_value(
-                                    p.schema(p.json(), "tool-" + name + "-arg-" + prop_name + "-schema", prop_schema, false)
-                                ) + p.tool_arg_close(p.literal("</param>"));
-                        }
-
-                        auto arg_rule = p.tool_arg(
-                            p.tool_arg_open(p.literal("<param name=\"") + p.tool_arg_name(p.literal(prop_name)) + p.literal("\">")) +
-                            value_parser
-                        );
-
-                        arg_choice |= arg_rule;
-                    }
-                    args = p.zero_or_more(arg_choice + p.space());
-                }
-
-                auto tool_parser = p.tool(
-                    p.tool_open(p.literal("<function name=\"") + p.tool_name(p.literal(name)) + p.literal("\">"))
-                    << p.tool_args(args)
-                    << p.tool_close(p.literal("</function>")));
-
-                tool_choice |= p.rule("tool-" + name, tool_parser);
-            });
-
-            auto max_calls  = inputs.parallel_tool_calls ? -1 : 1;
-            auto tool_calls = p.trigger_rule("tool-call", p.repeat(tool_choice + p.space(), 1, max_calls));
-
-            auto content = p.content(p.until("<function"));
-
-            return generation_prompt + reasoning + content + tool_calls + p.end();
-        }
-
-        return generation_prompt + reasoning + p.content(p.rest()) + p.end();
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function" },
-        };
-    }
-
-    return data;
-}

common/parsers/minimax-m3.cpp

@@ -1,259 +0,0 @@
-#include "parsers.h"
-
-common_chat_params common_chat_params_init_minimax_m3(const common_chat_template &          tmpl,
-                                                             const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format             = COMMON_CHAT_FORMAT_PEG_MINIMAX_M3;
-    data.supports_thinking  = true;
-    data.thinking_start_tag = "<mm:think>";
-    data.thinking_end_tags  = {"</mm:think>"};
-
-    // M3 prefixes every tool tag with the namespace token "]<]minimax[>[";
-    // params use the parameter name as the tag (<file_path>...</file_path>).
-    const std::string NS          = "]<]minimax[>[";
-    const std::string THINK_START = "<mm:think>";
-    const std::string THINK_END   = "</mm:think>";
-    const std::string FC_START    = NS + "<tool_call>";
-    const std::string FC_END      = NS + "</tool_call>";
-    const std::string INVOKE_END  = NS + "</invoke>";
-
-    data.preserved_tokens = {
-        NS,
-        "<tool_call>",
-        "</tool_call>",
-        THINK_START,
-        THINK_END,
-    };
-
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, "]~b]ai"        },
-        { COMMON_CHAT_ROLE_USER,      "]~b]user"      },
-        { COMMON_CHAT_ROLE_TOOL,      "]~b]tool"      },
-        { COMMON_CHAT_ROLE_SYSTEM,    "]~b]developer" },
-        { COMMON_CHAT_ROLE_SYSTEM,    "]~b]system"    },
-    };
-
-    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
-    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
-    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
-
-    const std::string GEN_PROMPT = data.generation_prompt;
-
-    using mm3 = common_chat_peg_minimax_m3_mapper;
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.prefix(GEN_PROMPT, THINK_START);
-        auto end = p.end();
-
-        auto reasoning = p.eps();
-        if (extract_reasoning) {
-            auto block = inputs.enable_thinking
-                             ? p.literal(THINK_START) + p.space() +
-                                   p.ac(p.reasoning(p.until(THINK_END)) + p.literal(THINK_END), THINK_END)
-                             : p.literal(THINK_START) + p.ac(p.until(THINK_END) + p.literal(THINK_END), THINK_END);
-
-            // A turn without reasoning is prefixed with a bare </mm:think>, written either by the
-            // generation prompt (thinking_mode = "disabled") or by the model itself.
-            reasoning = p.optional(p.choice({ block, p.literal(THINK_END) }));
-        }
-
-        if (has_response_format) {
-            auto response_format = p.rule("response-format",
-                p.literal("```json") + p.space() +
-                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)) +
-                p.space() + p.literal("```"));
-            return generation_prompt + reasoning + response_format + end;
-        }
-
-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-            return generation_prompt + reasoning + p.content(p.rest()) + end;
-        }
-
-        auto alternatives_of = [](const json & schema) -> std::optional<json> {
-            for (const auto * keyword : { "oneOf", "anyOf" }) {
-                if (schema.contains(keyword) && schema.at(keyword).is_array() && !schema.at(keyword).empty()) {
-                    return schema.at(keyword);
-                }
-            }
-            return std::nullopt;
-        };
-
-        auto tool_choice = p.choice();
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            std::string  name     = function.at("name");
-            auto         params   = function.contains("parameters") ? function.at("parameters") : json::object();
-
-            auto schema_info = common_schema_info();
-            schema_info.resolve_refs(params);
-
-            // The template expands argument values recursively in XML (see the to_xml() macro)
-            std::function<common_peg_parser(const json &, const std::string &, const std::string &)> value_of;
-            std::function<common_peg_parser(const json &, const std::string &)>                      members_of;
-
-            auto element_of = [&](const std::string & tag, const json & schema, const std::string & rule_name) {
-                const std::string close = NS + "</" + tag + ">";
-                return p.rule(rule_name,
-                    p.tool_arg(
-                        p.tool_arg_open(
-                            p.literal(NS + "<") +
-                            p.tool_arg_name(p.literal(tag)) +
-                            p.literal(">")) +
-                        value_of(schema, rule_name, close)));
-            };
-
-            value_of = [&](const json & schema,
-                           const std::string & rule_name,
-                           const std::string & close) -> common_peg_parser {
-                auto close_tag = p.tool_arg_close(p.literal(close));
-
-                // A string accepts anything, so a union with a string alternative is a string
-                if (schema_info.resolves_to_string(schema)) {
-                    return p.ac(p.tool_arg_string_value(p.until(close)) + close_tag, close);
-                }
-
-                if (auto alternatives = alternatives_of(schema)) {
-                    std::vector<common_peg_parser> choices;
-
-                    size_t index = 0;
-                    for (const auto & alternative : *alternatives) {
-                        const std::string alt_name = rule_name + "-" + std::to_string(index++);
-
-                        // There is a risk that this breaks streaming deltas, but that's a risk we
-                        // assume to provide tool arg streaming.
-                        choices.push_back(value_of(alternative, alt_name, close));
-                    }
-
-                    return p.choice(choices);
-                }
-
-                const std::string type = schema.contains("type") && schema.at("type").is_string()
-                                             ? schema.at("type").get<std::string>()
-                                             : "";
-
-                if (type == "object" && schema.contains("properties")) {
-                    return p.tag(mm3::TOOL_ARG_OBJECT, members_of(schema, rule_name)) + p.space() + close_tag;
-                }
-
-                if (type == "array" && schema.contains("items")) {
-                    const std::string item_close = NS + "</item>";
-                    auto item = p.rule(rule_name + "-item",
-                        p.tag(mm3::TOOL_ARG_ITEM,
-                              p.literal(NS + "<item>") +
-                                  value_of(schema.at("items"), rule_name + "-item", item_close)));
-                    return p.tag(mm3::TOOL_ARG_ARRAY, p.repeat(p.space() + item, 0, -1)) + p.space() + close_tag;
-                }
-
-                return p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", schema, false)) + close_tag;
-            };
-
-            // Required properties in schema order, then any number of optional ones in any order.
-            members_of = [&](const json & schema, const std::string & rule_prefix) -> common_peg_parser {
-                const auto & props = schema.at("properties");
-
-                std::set<std::string> required;
-                if (schema.contains("required")) {
-                    required = schema.at("required").get<std::set<std::string>>();
-                }
-
-                std::vector<common_peg_parser> required_elements;
-                std::vector<common_peg_parser> optional_elements;
-                for (const auto & [key, key_schema] : props.items()) {
-                    auto element = element_of(key, key_schema, rule_prefix + "-" + key);
-                    if (required.find(key) != required.end()) {
-                        required_elements.push_back(element);
-                    } else {
-                        optional_elements.push_back(element);
-                    }
-                }
-
-                common_peg_parser members = p.eps();
-                for (size_t i = 0; i < required_elements.size(); i++) {
-                    if (i > 0) {
-                        members = members + p.space();
-                    }
-                    members = members + required_elements[i];
-                }
-
-                if (!optional_elements.empty()) {
-                    common_peg_parser any_optional = p.choice();
-                    for (const auto & element : optional_elements) {
-                        any_optional |= element;
-                    }
-                    members = members + p.repeat(p.space() + any_optional, 0, -1);
-                }
-
-                return members;
-            };
-
-            common_peg_parser invoke_body =
-                params.contains("properties") ? members_of(params, "tool-" + name + "-arg") : p.eps();
-
-            auto func_parser = p.tool(
-                p.tool_open(p.literal(NS + "<invoke name=\"") +
-                            p.tool_name(p.literal(name)) + p.literal("\">")) +
-                p.space() + invoke_body + p.space() +
-                p.tool_close(p.literal(INVOKE_END)));
-
-            tool_choice |= p.rule("tool-" + name, func_parser);
-        });
-
-        auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-        common_peg_parser tool_calls = p.eps();
-        if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call",
-                p.literal(FC_START) + p.space() + tool_choice +
-                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
-        } else {
-            tool_calls = p.trigger_rule("tool-call",
-                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
-        }
-
-        if (!require_tools) {
-            tool_calls = p.optional(tool_calls);
-        }
-
-        auto content_before_tools = p.content(p.until(FC_START));
-        return generation_prompt + reasoning + content_before_tools + tool_calls + end;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, FC_START },
-        };
-    }
-
-    return data;
-}

common/parsers/ministral3.cpp

@@ -1,135 +0,0 @@
-#include "parsers.h"
-
-common_chat_params common_chat_params_init_ministral_3(const common_chat_template &    tmpl,
-                                                              const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    // Build up messages to follow the format: https://huggingface.co/mistralai/Ministral-3-14B-Reasoning-2512/blob/main/chat_template.jinja
-    auto adjusted_messages = json::array();
-    for (const auto & msg : inputs.messages) {
-        auto role = msg.value("role", "");
-        if (role != "system" && role != "assistant") {
-            // Only adjust system and assistant messages. Interestingly, the system message may contain thinking.
-            adjusted_messages.push_back(msg);
-            continue;
-        }
-
-        auto content = json::array();
-
-        // If message contains `reasoning_content`, add it as a block of type `thinking`
-        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
-            content.push_back({
-                { "type",     "thinking"                                     },
-                { "thinking", msg.at("reasoning_content").get<std::string>() },
-            });
-        }
-
-        // If message contains `content`, add it as a block of type `text`
-        if (msg.contains("content")) {
-            if (msg.at("content").is_string()) {
-                content.push_back({
-                    { "type", "text"                               },
-                    { "text", msg.at("content").get<std::string>() },
-                });
-            } else if (msg.at("content").is_array()) {
-                auto blocks = msg.at("content");
-                content.insert(blocks);
-            }
-        }
-
-        auto adjusted       = msg;
-        adjusted["content"] = content;
-        adjusted.erase("reasoning_content");
-        adjusted_messages.push_back(adjusted);
-    }
-
-    auto has_tools            = inputs.tools.is_array() && !inputs.tools.empty();
-    auto has_response_format  = inputs.json_schema.is_object() && !inputs.json_schema.empty();
-    auto extract_reasoning    = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-    auto include_grammar      = true;
-
-    data.supports_thinking  = true;
-    data.thinking_start_tag = "[THINK]";
-    data.thinking_end_tags  = {"[/THINK]"};
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs, /* messages_override = */ adjusted_messages);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs, /* messages_override = */ adjusted_messages);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.preserved_tokens  = {
-        "[THINK]",
-        "[/THINK]",
-        "[TOOL_CALLS]",
-        "[ARGS]",
-    };
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = "[THINK]" + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += "[/THINK]" + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.eps();
-        auto reasoning =
-            extract_reasoning ? p.optional("[THINK]" + p.reasoning(p.until("[/THINK]")) + "[/THINK]") : p.eps();
-
-        // Response format parser
-        if (has_response_format) {
-            // Ministral wants to emit json surrounded by code fences
-            return generation_prompt + (reasoning << "```json" << p.content(p.schema(p.json(), "response-format", inputs.json_schema)) << "```");
-        }
-
-        // Tool call parser
-        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
-            auto tool_choice = p.choice();
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                std::string  name     = function.at("name");
-                const auto & schema   = function.at("parameters");
-
-                tool_choice |=
-                    p.rule("tool-" + name, p.tool_open(p.tool_name(p.literal(name)) + "[ARGS]") +
-                                               p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema)));
-            });
-
-            auto min_calls  = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
-            auto max_calls  = inputs.parallel_tool_calls ? -1 : 1;
-            auto tool_calls = p.trigger_rule("tool-call", p.repeat("[TOOL_CALLS]" + tool_choice, min_calls, max_calls));
-
-            return generation_prompt + (reasoning << p.content(p.until("[TOOL_CALLS]")) << tool_calls);
-        }
-
-        // Content only parser
-        include_grammar = false;
-        return generation_prompt + (reasoning << p.content(p.rest()));
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
-
-        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "[TOOL_CALLS]" }
-        };
-    }
-
-    return data;
-}

common/parsers/muse-glimmer.cpp

@@ -1,148 +0,0 @@
-#include "parsers.h"
-
-// An assistant turn is rendered as one or more messages, each
-// "<|start|>assistant to=<recipient><|message|>{content}{END}" where END is
-// <|eom|> (more messages follow) or <|eot|> (end of turn):
-//   - chain-of-thought: to=self, terminated by <|eom|>
-//   - final answer:     to=user, terminated by <|eot|>
-// The generation prompt is just "<|start|>assistant"; the model emits its own
-// " to=...<|message|>".
-common_chat_params common_chat_params_init_muse_glimmer(const common_chat_template &          tmpl,
-                                                               const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = "<|start|>assistant";
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
-
-    data.preserved_tokens = {
-        "<|start|>", "<|message|>", "<|eom|>", "<|eot|>",
-        // ATEM tool-call markup emitted on " to=<tool>" turns.
-        "<atem:function_calls>", "<atem:invoke", "<atem:parameter", "</atem:parameter>",
-        "</atem:invoke>", "</atem:function_calls>",
-    };
-
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, "<|start|>assistant" },
-        { COMMON_CHAT_ROLE_USER,      "<|start|>user"      },
-        { COMMON_CHAT_ROLE_SYSTEM,    "<|start|>system"    },
-        { COMMON_CHAT_ROLE_TOOL,      "<|start|>tool"      },
-    };
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = "<|start|>assistant to=self<|message|>" + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += "<|eom|><|start|>assistant to=user<|message|>" + msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-
-    auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
-    // Constrained grammar whenever tools are offered.
-    auto include_grammar = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto start = p.rule("start", p.literal("<|start|>assistant"));
-
-        if (!extract_reasoning && !include_grammar) {
-            return start + p.content(p.rest());
-        }
-
-        if (extract_reasoning) {
-            p.rule("analysis", p.literal(" to=self<|message|>") + p.reasoning(p.until("<|eom|>")) + p.literal("<|eom|>"));
-        } else {
-            p.rule("analysis", p.literal(" to=self<|message|>") + p.content(p.until("<|eom|>")) + p.literal("<|eom|>"));
-        }
-        auto analysis = p.ref("analysis");
-
-        auto recipient  = p.optional(p.literal(" to=user"));
-        auto final_msg  = p.rule("final", recipient + p.literal("<|message|>") +
-                                              p.content(p.until_one_of({ "<|eot|>", "<|eom|>" })));
-
-        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
-            auto string_value = p.ac(
-                p.tool_arg_string_value(p.until("</atem:parameter>")) + p.tool_arg_close(p.literal("</atem:parameter>")),
-                "</atem:parameter>");
-
-            auto tool_choice = p.choice();
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto &      function = tool.at("function");
-                const std::string name     = function.at("name");
-                auto              params   = function.contains("parameters") ? function.at("parameters") : json::object();
-
-                auto args = p.eps();
-                if (params.contains("properties") && params.at("properties").is_object() && !params.at("properties").empty()) {
-                    auto schema_info = common_schema_info();
-                    schema_info.resolve_refs(params);
-
-                    auto arg_choice = p.choice();
-                    for (const auto & [prop_name, prop_schema] : params.at("properties").items()) {
-                        auto value_parser = p.eps();
-                        if (schema_info.resolves_to_string(prop_schema)) {
-                            value_parser = string_value;
-                        } else {
-                            value_parser = p.tool_arg_json_value(
-                                    p.schema(p.json(), "tool-" + name + "-arg-" + prop_name + "-schema", prop_schema, false))
-                                + p.tool_arg_close(p.literal("</atem:parameter>"));
-                        }
-
-                        auto arg_rule = p.tool_arg(
-                            p.tool_arg_open(p.literal("<atem:parameter name=\"") + p.tool_arg_name(p.literal(prop_name)) + p.literal("\">")) +
-                            value_parser);
-
-                        arg_choice |= arg_rule;
-                    }
-                    args = p.zero_or_more(arg_choice + p.space());
-                }
-
-                auto tool_parser = p.tool(
-                    p.tool_open(p.literal(" to=") + p.until("<|message|>") +
-                                p.literal("<|message|><atem:function_calls>") + p.space() +
-                                p.literal("<atem:invoke name=\"") + p.tool_name(p.literal(name)) + p.literal("\">") + p.space())
-                    << p.tool_args(args)
-                    << p.tool_close(p.literal("</atem:invoke>") + p.space() + p.literal("</atem:function_calls>")));
-
-                tool_choice |= p.rule("tool-" + name, tool_parser);
-            });
-
-            auto tool_calls = inputs.parallel_tool_calls
-                ? p.trigger_rule("tool-call", tool_choice + p.zero_or_more(p.literal("<|eom|>") + start + tool_choice))
-                : p.trigger_rule("tool-call", tool_choice);
-
-
-            if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
-                return p.zero_or_more(start + analysis) + start + tool_calls;
-            }
-            auto trailing_calls = p.optional(p.literal("<|eom|>") + start + tool_calls);
-            return p.zero_or_more(start + analysis) + start + (tool_calls | (final_msg + trailing_calls));
-        }
-
-        return p.zero_or_more(start + analysis) + start + final_msg;
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-                builder.resolve_refs(schema);
-            });
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-        data.grammar_triggers = {
-            { COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
-              "<\\|start\\|>assistant( to=(?!self<\\|message\\|>)(?!user<\\|message\\|>)[^<]*?<\\|message\\|>)" },
-        };
-    }
-
-    return data;
-}

common/parsers/parsers.cpp

@@ -1,34 +0,0 @@
-#include "parsers.h"
-
-#include "log.h"
-
-#include <set>
-
-void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
-    for (const auto & tool : tools) {
-        if (!tool.contains("type") || tool.at("type") != "function" || !tool.contains("function")) {
-            LOG_INF("Skipping tool without function: %s", tool.dump(2).c_str());
-            continue;
-        }
-        fn(tool);
-    }
-}
-
-void foreach_parameter(const json & function, const std::function<void(const std::string &, const json &, bool)> & fn) {
-    if (!function.contains("parameters") || !function.at("parameters").is_object()) {
-        return;
-    }
-    const auto & params = function.at("parameters");
-    if (!params.contains("properties") || !params.at("properties").is_object()) {
-        return;
-    }
-    const auto & props = params.at("properties");
-    std::set<std::string> required;
-    if (params.contains("required") && params.at("required").is_array()) {
-        required = params.at("required").get<std::set<std::string>>();
-    }
-    for (const auto & [name, prop] : props.items()) {
-        bool is_required = (required.find(name) != required.end());
-        fn(name, prop, is_required);
-    }
-}

common/parsers/parsers.h

@@ -1,77 +0,0 @@
-#pragma once
-
-#include "chat.h"
-#include "chat-auto-parser.h"
-#include "chat-auto-parser-helpers.h"
-#include "chat-peg-parser.h"
-#include "common.h"
-#include "ggml.h"
-#include "json-schema-to-grammar.h"
-#include "json.h"
-
-#include <functional>
-#include <optional>
-#include <set>
-#include <string>
-#include <vector>
-
-using json = common_json;
-
-// iterate over the function tools of an OpenAI-style tools array
-void foreach_function(const json & tools, const std::function<void(const json &)> & fn);
-
-// iterate over the parameters of a function tool, flagging the ones listed as required
-void foreach_parameter(const json & function, const std::function<void(const std::string &, const json &, bool)> & fn);
-
-// render a template; the override arguments let a parser feed in messages, tools or context it has rewritten
-std::string common_chat_template_direct_apply_impl(
-    const common_chat_template & tmpl,
-    const autoparser::generation_params & inputs,
-    const std::optional<json> & messages_override = std::nullopt,
-    const std::optional<json> & tools_override = std::nullopt,
-    const std::optional<json> & additional_context = std::nullopt);
-
-// the suffix a template appends when add_generation_prompt is set
-std::string common_chat_template_generation_prompt_impl(
-    const common_chat_template & tmpl,
-    const autoparser::generation_params & inputs,
-    const std::optional<json> & messages_override = std::nullopt,
-    const std::optional<json> & tools_override = std::nullopt,
-    const std::optional<json> & additional_context = std::nullopt);
-
-bool is_lfm2_template(const std::string & src);
-
-namespace workaround {
-
-void convert_tool_responses_gemma4(json & messages);
-
-}
-
-common_chat_params common_chat_params_init_cohere2moe(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
-common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
-common_chat_params common_chat_params_init_functionary_v3_2(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
-common_chat_params common_chat_params_init_gemma4(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
-common_chat_params common_chat_params_init_gigachat_v3(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
-common_chat_params common_chat_params_init_gpt_oss(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
-common_chat_params common_chat_params_init_kimi_k2(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
-common_chat_params common_chat_params_init_kimi_k3(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
-// tool_list_tokens preserves the LFM2 system tool-list markers; LFM2.5 renders without them
-common_chat_params common_chat_params_init_lfm2(const common_chat_template & tmpl, const autoparser::generation_params & inputs, bool tool_list_tokens);
-
-common_chat_params common_chat_params_init_minicpm5(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
-common_chat_params common_chat_params_init_minimax_m3(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
-common_chat_params common_chat_params_init_ministral_3(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
-common_chat_params common_chat_params_init_muse_glimmer(const common_chat_template & tmpl, const autoparser::generation_params & inputs);
-
-common_chat_params common_chat_params_init_qwen3_coder(const common_chat_template & tmpl, const autoparser::generation_params & inputs);

common/parsers/qwen3-coder.cpp

@@ -1,181 +0,0 @@
-#include "parsers.h"
-
-common_chat_params common_chat_params_init_qwen3_coder(const common_chat_template &          tmpl,
-                                                              const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    const std::string GEN_PREFIX = "<|im_start|>assistant\n";
-
-    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-
-    auto supports_reasoning = tmpl.source().find("<think>") != std::string::npos;
-
-    data.supports_thinking = supports_reasoning;
-    data.preserved_tokens  = {
-        "<tool_call>",
-        "</tool_call>",
-    };
-
-    auto is_qwen3_coder  = !supports_reasoning;
-
-    if (supports_reasoning) {
-        data.thinking_start_tag = "<think>";
-        // Support both </think> and <tool_call> as reasoning end sequences.
-        // <function= is omitted, as it is a workaround for Qwen3-Coder which is not a thinking model
-        data.thinking_end_tags = { "</think>", "<tool_call>" };
-        data.preserved_tokens.insert(data.preserved_tokens.end(), { "<think>", "</think>" });
-    }
-
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, "<|im_start|>assistant"             },
-        { COMMON_CHAT_ROLE_TOOL,      "<|im_start|>user\n<tool_response>" }, // Qwen3-Coder, Qwen3.5, Nemotron Nano 3
-        { COMMON_CHAT_ROLE_TOOL,      "<|im_start|>tool_response"         }, // StepFun-3.5-Flash
-        { COMMON_CHAT_ROLE_USER,      "<|im_start|>user"                  },
-        { COMMON_CHAT_ROLE_SYSTEM,    "<|im_start|>system"                },
-    };
-
-    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
-    auto has_response_format = inputs.json_schema.is_object() && !inputs.json_schema.empty();
-    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-
-        data.generation_prompt = GEN_PREFIX;
-        if (supports_reasoning) {
-            data.generation_prompt += "<think>\n" + msg.reasoning_content;
-            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-                data.generation_prompt += "\n</think>\n\n";
-            }
-        }
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += msg.render_content();
-        }
-
-        data.prompt += data.generation_prompt;
-    }
-
-    std::vector<std::string> tool_call_starts = { "<tool_call>" };
-
-    if (is_qwen3_coder) {
-        // Match complete <function=name> opener for Qwen3-Coder models that occasionally omit the
-        // starting <tool_call>. The model may hallucinate a tool name, but it is preferable over
-        // constraining on <function which may occur in valid content generation, e.g. #include <functional>
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const std::string name = tool.at("function").at("name");
-            tool_call_starts.push_back("<function=" + name + ">");
-        });
-    }
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.literal(GEN_PREFIX);
-
-        auto reasoning = p.eps();
-        if (supports_reasoning && extract_reasoning) {
-            reasoning = p.optional("<think>" + p.space() +
-                                   p.reasoning(p.until_one_of({ "</think>", "<tool_call>" })) +
-                                   (p.literal("</think>") | p.peek(p.literal("<tool_call>"))));
-        }
-
-        // Response format parser
-        if (has_response_format) {
-            return generation_prompt + (reasoning << p.content(p.schema(p.json(), "response-format", inputs.json_schema)));
-        }
-
-        // Tool call parser
-        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
-            auto arg_close  = p.tool_arg_close(p.literal("\n</parameter>\n"));
-            auto arg_string = p.rule("xml-arg-string",
-                p.ac(p.tool_arg_string_value(p.until("\n</parameter>\n")) + arg_close, "\n</parameter>\n"));
-
-            auto tool_choice = p.choice();
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function   = tool.at("function");
-                std::string  name       = function.at("name");
-                auto         parameters = function.contains("parameters") ? function.at("parameters") : json::object();
-
-                auto schema_info = common_schema_info();
-                schema_info.resolve_refs(parameters);
-
-                std::vector<common_peg_parser> required_args;
-                std::vector<common_peg_parser> optional_args;
-
-                foreach_parameter(function, [&](const std::string & param_name, const json & param_schema, bool is_required) {
-                    auto rule_name = "tool-" + name + "-arg-" + param_name;
-
-                    auto arg_open = p.tool_arg_open("<parameter=" + p.tool_arg_name(p.literal(param_name)) + ">\n");
-
-                    auto arg_value = schema_info.resolves_to_string(param_schema) ?
-                        arg_string :
-                        p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", param_schema)) + arg_close;
-
-                    auto arg_rule = p.rule(rule_name, p.tool_arg(arg_open + arg_value));
-
-                    (is_required ? required_args : optional_args).push_back(arg_rule);
-                });
-
-                // Accept required arguments in any order, as Qwen does not always adhere to the
-                // order provided.
-                auto args = p.permute("tool-" + name + "-args", required_args);
-                if (!optional_args.empty()) {
-                    args = args + p.zero_or_more(p.choice(optional_args));
-                }
-
-                auto func = p.tool(p.tool_open("<function=" + p.tool_name(p.literal(name)) + ">\n") +
-                                   p.tool_args(args) +
-                                   p.tool_close(p.literal("</function>\n")));
-
-                tool_choice |= p.rule("tool-" + name, func);
-            });
-
-            auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
-
-            auto tool_call_body = tool_choice + "</tool_call>" + p.space();
-            auto tool_call      = p.rule("tool-call", "<tool_call>\n" + tool_call_body);
-
-            // Qwen3-Coder models may occasionally omit the <tool_call> token.
-            auto tool_call_first = is_qwen3_coder ?
-                p.rule("tool-call-first", p.optional(p.literal("<tool_call>\n")) + tool_call_body) :
-                tool_call;
-
-            auto calls      = inputs.parallel_tool_calls ? tool_call_first + p.zero_or_more(tool_call) : tool_call_first;
-            auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(calls, min_calls, 1));
-
-            return generation_prompt +
-                   (reasoning << p.content(p.until_one_of(tool_call_starts)) << tool_calls);
-        }
-
-        // Content only parser
-        return generation_prompt + (reasoning << p.content(p.rest()));
-    });
-
-    data.parser = parser.save();
-
-    if (include_grammar) {
-        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
-
-        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-            foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool.at("function");
-                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
-                builder.resolve_refs(schema);
-            });
-            if (has_response_format) {
-                auto schema = inputs.json_schema;
-                builder.resolve_refs(schema);
-            }
-            parser.build_grammar(builder, data.grammar_lazy);
-        });
-
-        if (data.grammar_lazy) {
-            for (const auto & start : tool_call_starts) {
-                data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_WORD, start });
-            }
-        }
-    }
-
-    return data;
-}

common/parsers/sources.cmake

@@ -1,20 +0,0 @@
-# Specialized chat template parsers, listed explicitly so that adding or removing one re-runs CMake instead of leaving an incremental build stale.
-
-set(LLAMA_CHAT_PARSERS_SOURCES
-    ${CMAKE_CURRENT_LIST_DIR}/parsers.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/parsers.h
-    ${CMAKE_CURRENT_LIST_DIR}/cohere2moe.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/deepseek.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/functionary-v3-2.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/gemma4.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/gigachat-v3.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/gpt-oss.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/kimi-k2.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/kimi-k3.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/lfm2.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/minicpm5.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/minimax-m3.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/ministral3.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/muse-glimmer.cpp
-    ${CMAKE_CURRENT_LIST_DIR}/qwen3-coder.cpp
-)

common/speculative.cpp

@@ -14,7 +14,6 @@
 
 #include <algorithm>
 #include <cassert>
-#include <cmath>
 #include <cstring>
 #include <iomanip>
 #include <map>
@@ -139,7 +138,6 @@ struct common_speculative_impl {
     const common_speculative_type type;
 
     uint32_t n_seq;
-    int32_t n_max; // maximum draft length after implementation-specific limits
 
     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.
@@ -159,7 +157,7 @@ struct common_speculative_impl {
     int64_t t_draft_us  = 0; // total time spent in generating drafts in this implementation in microseconds.
     int64_t t_accept_us = 0; // total time spent in accumulation of this implementation in microseconds.
 
-    common_speculative_impl(common_speculative_type type, uint32_t n_seq, int32_t n_max) : type(type), n_seq(n_seq), n_max(n_max) {}
+    common_speculative_impl(common_speculative_type type, uint32_t n_seq) : type(type), n_seq(n_seq) {}
 
     virtual ~common_speculative_impl() = default;
 
@@ -184,7 +182,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
     std::vector<common_sampler_ptr> smpls;
 
     common_speculative_impl_draft_simple(const common_params_speculative & params, uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE, n_seq, params.draft.n_max)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE, n_seq)
         , params(params.draft)
     {
         auto * ctx_dft = this->params.ctx_dft;
@@ -454,7 +452,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
     std::vector<float> g_embd_buf;
 
     common_speculative_impl_draft_eagle3(const common_params_speculative & params, uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, n_seq, params.draft.n_max)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, n_seq)
         , params(params.draft)
     {
         SPC_TRC("%s", "adding speculative implementation 'draft-eagle3'\n");
@@ -925,25 +923,21 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
     int32_t     block_size    = 0;
     llama_token mask_token_id = 0;
 
-    bool    is_dflash2     = false;
-    bool    is_mrope       = false;
-    int32_t selector_top_k = 0;
-
     // draft-dspark: the draft carries a Markov head and uses an anchor-first block layout
     const bool is_dspark;
 
     // dspark speculators
     bool sample_from_anchor = true;
 
-    // block-internal attention
-    bool causal_attn = false;
-
     const int32_t * target_layer_ids   = nullptr; // model_dft's extract layer indices
     uint32_t        target_layer_ids_n = 0;
 
+    // scratch buffer for concatenated target features [n_tokens, n_embd_enc]
+    std::vector<float> features_buf;
+
     common_speculative_impl_draft_dflash(const common_params_speculative & params, uint32_t n_seq,
             common_speculative_type type = COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH)
-        : common_speculative_impl(type, n_seq, params.draft.n_max)
+        : common_speculative_impl(type, n_seq)
         , params(params.draft)
         , is_dspark(type == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK)
     {
@@ -972,25 +966,9 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             if (llama_model_meta_val_str(model_dft, "dflash.sample_from_anchor", buf, sizeof(buf)) >= 0) {
                 sample_from_anchor = std::strcmp(buf, "true") == 0;
             }
-            if (llama_model_meta_val_str(model_dft, "dflash.attention.causal", buf, sizeof(buf)) >= 0) {
-                causal_attn = std::strcmp(buf, "true") == 0;
-            }
         }
-
-        selector_top_k = llama_model_dflash_selector_top_k(model_dft);
-        is_dflash2     = selector_top_k > 0;
         mask_token_id = llama_vocab_mask(llama_model_get_vocab(model_dft));
 
-        if (is_dspark && this->params.p_min > 0.0f) {
-            char buf[16] = {};
-            const bool has_conf =
-                llama_model_meta_val_str(model_dft, "dflash.has_confidence_head", buf, sizeof(buf)) < 0 ||
-                std::strcmp(buf, "true") == 0;
-            if (!has_conf) {
-                throw std::runtime_error("DSpark draft has no confidence head: please set --spec-draft-p-min 0");
-            }
-        }
-
         LOG_INF("%s: adding speculative implementation '%s'\n", __func__, common_speculative_type_to_str(type).c_str());
         LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min);
         LOG_INF("%s: - block_size=%d, mask_token_id=%d, n_extract=%u, sample_from_anchor=%s\n", __func__,
@@ -1005,17 +983,9 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             this->params.n_max = std::min(this->params.n_max, n_draft_max);
             this->params.n_min = std::min(this->params.n_min, n_draft_max);
         }
-        this->n_max = this->params.n_max;
 
         batch        = llama_batch_init(llama_n_batch(ctx_dft), 0,          n_seq);
-        batch_inject = llama_batch_init(llama_n_ubatch(ctx_dft), n_embd_enc, n_seq);
-
-        // embd batches on an M-RoPE draft need 4 position rows per token
-        is_mrope = llama_model_rope_type(model_dft) == LLAMA_ROPE_TYPE_MROPE;
-        if (is_mrope) {
-            free(batch_inject.pos);
-            batch_inject.pos = (llama_pos *) malloc(sizeof(llama_pos) * 4 * llama_n_batch(ctx_dft));
-        }
+        batch_inject = llama_batch_init(llama_n_batch(ctx_dft), n_embd_dec, n_seq);
 
         smpls.resize(n_seq);
         for (auto & s : smpls) {
@@ -1028,7 +998,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
         // offload draft sampling to the backend
         backend_chains.assign(n_seq, nullptr);
-        if (this->params.backend_sampling && !is_dflash2) {
+        if (this->params.backend_sampling) {
             for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
                 llama_sampler * chain = llama_sampler_chain_init(llama_sampler_chain_default_params());
                 llama_sampler_chain_add(chain, llama_sampler_init_top_k(10));
@@ -1047,9 +1017,8 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             llama_set_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k], true);
         }
 
-        // DFlash2 reads its selector lattice from h_nextn and never consumes raw logits.
-        llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ !is_dflash2);
-        llama_set_causal_attn(ctx_dft, causal_attn); // DFlash needs non-causal attention unless the model says otherwise
+        llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ true);
+        llama_set_causal_attn(ctx_dft, false); // DFlash needs non-causal attention
     }
 
     ~common_speculative_impl_draft_dflash() override {
@@ -1134,34 +1103,52 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             for (int32_t offset = 0; offset < n_rows; offset += n_ubatch) {
                 const int32_t n_chunk = std::min(n_ubatch, n_rows - offset);
 
-                // gather target features per extract layer; the fused decode encodes and
-                // injects them into the K/V cache at the target positions
-                batch_inject.n_tokens = n_chunk;
+                // gather this chunk's target features, interleaved by extract layer
+                features_buf.resize((size_t) n_chunk * n_embd_enc);
                 for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
                     const float * layer = llama_get_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k]);
                     if (!layer) {
                         GGML_ABORT("DFlash: target layer %d input not extracted.", target_layer_ids[k]);
                     }
                     for (int32_t i = 0; i < n_chunk; ++i) {
-                        float       * dst = batch_inject.embd + (size_t) i * n_embd_enc + k * (size_t) n_embd_tgt;
+                        float       * dst = features_buf.data() + (size_t) i * n_embd_enc + k * (size_t) n_embd_tgt;
                         const float * src = layer + (size_t) (i_batch_beg[seq_id] + offset + i) * n_embd_tgt;
                         std::memcpy(dst, src, (size_t) n_embd_tgt * sizeof(float));
                     }
                 }
 
+                // fuse extracted features through DFlash encoder
+                llama_batch enc_batch = {
+                    /*.n_tokens =*/ n_chunk,
+                    /*.token    =*/ nullptr,
+                    /*.embd     =*/ features_buf.data(),
+                    /*.pos      =*/ nullptr,
+                    /*.n_seq_id =*/ nullptr,
+                    /*.seq_id   =*/ nullptr,
+                    /*.logits   =*/ nullptr,
+                };
+
+                int32_t rc = llama_encode(ctx_dft, enc_batch);
+                if (rc != 0) {
+                    LOG_ERR("%s: llama_encode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n",
+                            __func__, rc, (int) n_chunk, (int) offset);
+                    return false;
+                }
+
+                const float * inp_g = llama_get_embeddings_nextn(ctx_dft);
+                GGML_ASSERT(inp_g && "DFlash encoder produced no output.");
+
+                // inject the DFlash decoder K/V cache at the tokens' target positions
+                batch_inject.n_tokens = n_chunk;
+                std::memcpy(batch_inject.embd, inp_g, (size_t) n_chunk * n_embd_dec * sizeof(float));
+
                 for (int32_t i = 0; i < n_chunk; ++i) {
-                    const llama_pos p = batch_in.pos[i_batch_beg[seq_id] + offset + i];
-                    batch_inject.pos[i] = p;
-                    if (is_mrope) {
-                        batch_inject.pos[1 * n_chunk + i] = p;
-                        batch_inject.pos[2 * n_chunk + i] = p;
-                        batch_inject.pos[3 * n_chunk + i] = 0;
-                    }
+                    batch_inject.pos[i]       = batch_in.pos[i_batch_beg[seq_id] + offset + i];
                     batch_inject.n_seq_id[i]  = 1;
                     batch_inject.seq_id[i][0] = seq_id;
                     batch_inject.logits[i]    = false;
                 }
-                const int32_t rc = llama_decode(ctx_dft, batch_inject);
+                rc = llama_decode(ctx_dft, batch_inject);
                 if (rc != 0) {
                     LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n",
                             __func__, rc, (int) n_chunk, (int) offset);
@@ -1199,7 +1186,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             i_block_beg[seq_id] = batch.n_tokens;
             n_block    [seq_id] = n_block_tokens;
             for (int32_t i = 0; i < n_block_tokens; ++i) {
-                common_batch_add(batch, i == 0 ? dp.id_last : mask_token_id, n + i, { seq_id }, !is_dflash2);
+                common_batch_add(batch, i == 0 ? dp.id_last : mask_token_id, n + i, { seq_id }, true);
             }
         }
 
@@ -1227,36 +1214,6 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
             auto & result = *dp.result;
 
-            if (is_dflash2) {
-                const float * lattice = llama_get_embeddings_nextn(ctx_dft);
-                GGML_ASSERT(lattice && "DFlash2 selector produced no lattice");
-
-                int32_t predecessor = 0;
-                for (int32_t i = 1; i < n_block_tokens; ++i) {
-                    const float * row = lattice + (size_t) (beg + i) * n_embd_dec;
-                    const float * scores = row + selector_top_k + (size_t) predecessor * selector_top_k;
-
-                    predecessor = (int32_t) std::distance(scores,
-                            std::max_element(scores, scores + selector_top_k));
-                    if (params.p_min > 0.0f) {
-                        // softmax(scores) at the argmax, i.e. 1 / sum(exp(s_k - s_max))
-                        float sum = 0.0f;
-                        for (int32_t k = 0; k < selector_top_k; ++k) {
-                            sum += std::exp(scores[k] - scores[predecessor]);
-                        }
-                        if (1.0f / sum < params.p_min) {
-                            break;
-                        }
-                    }
-                    result.push_back((llama_token) row[predecessor]);
-                }
-
-                if (result.size() < (size_t) params.n_min) {
-                    result.clear();
-                }
-                continue;
-            }
-
             if (is_dspark) {
                 // DSpark: read from the first draft slot, truncate below the confidence threshold
                 const float * conf = params.p_min > 0.0f ? llama_get_embeddings_nextn(ctx_dft) : nullptr;
@@ -1358,7 +1315,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
     std::vector<std::vector<float>> chain_h;
 
     common_speculative_impl_draft_mtp(const common_params_speculative & params, uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, n_seq, params.draft.n_max)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, n_seq)
         , params(params.draft)
     {
         auto * ctx_tgt = this->params.ctx_tgt;
@@ -1425,7 +1382,6 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                 c.reserve((size_t) (this->params.n_max + 1) * n_embd);
             }
         }
-        this->n_max = this->params.n_max;
 
         pending_h.assign(n_seq, std::vector<float>(n_embd, 0.0f));
 
@@ -1770,7 +1726,7 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
     common_speculative_impl_ngram_simple(
             const common_params_speculative & params, uint32_t n_seq,
             common_ngram_simple_config config)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, n_seq, params.ngram_simple.size_m)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, n_seq)
         , params(params.ngram_simple)
         , config(config)
     {
@@ -1814,7 +1770,7 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
             const common_ngram_map & config,
             uint32_t n_seq)
         : common_speculative_impl(config.key_only ? COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K
-            : COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, n_seq, config.size_value)
+            : COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, n_seq)
     {
         for (uint32_t i = 0; i < n_seq; i++) {
             this->config.push_back(config);
@@ -1885,7 +1841,7 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
     common_speculative_impl_ngram_mod(
             const common_params_speculative & params,
             uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_MOD, n_seq, params.ngram_mod.n_max)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_MOD, n_seq)
         , params(params.ngram_mod)
         , mod(params.ngram_mod.n_match, 4*1024*1024)
         , verbose(std::getenv("LLAMA_TRACE") != nullptr) {
@@ -2061,7 +2017,7 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
             const std::string & path_dynamic,
             bool save_dynamic,
             bool save_static)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, n_seq, n_draft)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, n_seq)
         , params(params.ngram_cache)
         , n_draft(n_draft)
         , save_dynamic(save_dynamic)
@@ -2182,8 +2138,6 @@ struct common_speculative {
 
     // which implementaion was used for a given seq_id
     std::vector<common_speculative_impl *> impl_last;
-
-    std::vector<double> synth_probs;
 };
 
 static common_ngram_map get_common_ngram_map(
@@ -2362,101 +2316,6 @@ int32_t common_speculative_n_max(const common_params_speculative * spec) {
     return n_max;
 }
 
-int32_t common_speculative_n_max(const common_speculative * spec) {
-    int32_t n_max = 0;
-
-    if (spec == nullptr) {
-        return n_max;
-    }
-
-    for (const auto & impl : spec->impls) {
-        n_max = std::max(n_max, std::max(0, impl->n_max));
-    }
-
-    return n_max;
-}
-
-std::vector<double> common_speculative_synth_rates_resolve(const common_params_speculative * spec, int32_t n_max) {
-    const bool has_length = spec->synth_len != -1.0;
-    const bool has_rates  = !spec->synth_rates.empty();
-
-    if (!has_length && !has_rates) {
-        return {};
-    }
-    if (has_length && has_rates) {
-        throw std::invalid_argument("synthetic acceptance length and rates are mutually exclusive");
-    }
-
-    if (n_max <= 0) {
-        throw std::invalid_argument("synthetic acceptance requires at least one speculative token");
-    }
-
-    if (has_rates) {
-        const auto & rates = spec->synth_rates;
-        if (rates.size() != (size_t) n_max) {
-            throw std::invalid_argument(string_format(
-                    "synthetic acceptance rates must contain %d values, got %zu", n_max, rates.size()));
-        }
-
-        for (size_t i = 0; i < rates.size(); ++i) {
-            if (!std::isfinite(rates[i]) || rates[i] < 0.0 || rates[i] > 1.0) {
-                throw std::invalid_argument("synthetic acceptance rates must be finite and within [0, 1]");
-            }
-            if (i > 0 && rates[i] > rates[i - 1]) {
-                throw std::invalid_argument("synthetic acceptance rates must be monotonically non-increasing");
-            }
-        }
-
-        return rates;
-    }
-
-    const double length = spec->synth_len;
-    const double length_max = (double) n_max + 1.0;
-    if (!std::isfinite(length) || length < 1.0 || length > length_max) {
-        throw std::invalid_argument(string_format(
-                "synthetic acceptance length must be finite and within [1, %.0f]", length_max));
-    }
-
-    double p = 0.0;
-    if (length == length_max) {
-        p = 1.0;
-    } else if (length > 1.0) {
-        double p_min = 0.0;
-        double p_max = 1.0;
-        for (int i = 0; i < 32; ++i) {
-            const double p_mid = 0.5 * (p_min + p_max);
-            double sum = 0.0;
-            double term = p_mid;
-            for (int32_t j = 0; j < n_max; ++j) {
-                sum += term;
-                term *= p_mid;
-            }
-
-            if (sum < length - 1.0) {
-                p_min = p_mid;
-            } else {
-                p_max = p_mid;
-            }
-        }
-        p = 0.5 * (p_min + p_max);
-    }
-
-    std::vector<double> rates;
-    rates.reserve(n_max);
-    double rate = p;
-    for (int32_t i = 0; i < n_max; ++i) {
-        rates.push_back(rate);
-        rate *= p;
-    }
-
-    return rates;
-}
-
-const std::vector<double> & common_speculative_get_synth_probs(const common_speculative * spec) {
-    GGML_ASSERT(spec);
-    return spec->synth_probs;
-}
-
 common_params common_base_params_to_speculative(const common_params & params) {
     const bool has_draft = params.speculative.has_dft();
 
@@ -2467,22 +2326,11 @@ common_params common_base_params_to_speculative(const common_params & params) {
     result.pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED;
 
     if (has_draft) {
-        // default to global devices value
-        if (!params_spec.devices.empty()) {
-            result.devices           = params_spec.devices;
-        }
+        result.devices               = params_spec.devices;
         result.model                 = params_spec.mparams;
         result.n_gpu_layers          = params_spec.n_gpu_layers;
         result.tensor_buft_overrides = params_spec.tensor_buft_overrides;
 
-        // a draft pinned to a single device doesn't need the meta wrapper an inherited -sm tensor would give it
-        // (the device list is null-terminated, so a single device means size 2)
-        const size_t n_devs = std::count_if(params_spec.devices.begin(), params_spec.devices.end(),
-                [](ggml_backend_dev_t d) { return d != nullptr; });
-        if (n_devs == 1) {
-            result.split_mode = LLAMA_SPLIT_MODE_LAYER;
-        }
-
         if (params_spec.cpuparams.n_threads > 0) {
             result.cpuparams.n_threads       = params_spec.cpuparams.n_threads;
             result.cpuparams_batch.n_threads = params_spec.cpuparams_batch.n_threads;
@@ -2540,9 +2388,6 @@ common_speculative_init_result::common_speculative_init_result(
         cparams.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
     }
 
-    // the draft context holds as many tokens per sequence as the target context
-    cparams.n_ctx = llama_n_ctx(ctx_tgt);
-
     // note: for small models maybe we can set this to the maximum possible draft from all speculative types
     //       the extra memory for small models is likely negligible?
     cparams.n_rs_seq  = 0;
@@ -2720,39 +2565,13 @@ common_speculative * common_speculative_init(common_params_speculative & params,
         return nullptr;
     }
 
-    common_speculative_ptr result(new common_speculative {
-        /* .dparams     = */ common_speculative_draft_params_vec(n_seq),
-        /* .impls       = */ std::move(impls),
-        /* .impl_last   = */ std::vector<common_speculative_impl *>(n_seq, nullptr),
-        /* .synth_probs = */ {},
-    });
+    auto * result = new common_speculative {
+        /* .dparams   = */ common_speculative_draft_params_vec(n_seq),
+        /* .impls     = */ std::move(impls),
+        /* .impl_last = */ std::vector<common_speculative_impl *>(n_seq, nullptr)
+    };
 
-    const int32_t n_max_configured = common_speculative_n_max(&params);
-    const int32_t n_max_effective  = common_speculative_n_max(result.get());
-    const auto rates = common_speculative_synth_rates_resolve(&params, n_max_effective);
-
-    std::vector<std::string> rates_str;
-    rates_str.reserve(rates.size());
-    result->synth_probs.reserve(rates.size());
-    double rate_prev = 1.0;
-    double acceptance_length = 1.0;
-    for (const double rate : rates) {
-        result->synth_probs.push_back(rate_prev > 0.0 ? rate / rate_prev : 0.0);
-        rates_str.push_back(string_format("%.6g", rate));
-        rate_prev = rate;
-        acceptance_length += rate;
-    }
-    if (!result->synth_probs.empty()) {
-        SPC_WRN("%s", "synthetic speculative acceptance is enabled for benchmarking; generated output is not valid\n");
-        if (n_max_effective != n_max_configured) {
-            SPC_WRN("synthetic acceptance draft limit was reduced from %d to %d by the initialized speculative implementations\n",
-                    n_max_configured, n_max_effective);
-        }
-        SPC_INF("synthetic acceptance: n_max = %zu, mean length = %.6f, rates = [%s]\n",
-                rates.size(), acceptance_length, string_join(rates_str, ", ").c_str());
-    }
-
-    return result.release();
+    return result;
 }
 
 void common_speculative_free(common_speculative * spec) {

common/speculative.h

@@ -26,15 +26,6 @@ std::string common_speculative_type_to_str(enum common_speculative_type type);
 // return the max number of draft tokens based on the speculative parameters
 int32_t common_speculative_n_max(const common_params_speculative * spec);
 
-// return the max number of draft tokens from the initialized implementations
-int32_t common_speculative_n_max(const common_speculative * spec);
-
-// validate and resolve the unconditional synthetic acceptance rates
-std::vector<double> common_speculative_synth_rates_resolve(const common_params_speculative * spec, int32_t n_max);
-
-// return the conditional synthetic acceptance probabilities
-const std::vector<double> & common_speculative_get_synth_probs(const common_speculative * spec);
-
 common_params common_base_params_to_speculative(const common_params & params);
 
 struct common_speculative_output_limits {

conversion/__init__.py

@@ -54,7 +54,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "DeepseekV3ForCausalLM": "deepseek",
     "DeepseekV32ForCausalLM": "deepseek",
     "DFlashDraftModel": "qwen",
-    "DFlash2DraftModel": "qwen",
     "Qwen3DSparkModel": "qwen",
     "DSparkDraftModel": "qwen",
     "DSparkSpeculator": "qwen",
@@ -124,7 +123,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "HunYuanMoEV1ForCausalLM": "hunyuan",
     "HunYuanVLForConditionalGeneration": "hunyuan",
     "HYV3ForCausalLM": "hunyuan",
-    "HYV4ForCausalLM": "hy_v4",
     "IQuestCoderForCausalLM": "llama",
     "InternLM2ForCausalLM": "internlm",
     "InternLM3ForCausalLM": "internlm",
@@ -189,7 +187,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "NanbeigeForCausalLM": "nanbeige",
     "NemotronForCausalLM": "nemotron",
     "NemotronHForCausalLM": "nemotron",
-    "NemotronHPuzzleForCausalLM": "nemotron",
     "NeoBERT": "bert",
     "NeoBERTForSequenceClassification": "bert",
     "NeoBERTLMHead": "bert",
@@ -238,8 +235,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "Qwen3_5ForConditionalGeneration": "qwen",
     "Qwen3_5MoeForCausalLM": "qwen",
     "Qwen3_5MoeForConditionalGeneration": "qwen",
-    "Qwen4ExpForCausalLM": "qwen4exp",
-    "Qwen4ExpForConditionalGeneration": "qwen4exp",
     "RND1": "qwen",
     "RWForCausalLM": "falcon",
     "RWKV6Qwen2ForCausalLM": "rwkv",
@@ -255,7 +250,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "SeedOssForCausalLM": "olmo",
     "SmallThinkerForCausalLM": "smallthinker",
     "SmolLM3ForCausalLM": "llama",
-    "Spark2_5ForCausalLM": "spark2_5",
     "SolarOpenForCausalLM": "glm",
     "StableLMEpochForCausalLM": "stablelm",
     "StableLmForCausalLM": "stablelm",
@@ -289,7 +283,6 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "CogVLMForCausalLM": "cogvlm",
     "DeepseekOCR2ForCausalLM": "deepseek",
     "DeepseekOCRForCausalLM": "deepseek",
-    "DeepseekV4ForCausalLM": "deepseek",
     "Dots3NoteForCausalLM": "dots3",
     "Dots3NoteForConditionalGeneration": "dots3",
     "DotsOCRForCausalLM": "dotsocr",
@@ -339,7 +332,6 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
     "Qwen3_5ForConditionalGeneration": "qwen3vl",
     "Qwen3_5MoeForConditionalGeneration": "qwen3vl",
-    "Qwen4ExpForConditionalGeneration": "qwen4exp",
     "RADIOModel": "nemotron",
     "Sarashina2VisionForCausalLM": "sarashina2",
     "SmolVLMForConditionalGeneration": "smolvlm",

conversion/base.py

@@ -130,8 +130,7 @@ class ModelBase:
                  sentence_transformers_dense_modules: bool = False,
                  target_model_dir: Path | None = None,
                  fuse_gate_up_exps: bool = False,
-                 fp8_as_q8: bool = False,
-                 fuse_qkv: bool = False):
+                 fp8_as_q8: bool = False):
         if type(self) is ModelBase or \
                 type(self) is TextModel or \
                 type(self) is MmprojModel:
@@ -154,15 +153,6 @@ class ModelBase:
         self.fuse_gate_up_exps = fuse_gate_up_exps
         self._gate_exp_buffer: dict[int, Tensor] = {}
         self._up_exp_buffer: dict[int, Tensor] = {}
-        self.fuse_qkv = fuse_qkv
-        self._q_buffer: dict[int, Tensor] = {}
-        self._k_buffer: dict[int, Tensor] = {}
-        self._v_buffer: dict[int, Tensor] = {}
-        self._q_bias_buffer: dict[int, Tensor] = {}
-        self._k_bias_buffer: dict[int, Tensor] = {}
-        self._v_bias_buffer: dict[int, Tensor] = {}
-        self._fusable_qkv_weight_layers: set[int] = set()
-        self._fusable_qkv_bias_layers: set[int] = set()
         self.hparams = ModelBase.load_hparams(self.dir_model, self.is_mistral_format) if hparams is None else hparams
         self.model_tensors = self.index_tensors(remote_hf_model_id=remote_hf_model_id)
         self.metadata_override = metadata_override
@@ -627,43 +617,6 @@ class ModelBase:
             raise ValueError(f"Can not map tensor {name!r}")
         return new_name
 
-    def prepare_qkv_fusion(self) -> None:
-        self._fusable_qkv_weight_layers.clear()
-        self._fusable_qkv_bias_layers.clear()
-        if not self.fuse_qkv or gguf.MODEL_TENSOR.ATTN_QKV not in gguf.MODEL_TENSORS[self.model_arch]:
-            return
-
-        qkv_types = {
-            gguf.MODEL_TENSOR.ATTN_Q,
-            gguf.MODEL_TENSOR.ATTN_K,
-            gguf.MODEL_TENSOR.ATTN_V,
-        }
-        weights: dict[int, set[gguf.MODEL_TENSOR]] = {}
-        biases: dict[int, set[gguf.MODEL_TENSOR]] = {}
-
-        for name in self.model_tensors:
-            mapped = self.tensor_map.get_type_and_name(name, try_suffixes=(".weight", ".bias"))
-            if mapped is None:
-                continue
-            tensor_type, new_name = mapped
-            if tensor_type not in qkv_types:
-                continue
-
-            bid = next((int(part) for part in new_name.split(".") if part.isdecimal()), None)
-            if bid is None:
-                continue
-            if new_name.endswith(".weight"):
-                weights.setdefault(bid, set()).add(tensor_type)
-            elif new_name.endswith(".bias"):
-                biases.setdefault(bid, set()).add(tensor_type)
-
-        for bid, weight_types in weights.items():
-            bias_types = biases.get(bid, set())
-            if weight_types == qkv_types and (not bias_types or bias_types == qkv_types):
-                self._fusable_qkv_weight_layers.add(bid)
-                if bias_types:
-                    self._fusable_qkv_bias_layers.add(bid)
-
     def set_gguf_parameters(self):
         raise NotImplementedError("set_gguf_parameters() must be implemented in subclasses")
 
@@ -692,40 +645,6 @@ class ModelBase:
                self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_UP_EXP, bid):
                 return []
 
-        # Handle Q/K/V tensor fusion if enabled
-        qkv_bid = next((int(part) for part in new_name.split(".") if part.isdecimal()), None) if self.fuse_qkv else None
-        if qkv_bid is not None:
-            is_bias = new_name.endswith('.bias')
-            suffix = '.bias' if is_bias else '.weight'
-            fusable_layers = self._fusable_qkv_bias_layers if is_bias else self._fusable_qkv_weight_layers
-            if qkv_bid not in fusable_layers:
-                return [(new_name, data_torch)]
-
-            buf_q = self._q_bias_buffer if is_bias else self._q_buffer
-            buf_k = self._k_bias_buffer if is_bias else self._k_buffer
-            buf_v = self._v_bias_buffer if is_bias else self._v_buffer
-
-            if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_Q, qkv_bid, suffix):
-                buf_q[qkv_bid] = data_torch
-            elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_K, qkv_bid, suffix):
-                buf_k[qkv_bid] = data_torch
-            elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_V, qkv_bid, suffix):
-                buf_v[qkv_bid] = data_torch
-
-            if qkv_bid in buf_q and qkv_bid in buf_k and qkv_bid in buf_v:
-                q_data = buf_q.pop(qkv_bid)
-                k_data = buf_k.pop(qkv_bid)
-                v_data = buf_v.pop(qkv_bid)
-                fused_data = torch.cat([q_data, k_data, v_data], dim=0)
-                fused_name = self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_QKV, qkv_bid, suffix=suffix)
-                logger.info(f"Fused Q, K, V {suffix[1:]} into QKV for layer {qkv_bid}")
-                return [(fused_name, fused_data)]
-
-            if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_Q, qkv_bid, suffix) or \
-               self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_K, qkv_bid, suffix) or \
-               self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.ATTN_V, qkv_bid, suffix):
-                return []
-
         return [(new_name, data_torch)]
 
     def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
@@ -980,8 +899,6 @@ class ModelBase:
 
         self.dequant_model()
 
-        self.prepare_qkv_fusion()
-
         # Handle empty tensor_map for models with block_count=0 (like MobileNetV5)
         if self.tensor_map.mapping:
             max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
@@ -1089,16 +1006,12 @@ class ModelBase:
                     else:
                         raise ValueError(f"Unknown file type: {self.ftype.name}")
 
-                # a chunked tensor quantizes as one chunk at a time, while it is written
-                quantize = data.quantize if isinstance(data, gguf.LazyChunkedTensor) else (
-                    lambda qtype, d=data: gguf.quants.quantize(d, qtype))
-
                 try:
-                    data = quantize(data_qtype)
+                    data = gguf.quants.quantize(data, data_qtype)
                 except gguf.QuantError as e:
                     logger.warning("%s, %s", e, "falling back to F16")
                     data_qtype = gguf.GGMLQuantizationType.F16
-                    data = quantize(data_qtype)
+                    data = gguf.quants.quantize(data, data_qtype)
 
                 shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
 
@@ -1110,13 +1023,6 @@ class ModelBase:
 
                 self.gguf_writer.add_tensor(new_name, data, raw_dtype=data_qtype)
 
-        qkv_buffers = (
-            self._q_buffer, self._k_buffer, self._v_buffer,
-            self._q_bias_buffer, self._k_bias_buffer, self._v_bias_buffer,
-        )
-        if any(qkv_buffers):
-            raise ValueError("QKV fusion did not consume all buffered tensors")
-
     def set_type(self):
         self.gguf_writer.add_type(gguf.GGUFType.MODEL)
 
@@ -1597,9 +1503,6 @@ class TextModel(ModelBase):
         if chkhsh == "bba3b3366b646dbdded5dbc42d59598b849371afc42f7beafa914afaa5b70aa6":
             # ref: https://huggingface.co/tencent/Hunyuan-4B-Instruct
             res = "hunyuan-dense"
-        if chkhsh == "e6ddf9c6686791c12d698d34c31ab9be1fea9af5a3d9a6909783ab382198ae1c":
-            # ref: https://huggingface.co/tencent/Hy4-preview
-            res = "hy_v4"
         if chkhsh == "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6":
             # ref: https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base
             res = "falcon-h1"
@@ -1633,9 +1536,6 @@ class TextModel(ModelBase):
         if chkhsh == "9e454714343b69b99b71795c1d27a68c2a1d15dab111f4d353109f966af29da7":
             # ref: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B
             res = "lfm2"
-        if chkhsh == "0a766d034107bc736a3f2dc4968fd62e54a3570f1454443e0c5a4cc6bd7941ed":
-            # ref: https://huggingface.co/XHToken/Spark-X2.5-1.7B
-            res = "spark2_5"
         if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
             # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
             res = "llama-bpe"

conversion/deepseek.py

@@ -578,8 +578,6 @@ class DeepseekV4Model(TextModel):
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         name, gen = item
-        if name.startswith(("aligner.", "image_")):
-            return None
         if name.startswith("mtp."):
             if not cls.mtp_only:
                 cls._skipped_mtp_tensors += 1
@@ -855,7 +853,6 @@ class DeepseekV4Model(TextModel):
             "ffn_norm.weight": (gguf.MODEL_TENSOR.FFN_NORM, ".weight"),
             "ffn.gate.weight": (gguf.MODEL_TENSOR.FFN_GATE_INP, ".weight"),
             "ffn.gate.bias": (gguf.MODEL_TENSOR.FFN_EXP_PROBS_B, ".bias"),
-            "ffn.gate.bias_vl": (gguf.MODEL_TENSOR.FFN_EXP_PROBS_B_VL, ".bias"),
             "ffn.gate.tid2eid": (gguf.MODEL_TENSOR.FFN_GATE_TID2EID, ".weight"),
             "ffn.shared_experts.w1.weight": (gguf.MODEL_TENSOR.FFN_GATE_SHEXP, ".weight"),
             "ffn.shared_experts.w2.weight": (gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, ".weight"),
@@ -881,10 +878,6 @@ class DeepseekV4Model(TextModel):
         if re.match(r"layers\.\d+\.ffn\.experts\.\d+\.w[123]\.(weight|scale)$", name):
             return []
 
-        # hash layers route text tokens via tid2eid and image tokens via bias_vl; gate.bias is unused
-        if name.endswith(".ffn.gate.bias") and bid is not None and bid < self.hparams["num_hash_layers"]:
-            return []
-
         tensor_key, suffix = self._map_dsv4_tensor_name(name, bid)
         if tensor_key == gguf.MODEL_TENSOR.FFN_GATE_TID2EID:
             return []
@@ -1007,13 +1000,6 @@ class DeepseekV4DSparkModel(DeepseekV4Model):
             return self._DSPARK_ROOT_MAP[name]
         return super()._map_dsv4_tensor_name(name, bid)
 
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # the DFlash draft uses the plain exp-probs bias (ffn.gate.bias -> FFN_EXP_PROBS_B);
-        # the mtmd-only hash routing tensors (bias_vl, tid2eid) are not part of the DFLASH arch
-        if name.endswith(".ffn.gate.bias_vl"):
-            return
-        yield from super().modify_tensors(data_torch, name, bid)
-
     def set_vocab(self):
         if self.target_model_dir is None:
             raise ValueError("DeepSeek-V4 DSpark requires --target-model-dir with the target tokenizer")
@@ -1032,73 +1018,3 @@ class DeepseekV4DSparkModel(DeepseekV4Model):
 
         self.gguf_writer.add_block_size(self.hparams["dspark_block_size"])
         self.gguf_writer.add_target_layers([layer + 1 for layer in self.hparams["dspark_target_layer_ids"]])
-
-
-@ModelBase.register("DeepseekV4ForCausalLM")
-@ModelBase.example("deepseek-ai/DeepSeek-V4-Flash-Vision-Exp")
-class DeepseekV4FlashVisionModel(MmprojModel):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        assert self.hparams_vision is not None
-        # no preprocessor_config.json in the repo; normalization is (x/255 - 0.5) / 0.5
-        # ref: inference/image_processor.py (load_image)
-        self.preprocessor_config = {
-            "image_mean": [0.5, 0.5, 0.5],
-            "image_std":  [0.5, 0.5, 0.5],
-            **self.preprocessor_config,
-        }
-
-    def get_vision_config(self) -> dict[str, Any] | None:
-        cfg = self.global_config
-        if cfg.get("vision_n_layers", 0) == 0:
-            raise ValueError("DeepseekV4FlashVisionModel requires vision_n_layers > 0 in the model config")
-        return {
-            "num_hidden_layers":   cfg["vision_n_layers"],
-            "hidden_size":         cfg["vision_dim"],
-            "num_attention_heads": cfg["vision_n_heads"],
-            "intermediate_size":   cfg["vision_inter_dim"],
-            "patch_size":          cfg["vision_patch_size"],
-            # dynamic resolution; only used for compat / warmup
-            "image_size":          cfg["vision_patch_size"] * cfg["vision_downsample_ratio"] * 16,
-            "rope_theta":          cfg.get("vision_rope_theta", 10000.0),
-            "downsample_ratio":    cfg["vision_downsample_ratio"],
-            "min_pixels":          cfg["vision_min_pixels"],
-        }
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        assert self.hparams_vision is not None
-        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.DEEPSEEK4V)
-        # vision RMSNorm eps is the pytorch default, NOT the LLM's rms_norm_eps (1e-20)
-        # ref: inference/vision.py (RMSNorm)
-        self.gguf_writer.add_vision_attention_layernorm_eps(1e-6)
-        self.gguf_writer.add_vision_use_silu(True) # SwiGLU MLP
-        self.gguf_writer.add_vision_projector_scale_factor(self.hparams_vision["downsample_ratio"])
-        self.gguf_writer.add_vision_min_pixels(self.hparams_vision["min_pixels"])
-        # hardcoded on the C++ side (see PROJECTOR_TYPE_DEEPSEEK4V in clip.cpp)
-        # if future models use different values, add GGUF keys for those
-        assert self.global_config["vision_max_n_token"] == 384
-        assert self.global_config["vision_max_wh_ratio"] == 8
-
-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, _ = item
-        if not (name.startswith(("vision.", "aligner.", "image_"))):
-            return None
-        return super().filter_tensors(item)
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        assert self.hparams_vision is not None
-        if name == "vision.patch_embed.proj.weight":
-            # nn.Linear over flattened (3, p, p) patches == conv2d weight
-            p = self.hparams_vision["patch_size"]
-            data_torch = data_torch.reshape(data_torch.shape[0], 3, p, p)
-
-        if ".mlp.w1." in name:
-            # fused SwiGLU gate+up
-            gate, up = data_torch.chunk(2, dim=0)
-            yield from super().modify_tensors(gate, name.replace("w1", "w1_gate"), bid)
-            yield from super().modify_tensors(up, name.replace("w1", "w1_up"), bid)
-            return
-
-        yield from super().modify_tensors(data_torch, name, bid)

conversion/glm.py

@@ -112,38 +112,12 @@ class GlmOCRModel(Glm4Model):
 @ModelBase.example("zai-org/GLM-4.5-Air")
 class Glm4MoeModel(TextModel):
     model_arch = gguf.MODEL_ARCH.GLM4_MOE
-    supports_mtp_export = True
-    _n_main_layers: int | None = None
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        if not self.no_mtp:
-            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
-            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
-
-    def index_tensors(self, remote_hf_model_id: str | None = None):
-        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
-        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
-        type(self)._n_main_layers = hparams.get(key)
-        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
-
-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        if (titem := super().filter_tensors(item)) is None:
-            return None
-        name, gen = titem
-
-        assert cls._n_main_layers is not None
-        is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers
-
-        if is_mtp and cls.no_mtp:
-            return None
-        if cls.mtp_only and not is_mtp and name not in (
-            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
-        ):
-            return None
-
-        return name, gen
+        # GLM4_MOE has num_hidden_layers + 1 actual layers (including NextN layer)
+        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
 
     def set_vocab(self):
         return self._set_vocab_glm()
@@ -179,22 +153,10 @@ class Glm4MoeModel(TextModel):
         if (norm_topk_prob := self.hparams.get("norm_topk_prob")) is not None:
             self.gguf_writer.add_expert_weights_norm(norm_topk_prob)
 
-        if not self.no_mtp and (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+        # NextN/MTP prediction layers
+        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
             self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
 
-    def prepare_metadata(self, vocab_only: bool):
-        from_dir = self.fname_out.is_dir()
-        super().prepare_metadata(vocab_only=vocab_only)
-
-        if not self.mtp_only or not from_dir:
-            return
-
-        output_type: str = self.ftype.name.partition("_")[2]
-        fname_default: str = gguf.naming_convention(
-            self.metadata.name, self.metadata.basename, self.metadata.finetune,
-            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
-        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
-
     _experts: list[dict[str, Tensor]] | None = None
 
     # note: unlike GLM4V non-MoE, we don't need to permute Q/K here since GLM4V_MOE uses Neox ordering already
@@ -386,7 +348,6 @@ class GlmMoeDsaModel(DeepseekV2Model):
 @ModelBase.example("upstage/Solar-Open-100B")
 class SolarOpenModel(Glm4MoeModel):
     model_arch = gguf.MODEL_ARCH.GLM4_MOE
-    supports_mtp_export = False
 
     def set_vocab(self):
         from transformers import AutoTokenizer

conversion/hy_v4.py

@@ -1,244 +0,0 @@
-from __future__ import annotations
-
-import re
-from typing import Iterable
-
-import torch
-
-from .base import ModelBase, gguf, logger
-from .deepseek import DeepseekV2Model
-
-
-def split_gate_up(weight: torch.Tensor, moe_intermediate_size: int):
-    """Split a fused stacked gate_up expert tensor into (gate, up).
-
-    weight: [n_expert, 2*moe_intermediate_size, hidden] (gate first, up second).
-    Returns (gate, up) each [n_expert, moe_intermediate_size, hidden].
-    """
-    assert weight.shape[1] == 2 * moe_intermediate_size, f"{weight.shape[1]} != 2*{moe_intermediate_size}"
-    gate = weight[:, :moe_intermediate_size, :].contiguous()
-    up = weight[:, moe_intermediate_size:, :].contiguous()
-    return gate, up
-
-
-@ModelBase.register("HYV4ForCausalLM")
-@ModelBase.example("tencent/Hy4-preview")
-class HYV4Model(DeepseekV2Model):
-    """HY_V4: DeepSeek-V3 style MLA + MoE with iHC, a gated MLA output and a learnable sink.
-
-    Reuses DeepseekV2Model for the vocab and the MLA metadata, but overrides the tensor mapping
-    because HY_V4 ships pre-stacked / fused experts plus extra iHC, gate and sink tensors. The
-    rope rows are mapped straight through (no permute) - the graph rotates consecutive pairs.
-
-    DSA is supported: indexer weights are exported for the layers marked "full" in indexer_types.
-    "shared" layers reuse the top-k of the last preceding full layer at inference time, so they
-    carry no indexer weights.
-
-    MTP (num_nextn_predict_layers) is dropped, so the GGUF cannot be used for speculative
-    decoding. The reference only runs the MTP layers while training or while speculating, so they
-    cannot change single-token logits.
-    """
-
-    model_arch = gguf.MODEL_ARCH.HY_V4
-
-    merge_expert = False
-
-    # tensors a "full" indexer layer must carry
-    INDEXER_SUFFIXES = frozenset({
-        "self_attn.indexer.wq_b.weight",
-        "self_attn.indexer.wk.weight",
-        "self_attn.indexer.k_norm.weight",
-        "self_attn.indexer.k_norm.bias",
-        "self_attn.indexer.weights_proj.weight",
-    })
-
-    @classmethod
-    def filter_tensors(cls, item):
-        # drop MTP here, not in modify_tensors, so the weights are never read
-        if item[0].startswith("model.mtp_layers."):
-            return None
-        return super().filter_tensors(item)
-
-    def _check_indexer_hparams(self):
-        for key in ("index_n_heads", "index_head_dim", "index_topk"):
-            if key not in self.hparams:
-                raise ValueError(f"HY_V4 has DSA layers but no {key}")
-
-    def indexer_is_full(self) -> list[bool] | None:
-        """Per-layer indexer ownership, or None when the checkpoint has no DSA.
-
-        indexer_types entries are "full" (owns an indexer) or "shared" (reuses the preceding
-        full layer's top-k). Missing indexer_types with sparse layers means every sparse layer
-        owns one.
-        """
-        hparams = self.hparams
-        n_layer = hparams["num_hidden_layers"]
-        indexer_types = hparams.get("indexer_types")
-
-        # the reference drives DSA off indexer_types alone; layer_types is only a fallback for
-        # checkpoints predating it (it was renamed to deepseek_sparse_attention upstream)
-        if indexer_types is None:
-            layer_types = hparams.get("layer_types") or []
-            sparse = {"sparse_attention", "deepseek_sparse_attention"}
-            if not any(t in sparse for t in layer_types):
-                return None
-            if len(layer_types) < n_layer:
-                raise ValueError(f"HY_V4 layer_types has {len(layer_types)} entries, need {n_layer}")
-            self._check_indexer_hparams()
-            return [t in sparse for t in layer_types[:n_layer]]
-
-        self._check_indexer_hparams()
-
-        if len(indexer_types) < n_layer:
-            raise ValueError(f"HY_V4 indexer_types has {len(indexer_types)} entries, need {n_layer}")
-        unknown = {t for t in indexer_types[:n_layer]} - {"full", "shared"}
-        if unknown:
-            raise ValueError(f"HY_V4 unknown indexer_types values: {sorted(unknown)}")
-        is_full = [t == "full" for t in indexer_types[:n_layer]]
-        if is_full and not is_full[0]:
-            raise ValueError("HY_V4 layer 0 must be indexer_types 'full' (nothing precedes it to share)")
-        return is_full
-
-    def set_gguf_parameters(self):
-        hparams = self.hparams
-
-        # HY4 has n_group == topk_group == 1 (no group routing). Drop the keys so the base does
-        # not emit expert_group_count/used; llama.cpp then takes the ungrouped MoE path.
-        if hparams.get("n_group") == 1 and hparams.get("topk_group") == 1:
-            hparams.pop("n_group", None)
-            hparams.pop("topk_group", None)
-
-        # HY_V4 config expresses dense/sparse layers via mlp_layer_types, but DeepseekV2Model
-        # needs first_k_dense_replace. Derive it as the contiguous leading "dense" block
-        # (the real config.json also carries first_k_dense_replace; prefer it when present,
-        # but assert the two agree so a mismatch fails loudly).
-        mlp_types = hparams.get("mlp_layer_types")
-        explicit = hparams.get("first_k_dense_replace")
-        derived = None
-        if mlp_types is not None:
-            lead = 0
-            for t in mlp_types:
-                if t == "dense":
-                    lead += 1
-                else:
-                    break
-            if any(t == "dense" for t in mlp_types[lead:]):
-                raise NotImplementedError("HY_V4 converter expects a contiguous leading dense block")
-            derived = lead
-        if explicit is not None and derived is not None and explicit != derived:
-            raise ValueError(
-                f"HY_V4 first_k_dense_replace ({explicit}) disagrees with mlp_layer_types "
-                f"leading-dense count ({derived})"
-            )
-        if explicit is None:
-            if derived is None:
-                raise ValueError("HY_V4 needs first_k_dense_replace or mlp_layer_types to place dense layers")
-            hparams["first_k_dense_replace"] = derived
-
-        # reuse DeepseekV2 MLA + MoE metadata (forces num_key_value_heads=1, writes q/kv lora,
-        # key/value lengths, expert counts, weights scale/norm, rope dims, etc.)
-        super().set_gguf_parameters()
-
-        # HY4 uses DeepSeek-V3 sigmoid routing with e_score_correction_bias. The config has no
-        # scoring_func key, so the base does not write a gating func; set it explicitly.
-        self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
-
-        # routed-expert SwiGLU logits clamp (only routed experts; shared/dense are not clamped,
-        # so swiglu_clamp_shexp is intentionally not written). 0.0 disables the clamp.
-        swiglu_limit = float(hparams.get("swiglu_limit", 0.0) or 0.0)
-        if swiglu_limit > 0.0:
-            self.gguf_writer.add_swiglu_clamp_exp([swiglu_limit] * self.block_count)
-
-        # iHC (independent Hyper-Connections)
-        self.gguf_writer.add_hyper_connection_count(hparams["hc_mult"])
-        self.gguf_writer.add_hyper_connection_epsilon(hparams["hc_eps"])
-        self.gguf_writer.add_hyper_connection_magnitude(hparams["hc_magnitude"])
-
-        # is_full is written explicitly; the graph must not infer it from tensor presence
-        is_full = self.indexer_is_full()
-        if is_full is not None:
-            self.gguf_writer.add_indexer_head_count(hparams["index_n_heads"])
-            self.gguf_writer.add_indexer_key_length(hparams["index_head_dim"])
-            self.gguf_writer.add_indexer_top_k(hparams["index_topk"])
-            self.gguf_writer.add_indexer_types(is_full)
-            logger.info(
-                "HY_V4 DSA: %d/%d layers own an indexer (top_k=%d, n_heads=%d, head_dim=%d)",
-                sum(is_full), len(is_full), hparams["index_topk"],
-                hparams["index_n_heads"], hparams["index_head_dim"],
-            )
-
-        if hparams.get("num_nextn_predict_layers", 0):
-            logger.warning(
-                "HY_V4: dropping %d MTP (nextn) layer(s) - the reference runs them only under "
-                "training / speculative decoding. This GGUF cannot be used for speculative decoding.",
-                hparams["num_nextn_predict_layers"],
-            )
-
-    def prepare_tensors(self):
-        # Hy4-preview for some reason has num_key_value_heads equal to 8, so override it here
-        # without this conversion/deepseek.py fails on assert
-        self.hparams["num_key_value_heads"] = self.hparams["num_attention_heads"]
-
-        # validate before the base materializes tensors, so a mismatch fails early
-        is_full = self.indexer_is_full()
-        if is_full is not None:
-            present: dict[int, set[str]] = {}
-            for name in self.model_tensors:
-                m = re.match(r"model\.layers\.(\d+)\.(self_attn\.indexer\..+)$", name)
-                if m:
-                    present.setdefault(int(m.group(1)), set()).add(m.group(2))
-            for il, expect_full in enumerate(is_full):
-                seen = present.get(il, set())
-                if expect_full and seen != self.INDEXER_SUFFIXES:
-                    raise ValueError(
-                        f"HY_V4 layer {il} is indexer_types 'full' but is missing indexer tensors: "
-                        f"{sorted(self.INDEXER_SUFFIXES - seen)}"
-                    )
-                if not expect_full and seen:
-                    raise ValueError(
-                        f"HY_V4 layer {il} is indexer_types 'shared' but carries indexer tensors: "
-                        f"{sorted(seen)}"
-                    )
-
-        super().prepare_tensors()
-
-    def tensor_force_quant(self, name, new_name, bid, n_dims):
-        # iHC mixing matrices are 2D .weight tensors that the reference keeps in fp32
-        # (_keep_in_fp32_modules_strict). 1D tensors (hc_base/scale, attn_sinks,
-        # e_score_correction_bias) and the router (FFN_GATE_INP) are already forced F32 by the
-        # base rules. Force the HC *_fn matrices here.
-        if new_name.endswith(("hc_attn_fn.weight", "hc_ffn_fn.weight", "output_hc_fn.weight")):
-            return gguf.GGMLQuantizationType.F32
-        # indexer k_norm is fp32 in the reference; the base rules already cover
-        # *_norm.weight and INDEXER_PROJ, but not this bias
-        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.INDEXER_K_NORM, bid, suffix=".bias"):
-            return gguf.GGMLQuantizationType.F32
-        # enable_lm_head_fp32: mirror the reference fp32 LM-head matmul by keeping output F32.
-        if new_name == "output.weight" and self.hparams.get("enable_lm_head_fp32", False):
-            return gguf.GGMLQuantizationType.F32
-        return super().tensor_force_quant(name, new_name, bid, n_dims)
-
-    def modify_tensors(self, data_torch: torch.Tensor, name: str, bid: int | None) -> Iterable[tuple[str, torch.Tensor]]:
-        hparams = self.hparams
-        moe_inter = hparams["moe_intermediate_size"]
-
-        tn = self.format_tensor_name
-
-        # fused stacked experts: split gate_up into gate/up
-        if name.endswith("mlp.experts.gate_up_proj"):
-            gate, up = split_gate_up(data_torch, moe_inter)
-            yield from super().modify_tensors(gate, tn(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), bid)
-            yield from super().modify_tensors(up,   tn(gguf.MODEL_TENSOR.FFN_UP_EXP,   bid), bid)
-            return
-
-        # add .weight suffixes
-        if name.endswith("mlp.experts.down_proj") or name.endswith(".self_attn.learnable_sink_param"):
-            name += ".weight"
-
-        if re.search(r"\.hc_head\.hc_head_(?:fn|base|scale)$", name):
-            name += ".weight"
-
-        if re.search(r"\.hc_(?:attn|mlp)_layer\.hc_pre\.hc_(?:fn|base|scale)$", name):
-            name += ".weight"
-
-        yield from super().modify_tensors(data_torch, name, bid)

conversion/minimax.py

@@ -25,7 +25,7 @@ class MiniMaxText01Model(TextModel):
         # they get in the way of the token sampling process and must be suppressed
 
         tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
-        tokenizer_vocab_size = tokenizer.vocab_size  # ty: ignore[unresolved-attribute]
+        tokenizer_vocab_size = tokenizer.vocab_size
 
         with open(self.dir_model / "model.safetensors.index.json", "r", encoding="utf-8") as f:
             weight_map = json.load(f)["weight_map"]

conversion/muse_glimmer.py

@@ -37,7 +37,7 @@ class MuseGlimmerModel(TextModel):
 
         from transformers import AutoTokenizer
         tok = AutoTokenizer.from_pretrained(self.dir_model)
-        eot_id = tok.convert_tokens_to_ids("<|eot|>")  # ty: ignore[unresolved-attribute]
+        eot_id = tok.convert_tokens_to_ids("<|eot|>")
         if isinstance(eot_id, int) and eot_id >= 0:
             self.gguf_writer.add_eot_token_id(eot_id)
 

conversion/nemotron.py

@@ -5,7 +5,6 @@ from typing import Any, Callable, Iterable, TYPE_CHECKING
 import torch
 
 if TYPE_CHECKING:
-    from pathlib import Path
     from torch import Tensor
 
 from .base import MmprojModel, ModelBase, TextModel, gguf, logger
@@ -202,11 +201,6 @@ class NemotronHModel(GraniteHybridModel):
     model_arch = gguf.MODEL_ARCH.NEMOTRON_H
     is_moe: bool = False
     supports_mtp_export = True
-    _experts: list[dict[str, Tensor]] | None = None
-
-    _SSM_LAYER_TYPES = {"mamba", "linear_attention"}
-    _ATTN_LAYER_TYPES = {"attention", "full_attention"}
-    _MLP_LAYER_TYPES = {"moe"}
 
     def __init__(self, *args, **kwargs):
         # We have to determine the correct model architecture (MoE vs non-MoE) before
@@ -248,8 +242,8 @@ class NemotronHModel(GraniteHybridModel):
             self._ssm_layers = [i for i, val in enumerate(pattern) if val == "M"]
             self._mlp_layers = [i for i, val in enumerate(pattern) if val == ("E" if self.is_moe else "-")]
         else:
-            self._ssm_layers = [i for i, val in enumerate(pattern) if val in self._SSM_LAYER_TYPES]
-            self._mlp_layers = [i for i, val in enumerate(pattern) if val in self._MLP_LAYER_TYPES]
+            self._ssm_layers = [i for i, val in enumerate(pattern) if val == "mamba"]
+            self._mlp_layers = [i for i, val in enumerate(pattern) if val == "moe"]
 
         # `--no-mtp` drops it entirely; `--mtp` exports only the MTP head
         self._mtp_bid: int | None = None
@@ -278,7 +272,7 @@ class NemotronHModel(GraniteHybridModel):
         if isinstance(pattern, str):
             return [i for i, val in enumerate(pattern) if val == "*"]
 
-        return [i for i, val in enumerate(pattern) if val in self._ATTN_LAYER_TYPES]
+        return [i for i, val in enumerate(pattern) if val == "attention"]
 
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
@@ -304,10 +298,6 @@ class NemotronHModel(GraniteHybridModel):
             )
             if not keep:
                 return None
-        # PEFT names adapter tensors using model.layers.*, while Nemotron-H checkpoints
-        # and the GGUF tensor map use backbone.layers.*
-        if name.startswith("model.layers.") and ".mixer." in name:
-            name = name.replace("model.layers.", "backbone.layers.", 1)
         return super().filter_tensors((name, gen))
 
     def prepare_metadata(self, vocab_only: bool):
@@ -515,88 +505,3 @@ class NemotronHModel(GraniteHybridModel):
             experts = [k for d in self._experts for k in d.keys()]
             if len(experts) > 0:
                 raise ValueError(f"Unprocessed experts: {experts}")
-
-
-@ModelBase.register("NemotronHPuzzleForCausalLM")
-@ModelBase.example("nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16")
-class NemotronHPuzzleModel(NemotronHModel):
-    """NVIDIA Puzzle: NemotronH with a per-block MoE config (block_configs).
-
-    The checkpoint also ships an MTP draft head (mtp.safetensors). It is skipped
-    here: there is no Puzzle MTP inference path in tree, and the head is laid out
-    by mtp_block_configs rather than the mtp.layers.* form NemotronHModel maps."""
-
-    model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
-    is_moe: bool = True
-    supports_mtp_export = False
-
-    def __init__(self, dir_model: "Path", *args, **kwargs):
-        hparams = dict(kwargs.pop("hparams", None) or ModelBase.load_hparams(dir_model, self.is_mistral_format))
-
-        self.block_configs: list[dict] = hparams["block_configs"]
-        self.n_layer_trunk = len(self.block_configs)
-
-        # block_configs carries the per-block MoE shape, and is the authority on the
-        # block pattern too: the layers_block_type the HF config wrapper computes is
-        # not sized to it.
-        hparams["num_hidden_layers"] = self.n_layer_trunk
-        hparams["layers_block_type"] = [bc["block_type"] for bc in self.block_configs]
-
-        self.model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
-
-        # Bypass NemotronHModel.__init__: it assumes a flat num_experts_per_tok /
-        # moe_intermediate_size and a layers_block_type sized to block_count, neither
-        # of which hold for Puzzle's per-block config.
-        GraniteHybridModel.__init__(self, dir_model, *args, hparams=hparams, **kwargs)
-
-        self.head_dim = self.find_hparam(["head_dim", "attention_head_dim"])
-        self.d_inner = self.find_hparam(["num_heads"]) * self.d_model
-
-        # NemotronHModel.__init__ folds an MTP block into block_count when the
-        # config carries num_nextn_predict_layers; Puzzle's config does, but its
-        # head has a different layout and no inference path, so stay opted out.
-        self._mtp_bid = None
-
-    def set_gguf_parameters(self):
-        GraniteHybridModel.set_gguf_parameters(self)
-
-        head_dim = self.head_dim
-        if head_dim is None:
-            raise ValueError("Could not find the attention head dim in config")
-        self.gguf_writer.add_key_length(head_dim)
-        self.gguf_writer.add_value_length(head_dim)
-
-        ffn_lengths = [bc.get("moe_intermediate_size") or 0 for bc in self.block_configs]
-        experts_used = [bc.get("num_experts_per_tok") or 0 for bc in self.block_configs]
-
-        self.gguf_writer.add_feed_forward_length(ffn_lengths)
-        self.gguf_writer.add_expert_feed_forward_length(ffn_lengths)
-        self.gguf_writer.add_expert_used_count(experts_used)
-
-        self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["moe_shared_expert_intermediate_size"])
-        self.gguf_writer.add_expert_count(self.hparams["n_routed_experts"])
-        self.gguf_writer.add_expert_shared_count(self.hparams["n_shared_experts"])
-        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
-        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
-        self.gguf_writer.add_expert_group_count(self.hparams["n_group"])
-        self.gguf_writer.add_moe_latent_size(self.hparams["moe_latent_size"])
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # The official BF16 checkpoint (NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16)
-        # names the trunk "model.*" (model.layers.*, model.embeddings, model.norm_f)
-        # where the original release used the NemotronH-style "backbone.*", and spells
-        # the router bias "e_score_correction_bias" instead of "e_score_correction.bias";
-        # normalize so both convert identically.
-        if name.startswith("model."):
-            name = "backbone." + name[len("model."):]
-        if name.endswith("mixer.gate.e_score_correction_bias"):
-            name = name[: -len("e_score_correction_bias")] + "e_score_correction.bias"
-
-        yield from super().modify_tensors(data_torch, name, bid)
-
-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        # Drop the MTP head unconditionally; see the class docstring.
-        if item[0].startswith("mtp."):
-            return None
-        return super().filter_tensors(item)

conversion/qwen.py

@@ -379,13 +379,6 @@ class Qwen3NextModel(_QwenMtpMixin, Qwen2MoeModel):
         self.gguf_writer.add_ssm_group_count(self.hparams["linear_num_key_heads"])
         self.gguf_writer.add_ssm_time_step_rank(self.hparams["linear_num_value_heads"])
         self.gguf_writer.add_ssm_inner_size(self.hparams["linear_value_head_dim"] * self.hparams["linear_num_value_heads"])
-        if (layer_types := self.hparams.get("layer_types")) is not None:
-            n_layer = self.hparams["num_hidden_layers"]
-            if len(layer_types) != n_layer:
-                raise ValueError(f"layer_types has {len(layer_types)} entries, expected num_hidden_layers ({n_layer})")
-            recurrent = [t == "linear_attention" for t in layer_types]
-            recurrent += [False] * (self.block_count - n_layer)
-            self.gguf_writer.add_recurrent_layers(recurrent)
         self.gguf_writer.add_full_attention_interval(self.hparams.get("full_attention_interval", 4))
         if (rope_dim := self.hparams.get("head_dim")) is None:
             rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
@@ -646,7 +639,7 @@ class Qwen3_5MoeTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
     model_arch = gguf.MODEL_ARCH.QWEN35MOE
 
 
-@ModelBase.register("DFlashDraftModel", "DFlash2DraftModel")
+@ModelBase.register("DFlashDraftModel")
 @ModelBase.example("z-lab/Qwen3.5-9B-DFlash")
 class DFlashModel(Qwen3Model):
     model_arch = gguf.MODEL_ARCH.DFLASH
@@ -685,98 +678,34 @@ class DFlashModel(Qwen3Model):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
 
-        dflash_config = self.hparams.get("dflash_config", {})
-        block_size = dflash_config.get("block_size", self.hparams.get("block_size", 16))
+        block_size = self.hparams.get("block_size", 16)
         self.gguf_writer.add_block_size(block_size)
-
-        if "conv_kernel_size" in dflash_config:
-            self.gguf_writer.add_conv_kernel_size(int(dflash_config["conv_kernel_size"]))
-            self.gguf_writer.add_conv_group_size(int(dflash_config["conv_group_size"]))
-            self.gguf_writer.add_selector_rank(int(dflash_config["selector_rank"]))
-            self.gguf_writer.add_selector_top_k(int(dflash_config["selector_top_k"]))
-
-        output_multiplier = dflash_config.get(
-            "output_multiplier", self.hparams.get("output_multiplier")
-        )
-        if output_multiplier is not None:
-            self.gguf_writer.add_logit_scale(float(output_multiplier))
-        softcap = dflash_config.get(
-            "final_logit_softcapping", self.hparams.get("final_logit_softcapping")
-        )
-        if softcap is not None and float(softcap) > 0:
-            self.gguf_writer.add_final_logit_softcapping(float(softcap))
-        embedding_scale = dflash_config.get(
-            "input_embedding_scale", self.hparams.get("input_embedding_scale")
-        )
-        if embedding_scale is not None:
-            self.gguf_writer.add_embedding_scale(float(embedding_scale))
+        dflash_config = self.hparams.get("dflash_config", {})
 
         target_layer_ids = dflash_config.get("target_layer_ids", [])
         if target_layer_ids:
             extract_layer_ids = [i + 1 for i in target_layer_ids]
             self.gguf_writer.add_target_layers(extract_layer_ids)
 
-        use_sliding_window = self.hparams.get("use_sliding_window", False) or dflash_config.get("use_swa", False)
-        sliding_window = dflash_config.get("swa_window_size") or self.hparams.get("sliding_window")
+        use_sliding_window = self.hparams.get("use_sliding_window", False)
+        sliding_window = self.hparams.get("sliding_window")
         layer_types = self.hparams.get("layer_types")
         if use_sliding_window and sliding_window and layer_types:
             is_swa = [lt == "sliding_attention" for lt in layer_types]
             self.gguf_writer.add_sliding_window(sliding_window)
             self.gguf_writer.add_sliding_window_pattern(is_swa)
 
-        causal = self.hparams.get("is_causal")
-        if causal is None:
-            causal = dflash_config.get("causal")
-        if causal is not None:
-            self.gguf_writer.add_causal_attention(bool(causal))
-
-        # M-RoPE target: the draft ropes on the temporal dim only, so write
-        # degenerate sections [n_rot/2, 0, 0, 0]
-        if self._target_uses_mrope():
-            head_dim = self.hparams.get("head_dim") or self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
-            self.gguf_writer.add_rope_dimension_sections([head_dim // 2, 0, 0, 0])
-
-    def _target_uses_mrope(self) -> bool:
-        if self.target_model_dir is None:
-            return False
-        with open(self.target_model_dir / "config.json", "r", encoding="utf-8") as f:
-            cfg = json.load(f)
-        cfg = cfg.get("text_config", cfg)
-        rope = cfg.get("rope_parameters") or cfg.get("rope_scaling") or {}
-        return "mrope_section" in rope
-
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         name, gen = item
         if not name.startswith("model."):
             name = "model." + name
-        if "sink" in name and not name.endswith(".weight"):
-            name += ".weight"
         return super().filter_tensors((name, gen))
 
-    _ROPE_PERMUTE_SUFFIXES = (
-        "self_attn.q_proj.weight",
-        "self_attn.k_proj.weight",
-        "self_attn.q_norm.weight",
-        "self_attn.k_norm.weight",
-    )
-
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if name == "model.embed_tokens.weight" and not self.hparams.get("has_embed_tokens", True):
             return
 
-        # interleaved-rope checkpoints (rope_is_neox_style = false) -> NeoX layout: per head, even dims first then odd
-        if not self.hparams.get("rope_is_neox_style", True) and name.endswith(self._ROPE_PERMUTE_SUFFIXES):
-            head_dim = self.hparams["head_dim"]
-            shape = data_torch.shape
-            data_torch = data_torch.reshape(-1, head_dim // 2, 2, *shape[1:]).transpose(1, 2).reshape(shape)
-
-        if name in (
-            "model.candidate_selector.predecessor_codebook",
-            "model.candidate_selector.successor_codebook",
-        ):
-            name += ".weight"
-
         yield from super().modify_tensors(data_torch, name, bid)
 
 
@@ -830,10 +759,6 @@ class DSparkModel(DFlashModel):
         super().set_gguf_parameters()
         self.gguf_writer.add_sample_from_anchor(self._sample_from_anchor)
 
-        # confidence head is optional: vanilla-markov exports ship without it
-        has_conf = any("confidence_head.proj" in name for name in self.model_tensors)
-        self.gguf_writer.add_has_confidence_head(has_conf)
-
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         if item[0] == "t2d":  # not used at runtime
@@ -852,7 +777,7 @@ class DSparkModel(DFlashModel):
             self._d2t = data_torch
             return
 
-        if self._n_vocab_draft == self.hparams["vocab_size"] and name.endswith("lm_head.weight"):
+        if self._n_vocab_draft == self.hparams["vocab_size"] and name.endswith(("embed_tokens.weight", "lm_head.weight")):
             return
 
         # interleaved-rope checkpoints (rope_is_neox_style = false) -> NeoX layout: per head, even dims first then odd

conversion/qwen3tts.py

@@ -276,10 +276,6 @@ class Qwen3TTSSpeakerEncoderModel(MmprojModel):
         # ConvTranspose1d kernels: only F16/F32 are implemented, no BF16
         if new_name.endswith(".conv.weight") and (".up.blk." in new_name or ".dac.blk." in new_name):
             return gguf.GGMLQuantizationType.F32
-        # the code predictor FFN intermediate peaks around 1.5e5, above the F16 range, and mul_mat
-        # casts its input to the weight type
-        if new_name.startswith("a.gen.code.blk.") and new_name.endswith(".ffn_down.weight"):
-            return gguf.GGMLQuantizationType.F32
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     @classmethod

conversion/qwen4exp.py

@@ -1,195 +0,0 @@
-from __future__ import annotations
-
-from typing import Iterable, cast
-
-import torch
-from torch import Tensor
-
-import gguf
-import numpy as np
-
-from .base import ModelBase
-from .qwen import _LinearAttentionVReorderBase, _Qwen35MRopeMixin
-from .qwen3vl import Qwen3VLVisionModel
-
-
-@ModelBase.register("Qwen4ExpForConditionalGeneration", "Qwen4ExpForCausalLM")
-@ModelBase.example("Qwen/Qwen3.8-Flash-Next")
-class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
-    """Qwen3.8-Flash-Next.
-
-    Shares the Qwen3.5 gated delta net and interleaved mrope, and adds three things:
-    hyper-connections in place of every layer norm, QSA sparse attention on the full
-    attention layers, and PLE n-gram hash embeddings on a single layer.
-    """
-
-    model_arch = gguf.MODEL_ARCH.QWEN4EXP
-
-    # the MTP block is a separate draft head; vLLM drops it too
-    supports_mtp_export = False
-    no_mtp = True
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        # only the shard names, so the table itself is never held
-        self._ple_shards: dict[int, str] = {}
-        self._ple_row_dim: int | None = None
-
-    def _read_hash_constants(self, suffix: str) -> list[int]:
-        """Read an int64 PLE constant straight from the checkpoint.
-
-        prepare_tensors() casts every non-float dtype to float32 before
-        modify_tensors() sees it (base.py), which would silently round these
-        45-bit multipliers. Reading the lazy tensor here bypasses that.
-        """
-        for name, gen in self.model_tensors.items():
-            if name.endswith(suffix):
-                t = gen()
-                if t.dtype != torch.int64:
-                    t = t.to(torch.int64)
-                return [int(x) for x in t.tolist()]
-        raise ValueError(f"PLE constant {suffix!r} missing from the checkpoint")
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        hp = self.hparams
-
-        self.gguf_writer.add_hyper_connection_count(hp["hc_count"])
-        self.gguf_writer.add_hyper_connection_low_rank(hp["hc_lowrank"])
-
-        n_layer = hp["num_hidden_layers"]
-        self.gguf_writer.add_indexer_head_count(hp["indexer_n_heads"])
-        self.gguf_writer.add_indexer_key_length(hp["indexer_head_dim"])
-        self.gguf_writer.add_indexer_top_k(hp["indexer_budget"])
-        ratio = hp["indexer_compress_ratio"]
-        layer_types = hp["layer_types"]
-        self.gguf_writer.add_attention_compress_ratios(
-            [ratio if layer_types[i] == "full_attention" else 0 for i in range(n_layer)]
-        )
-
-        # ple_layer_ids is 1-based in the HF config; empty means no n-gram table,
-        # so emit no PLE keys rather than optional ones
-        ple_layers = [i - 1 for i in hp["ple_layer_ids"]]
-        if not ple_layers:
-            return
-        self.gguf_writer.add_ple_layers(ple_layers)
-        self.gguf_writer.add_ple_ngram_size(hp["ngram_size"])
-        self.gguf_writer.add_ple_heads_per_ngram(hp["heads_per_ngram"])
-        self.gguf_writer.add_ple_conv_kernel(hp["ple_conv_kernel_size"])
-        self.gguf_writer.add_ple_eos_token_id(self._eos_token_id())
-        # an image is decoded as an embeddings-only batch, so the graph has no placeholder
-        # ids to hash; carry the id and let it stand in for those positions
-        _img = self._image_token_id()
-        if _img is not None:
-            self.gguf_writer.add_ple_image_token_id(int(_img))
-        if self._ple_row_dim is not None:
-            self.gguf_writer.add_embedding_length_per_layer_input(self._ple_row_dim)
-
-        self.gguf_writer.add_ple_layer_multipliers(
-            self._read_hash_constants("ple_embedding.layer_multipliers"))
-        self.gguf_writer.add_ple_head_offsets(
-            self._read_hash_constants("ple_embedding.ngram_heads_offsets"))
-        self.gguf_writer.add_ple_head_vocab_sizes(
-            self._read_hash_constants("ple_embedding.ngram_heads_vocab_sizes"))
-
-    def _image_token_id(self) -> int | None:
-        img = self.hparams.get("image_token_id")
-        return None if img is None else int(img)
-
-    def _eos_token_id(self) -> int:
-        eos = self.hparams.get("eos_token_id")
-        if isinstance(eos, list):
-            # the PLE hash resets n-grams on the primary EOS
-            return int(eos[-1])
-        if eos is None:
-            raise ValueError("eos_token_id is required: the PLE hash resets its n-grams on it")
-        return int(eos)
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # int64 hash constants must stay exact; 1-D tensors force F32, so use KV
-        if name.endswith("ple_embedding.layer_multipliers"):
-            self._ple_multipliers = [int(x) for x in data_torch.tolist()]
-            return []
-        if name.endswith("ple_embedding.ngram_heads_offsets"):
-            self._ple_head_offsets = [int(x) for x in data_torch.tolist()]
-            return []
-        if name.endswith("ple_embedding.ngram_heads_vocab_sizes"):
-            self._ple_head_vocab_sizes = [int(x) for x in data_torch.tolist()]
-            return []
-
-        if ".ngram_embedding.shard_" in name:
-            return self._place_ple_shard(data_torch, name)
-
-        # one projection feeds indexer q and k; split it, as minimax-m3 does
-        if ".indexer.index_qk_proj.weight" in name:
-            n_q = self.hparams["indexer_n_heads"] * self.hparams["indexer_head_dim"]
-            q = data_torch[:n_q]
-            k = data_torch[n_q:]
-            return [
-                (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_Q_PROJ, bid, ".weight"), q),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_K_PROJ, bid, ".weight"), k),
-            ]
-
-        # Gemma zero-centred gammas the inherited norm.weight rule misses
-        if name.endswith((".ple.norm_key.weight", ".ple.norm_query.weight", ".ple.norm_conv.weight",
-                          ".indexer.q_layernorm.weight", ".indexer.k_layernorm.weight")):
-            return [(self.map_tensor_name(name), data_torch + 1)]
-
-        if name.endswith(".ple.conv1d.weight"):
-            return [(self.map_tensor_name(name), data_torch.squeeze())]
-
-        return super().modify_tensors(data_torch, name, bid)
-
-    # the shards concatenate into a tensor of well over 100 GB
-    # use LazyChunkedTensor here, a single shard resident at a time
-    def _place_ple_shard(self, data_torch: Tensor, name: str) -> Iterable[tuple[str, Tensor]]:
-
-        idx = int(name.rpartition(".shard_")[2].partition(".")[0])
-        n_parts = self.hparams["split_ngram_parts"]
-
-        self._ple_shards[idx] = name
-        self._ple_row_dim = int(data_torch.shape[-1])
-
-        if len(self._ple_shards) < n_parts:
-            return []
-
-        # the checkpoint may yield the shards in any order, the row order is by index
-        shards = [self._ple_shards[i] for i in sorted(self._ple_shards)]
-        rows = 0
-        for shard in shards:
-            shape = self.model_tensors[shard]().shape
-            if int(shape[-1]) != self._ple_row_dim:
-                raise ValueError(
-                    f"PLE shard {shard} has row dim {int(shape[-1])}, expected {self._ple_row_dim}")
-            rows += int(shape[0])
-
-        table = gguf.LazyChunkedTensor(
-            [self._load_ple_shard(shard) for shard in shards],
-            shape=(rows, self._ple_row_dim),
-            dtype=np.float32,
-        )
-        gguf_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.PER_LAYER_TOKEN_EMBD]
-        return [(gguf_name + ".weight", cast(Tensor, table))]
-
-    def _load_ple_shard(self, name: str):
-        def load() -> np.ndarray:
-            from .base import LazyTorchTensor
-
-            # a fresh lazy tensor every call, or to_eager() memoizes every shard
-            eager = LazyTorchTensor.to_eager(self.model_tensors[name]())
-            return eager.to(torch.float32).contiguous().numpy()
-        return load
-
-    def prepare_tensors(self):
-        super().prepare_tensors()
-        n_parts = self.hparams.get("split_ngram_parts", 0)
-        if self._ple_shards and len(self._ple_shards) != n_parts:
-            raise ValueError(
-                f"got {len(self._ple_shards)} PLE embedding shards, expected {n_parts}"
-            )
-
-
-@ModelBase.register("Qwen4ExpForConditionalGeneration")
-@ModelBase.example("Qwen/Qwen3.8-Flash-Next")
-class Qwen4ExpVisionModel(Qwen3VLVisionModel):
-    """The vision tower is an unmodified Qwen3-VL ViT."""

conversion/spark2_5.py

@@ -1,65 +0,0 @@
-from __future__ import annotations
-
-from collections.abc import Iterable
-from typing import TYPE_CHECKING
-
-if TYPE_CHECKING:
-    from torch import Tensor
-
-from .base import ModelBase, TextModel, gguf
-
-
-@ModelBase.register("Spark2_5ForCausalLM")
-@ModelBase.example("XHToken/Spark-X2.5-1.7B")
-class Spark2_5Model(TextModel):
-    model_arch = gguf.MODEL_ARCH.SPARK2_5
-
-    def set_gguf_parameters(self) -> None:
-        super().set_gguf_parameters()
-
-        hparams = self.hparams
-        layer_types = hparams["layer_types"]
-        if len(layer_types) != self.block_count:
-            raise ValueError(
-                f"Spark2_5 layer_types length {len(layer_types)} != num_hidden_layers {self.block_count}"
-            )
-        if any(layer_type not in ("sliding_attention", "full_attention") for layer_type in layer_types):
-            raise ValueError(f"Spark2_5 has unsupported layer_types: {layer_types}")
-        if hparams.get("gate_attn_act_mode") != "sigmoid" or hparams.get("headwise_attn_output_gate") is not True:
-            raise ValueError("Spark2_5 conversion requires head-wise sigmoid attention gates")
-        if hparams.get("hidden_act") != "gelu":
-            raise ValueError(f"Spark2_5 conversion requires GELU, got {hparams.get('hidden_act')!r}")
-
-        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
-        self.gguf_writer.add_sliding_window(hparams["sliding_window"])
-        self.gguf_writer.add_sliding_window_pattern(
-            [layer_type == "sliding_attention" for layer_type in layer_types]
-        )
-
-        head_dim = hparams["head_dim"]
-        full_rope = self.rope_parameters["full_attention"]
-        swa_rope = self.rope_parameters["sliding_attention"]
-        self.gguf_writer.add_rope_dimension_count(
-            int(head_dim * float(full_rope["partial_rotary_factor"]))
-        )
-        self.gguf_writer.add_rope_dimension_count_swa(
-            int(head_dim * float(swa_rope["partial_rotary_factor"]))
-        )
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        if name.endswith(".self_attn.q_k_v_proj.weight"):
-            if bid is None:
-                raise ValueError(f"Spark2_5 fused QKV tensor has no block id: {name}")
-            yield self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_QKV, bid), data_torch
-            return
-
-        if name.endswith(".self_attn.g_proj.weight"):
-            if bid is None:
-                raise ValueError(f"Spark2_5 attention gate tensor has no block id: {name}")
-            expected = self.hparams["num_attention_heads"]
-            if data_torch.shape[0] != expected:
-                raise ValueError(
-                    f"Spark2_5 layer {bid} attention gate width {data_torch.shape[0]} != head count {expected}"
-                )
-
-        yield from super().modify_tensors(data_torch, name, bid)

convert_hf_to_gguf.py

@@ -157,10 +157,6 @@ def parse_args() -> argparse.Namespace:
         help="Store tensors dequantized from FP8 as Q8_0 instead of BF16/F16.",
     )
 
-    parser.add_argument(
-        "--fuse-qkv", action="store_true",
-        help="Fuse separate Q, K, V weight tensors into a single QKV tensor.",
-    )
     parser.add_argument(
         "--target-model-dir", type=str, default=None,
         help=(
@@ -294,7 +290,6 @@ def main() -> None:
                                      target_model_dir=Path(args.target_model_dir) if args.target_model_dir else None,
                                      fuse_gate_up_exps=args.fuse_gate_up_exps,
                                      fp8_as_q8=args.fp8_as_q8,
-                                     fuse_qkv=args.fuse_qkv,
                                      )
 
         if args.vocab_only:

convert_hf_to_gguf_update.py

@@ -176,7 +176,6 @@ pre_computed_hashes = [
     {"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"},
-    {"name": "hy_v4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hy4-preview", "chkhsh": "e6ddf9c6686791c12d698d34c31ab9be1fea9af5a3d9a6909783ab382198ae1c"},
     # falcon-h1 series uses 4 different tokenizers across model sizes (0.5b - 34b), hence we need to define 4 different hashes
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base", "chkhsh": "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6"},
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-1B-Base", "chkhsh": "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86"},
@@ -191,7 +190,6 @@ pre_computed_hashes = [
     {"name": "gpt-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/evilfreelancer/ruGPT3XL", "chkhsh": "0fe1cf6eda062318a1af7270f3331a85c539a01778ff948e24388e949c5282f4"},
     # lfm2 variants
     {"name": "lfm2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LiquidAI/LFM2.5-8B-A1B", "chkhsh": "9e454714343b69b99b71795c1d27a68c2a1d15dab111f4d353109f966af29da7"},
-    {"name": "spark2_5", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/XHToken/Spark-X2.5-1.7B", "chkhsh": "0a766d034107bc736a3f2dc4968fd62e54a3570f1454443e0c5a4cc6bd7941ed"},
 ]
 
 

docs/android.md

@@ -53,7 +53,7 @@ To see what it might look like visually, here's an old demo of an interactive se
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
 
 ## Cross-compile CLI using Android NDK
-It's possible to build `llama.cpp` for Android on your host system via CMake and the Android NDK. If you are interested in this path, ensure you already have an environment prepared to cross-compile programs for Android (i.e., install the Android SDK/NDK and set `ANDROID_NDK` to the NDK root). Note that, unlike desktop environments, the Android environment ships with a limited set of native libraries, and so only those libraries are available to CMake when building with the Android NDK (see: https://developer.android.com/ndk/guides/stable_apis.)
+It's possible to build `llama.cpp` for Android on your host system via CMake and the Android NDK. If you are interested in this path, ensure you already have an environment prepared to cross-compile programs for Android (i.e., install the Android SDK). Note that, unlike desktop environments, the Android environment ships with a limited set of native libraries, and so only those libraries are available to CMake when building with the Android NDK (see: https://developer.android.com/ndk/guides/stable_apis.)
 
 Once you're ready and have cloned `llama.cpp`, invoke the following in the project directory:
 
@@ -62,22 +62,18 @@ $ cmake \
   -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
   -DANDROID_ABI=arm64-v8a \
   -DANDROID_PLATFORM=android-28 \
-  -DGGML_NATIVE=OFF \
+  -DCMAKE_C_FLAGS="-march=armv8.7a" \
+  -DCMAKE_CXX_FLAGS="-march=armv8.7a" \
   -DGGML_OPENMP=OFF \
   -DGGML_LLAMAFILE=OFF \
-  -DLLAMA_OPENSSL=OFF \
   -B build-android
 ```
 
 Notes:
-  - `GGML_NATIVE=OFF` is required for cross-compilation because the host CPU is not the Android target CPU
   - While later versions of Android NDK ship with OpenMP, it must still be installed by CMake as a dependency, which is not supported at this time
   - `llamafile` does not appear to support Android devices (see: https://github.com/Mozilla-Ocho/llamafile/issues/325)
-  - `LLAMA_OPENSSL=OFF` avoids depending on OpenSSL, which is not part of the Android NDK stable native API set
 
-The above command configures a portable Android `arm64-v8a` build. Do not add a global `-march` flag unless you intentionally want to raise the baseline instruction set for every compiled source.
-
-For optional KleidiAI acceleration on Android `arm64-v8a`, see the [Arm KleidiAI section in build.md](./build.md#arm-kleidiai).
+The above command should configure `llama.cpp` with the most performant options for modern devices. Even if your device is not running `armv8.7a`, `llama.cpp` includes runtime checks for available CPU features it can use.
 
 Feel free to adjust the Android ABI for your target. Once the project is configured:
 

docs/autoparser.md

@@ -443,21 +443,21 @@ Each returned parser is wrapped by `wrap_for_generation_prompt()`, which prepend
 |                                           | `wrap_for_generation_prompt()`, string helpers                                  |
 | `common/chat-peg-parser.h/cpp`            | `common_chat_peg_builder`, `common_chat_peg_mapper`, and helpers                |
 | `common/chat.cpp`                         | Entry point: `common_chat_templates_apply_jinja()`                              |
-| `tests/test-chat-auto-parser.cpp`         | Auto-parser unit tests; also a debug tool when given a template path            |
-| `tests/test-chat-analysis.cpp`            | Template differential analysis debug tool                                       |
+| `tools/parser/debug-template-parser.cpp`  | Debug tool for template analysis                                                |
+| `tools/parser/template-analysis.cpp`      | Template analysis tool                                                          |
 
 ## Testing & Debugging
 
 ### Debug Tools
 
-**Template Debugger**: `tests/test-chat-auto-parser.cpp`
+**Template Debugger**: `tools/parser/debug-template-parser.cpp`
 
-- Usage: `./bin/test-chat-auto-parser path/to/template.jinja` (without a path, it runs the automated tests)
+- Usage: `./bin/llama-debug-template-parser path/to/template.jinja`
 - Shows detected format, markers, generated parser, and GBNF grammar
 
-**Template Analysis**: `tests/test-chat-analysis.cpp`
+**Template Analysis**: `tools/parser/template-analysis.cpp`
 
-- Usage: `./bin/test-chat-analysis --template-file path/to/template.jinja` (without arguments, it runs on all templates from the test suite)
+- Usage: `./bin/llama-template-analysis path/to/template.jinja`
 
 **Debug Logging**: Enable with `LLAMA_ARG_LOG_VERBOSITY=2`
 
@@ -514,13 +514,12 @@ The following templates have active tests in `tests/test-chat.cpp`:
 | Mistral Small 3.2 | JSON_NATIVE | `[TOOL_CALLS]func[ARGS]{...}` with call ID |
 | Devstral | JSON_NATIVE | `[TOOL_CALLS]func[ARGS]{...}` without call ID |
 | StepFun 3.5 Flash | TAG_WITH_TAGGED | `<function=X><parameter=Y>` format |
-| Spark2.5 | TAG_WITH_TAGGED | `<tool_call>name<arg_key>...<arg_value>...` format |
 
 ## Adding Support for New Templates
 
 To support a new template format:
 
-1. **If it follows standard patterns** — The auto-parser should detect it automatically. Run `test-chat-auto-parser <template_path>` to verify markers are correctly extracted.
+1. **If it follows standard patterns** — The auto-parser should detect it automatically. Run `llama-debug-template-parser` to verify markers are correctly extracted.
 2. **If differential analysis extracts incorrect markers** — Add a workaround lambda to the `workarounds` vector in `common/chat-diff-analyzer.cpp`. Inspect the template source for a unique identifying substring.
 3. **If it needs fundamentally different handling** — Add a dedicated handler function in `chat.cpp` before the auto-parser block (as done for GPT-OSS, Functionary v3.2, and Ministral).
 

docs/backend/OPENVINO.md

@@ -22,8 +22,8 @@ The OpenVINO backend is implemented in `ggml/src/ggml-openvino` and provides a t
   - [0. Prerequisites](#0-prerequisites)
   - [1. Install OpenVINO Runtime](#1-install-openvino-runtime)
   - [2. Build llama.cpp with OpenVINO Backend](#2-build-llamacpp-with-openvino-backend)
-    - [Ubuntu Build Script](#ubuntu-build-script)
-    - [Windows Build Script](#windows-build-script)
+    - [Automated Ubuntu Build Script](#automated-ubuntu-build-script)
+    - [Automated Windows Build Script](#automated-windows-build-script)
   - [3. Download Sample Model](#3-download-sample-model)
   - [4. Run Inference with OpenVINO Backend](#4-run-inference-with-openvino-backend)
   - [5. Docker Build](#5-docker-build)
@@ -96,7 +96,7 @@ Although, the validated models below were tested with `llama-cli` using the `Q4_
   - **SL** = Stateless (`GGML_OPENVINO_STATEFUL_EXECUTION=0`)
   - **SF** = Stateful (`GGML_OPENVINO_STATEFUL_EXECUTION=1`)
   - Note: The NPU operates in stateless mode only.
-- **Validation system:** Intel® Core™ Ultra 5 238V (Lunar Lake) | 32 GB RAM | Ubuntu 24.04 | Intel OpenCL GPU Driver 26.31.39395.13-0 | Intel NPU Driver 1.35.0.
+- **Validation system:** Intel® Core™ Ultra 5 238V (Lunar Lake) | 32 GB RAM | Ubuntu 24.04 | Intel OpenCL GPU Driver 26.18.38308.1 | Intel NPU Driver 1.33.0.
 - See [Known Limitations](#known-limitations) for context on observed failures.
 
 | Model | CPU (SL / SF) | GPU (SL / SF) | NPU (SL) |
@@ -105,32 +105,27 @@ Although, the validated models below were tested with `llama-cli` using the `Q4_
 | [bartowski/Llama-3.2-3B-Instruct-Q4_K_M](https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 | [bartowski/Meta-Llama-3.1-8B-Instruct-Q4_K_M](https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 |  |  |  |  |
-| [Qwen/qwen2.5-1.5b-instruct-q4_k_m](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [Qwen/qwen2.5-coder-7b-instruct-q4_k_m](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/Qwen_Qwen3-0.6B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3-0.6B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/Qwen_Qwen3-1.7B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3-1.7B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [Qwen/Qwen3-4B-Q4_K_M](https://huggingface.co/Qwen/Qwen3-4B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [lm-kit/Qwen3-8B-Q4_K_M](https://huggingface.co/lm-kit/qwen-3-8b-instruct-gguf) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/Qwen_Qwen3.5-0.8B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3.5-0.8B-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
-| [bartowski/Qwen_Qwen3.5-2B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3.5-2B-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
-| [bartowski/Qwen_Qwen3.5-4B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3.5-4B-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
-| [lmstudio-community/Qwen3.5-9B-Q4_K_M](https://huggingface.co/lmstudio-community/Qwen3.5-9B-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
+| [Qwen/qwen2.5-1.5b-instruct-q4_k_m](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [Qwen/qwen2.5-coder-7b-instruct-q4_k_m](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [bartowski/Qwen_Qwen3-0.6B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3-0.6B-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [bartowski/Qwen_Qwen3-1.7B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3-1.7B-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [Qwen/Qwen3-4B-Q4_K_M](https://huggingface.co/Qwen/Qwen3-4B-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [lm-kit/Qwen3-8B-Q4_K_M](https://huggingface.co/lm-kit/qwen-3-8b-instruct-gguf) | ✓ / ✓ | ✓ / ✗ | ✓ |
 |  |  |  |  |
-| [unsloth/gemma-3-4b-it-Q4_K_M](https://huggingface.co/unsloth/gemma-3-4b-it-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/google_gemma-4-E2B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E2B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
+| [unsloth/gemma-3-4b-it-Q4_K_M](https://huggingface.co/unsloth/gemma-3-4b-it-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [bartowski/google_gemma-4-E2B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E2B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✓ |
 | [bartowski/google_gemma-4-E4B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E4B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✓ |
-| [bartowski/gemma-4-12B-it-Q4_K_M](https://huggingface.co/bartowski/gemma-4-12B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✓ |
+| [bartowski/gemma-4-12B-it-Q4_K_M](https://huggingface.co/bartowski/gemma-4-12B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
 |  |  |  |  |
-| [bartowski/Phi-3-mini-4k-instruct-Q4_K_M](https://huggingface.co/bartowski/Phi-3-mini-4k-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/Phi-3.5-mini-instruct-Q4_K_M](https://huggingface.co/bartowski/Phi-3.5-mini-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/microsoft_Phi-4-mini-instruct-Q4_K_M](https://huggingface.co/bartowski/microsoft_Phi-4-mini-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
+| [bartowski/Phi-3-mini-4k-instruct-Q4_K_M](https://huggingface.co/bartowski/Phi-3-mini-4k-instruct-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [bartowski/Phi-3.5-mini-instruct-Q4_K_M](https://huggingface.co/bartowski/Phi-3.5-mini-instruct-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
 |  |  |  |  |
 | [bartowski/Mistral-7B-Instruct-v0.3-Q4_K_M](https://huggingface.co/bartowski/Mistral-7B-Instruct-v0.3-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 | [QuantFactory/Ministral-3b-instruct.Q4_K_M](https://huggingface.co/QuantFactory/Ministral-3b-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 | [bartowski/Ministral-8B-Instruct-2410-Q4_K_M](https://huggingface.co/bartowski/Ministral-8B-Instruct-2410-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 |  |  |  |  |
 | [bartowski/DeepSeek-R1-Distill-Llama-8B-Q4_K_M](https://huggingface.co/bartowski/DeepSeek-R1-Distill-Llama-8B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/DeepSeek-R1-Distill-Qwen-7B-Q4_K_M](https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
+| [bartowski/DeepSeek-R1-Distill-Qwen-7B-Q4_K_M](https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
 |  |  |  |  |
 | [ibm-granite/granite-4.0-350m-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-350m-GGUF) | ✓ / ✓ | ✗ / ✗ | ✓ |
 | [ibm-granite/granite-4.0-micro-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-micro-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
@@ -138,10 +133,10 @@ Although, the validated models below were tested with `llama-cli` using the `Q4_
 | [ibm-research/granite-3.2-8b-instruct-Q4_K_M](https://huggingface.co/ibm-research/granite-3.2-8b-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 |  |  |  |  |
 | [HuggingFaceTB/smollm2-1.7b-instruct-q4_k_m](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [openbmb/MiniCPM-V-2_6-Q4_K_M](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/tencent_Hunyuan-7B-Instruct-Q4_K_M](https://huggingface.co/bartowski/tencent_Hunyuan-7B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct-Q4_K_M](https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/prism-ml_Bonsai-8B-unpacked-Q4_K_M](https://huggingface.co/bartowski/prism-ml_Bonsai-8B-unpacked-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
+| [openbmb/MiniCPM-V-2_6-Q4_K_M](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [bartowski/tencent_Hunyuan-7B-Instruct-Q4_K_M](https://huggingface.co/bartowski/tencent_Hunyuan-7B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct-Q4_K_M](https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [bartowski/prism-ml_Bonsai-8B-unpacked-Q4_K_M](https://huggingface.co/bartowski/prism-ml_Bonsai-8B-unpacked-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
 |  |  |  |  |
 | [gpustack/bge-m3-Q4_K_M.gguf](https://huggingface.co/gpustack/bge-m3-GGUF) | ✓ | ✗ | ✗ |
 
@@ -222,18 +217,18 @@ cmake --build build\ReleaseOV --parallel
 > [!NOTE]
 > The Windows install path is `C:\Intel\openvino` (no spaces) to avoid quoting problems some CMake/Ninja toolchains have with `C:\Program Files (x86)\...`. Adjust to wherever you installed OpenVINO Runtime. From `cmd`, run `C:\Intel\openvino\setupvars.bat`; from PowerShell, run `& "C:\Intel\openvino\setupvars.ps1"` instead. Once the build is finished you can launch the binaries from any `cmd` or `PowerShell` window after sourcing the matching `setupvars` script for that shell.
 
-#### Ubuntu Build Script
+#### Automated Ubuntu Build Script
 
 For Ubuntu24 users, the following shell script automates the prerequisite installs (build tools, OpenCL ICD), the OpenVINO Runtime download/extract/setup, and the Ninja-based llama.cpp build.
-Save the following as `build-llamacpp-ov.sh` next to where you want the `llama.cpp` folder to land, then run it:
+Save the following as `ubuntu-llamacpp-ov-install.sh` next to where you want the `llama.cpp` folder to land, then run it:
 
 ```bash
-chmod +x build-llamacpp-ov.sh
-./build-llamacpp-ov.sh
+chmod +x ubuntu-llamacpp-ov-install.sh
+./ubuntu-llamacpp-ov-install.sh
 ```
 
 <details>
-<summary>Click to expand <code>build-llamacpp-ov.sh</code></summary>
+<summary>Click to expand <code>ubuntu-llamacpp-ov-install.sh</code></summary>
 
 ```bash
 #!/usr/bin/env bash
@@ -242,8 +237,8 @@ chmod +x build-llamacpp-ov.sh
 # ============================================
 set -euo pipefail
 
-OPENVINO_VERSION_MAJOR="2026.3.1"
-OPENVINO_VERSION_FULL="2026.3.1.22476.56d9685302d"
+OPENVINO_VERSION_MAJOR="2026.3"
+OPENVINO_VERSION_FULL="2026.3.0.22451.bd8d6542e3c"
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 OPENVINO_INSTALL_DIR="/opt/intel/openvino_${OPENVINO_VERSION_MAJOR}"
@@ -318,9 +313,8 @@ fi
 echo "============================================"
 echo "Configuring with CMake..."
 echo "============================================"
-set +u
+# shellcheck disable=SC1091
 source "${OPENVINO_ROOT}/setupvars.sh"
-set -u
 
 cmake -B build/ReleaseOV -G Ninja \
     -DCMAKE_BUILD_TYPE=Release \
@@ -340,27 +334,27 @@ echo "  ./build/ReleaseOV/bin/llama-cli -m model.gguf"
 ```
 
 > [!NOTE]
-> The script pins OpenVINO `2026.3.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
+> The script pins OpenVINO `2026.3` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
 
 </details>
 
-#### Windows Build Script
+#### Automated Windows Build Script
 
 For Windows users, the following `.bat` script automates the prerequisite installs (Git, Ninja, CMake, Visual Studio 2022 Build Tools, vcpkg + OpenCL), the OpenVINO Runtime download/extract, and the Ninja-based llama.cpp build.
-Save the following as `build-llamacpp-ov.bat` next to where you want the `llama.cpp` to land, then run it from either **Command Prompt** or **PowerShell**:
+Save the following as `windows-llamacpp-ov-install.bat` next to where you want the `llama.cpp` to land, then run it from either **Command Prompt** or **PowerShell**:
 
 ```cmd
 :: Command Prompt
-build-llamacpp-ov.bat
+windows-llamacpp-ov-install.bat
 ```
 
 ```powershell
 # PowerShell
-.\build-llamacpp-ov.bat
+.\windows-llamacpp-ov-install.bat
 ```
 
 <details>
-<summary>Click to expand <code>build-llamacpp-ov.bat</code></summary>
+<summary>Click to expand <code>windows-llamacpp-ov-install.bat</code></summary>
 
 ```bat
 @echo off
@@ -370,8 +364,8 @@ REM ============================================
 REM llama.cpp OpenVINO Build Script (Ninja)
 REM ============================================
 
-set "OPENVINO_VERSION_MAJOR=2026.3.1"
-set "OPENVINO_VERSION_FULL=2026.3.1.22476.56d9685302d"
+set "OPENVINO_VERSION_MAJOR=2026.3"
+set "OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c"
 
 set "SCRIPT_DIR=%~dp0"
 set "VCPKG_DIR=C:\vcpkg"
@@ -459,6 +453,9 @@ if exist "%OPENVINO_INSTALL_DIR%\setupvars.bat" (
     )
 
     REM Move the single top-level folder contents into the versioned install dir.
+    REM NOTE: delayed expansion (!VAR!) is required because the surrounding else( ... )
+    REM block is parsed once up-front, so %OPENVINO_EXTRACTED% would expand to "" here
+    REM and xcopy would then treat "\*" as C:\* and fail with "Cannot perform a cyclic copy".
     set "OPENVINO_EXTRACTED="
     for /d %%i in ("%OPENVINO_EXTRACT_TMP%\*") do set "OPENVINO_EXTRACTED=%%i"
     if not defined OPENVINO_EXTRACTED (
@@ -550,7 +547,7 @@ endlocal
 ```
 
 > [!NOTE]
-> The script pins OpenVINO `2026.3.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
+> The script pins OpenVINO `2026.3` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
 
 </details>
 
@@ -715,7 +712,6 @@ Boolean flags follow a uniform convention: set to a **positive integer** (e.g. `
 | `GGML_OPENVINO_CACHE_DIR`         | String    | `not set`  | Directory for OpenVINO model caching (recommended: `/tmp/ov_cache`). Enables model caching when set. **Not supported on NPU devices.** |
 | `GGML_OPENVINO_COMPILED_MODEL_CACHE_DIR` | String | `not set` | Directory for the frontend compiled-model cache. When set, OpenVINO compiled models are exported as blobs and imported on later runs to skip weight requantization, graph conversion, and compilation for matching single-graph models. |
 | `GGML_OPENVINO_PREFILL_CHUNK_SIZE`| Integer   | `256`      | Token chunk size for **NPU** prefill (NPU-only; ignored on CPU/GPU). Must be a positive integer; otherwise the default is used. |
-| `GGML_OPENVINO_NPU_COMPILE_CONFIG` | String | `not set` | NPU-only compiler mode parameters forwarded to OpenVINO as `NPU_COMPILATION_MODE_PARAMS`, for example `optimization-level=3`. |
 | `GGML_OPENVINO_STATEFUL_EXECUTION`| Boolean   | `0`        | Enable stateful KV cache for better performance. Recommended on CPU, GPU.                                   |
 | `GGML_OPENVINO_DISABLE_CACHE`     | Boolean   | `0`        | Disable the in-process compiled-model / decoder cache (cache is on by default). Set to `1` to disable.      |
 | `GGML_OPENVINO_DISABLE_KV_SLICE`  | Boolean   | `0`        | Disable the KV-cache input-tensor slicing optimization (slicing is on by default on CPU/GPU). Set to `1` to disable. |
@@ -729,11 +725,9 @@ Boolean flags follow a uniform convention: set to a **positive integer** (e.g. `
 | `GGML_OPENVINO_DEBUG_INPUT`       | Boolean   | `0`        | Enable input debugging and print input tensor info.                                                         |
 | `GGML_OPENVINO_DEBUG_OUTPUT`      | Boolean   | `0`        | Enable output debugging and print output tensor info.                                                       |
 | `GGML_OPENVINO_PRINT_CGRAPH_TENSOR_ADDRESS` | Boolean | `0` | Print tensor address map once.                                                                           |
-| `GGML_OPENVINO_LOG_UNSUPPORTED_OPS`| Boolean   | `0`        | Log warning messages with tensor details and rejection reasons for any ops not supported by the OpenVINO backend. Emits at `WARN` level (requires `--log-verbosity >= 2`, enabled by default). |
 
 > [!NOTE]
-> - `GGML_OPENVINO_STATEFUL_EXECUTION` is an **Experimental** feature to allow stateful execution for managing the KV cache internally inside the OpenVINO model, improving performance on CPUs and GPUs. Stateful execution is not effective on NPUs, and not all models currently support this feature. This feature is experimental and has been validated only with the llama-simple, llama-cli, llama-bench, and llama-run applications and is recommended to enable for the best performance. Other applications, such as llama-server and llama-perplexity, are not yet supported.
-> - `GGML_OPENVINO_LOG_UNSUPPORTED_OPS` emits logs at `WARN` level (`GGML_LOG_WARN`), which requires application log verbosity `--log-verbosity >= 2` (or `-lv 2`).
+>`GGML_OPENVINO_STATEFUL_EXECUTION` is an **Experimental** feature to allow stateful execution for managing the KV cache internally inside the OpenVINO model, improving performance on CPUs and GPUs. Stateful execution is not effective on NPUs, and not all models currently support this feature. This feature is experimental and has been validated only with the llama-simple, llama-cli, llama-bench, and llama-run applications and is recommended to enable for the best performance. Other applications, such as llama-server and llama-perplexity, are not yet supported.
 
 ### Example Usage
 

docs/backend/SYCL.md

@@ -795,9 +795,7 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | GGML_SYCL_ENABLE_FLASH_ATTN | 1 (default) or 0| Enable Flash-Attention. It can reduce memory usage. The performance impact depends on the LLM.|
 | GGML_SYCL_ENABLE_OPT | 0 or 1 (default)| Enable optimize features for Intel GPUs. (Recommended to 0 for Intel devices older than Gen 10) |
 | GGML_SYCL_ENABLE_GRAPH | 0 (default) or 1 | Enable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
-| GGML_SYCL_ENABLE_HOST_PINNED_MEM | 0 or 1 (default) | Enable host pinned memory to speed up copy data from host to device. When disable it, host memory will common malloc() on CPU. Disable it when use `--load-model mlock`.|
-| GGML_SYCL_HOST_PINNED_MEM_2G | 0 (default) or 1 | Limit the max memory allocation to be no more than 2GB when enable host pinned memory. USM allocations above 2 GiB take the relaxed/large-allocation path, which serializes H2D copies with compute and prevents copy/compute overlap. It will impact the startup time. Need more test. Depend on `GGML_SYCL_ENABLE_HOST_PINNED_MEM=1`.|
-| GGML_SYCL_GET_MEM_API | 0 (default) or 1  | Set to get memory info (free, total) by Level Zero or SYCL API:<br>0 - Level Zero API: support more GPUs, only run on Level Zero running time. When there is an error, fallback to call SYCL API. Depend on GGML_SYCL_SUPPORT_LEVEL_ZERO_API.<br>1 - SYCL API: legacy, support more running time, it can't get the free size of some GPUs (like Arc770). In such case, return total size for free size.|
+| GGML_SYCL_ENABLE_HOST_PINNED_MEM | 0 or 1 (default) | Enable host pinned memory to speed up copy data from host to device. When disable it, host memory will common malloc() on CPU.|
 | GGML_SYCL_USE_LEVEL_ZERO_API | 1 (default) or 0 | Use Level Zero API for device memory allocation instead of SYCL. Reduces system RAM usage on Intel dGPUs by avoiding DMA-buf/TTM host memory staging. Requires GGML_SYCL_SUPPORT_LEVEL_ZERO_API=ON at build time. SYCL backend always runs on Level Zero running time even if it's set as OFF (The SYCL api will be usage for memory allocation).|
 | GGML_SYCL_ENABLE_DNN | 0 or 1 (default)| Enable running computations through oneDNN and always use oneMKL. |
 | GGML_SYCL_FA_ONEDNN | 1 (default) or 0 | Enable the oneDNN fused SDPA (flash-attention) path on supported GPUs. Set to 0 to always use the native SYCL flash-attention kernel. |
@@ -805,10 +803,8 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | GGML_SYCL_ENABLE_VMM | 0 or 1 (default) | Enable the virtual-memory device pool. |
 | GGML_SYCL_ENABLE_MKL_FA | 1 (default) or 0 | Enable oneMKL GEMM flash attention for XMX-accelerated prompt processing with quantized KV cache. Automatically activates during prefill (prompt processing) when all conditions are met: (1) flash-attn enabled (`-fa` or `--flash-attn on`), (2) KV cache quantized (`--cache-type-k q8_0 --cache-type-v q8_0` or other `*_0/*_1` types), (3) batch size ≥ 1024 (`--batch-size 1024`), (4) prompt length ≥ 1024 tokens. Set to 0 to force the TILE kernel for A/B testing. Example minimum command: `llama-cli -m model.gguf -fa -ngl 99 --cache-type-k q8_0 --cache-type-v q8_0 --batch-size 1024 -p "your prompt"` |
 | GGML_SYCL_MKL_FA_DEBUG | 0 (default) or 1 | Enable per-call diagnostic logging for MKL flash attention: GEMM/softmax timings, interleaved-head detection, and buffer memory usage. |
-| GGML_SYCL_MEMTRACE | 0 (default), 1, 2 | Enable record and output memory allocation diagnostics. Requires `-lv 4`. <br>0 -  Disable<br>1 - Basic memory info, including current and peak allocations, as well allocations from other sources, around 50 lines per model load.<br>2 - More verbose, logging around 900 specific allocations and deallocations. |
-| GGML_SYCL_MEMTRACE_STEP | 64 (default) or positive integer | With GGML_SYCL_MEMTRACE=1, the minimum growth in memory usage to trigger another log record. |
 | GGML_SYCL_MKL_FA_DIAG | 0 (default) or 1 | Enable output fingerprinting for MKL flash attention. Dumps the first 64 float output values for the first 6 FA calls with n_kv ≥ 1024, labeled with kernel type (MKL/TILE/VEC) for cross-kernel comparison. |
-| GGML_SYCL_ENABLE_FUSION | 0 or 1 (default) | Enable fused-kernel dispatch in graph compute. Unsupported types and layouts fall back to the standalone op kernels. See `ggml_sycl_can_fuse()`. |
+| GGML_SYCL_ENABLE_FUSION | 0 or 1 (default) | Enable fused-kernel dispatch in graph compute. |
 | GGML_SYCL_ENABLE_ESIMD | 0 or 1 (default)| Enable ESIMD kernels when available. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Allow SYCL/Unified Runtime Level Zero device allocations larger than 4 GiB. llama.cpp's direct Level Zero allocation path requests the relaxed maximum-size limit itself when GGML_SYCL_ENABLE_LEVEL_ZERO=1. |

docs/backend/snapdragon/CMakeUserPresets.json

@@ -8,7 +8,7 @@
         "toolset":      { "value": "host=x86_64", "strategy": "external" },
         "cacheVariables": {
             "ANDROID_ABI":      "arm64-v8a",
-            "ANDROID_PLATFORM": "android-34",
+            "ANDROID_PLATFORM": "android-31",
             "CMAKE_TOOLCHAIN_FILE": "$env{ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake",
             "CMAKE_C_FLAGS":   "-march=armv8.7a+fp16+dotprod+i8mm -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
             "CMAKE_CXX_FLAGS": "-march=armv8.7a+fp16+dotprod+i8mm -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",

docs/backend/snapdragon/README.md

@@ -2,47 +2,39 @@
 
 ## Setup
 
-The cross-compilation toolchain images are provided by the
-[Qualcomm Snapdragon Toolchain registry](https://github.com/snapdragon-toolchain).
-These Docker images include the Android NDK, OpenCL SDK, Hexagon SDK, CMake, and the necessary cross-compilers:
+### Android
 
-* **Android toolchain**: `ghcr.io/snapdragon-toolchain/arm64-android:v0.7`
-* **Linux toolchain**: `ghcr.io/snapdragon-toolchain/arm64-linux:v0.7`
+The easiest way to build llama.cpp for a Snapdragon-based Android device is using the toolchain Docker image (see github.com/snapdragon-toolchain).
+This image includes Android NDK, OpenCL SDK, Hexagon SDK, CMake, etc.
 
-The unified build utility (`scripts/snapdragon/build.py`) automatically pulls
-and orchestrates these containers to perform target compilation.
-You only need to ensure that Docker (or Docker Desktop on macOS/Windows) is running on your host machine.
-Specific setup, build, and installation details for Linux and Windows on Snapdragon platforms are documented in:
-* [Linux on Snapdragon guide](linux.md)
-* [Windows on Snapdragon guide](windows.md)
+This method works on Linux, macOS, and Windows. macOS and Windows users should install Docker Desktop.
+
+```
+~/src/llama.cpp$ docker run -it -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-android:v0.7
+[d]/> cd /workspace
+```
+
+Note: The rest of the **Android** build process assumes that you're running inside the toolchain container.
+
+### Windows On Snapdragon
+
+Native Windows 11 arm64 builds has the following tools dependencies:
+- MS Visual Studio 2026 (Community Edition or Pro)
+  - MSVC arm64 standard and runtime libraries
+  - UCRT and Driver Kit
+- LLVM core libraries and Clang compiler (winget)
+- CMake, Git, Python (winget)
+- Hexagon SDK Community Edition 6.6 or later (see windows.md)
+- OpenCL SDK 2.3 or later (see windows.md)
+
+Note: The rest of the **Windows** build process assumes that you're running natively in Powershell.
+Adapt below build commands accordingly.
 
 ## How to Build
 
-### Using build.py script (Recommended)
+Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
 
-The easiest way to build llama.cpp is by using the `scripts/snapdragon/build.py` script. It automatically copies the CMake presets,
-launches the correct compilation Docker container, builds the libraries and tools,
-installs them, and optionally pushes them to your ADB device.
-
-Build and deploy for Android target (accepts `android` or `adb` alias):
 ```
-$ ./scripts/snapdragon/build.py --target adb --push
-```
-
-Build and deploy for Linux target (accepts `linux` or `lnx` alias):
-```
-$ ./scripts/snapdragon/build.py --target linux:user@host --push
-```
-
-### Manual CMake Build
-
-Alternatively, you can build llama.cpp manually by entering the cross-compilation Docker container and running the CMake commands:
-
-```bash
-# Start the cross-compilation container manually:
-~/src/llama.cpp$ docker run -it --rm -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-android:v0.7
-
-# Inside the container, build the project using presets:
 [d]/workspace> cp docs/backend/snapdragon/CMakeUserPresets.json .
 
 [d]/workspace> cmake --preset arm64-android-snapdragon-release -B build-snapdragon
@@ -76,19 +68,19 @@ Preset CMake variables:
 To generate an installable "package" simply use cmake --install:
 
 ```
-[d]/workspace> cmake --install build-snapdragon --prefix pkg-android/llama.cpp
+[d]/workspace> cmake --install build-snapdragon --prefix pkg-snapdragon/llama.cpp
 -- Install configuration: "Release"
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-cpu.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-opencl.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-hexagon.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-htp-v73.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-htp-v75.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-htp-v79.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-htp-v81.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-cpu.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-opencl.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-hexagon.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-htp-v73.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-htp-v75.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-htp-v79.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-htp-v81.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml.so
 ...
--- Installing: /workspace/pkg-android/llama.cpp/bin/llama-bench
--- Installing: /workspace/pkg-android/llama.cpp/bin/llama-cli
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/bin/llama-bench
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/bin/llama-cli
 ...
 ```
 
@@ -99,14 +91,14 @@ To generate an installable "package" simply use cmake --install:
 For this step, your device needs to be configured for on-device development.
 Please see https://developer.android.com/studio/debug/dev-options for details.
 
-Once ADB is enabled, use `adb push` to install `pkg-android` on the device.
+Once ADB is enabled, use `adb push` to install `pkg-snapdragon` on the device.
 **Note that the toolchain Docker image doesn't have ADB and doesn't set up the ADB bridge. Please use native ADB on the host.**
 
 ```
-~/src/llama.cpp$ adb push pkg-android/llama.cpp /data/local/tmp/
-pkg-android/llama.cpp/bin/: 67 files pushed, 0 skipped. 190.2 MB/s (919095042 bytes in 4.607s)
-pkg-android/llama.cpp/include/: 19 files pushed, 0 skipped. 20.5 MB/s (255173 bytes in 0.012s)
-pkg-android/llama.cpp/lib/: 16 files pushed, 0 skipped. 144.4 MB/s (43801382 bytes in 0.289s)
+~/src/llama.cpp$ adb push pkg-snapdragon/llama.cpp /data/local/tmp/
+pkg-snapdragon/llama.cpp/bin/: 67 files pushed, 0 skipped. 190.2 MB/s (919095042 bytes in 4.607s)
+pkg-snapdragon/llama.cpp/include/: 19 files pushed, 0 skipped. 20.5 MB/s (255173 bytes in 0.012s)
+pkg-snapdragon/llama.cpp/lib/: 16 files pushed, 0 skipped. 144.4 MB/s (43801382 bytes in 0.289s)
 102 files pushed, 0 skipped. 186.9 MB/s (963151597 bytes in 4.914s)
 ```
 
@@ -123,44 +115,24 @@ 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-wos` folder.
-To run, you can use the `scripts/snapdragon/run.py` runner script (see details below).
+All artifacts are already installed in the `pkg-snapdragon` folder.
+To run, adapt below instructions to use Powershell scripts in `scripts/snapdragon/windows`.
 
 ## How to Run
 
-The easiest way to run llama.cpp cli tools is using the provided `scripts/snapdragon/run.py` wrapper script. This script automatically
-maps CLI options to environment variables, resolves executable paths, and runs the command locally, via ADB, or remotely via SSH on the
-target device.
+The easiest way to run llama.cpp cli tools is using provided wrapper scripts that properly set up all required environment variables.
 
-llama.cpp supports three backends on Snapdragon-based devices: CPU, Adreno GPU (GPUOpenCL), and Hexagon NPU.
-You can select which backend(s) to run the model on using the `--device` option of the tool (or `--devices` option in `run.py`).
+llama.cpp supports three backends on Snapdragon-based devices: CPU, Adreno GPU (GPUOpenCL), and Hexagon NPU (HTP0-4).
+You can select which backend to run the model on using the `D=` variable, which maps to the `--device` option.
 
 Hexagon NPU behaves as a "GPU" device when it comes to `-ngl` and other offload-related options.
 
-Here are some examples of running various llama.cpp tools.
+Here are some examples of running various llama.cpp tools via ADB.
 
-Generating a completion with Gemma on Android (relying on default `HTP0:0` device and default thread count `-t 6`):
+Simple question for Llama-3.2-1B
 
 ```
-~/src/llama.cpp$ ./scripts/snapdragon/run.py --target adb -- llama-completion -m models/gemma-2-2b-it-Q4_0.gguf -f prompts/sample_prompt_1024.txt --jinja -st
-...
-ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
-ggml-hex: Hexagon Arch version v79
-ggml-hex: allocating new session: HTP0:0
-...
-load_tensors: offloading output layer to GPU
-load_tensors: offloaded 27/27 layers to GPU
-load_tensors:          CPU model buffer size =   300.00 MiB
-load_tensors:      HTP0:0 model buffer size  =  1400.26 MiB
-...
-llama_perf_context_print: prompt eval time =     320.00 ms /  1024 tokens (    0.31 ms per token,  3200.00 tokens per second)
-llama_perf_context_print:        eval time =     2100.00 ms /   100 runs   (   21.00 ms per token,    47.62 tokens per second)
-```
-
-Simple question for Llama-3.2-1B:
-
-```
-~/src/llama.cpp$ ./scripts/snapdragon/run.py --target android --devices HTP0 -- llama-cli -m Llama-3.2-1B-Instruct-Q4_0.gguf -p "what is the most popular cookie in the world?"
+~/src/llama.cpp$ M=Llama-3.2-1B-Instruct-Q4_0.gguf D=HTP0 ./scripts/snapdragon/adb/run-completion.sh -p "what is the most popular cookie in the world?"
 ...
 ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
 ggml-hex: Hexagon Arch version v79
@@ -170,7 +142,8 @@ ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v
 load_tensors: offloading output layer to GPU
 load_tensors: offloaded 17/17 layers to GPU
 load_tensors:          CPU model buffer size =   225.49 MiB
-load_tensors:         HTP0 model buffer size =   504.26 MiB
+load_tensors:         HTP0 model buffer size =     0.26 MiB
+load_tensors:  HTP0-REPACK model buffer size =   504.00 MiB
 ...
 I hope this helps you understand the world's most popular cookies! [end of text]
 ...
@@ -183,25 +156,60 @@ llama_perf_context_print:    graphs reused =        473
 llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
 llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
 llama_memory_breakdown_print: |   - Host               |                  439 =   225 +     136 +      77                |
+llama_memory_breakdown_print: |   - HTP0-REPACK        |                  504 =   504 +       0 +       0                |
 ```
 
-Op test for MUL_MAT:
+Summary request for OLMoE-1B-7B. This is a large model that requires two HTP sessions/devices
 
 ```
-~/src/llama.cpp$ ./scripts/snapdragon/run.py --target adb --hex-hostbuf 0 --devices HTP0:0 -- test-backend-ops -b HTP0:0 -o MUL_MAT
+~/src/llama.cpp$ M=OLMoE-1B-7B-0125-Instruct-Q4_0.gguf NDEV=2 D=HTP0,HTP1 ./scripts/snapdragon/adb/run-completion.sh -f surfing.txt
 ...
-Backend 2/3: HTP0:0
+ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
+ggml-hex: Hexagon Arch version v81
+ggml-hex: allocating new session: HTP0
+ggml-hex: allocating new session: HTP1
+...
+load_tensors: offloading output layer to GPU
+load_tensors: offloaded 17/17 layers to GPU
+load_tensors:          CPU model buffer size =   143.86 MiB
+load_tensors:         HTP1 model buffer size =     0.23 MiB
+load_tensors:  HTP1-REPACK model buffer size =  1575.00 MiB
+load_tensors:         HTP0 model buffer size =     0.28 MiB
+load_tensors:  HTP0-REPACK model buffer size =  2025.00 MiB
+...
+llama_context:        CPU  output buffer size =     0.19 MiB
+llama_kv_cache:       HTP1 KV buffer size =   238.00 MiB
+llama_kv_cache:       HTP0 KV buffer size =   306.00 MiB
+llama_kv_cache: size =  544.00 MiB (  8192 cells,  16 layers,  1/1 seqs), K (q8_0):  272.00 MiB, V (q8_0):  272.00 MiB
+llama_context:       HTP0 compute buffer size =    15.00 MiB
+llama_context:       HTP1 compute buffer size =    15.00 MiB
+llama_context:        CPU compute buffer size =    24.56 MiB
+...
+llama_perf_context_print: prompt eval time =    1730.57 ms /   212 tokens (    8.16 ms per token,   122.50 tokens per second)
+llama_perf_context_print:        eval time =    5624.75 ms /   257 runs   (   21.89 ms per token,    45.69 tokens per second)
+llama_perf_context_print:       total time =    7377.33 ms /   469 tokens
+llama_perf_context_print:    graphs reused =        255
+llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
+llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP1 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - Host               |                  742 =   144 +     544 +      54                |
+llama_memory_breakdown_print: |   - HTP1-REPACK        |                 1575 =  1575 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP0-REPACK        |                 2025 =  2025 +       0 +       0                |
+```
+
+Op test for MUL_MAT
+
+```
+~/src/llama.cpp$ HB=0 ./scripts/snapdragon/adb/run-tool.sh test-backend-ops -b HTP0 -o MUL_MAT
+...
+Backend 2/3: HTP0
 Device description: Hexagon
 Device memory: 2048 MB (2048 MB free)
 MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
 MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
 MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
-```
 
-Llama benchmark:
-
-```
-~/src/llama.cpp$ ./scripts/snapdragon/run.py --target adb --devices HTP0 -- llama-bench -p 128 -n 64 -m Llama-3.2-1B-Instruct-Q4_0.gguf
+~/src/llama.cpp-hexagon$ M=Llama-3.2-1B-Instruct-Q4_0.gguf ./scripts/snapdragon/adb/run-bench.sh -p 128 -n 64
 ...
 ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
 ggml-hex: Hexagon Arch version v79
@@ -211,20 +219,15 @@ ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v
 | ---------------| ---------: | -----: | ---------- | --: | ------: | ------: | ---: | ----: | ------------: |
 | llama 1B Q4_0  | 729.75 MiB | 1.24 B | HTP        |  99 |       4 |     128 |    0 | pp128 | 169.42 ± 1.75 |
 | llama 1B Q4_0  | 729.75 MiB | 1.24 B | HTP        |  99 |       4 |     128 |    0 |  tg64 |  51.54 ± 1.13 |
+
+build: 6a8cf8914 (6733)
 ```
 
 ## Environment variables
 
-- `GGML_HEXAGON_DEVICES` (default: not set, defaults to HTP0 session)
-  Controls which NPU devices and sessions to allocate. Can be configured as:
-  - A single integer `N`: Allocates `N` sessions named `HTP0`, `HTP1`, ..., `HTP<N-1>` (behaves identically to `GGML_HEXAGON_NDEV=N`).
-  - A comma-separated list of device names in `HTP<physical_idx>:<virtual_idx>` format (or legacy `HTP<idx>` format). For example, `HTP0:0,HTP0:1` creates two virtual
-    sessions on the first physical NPU (useful for memory limits). `HTP0:0,HTP1:0` allocates one session on each of the two physical NPUs
-    on a dual-NPU device.
-
-- `GGML_HEXAGON_NDEV` (deprecated)
-  Replaced by `GGML_HEXAGON_DEVICES`. Controls the number of virtual sessions to allocate on physical NPU `0`.
-  Allocates sessions named `HTP0`, `HTP1`, etc.
+- `GGML_HEXAGON_NDEV=1`
+  Controls the number of devices/sessions to allocate. The default is 1.
+  Most quantized models under 4B fit into a single session; an 8B model needs two, and a 20B model needs four.
 
 - `GGML_HEXAGON_NHVX=0`
   Controls the number of HVX hardware threads to use. The default is all (actual number varies depending on the hardware version).
@@ -252,17 +255,26 @@ ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v
   - `2` Extended profile with per-op `usecs`, `cycles` and default PMU counter data
   - `0x1,...,0x8` Extended profile with per-op `usecs`, `cycles` and custom PMU counter data
 
-  The logging output can be either saved into a file for post-processing or it can be piped directly into the post-processing tool
-  to generate the report.
+  The logging output can be either saved into a file for post-processing or it can be piped directly into the post-processing tool to generate the report.
   Examples:
 
-      `GGML_HEXAGON_PROFILE=1 ./scripts/snapdragon/run.py --target adb -- llama-cli ... |& ./scripts/snapdragon/ggml-hexagon-profile.py -`
+      `GGML_HEXAGON_PROFILE=1 llama-completion ... |& ./scripts/snapdragon/ggml-hexagon-profile.py -`
+
+- `GGML_HEXAGON_OPSTAGE=0x0`
+  Allows enabling specific stages of the Op processing pipeline:
+
+  - `0x1` Enable Op Queue (i.e., queuing Ops into NPU)
+  - `0x2` Enable Op Compute (MUL_MAT, etc.)
+
+  Examples:
+
+      `GGML_HEXAGON_OPSTAGE=0x1 llama-completion ...` - Ops are enqueued to the NPU but dma & compute are disabled
+      `GGML_HEXAGON_OPSTAGE=0x3 llama-completion ...` - Full queuing and processing of Ops (default)
 
 - `GGML_HEXAGON_OPFILTER=regex`
   Allows filtering (disabling) Ops that match the regex pattern:
 
   Examples:
 
-      `GGML_HEXAGON_OPFILTER="FLASH_ATTN_EXT" ./scripts/snapdragon/run.py --target adb -- llama-cli ...` - Disable Flash Attention on Hexagon (falls back to CPU or GPU)
-      `GGML_HEXAGON_OPFILTER="ADD\|SUB" ./scripts/snapdragon/run.py --target adb -- llama-cli ...` - Disable ADD and SUB on Hexagon (fall back to CPU or GPU)
-
+      `GGML_HEXAGON_OPFILTER="FLASH_ATTN_EXT" llama-completion ...` - Disable Flash Attention on Hexagon (falls back to CPU or GPU)
+      `GGML_HEXAGON_OPFILTER="ADD\|SUB" llama-completion ...` - Disable ADD and SUB on Hexagon (fall back to CPU or GPU)

docs/backend/snapdragon/developer.md

@@ -39,21 +39,22 @@ the repacking.
 
 ## Large model handling
 
-Hexagon NPU sessions (aka Process Domains (PD) in the Hexagon SDK) are limited to a maximum memory mapping window of around 3.5GB.
-In llama.cpp/GGML, each Hexagon session is mapped to a single GGML backend device (e.g., `HTP0:0`, `HTP0:1`, etc. when using
-`GGML_HEXAGON_DEVICES`, or `HTP0`, `HTP1` in legacy mode).
+Hexagon NPU session (aka Process Domain (PD) in the Hexagon docs) is limited to a memory mapping of around 3.5GB.
+In llama.cpp/GGML the Hexagon session is mapped to a single GGML backend device (HTP0, HTP1, etc).
 
-To support running models larger than 3.5GB on a single device, the Hexagon backend dynamically maps and unmaps execution buffers
-during the graph execution cycle to stay within the Process Domain window. This enables large models to run successfully on a single
-NPU device.
+In order to map models larger than 3.5GB we need to allocate multiple devices and split the model.
+For this we're taking advantage of the llama.cpp/GGML multi-GPU layer-splitting support.
+Each Hexagon device behaves like a GPU from the offload and model splitting perspective.
 
-Alternatively, users can choose to use standard llama.cpp/GGML layer-splitting mode to partition and split the model across
-multiple Hexagon devices or virtual sessions (which behave like multiple GPUs from the offload and splitting perspective).
-
-Here is an example of running GPT-OSS-20B model on a Snapdragon device using 4 virtual sessions on a single NPU (physical index 0).
+Here is an example of running GPT-OSS-20B model on a newer Snapdragon device with 16GB of DDR.
 
 ```
-~/src/llama.cpp$ ./scripts/snapdragon/run.py --target adb --devices HTP0:0,HTP0:1,HTP0:2,HTP0:3 -- llama-cli --load-mode none -m /data/local/tmp/gguf/gpt-oss-20b-Q4_0.gguf -t 4 --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on -ngl 99 -no-cnv -f surfing.txt
+M=gpt-oss-20b-Q4_0.gguf NDEV=4 D=HTP0,HTP1,HTP2,HTP3 P=surfing.txt scripts/snapdragon/adb/run-completion.sh -f surfing.txt -n 32
+...
+LD_LIBRARY_PATH=/data/local/tmp/llama.cpp/lib
+ADSP_LIBRARY_PATH=/data/local/tmp/llama.cpp/lib
+GGML_HEXAGON_NDEV=4 ./bin/llama-cli --load-mode none -m /data/local/tmp/llama.cpp/../gguf/gpt-oss-20b-Q4_0.gguf
+      -t 4 --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on -ngl 99 --device HTP0,HTP1,HTP2,HTP3 -no-cnv -f surfing.txt
 ...
 llama_model_loader: - type  f32:  289 tensors
 llama_model_loader: - type q4_0:   96 tensors
@@ -62,29 +63,33 @@ llama_model_loader: - type mxfp4:  72 tensors
 ...
 load_tensors: offloaded 25/25 layers to GPU
 load_tensors:          CPU model buffer size =  1182.09 MiB
-load_tensors:       HTP0:1 model buffer size =  2512.58 MiB
-load_tensors:       HTP0:3 model buffer size =  2093.83 MiB
-load_tensors:       HTP0:0 model buffer size =  2931.34 MiB
-load_tensors:       HTP0:2 model buffer size =  2512.58 MiB
+load_tensors:         HTP1 model buffer size =     6.64 MiB
+load_tensors:  HTP1-REPACK model buffer size =  2505.94 MiB
+load_tensors:         HTP3 model buffer size =     5.55 MiB
+load_tensors:  HTP3-REPACK model buffer size =  2088.28 MiB
+load_tensors:         HTP0 model buffer size =     7.75 MiB
+load_tensors:  HTP0-REPACK model buffer size =  2923.59 MiB
+load_tensors:         HTP2 model buffer size =     6.64 MiB
+load_tensors:  HTP2-REPACK model buffer size =  2505.94 MiB
 ...
 llama_context: n_ctx_per_seq (8192) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
 llama_context:        CPU  output buffer size =     0.77 MiB
 llama_kv_cache_iswa: creating non-SWA KV cache, size = 8192 cells
-llama_kv_cache:     HTP0:1 KV buffer size =    25.50 MiB
-llama_kv_cache:     HTP0:3 KV buffer size =    25.50 MiB
-llama_kv_cache:     HTP0:0 KV buffer size =    25.50 MiB
-llama_kv_cache:     HTP0:2 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP1 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP3 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP0 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP2 KV buffer size =    25.50 MiB
 llama_kv_cache: size =  102.00 MiB (  8192 cells,  12 layers,  1/1 seqs), K (q8_0):   51.00 MiB, V (q8_0):   51.00 MiB
 llama_kv_cache_iswa: creating     SWA KV cache, size = 256 cells
-llama_kv_cache:     HTP0:1 KV buffer size =     0.80 MiB
-llama_kv_cache:     HTP0:3 KV buffer size =     0.53 MiB
-llama_kv_cache:     HTP0:0 KV buffer size =     1.06 MiB
-llama_kv_cache:     HTP0:2 KV buffer size =     0.80 MiB
+llama_kv_cache:       HTP1 KV buffer size =     0.80 MiB
+llama_kv_cache:       HTP3 KV buffer size =     0.53 MiB
+llama_kv_cache:       HTP0 KV buffer size =     1.06 MiB
+llama_kv_cache:       HTP2 KV buffer size =     0.80 MiB
 llama_kv_cache: size =    3.19 MiB (   256 cells,  12 layers,  1/1 seqs), K (q8_0):    1.59 MiB, V (q8_0):    1.59 MiB
-llama_context:     HTP0:0 compute buffer size =    16.06 MiB
-llama_context:     HTP0:1 compute buffer size =    16.06 MiB
-llama_context:     HTP0:2 compute buffer size =    16.06 MiB
-llama_context:     HTP0:3 compute buffer size =    16.06 MiB
+llama_context:       HTP0 compute buffer size =    16.06 MiB
+llama_context:       HTP1 compute buffer size =    16.06 MiB
+llama_context:       HTP2 compute buffer size =    16.06 MiB
+llama_context:       HTP3 compute buffer size =    16.06 MiB
 llama_context:        CPU compute buffer size =    98.19 MiB
 ...
 llama_perf_context_print: prompt eval time =    3843.67 ms /   197 tokens ( 19.51 ms per token, 51.25 tokens per second)
@@ -92,9 +97,13 @@ llama_perf_context_print:        eval time =    1686.13 ms /    31 runs   ( 54.3
 llama_perf_context_print:       total time =    6266.30 ms /   228 tokens
 llama_perf_context_print:    graphs reused =         30
 llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
-llama_memory_breakdown_print: |   - HTP0:0 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP0:1 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP0:2 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP0:3 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP1 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP2 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP3 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
 llama_memory_breakdown_print: |   - Host               |                 1476 =  1208 +     105 +     162                |
+llama_memory_breakdown_print: |   - HTP1-REPACK        |                 2505 =  2505 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP3-REPACK        |                 2088 =  2088 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP0-REPACK        |                 2923 =  2923 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP2-REPACK        |                 2505 =  2505 +       0 +       0                |
 ```