llama-cpp-rs-check.yml

name: Llama Cpp Rs Check

on:
  pull_request:
    branches:
      - main
  workflow_dispatch:

concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}
  cancel-in-progress: true

permissions: read-all

jobs:
  check:
    name: Run tests on llama-cpp-rs
    runs-on: ubuntu-latest
    env:
      RUSTC_WRAPPER: sccache
      SCCACHE_GHA_ENABLED: "true"
      # `RUSTFLAGS` replaces .cargo/config.toml rustflags rather than appending,
      # so the $ORIGIN rpath that lets directly executed binaries find the
      # llama.cpp shared libraries must be repeated here.
      RUSTFLAGS: '-C link-arg=-fuse-ld=mold -C link-arg=-Wl,-rpath,$ORIGIN'
    steps:
      - name: Checkout
        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1
        with:
          submodules: recursive
      - name: Install Compile Deps
        env:
          DEBIAN_FRONTEND: noninteractive
        run:
          sudo apt-get update && sudo apt-get install -y build-essential curl libssl-dev libclang-dev pkg-config cmake git mold
      - name: Setup sccache
        uses: mozilla-actions/sccache-action@v0.0.11
      - uses: dtolnay/rust-toolchain@stable
        with:
          components: clippy, rustfmt
      - name: Clippy
        run: cargo clippy --workspace --all-targets -- -D warnings
      # `cargo fmt` on its own rewrites files in the runner and always exits 0,
      # so it never gated anything. `--check` reports and fails instead.
      - name: Fmt
        run: cargo fmt --all -- --check
      - name: Test
        run: cargo test --workspace --exclude openai-server --exclude llama-jni
      - name: sccache stats
        run: sccache --show-stats || true
  linux:
    name: Check that it builds on linux (native)
    runs-on: ubuntu-latest
    env:
      RUSTC_WRAPPER: sccache
      SCCACHE_GHA_ENABLED: "true"
      # `RUSTFLAGS` replaces .cargo/config.toml rustflags rather than appending,
      # so the $ORIGIN rpath that lets directly executed binaries find the
      # llama.cpp shared libraries must be repeated here.
      RUSTFLAGS: '-C link-arg=-fuse-ld=mold -C link-arg=-Wl,-rpath,$ORIGIN'
    steps:
      - name: checkout
        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1
        with:
          submodules: recursive
      - name: Install Compile Deps
        env:
          DEBIAN_FRONTEND: noninteractive
        run:
          sudo apt-get update && sudo apt-get install -y build-essential curl libssl-dev libclang-dev pkg-config cmake git mold
      - name: Setup sccache
        uses: mozilla-actions/sccache-action@v0.0.11
      - name: Setup Rust
        uses: dtolnay/rust-toolchain@stable
      - name: Build
        run: cargo build --workspace --all-targets --exclude llama-jni
      # `cargo run` and `cargo test` set LD_LIBRARY_PATH to the target directory,
      # so they cannot catch a missing rpath — only running the binary directly
      # exercises how a user (or a shipped artifact) actually loads the
      # llama.cpp shared libraries.
      - name: Verify a dynamically linked binary runs without cargo
        run: |
          output=$(./target/debug/simple --help 2>&1)
          echo "$output" | head -5
          if echo "$output" | grep -qiE 'error while loading shared libraries|cannot open shared object|no such file or directory'; then
            echo "::error::binary could not locate the llama.cpp shared libraries — is the \$ORIGIN rpath still applied?"
            exit 1
          fi
      - name: sccache stats
        run: sccache --show-stats || true
  package:
    # Builds the crate exactly as crates.io would receive it and compiles that
    # tree. Nothing else checks the published artifact: the workspace builds
    # from the git checkout, which has files the `include` list in
    # llama-cpp-sys-4/Cargo.toml does not ship. A llama.cpp bump that starts
    # requiring a new file type — `ggml-version.h.in` did, in upstream #28364 —
    # breaks the published crate while every other job stays green.
    #
    # Only llama-cpp-sys-4 can be verified here: llama-cpp-4 depends on it by
    # version, so `cargo package` cannot resolve until the sys crate is on
    # crates.io.
    name: Verify the publishable package builds
    runs-on: ubuntu-latest
    env:
      RUSTC_WRAPPER: sccache
      SCCACHE_GHA_ENABLED: "true"
      RUSTFLAGS: '-C link-arg=-fuse-ld=mold -C link-arg=-Wl,-rpath,$ORIGIN'
    steps:
      - name: checkout
        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1
        with:
          submodules: recursive
      - name: Install Compile Deps
        env:
          DEBIAN_FRONTEND: noninteractive
        run:
          sudo apt-get update && sudo apt-get install -y build-essential curl libssl-dev libclang-dev pkg-config cmake git mold
      - name: Setup sccache
        uses: mozilla-actions/sccache-action@v0.0.11
      - uses: dtolnay/rust-toolchain@stable
      # `--allow-dirty` because the submodule checkout leaves the tree dirty in
      # CI; the packaging and verification are what matter here, not cleanliness.
      - name: cargo package (with verification build)
        run: cargo package -p llama-cpp-sys-4 --allow-dirty
      - name: sccache stats
        run: sccache --show-stats || true

  feature-combos:
    # Every other job builds the default feature set
    # (openmp,mtmd,dynamic-link), so a break under any other combination stays
    # hidden until a release tag fires prebuilt-llama.yml.
    #
    # These jobs *lint* as well as build. `ggml` had accumulated 47 clippy
    # warnings precisely because nothing here ever turned the feature on — code
    # behind an off-by-default feature is not covered by the default-feature
    # jobs, so it drifts silently.
    name: Check non-default feature combos (${{ matrix.label }})
    runs-on: ubuntu-latest
    strategy:
      fail-fast: false
      matrix:
        include:
          - label: static + mtmd
            features: mtmd
          - label: rpc
            features: rpc
          # Raw ggml tensor/graph API. Off by default, so only this job sees it.
          - label: ggml
            features: ggml
          # Q1 quantization; gates a build-time patch as well as Rust code.
          - label: q1
            features: q1
          # No features at all: static linkage, no mtmd/openmp. The floor every
          # other combination builds on.
          - label: no-default-features
            features: ""
    env:
      RUSTC_WRAPPER: sccache
      SCCACHE_GHA_ENABLED: "true"
      # `RUSTFLAGS` replaces .cargo/config.toml rustflags rather than appending,
      # so the $ORIGIN rpath that lets directly executed binaries find the
      # llama.cpp shared libraries must be repeated here.
      RUSTFLAGS: '-C link-arg=-fuse-ld=mold -C link-arg=-Wl,-rpath,$ORIGIN'
    steps:
      - name: checkout
        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1
        with:
          submodules: recursive
      - name: Install Compile Deps
        env:
          DEBIAN_FRONTEND: noninteractive
        run:
          sudo apt-get update && sudo apt-get install -y build-essential curl libssl-dev libclang-dev pkg-config cmake git mold
      - name: Setup sccache
        uses: mozilla-actions/sccache-action@v0.0.11
      - name: Setup Rust
        uses: dtolnay/rust-toolchain@stable
      - name: Build
        run: |
          if [ -z "${{ matrix.features }}" ]; then
            cargo build -p llama-cpp-4 --no-default-features
          else
            cargo build -p llama-cpp-4 --no-default-features --features ${{ matrix.features }}
          fi
      # Lint too, not just build. A feature-gated module that compiles can still
      # be accumulating warnings nobody sees.
      - name: Clippy
        run: |
          if [ -z "${{ matrix.features }}" ]; then
            cargo clippy -p llama-cpp-4 --all-targets --no-default-features -- -D warnings
          else
            cargo clippy -p llama-cpp-4 --all-targets --no-default-features \
              --features ${{ matrix.features }} -- -D warnings
          fi
      # `ggml` and `q1` carry their own test files, which nothing else runs.
      - name: Test
        if: matrix.features == 'ggml' || matrix.features == 'q1'
        run: cargo test -p llama-cpp-4 --no-default-features --features ${{ matrix.features }}
      # examples/rpc is its own workspace (see its Cargo.toml), so no other job
      # ever compiles it — which is how it silently went stale.
      - name: Build rpc example
        if: matrix.features == 'rpc'
        run: cargo build --manifest-path examples/rpc/Cargo.toml
      - name: sccache stats
        run: sccache --show-stats || true
  mac:
    name: Check that it builds on mac
    runs-on: macos-latest
    env:
      RUSTC_WRAPPER: sccache
      SCCACHE_GHA_ENABLED: "true"
    steps:
      - name: checkout
        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1
        with:
          submodules: recursive
      - name: Setup sccache
        uses: mozilla-actions/sccache-action@v0.0.11
      - name: Setup Rust
        uses: dtolnay/rust-toolchain@stable
      - name: Build
        run: cargo build --workspace --exclude llama-jni
      - name: sccache stats
        run: sccache --show-stats || true
  windows:
    name: Check that it builds on windows
    runs-on: windows-latest
    env:
      RUSTC_WRAPPER: sccache
      SCCACHE_GHA_ENABLED: "true"
    steps:
      - name: checkout
        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1
        with:
          submodules: recursive
      - name: Ensure patch.exe is discoverable
        shell: pwsh
        run: |
          $gitUsrBin = Join-Path $env:ProgramFiles 'Git\usr\bin'
          if (Test-Path $gitUsrBin) {
            if (-not ($env:PATH -split ';' | Where-Object { $_ -ieq $gitUsrBin })) {
              $gitUsrBin | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append
            }
          }
      - name: Setup sccache
        uses: mozilla-actions/sccache-action@v0.0.11
      - name: Setup Rust
        uses: dtolnay/rust-toolchain@stable
      - name: Build
        run: cargo build --workspace --exclude llama-jni
      - name: Setup Vulkan SDK
        uses: ./.github/actions/setup-vulkan-sdk-windows
      - name: Build Vulkan
        shell: pwsh
        run: cargo build --features vulkan
      - name: Test
        run: cargo test --workspace --exclude openai-server --exclude llama-jni
      - name: sccache stats
        shell: pwsh
        run: sccache --show-stats
  llama-cpp-4-integration:
    name: llama-cpp-4 integration tests (GGUF)
    runs-on: ubuntu-latest
    env:
      RUSTC_WRAPPER: sccache
      SCCACHE_GHA_ENABLED: "true"
      # `RUSTFLAGS` replaces .cargo/config.toml rustflags rather than appending,
      # so the $ORIGIN rpath that lets directly executed binaries find the
      # llama.cpp shared libraries must be repeated here.
      RUSTFLAGS: '-C link-arg=-fuse-ld=mold -C link-arg=-Wl,-rpath,$ORIGIN'
    steps:
      - name: Checkout
        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1
        with:
          submodules: recursive
      - name: Install Compile Deps
        env:
          DEBIAN_FRONTEND: noninteractive
        run:
          sudo apt-get update && sudo apt-get install -y build-essential curl libssl-dev libclang-dev pkg-config cmake git mold
      - name: Setup sccache
        uses: mozilla-actions/sccache-action@v0.0.11
      - uses: dtolnay/rust-toolchain@stable
      - name: Fetch test model
        run: ./scripts/fetch-test-model.sh
      - name: Integration tests
        env:
          LLAMA_TEST_MODEL: ${{ github.workspace }}/target/test-models/stories260K.gguf
        run: cargo test -p llama-cpp-4 --test test_integration -- --test-threads=1
      - name: sccache stats
        run: sccache --show-stats || true