ci.yml

name: CI

on:
  push:
    branches: [master]
  pull_request:

jobs:
  # The default build is CPU-only, so runners need no CUDA toolkit; the full
  # finite-difference gradient-check suite exercises the CPU maths, and the
  # kernels are checked against it locally by tests/cuda_parity.rs.
  cpu:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - name: Install Rust
        uses: dtolnay/rust-toolchain@stable
        with:
          components: rustfmt, clippy

      - name: Cache cargo
        uses: actions/cache@v4
        with:
          path: |
            ~/.cargo/registry
            ~/.cargo/git
            target
          key: ${{ runner.os }}-cargo-${{ hashFiles('**/Cargo.lock') }}

      - name: Build (CPU-only)
        run: cargo build --all-targets

      - name: Gradient checks (finite difference)
        run: cargo test --test gradcheck -- --nocapture

      - name: Robustness (resume, anomalies, freezing)
        run: cargo test --test robustness

      - name: Generation (kv cache, sampling)
        run: cargo test --test generation

      - name: Unit + integration tests
        run: cargo test

      - name: Clippy
        run: cargo clippy --all-targets -- -D warnings

  # Hosted runners have no GPU, so the kernels cannot run here. They still
  # compile: nvcc parses kernels.cu, the linker resolves every FFI symbol
  # against the real libraries, and a signature drift between ffi.rs and the
  # kernels fails in CI instead of on the next machine with a GPU. The parity
  # suite notices there is no device at runtime and skips itself.
  cuda-build:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - name: Install Rust
        uses: dtolnay/rust-toolchain@stable

      # No version pin: the action's default tracks what the runner image can
      # actually install, and any recent nvcc compiles these kernels.
      - name: Install CUDA toolkit (no driver, compile only)
        uses: Jimver/cuda-toolkit@v0.2.36
        with:
          method: 'network'

      - name: Cache cargo
        uses: actions/cache@v4
        with:
          path: |
            ~/.cargo/registry
            ~/.cargo/git
            target
          key: ${{ runner.os }}-cargo-cuda-${{ hashFiles('**/Cargo.lock') }}

      - name: Build with the cuda feature
        run: cargo build --features cuda --all-targets

      - name: Parity tests (self-skip without a device)
        run: cargo test --features cuda --test cuda_parity -- --test-threads=1