feat(server): close the b10621 context, batching and YaRN runtime gaps #56
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| name: Pipeline Parallel CI | |
| # Multi-host pipeline-parallel (PP) regression harness. | |
| # | |
| # The workflow has two jobs: | |
| # | |
| # - `two-host-logical`: simulates two PP stages as two processes on a | |
| # single GitHub-hosted Linux runner. Exercises the full production | |
| # code path under `src/distributed/pipeline/` — coordinator, remote | |
| # stage service, transport, and the heterogeneous-memory partition | |
| # + admission-control regressions. Runs on every PR that touches | |
| # PP-related paths. | |
| # | |
| # - `three-host-real-model`: targets a self-hosted runner lab with | |
| # three real machines. Gated by the `ci:pp-three-host` label so it | |
| # only runs on demand (manual trigger) for PRs that need a full | |
| # 3-stage real-model parity sweep. The model checkout and the | |
| # runner lab are expected to be available on hosts tagged with | |
| # `pp-three-host`. | |
| # | |
| # Out of scope: | |
| # - running the 3-host job on every PR | |
| # - cross-OS matrices on the self-hosted path | |
| # | |
| # The workflow intentionally drives the same scripts that operators run | |
| # locally (`scripts/ci/run-pp-*.sh`) so the CI result and the local | |
| # developer experience never drift. | |
| on: | |
| pull_request: | |
| paths: | |
| - "src/distributed/pipeline/**" | |
| - "src/distributed/tcp_transport.rs" | |
| - "src/distributed/thunderbolt_transport.rs" | |
| - "src/distributed/rdma_transport.rs" | |
| - "src/distributed/rdma_capabilities.rs" | |
| - "src/distributed/transport.rs" | |
| - "src/distributed/transport_factory.rs" | |
| - "src/distributed/cluster_init.rs" | |
| - "src/distributed/mod.rs" | |
| - "tests/pipeline_ci_multi_stage_real_models.rs" | |
| - "tests/pipeline_cli_real_models.rs" | |
| - "tests/pipeline_server_real_models.rs" | |
| - "tests/pipeline_server_remote_real_models.rs" | |
| - "tests/pipeline_remote_real_models.rs" | |
| - "tests/pipeline_stage_executor_real_models.rs" | |
| - "tests/zero_config_cluster_init.rs" | |
| - "tests/rdma_transport_bench.rs" | |
| - "scripts/ci/run-pp-*.sh" | |
| - "scripts/ci/_common.sh" | |
| - ".github/workflows/pipeline-parallel-ci.yml" | |
| # Manual dispatch for ad-hoc runs of the 3-host job on a PR that did | |
| # not yet carry the `ci:pp-three-host` label at PR-open time. | |
| workflow_dispatch: | |
| inputs: | |
| run_three_host: | |
| description: "Also run the 3-host real-model parity job (requires self-hosted runner)" | |
| required: false | |
| default: "false" | |
| permissions: | |
| contents: read | |
| concurrency: | |
| group: pp-ci-${{ github.event.pull_request.number || github.ref }} | |
| cancel-in-progress: true | |
| env: | |
| CARGO_TERM_COLOR: always | |
| # The Linux-hosted logical job compiles without CUDA / Metal because | |
| # it runs in a fully CPU-bound transport + scheduler regression mode. | |
| # Model weights are not present on GitHub-hosted runners, so the | |
| # ignored multi-stage smoke self-skips — this is expected and the | |
| # script exits 0 in that case. | |
| RUST_BACKTRACE: short | |
| jobs: | |
| two-host-logical: | |
| name: 2-host logical (${{ matrix.os }}) | |
| runs-on: ${{ matrix.os }} | |
| # Public-mirror guard: skip on forks. PRs opened from a fork would | |
| # otherwise execute fork-author code against this canonical repository's | |
| # GITHUB_TOKEN and CI fixture release, which we intentionally disallow on | |
| # the public release mirror. Internal PRs (head repo == this repo) and | |
| # `workflow_dispatch` still run normally. | |
| if: >- | |
| github.repository == 'lablup/mlxcel' && | |
| (github.event_name != 'pull_request' || | |
| github.event.pull_request.head.repo.full_name == github.repository) | |
| strategy: | |
| fail-fast: false | |
| matrix: | |
| os: | |
| - ubuntu-latest | |
| timeout-minutes: 45 | |
| steps: | |
| - name: Checkout code | |
| uses: actions/checkout@v7 | |
| with: | |
| # Don't leave GITHUB_TOKEN in .git/config; CI jobs only fetch. | |
| persist-credentials: false | |
| - name: Install Rust toolchain | |
| uses: dtolnay/rust-toolchain@stable | |
| with: | |
| components: clippy, rustfmt | |
| - name: Cache cargo registry | |
| uses: actions/cache@v6 | |
| with: | |
| path: | | |
| ~/.cargo/registry/index/ | |
| ~/.cargo/registry/cache/ | |
| ~/.cargo/git/db/ | |
| target/ | |
| key: cargo-pp-ci-${{ runner.os }}-${{ hashFiles('Cargo.lock') }} | |
| restore-keys: | | |
| cargo-pp-ci-${{ runner.os }}- | |
| - name: Install Linux build dependencies | |
| if: runner.os == 'Linux' | |
| run: | | |
| sudo apt-get update | |
| # liblapacke-dev ships lapacke.h, which MLX's CMake consumes via | |
| # find_path(LAPACK_INCLUDE_DIRS lapacke.h ...). liblapack-dev alone | |
| # only provides the Fortran LAPACK library, so without lapacke the | |
| # header lookup returns NOTFOUND and the MLX configure step fails. | |
| sudo apt-get install -y --no-install-recommends \ | |
| build-essential \ | |
| cmake \ | |
| ninja-build \ | |
| libopenblas-dev \ | |
| liblapack-dev \ | |
| liblapacke-dev \ | |
| pkg-config | |
| - name: Validate workflow YAML | |
| run: | | |
| python3 -c "import yaml; yaml.safe_load(open('.github/workflows/pipeline-parallel-ci.yml'))" | |
| # The multi-stage smoke test needs real model weights. We ship them | |
| # out-of-band via a GitHub Release asset (pinned HF snapshot packaged | |
| # by `scripts/ci/build-ci-fixture.sh`) so CI never calls HuggingFace | |
| # during a build and the weights are byte-identical across runs. | |
| # | |
| # `CI_FIXTURE_TAG` is both the release tag and the cache key; bump | |
| # the version suffix when the fixture is rotated. | |
| - name: Cache CI fixture weights | |
| id: ci-fixture-cache | |
| uses: actions/cache@v6 | |
| with: | |
| path: models/qwen3-0.6b-4bit | |
| key: ${{ runner.os }}-ci-fixture-qwen3-0.6b-4bit-v1 | |
| - name: Download CI fixture | |
| if: steps.ci-fixture-cache.outputs.cache-hit != 'true' | |
| env: | |
| GH_TOKEN: ${{ secrets.GITHUB_TOKEN }} | |
| CI_FIXTURE_TAG: ci-fixtures/qwen3-0.6b-4bit-v1 | |
| CI_FIXTURE_ASSET: qwen3-0.6b-4bit.tar.gz | |
| run: | | |
| mkdir -p models | |
| tmp_dir="$(mktemp -d)" | |
| gh release download "$CI_FIXTURE_TAG" \ | |
| --pattern "$CI_FIXTURE_ASSET" \ | |
| --dir "$tmp_dir" | |
| tar -xzf "$tmp_dir/$CI_FIXTURE_ASSET" -C models/ | |
| rm -rf "$tmp_dir" | |
| ls -la models/qwen3-0.6b-4bit | |
| - name: Run heterogeneous-memory partition regressions | |
| env: | |
| CARGO_PROFILE: debug | |
| run: ./scripts/ci/run-pp-heterogeneous-memory.sh | |
| - name: Run 2-host logical pipeline-parallel job | |
| env: | |
| CARGO_PROFILE: debug | |
| # The fixture step above populates `models/qwen3-0.6b-4bit/`, so | |
| # the multi-stage smoke actually exercises coordinator + transport | |
| # + remote stage service against real weights — instead of | |
| # self-skipping as it did before the fixture existed. | |
| MLXCEL_CI_PP_MODEL: qwen3-0.6b-4bit | |
| TEST_MODEL: qwen3-0.6b-4bit | |
| run: ./scripts/ci/run-pp-two-host-logical.sh | |
| - name: Clippy on PP modules | |
| run: | | |
| cargo clippy \ | |
| -p mlxcel \ | |
| --lib \ | |
| --tests \ | |
| -- -D warnings | |
| three-host-real-model: | |
| name: 3-host real-model parity (self-hosted) | |
| # Runs on either: | |
| # 1. PRs that explicitly carry the `ci:pp-three-host` label; or | |
| # 2. Manual `workflow_dispatch` with `run_three_host=true`. | |
| # | |
| # The job targets a self-hosted runner group labelled `pp-three-host` | |
| # that owns the 3-machine lab. The runner group is expected to have | |
| # model weights mounted at `$MODEL_ROOT` so the ignored | |
| # `pipeline_multi_stage_three_host_real_model_parity` test can | |
| # exercise the full 3-stage path. If the lab is offline, the job | |
| # stays queued until the runner becomes available or the PR is | |
| # updated — it does not auto-fail. | |
| if: | | |
| github.repository == 'lablup/mlxcel' && ( | |
| (github.event_name == 'pull_request' && | |
| github.event.pull_request.head.repo.full_name == github.repository && | |
| contains(github.event.pull_request.labels.*.name, 'ci:pp-three-host')) || | |
| (github.event_name == 'workflow_dispatch' && | |
| github.event.inputs.run_three_host == 'true') | |
| ) | |
| runs-on: | |
| - self-hosted | |
| - pp-three-host | |
| timeout-minutes: 120 | |
| needs: two-host-logical | |
| env: | |
| CARGO_PROFILE: release | |
| PP_THREE_HOST_MODEL: llama-3.2-1b-4bit | |
| PP_THREE_HOST_TOKENS: "16" | |
| PP_THREE_HOST_PROMPT: "Hello from the 3-host pipeline-parallel CI validation." | |
| steps: | |
| - name: Checkout code | |
| uses: actions/checkout@v7 | |
| with: | |
| # Don't leave GITHUB_TOKEN in .git/config; CI jobs only fetch. | |
| persist-credentials: false | |
| - name: Resolve model root | |
| id: model | |
| run: | | |
| # The self-hosted runner group is expected to mount model | |
| # weights at either $MODEL_ROOT or the repo-local models/ | |
| # path. The 3-host script fails loudly when STRICT_MODEL=1 | |
| # but weights are absent, so pass that through. | |
| if [[ -z "${MODEL_ROOT:-}" && -d "$HOME/mlxcel-models" ]]; then | |
| echo "MODEL_ROOT=$HOME/mlxcel-models" >> "$GITHUB_ENV" | |
| fi | |
| - name: Run 5 rollout-checklist tests | |
| run: ./scripts/ci/run-pp-rollout-tests.sh | |
| - name: Run 3-host real-model parity sweep | |
| env: | |
| STRICT_MODEL: "1" | |
| REPORT_PATH: "${{ github.workspace }}/docs_internal/performance/pipeline_parallel_3stage_ci_runs.md" | |
| run: ./scripts/ci/run-pp-three-host.sh | |
| - name: Upload captured rollout CSV | |
| if: always() | |
| uses: actions/upload-artifact@v7 | |
| with: | |
| name: pp-3host-benchmarks-${{ github.run_id }} | |
| path: | | |
| benchmarks/pipeline_parallel_*.csv | |
| docs_internal/performance/pipeline_parallel_3stage_ci_runs.md | |
| if-no-files-found: ignore | |
| retention-days: 30 |