Megatron Core Testing Guide

Testing Guide

Answer-First Testing Facts

For questions about disabling tests without deleting them:

• Functional recipe entries stay in YAML; disable by suffixing scope with -broken, for example scope: [mr-github] -> scope: [mr-github-broken].
• Unit-test skips use pytest markers instead: @pytest.mark.flaky_in_dev skips in the default dev environment, and @pytest.mark.flaky skips in LTS.
• Do not delete the test case or recipe entry when the goal is discoverability and easy re-enable.

Test Layout

tests/
├── unit_tests/          # pytest, 1 node × 8 GPUs, torch.distributed runner
├── functional_tests/    # end-to-end shell + training scripts
│   └── test_cases/
│       └── {model}/{test_case}/
│           ├── model_config.yaml          # training args
│           └── golden_values_{env}_{platform}.json
└── test_utils/
    ├── recipes/
    │   ├── h100/        # YAML recipes for H100 jobs
    │   └── gb200/       # YAML recipes for GB200 jobs
    └── python_scripts/  # helpers (recipe_parser, golden-value download, …)

How Tests Execute

The GitHub Actions runner invokes launch_nemo_run_workload.py, which uses nemo-run to launch a DockerExecutor container. The repo is bind-mounted at /opt/megatron-lm; training data is mounted at /mnt/artifacts.

Unit tests are dispatched through torch.distributed.run:

• Ranks 0 and 3 are tee-d to stdout; all other ranks write only to log files.
• Per-rank log files land at {assets_dir}/logs/1/ and are uploaded as a GitHub artifact after the run.

Functional tests are driven by tests/functional_tests/shell_test_utils/run_ci_test.sh. Only rank 0 runs the pytest validation step; training output from all ranks is uploaded as an artifact.

Flaky-failure auto-retry: launch_nemo_run_workload.py retries up to 3 times for known transient patterns (NCCL timeout, ECC error, segfault, HuggingFace connectivity, …) before declaring a genuine failure.

Recipe YAML Structure

Recipes live in tests/test_utils/recipes/ and are parsed by tests/test_utils/python_scripts/recipe_parser.py. Each file expands a cartesian products block into individual workload specs:

type: basic
format_version: 1
maintainers: [mcore]
loggers: [stdout]
spec:
  name: "{test_case}_{environment}_{platforms}"
  model: gpt              # maps to tests/functional_tests/test_cases/{model}/
  build: mcore-pyt-{environment}
  nodes: 1
  gpus: 8
  n_repeat: 5
  platforms: dgx_h100
  time_limit: 1800
  script_setup: |
    ...
  script: |-
    bash tests/functional_tests/shell_test_utils/run_ci_test.sh ...
products:
  - test_case: [my_test]
    products:
      - environment: [dev, lts]
        scope: [mr-github]
        platforms: [dgx_h100]

Key runtime placeholders: {assets_dir}, {artifacts_dir}, {test_case}, {environment}, {platforms}, {n_repeat}.

Disabling a Test Without Deleting It

To temporarily disable a test case in a recipe YAML, suffix its scope value with -broken — do not delete the entry:

# before (test runs in CI)
scope: [mr-github]

# after (test is skipped; entry preserved for easy re-enable)
scope: [mr-github-broken]

Running Unit Tests Locally

All unit tests initialize a torch.distributed group, so every invocation requires GPU access and must go through torch.distributed.run:

# Full suite
uv run python -m torch.distributed.run --nproc-per-node 8 -m pytest -q \
  tests/unit_tests

# Single file
uv run python -m torch.distributed.run --nproc-per-node 8 -m pytest -q \
  tests/unit_tests/models/test_gpt_model.py

# Single test
uv run python -m torch.distributed.run --nproc-per-node 8 -m pytest -q \
  tests/unit_tests/models/test_gpt_model.py::TestGPTModel::test_constructor

# Filter by name substring
uv run python -m torch.distributed.run --nproc-per-node 8 -m pytest -q \
  tests/unit_tests -k optimizer

Marker filters

# Exclude flaky tests during development
uv run python -m torch.distributed.run --nproc-per-node 8 -m pytest -q \
  tests/unit_tests -m "not flaky and not flaky_in_dev"

# Include experimental tests
uv run python -m torch.distributed.run --nproc-per-node 8 -m pytest -q \
  tests/unit_tests --experimental

CI parity

Use tests/unit_tests/run_ci_test.sh to reproduce a CI bucket failure exactly. For ad-hoc runs, prefer the direct torch.distributed.run invocations above.

Gotchas

• pyproject.toml sets addopts = --durations=15 -s -rA — stdout is not captured (-s), so ranks interleave during multi-rank runs. Override with --capture=fd when debugging a specific rank.
• tests/unit_tests/conftest.py looks for test data under /opt/data and attempts a download if missing. Supply it manually or skip data-dependent tests when running outside the canonical container.

Adding a Unit Test

1. Create tests/unit_tests/<category>/test_<name>.py.
2. Use fixtures from tests/unit_tests/conftest.py.
3. Apply markers as needed:
   • @pytest.mark.internal — skipped on legacy tag
   • @pytest.mark.flaky_in_dev — skipped in dev environment (CI default; use this to disable a flaky test without blocking the standard pipeline)
   • @pytest.mark.flaky — skipped in lts environment
   • @pytest.mark.experimental — latest tag only
4. Verify locally (see Running Unit Tests Locally above).
5. If the test needs a dedicated CI bucket, add an entry to tests/test_utils/recipes/h100/unit-tests.yaml.
6. If the change adds or modifies a GPU kernel (Triton, jit_fuser / torch.compile, CUDA extension, TE or external-library dispatch, or a scatter/index accumulation), add or update its bit-exact replay test under tests/unit_tests/determinism/kernels/ and register it in tests/unit_tests/determinism/kernels/manifest.py. The linting CI job (tools/check_kernel_determinism_coverage.py) fails kernel PRs without this; the determinism-exempt label overrides it for non-numeric edits. See docs/developer/determinism/testing.md.