CANN torchtitan-npu测试指南

Test Guide

【免费下载链接】torchtitan-npuAscend Extension for torchtitan项目地址: https://gitcode.com/cann/torchtitan-npu

Core Commands

Unit Tests

# Run all unit tests and generate reports sh build.sh -u --generate-report # Run only local `torchtitan-npu` unit tests RUN_TORCHTITAN_UT=false sh build.sh -u --generate-report

Smoke Tests

# Run the default smoke suite (core + extended) sh build.sh -s --generate-report # Run only core smoke ONLY_CORE_SMOKE=true sh build.sh -s --generate-report # Run only extended smoke ONLY_EXTENDED_SMOKE=true sh build.sh -s --generate-report # Run only upstream smoke ONLY_UPSTREAM_SMOKE=true sh build.sh -s --generate-report

Integration Test

tests/smoke_tests/integration_test.pyis the entry point for end-to-end integration tests, used to validate:

• New model functionality support
• Feature compatibility
• Parallelism strategy compatibility

Running

# Via build.sh (runs core + extended smoke by default) ONLY_CORE_SMOKE=true sh build.sh -s --generate-report # Run integration_test.py directly python tests/smoke_tests/integration_test.py output_dir \ --test_name all \ --ngpu 2

Command-line Arguments

OverrideDefinitions Usage

OverrideDefinitionsis the configuration class for defining integration test cases:

OverrideDefinitions( override_args=[[...]], # Required: command-line argument list test_descr="...", # Required: test description test_name="...", # Required: test name ngpu=2, # Optional: required GPU count disabled=False, # Optional: whether disabled )

Steps to Add a New Test Case

1. Opentests/smoke_tests/integration_test.py
2. Add a new configuration to thesmoke_caseslist ingenerate_smoke_tests():

OverrideDefinitions( [ [ "--model.name your_model", "--model.flavor your_flavor", "--parallelism.tensor_parallel_degree 2", ], ], "Your Model TP Test", "your_model_tp", ngpu=2, )

3. Run tests to verify:

python tests/smoke_tests/integration_test.py ./outputs --test_name your_model_tp

base_test.toml Configuration

tests/smoke_tests/base_test.tomlis the base configuration for integration tests. All tests run based on this configuration file, and parameters inoverride_argsoverride identically-named parameters in the base configuration.

Model Parallel Commands

# Basic model-parallel smoke python3 -m pytest -v tests/smoke_tests/model_parallel/ # Multi-rank model-parallel smoke RUN_MODEL_PARALLEL_MULTI_RANK=true torchrun --nproc_per_node=4 -m pytest -v tests/smoke_tests/model_parallel/

When to Use Which Command

Quick Decision Rule

• Changed only hardware-independent logic: start withbuild.sh -u
• Changed NPU feature paths or wrappers: runbuild.sh -s
• Changed training-path wiring: at least runONLY_CORE_SMOKE=true build.sh -s
• Changed model-parallel behavior: runONLY_EXTENDED_SMOKE=true build.sh -s
• Upstream integration compatibility needs a separate check: runONLY_UPSTREAM_SMOKE=true build.sh -s

Test Reports

• Output directory:test_reports/
• Common artifacts:
  • *.xml: JUnit results
  • *.html: HTML reports when--generate-reportis enabled
  • coverage/: UT coverage reports
  • README.md: generated index of report artifacts

Quick Tips

1. Start with the smallest command that matches your change.
2. Preferbuild.sh -uwhen NPU is not required.
3. Use targeted smoke variants instead of full smoke when possible.
4. Update docs when test layout or execution changes.

【免费下载链接】torchtitan-npuAscend Extension for torchtitan项目地址: https://gitcode.com/cann/torchtitan-npu