How PyTorch's Test Infrastructure Works

PyTorch's tests have the characteristic of being dynamically generated at import time, which means test names that appear in the CI (Continuous Integration) environment can differ from the class or method names in the source code.

The core operating principles are as follows:

• Dynamic Expansion: By leveraging decorators and OpInfo (operator metadata), a single test method is automatically expanded into multiple combinations of Device and Dtype.
• Device-Independent Tests: Tools like instantiate_device_type_tests() efficiently manage tests that are not tied to a specific device.
• CI Sharding: Large-scale tests are split up to run in parallel.

To reproduce a test failure locally, using the pytest -k option or test/run_test.py is the fastest approach, and contributors can debug and contribute more effectively by understanding how this infrastructure works.