The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes a script (download_golden_values.py) to fetch test data from GitHub. This is a standard operation for maintaining test parity in the Megatron-LM repository and targets a well-known service.
[COMMAND_EXECUTION]: Provides instructions for executing tests locally using torch.distributed.run and pytest. These commands are standard for distributed deep learning development and are used within the context of the project's testing suite.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external configuration files, such as YAML recipes and JSON golden values. While this represents a theoretical data ingestion surface, it is an essential part of the functional testing workflow for the Megatron-LM framework.
Ingestion points: model_config.yaml, golden_values_{env}_{platform}.json, and recipe YAMLs.
Boundary markers: Not explicitly defined in the instructions.
Capability inventory: Execution of pytest and torch.distributed.run via shell commands.
Sanitization: Not specified, though the skill relies on standard Python parsers (YAML/JSON).

mcore-testing