The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for executing distributed training workloads using CLI tools including ft_launcher, srun, and uv run. These commands are standard for managing multi-node GPU training and performance testing in high-performance computing (HPC) environments.
[EXTERNAL_DOWNLOADS]: The instructions identify a dependency on nvidia-resiliency-ext (version approximately 0.5.0). This is a specialized library provided by NVIDIA to support training stability and recovery features like rank monitoring and local checkpointing.
[DATA_EXFILTRATION]: The skill configures network communication for distributed ranks using standard environment variables such as MASTER_ADDR and MASTER_PORT. It does not contain patterns associated with exfiltrating sensitive local data to external or untrusted domains.
[SAFE]: The technical content, including Python code snippets for NeMo Run plugins and Megatron Bridge configurations, aligns with the stated purpose of improving training resiliency. The use of state files like ft_state.json and local checkpoint directories follows established patterns for distributed training state management.

resiliency