distributed-llm-pretraining-torchtitan
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches source code and requirements from the official PyTorch torchtitan and torchao GitHub repositories. These are established repositories for high-performance machine learning utilities.
- [COMMAND_EXECUTION]: Provides instructions to execute shell scripts (run_train.sh), distributed training launchers (torchrun, srun), and utility scripts for downloading model assets and converting checkpoints. These commands are standard for large-scale model training workflows.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest large-scale datasets (such as "c4") and external model assets from Hugging Face for pretraining purposes.
- Ingestion points: Dataset configurations in TOML files and model assets fetched via download_hf_assets.py.
- Boundary markers: None explicitly mentioned in the skill instructions; data processing is handled by the underlying training framework.
- Capability inventory: The training environment supports shell command execution, local file system access for checkpointing, and network access for asset synchronization.
- Sanitization: Relies on standard data loading and tokenization routines within the PyTorch ecosystem.
Audit Metadata