fastgen-runtime

name: fastgen-runtimedescription: >-  Resolve the execution context for an NVIDIA FastGen command: checkout and interpreter, config path and override syntax, output/cache/data roots, run identity, and distributed launch variables. Use it when preparing or debugging execution; method selection belongs to fastgen-configure.license: Apache-2.0metadata:   author: "Julius Berner <jberner@nvidia.com>"  domain: ml

FastGen runtime

Shared prerequisite for every FastGen command. The checkout's code and READMEs are authoritative; this skill says where to look and what to verify. The configuration README owns syntax, environment variables and credentials; the troubleshooting guide owns failure diagnosis.

1. Checkout and interpreter. Locate the checkout separately from this skill directory and read its READMEs rather than assuming another version matches. Apply fastgen-setup only for missing work. Select the interpreter and devices the task actually uses and obey local scheduler, container and storage rules; importing GPU dependencies or resolving a config may already need an allocation.
2. Config and overrides. Run from the repository root (or put it on PYTHONPATH): training enters through the root train.py. --config is a relative .py path exposing create_config(), and the loader in config_utils.py rejects absolute paths and ... Script arguments precede the single - required by override_config_with_opts to introduce Hydra overrides (command-line arguments). --dryrun resolves the config and exits before the process group is initialised and before the global batch is rounded to accumulation steps; it tests neither.
3. Roots. Set an absolute, persistent FASTGEN_OUTPUT_ROOT and respect the data, checkpoint and Hugging Face cache roots derived from it (.env is loaded on import; see environment variables). Check the documented cache layout before pre-downloading and enable offline mode only once every needed file is cached. Keep credentials out of snapshots and logs.
4. Run identity. Independent runs need distinct log_config.project/group/name: the run directory holds the resolved config, checkpoints and the W&B id, and resume is on by default. Also check the checkpoint destination: resolve_save_dir preserves absolute paths and S3 URIs, so those must be distinct for independent runs.
5. Launch. Check the entry point's distributed support before building a launcher: training distributes only with trainer.ddp or trainer.fsdp, and the standard inference scripts do not shard models. Match RANK, LOCAL_RANK and WORLD_SIZE to the processes launched; stale launcher variables make setup wait for peers (distributed startup). Verify actual world size, wrapping, accumulation and effective batch in startup logs; a review-only plan should identify these checks as pending.

Report the selected checkout, interpreter, config, roots and launch context. For import failures check the imported package path and working directory first; for access failures distinguish missing files, authentication and model terms, and never substitute another model silently. Linked skills live in the installed catalog or in the checkout's skills/ directory.