hf jobs uv run supports multi-GPU launchers and local config files
Introduced 2026-02-03 · score 0.95
HOW TO USE IT
When running multi-GPU training or passing a local config file through Hub Jobs, use `hf jobs uv run` with a supported distributed runner or put the config after the script as a script argument.
hf jobs uv run --with torch -- torchrun <script.py>hf jobs uv run --with accelerate -- accelerate launch <script.py>hf jobs uv run [--with <package>] <script.py> <config-file>WHY THIS HELPS
When developing or testing multi-GPU training code through Hub Jobs, launch it with a supported distributed runner and pass a local config file alongside the script.
REQUIREMENTS AND LIMITATIONS
- Version: v1.4.0
- SCRIPT (required): UV script to run; may be a local file or URL.
- SCRIPT_ARGS: Arguments for the script, following the script path.
- --with <package>: Run with the given package installed. Repeated use or mutual exclusivity is not specified in the cited docs.
- --flavor <flavor>: Select hardware flavor; defaults to `cpu-basic`. Use a GPU flavor for GPU training; `hf jobs hardware` lists available flavors.
- Prerequisite: Install the Hugging Face CLI and log in to a Hugging Face account before running Jobs.
FIRST-PARTY EVIDENCE
- https://github.com/huggingface/huggingface_hub/releases/tag/v1.4.0
- https://github.com/huggingface/huggingface_hub/blob/v1.4.0/docs/source/en/package_reference/cli.md
- https://huggingface.co/docs/huggingface_hub/en/jobs-quickstart
Check the installed version and the source before using it. Entry JSON