README.md

# HF Jobs Development

Use a Hugging Face GPU Job as the execution and debugging target for a local Code OSS workspace.

1. Run **HF Dev: Start or Reconnect GPU Job** and select the hardware and timeout.
2. Edit and save locally; files synchronize automatically.
3. Press F5 in a Python file to run it remotely and attach the debugger to local breakpoints. Remote prints and Hub
   progress appear automatically in the **HF Dev** Output channel.
4. Run **HF Dev: Stop GPU Job** when finished.

Only one HF Dev debugger runs per workspace. Ending the local attach stops its remote process, and starting another
debug run replaces any older one so concurrent model downloads do not silently duplicate work.

The local checkout is authoritative, and Transformers is installed editable on the Job.

**The MVP creates no cache Bucket and mounts no Bucket by default. Hub downloads use ephemeral Job-local storage: they
are reused within that Job and discarded when it ends. A persistent Bucket cache is an explicit CLI/TOML opt-in and
can make large-model loading slower because weights are read through the remote mount.**

On macOS, use `fn`/Globe + F5 unless function keys are standard.

Requires `hf >= 1.28.0`, Hub authentication and SSH key, `ssh`, Mutagen, `hf-dev`, Python, and debugpy.

See `AGENTS.md` for the architecture, CLI primitives, direct SSH, automation, configuration, safety, troubleshooting,
and release instructions.