huggingface_hub 2.2.0: hf jobs stats prints a snapshot, and parallel downloads always use hf_xet

huggingface_hub 2.2.0 changes hf jobs stats. Run it on a job and you get one reading of CPU, memory and GPU, then your prompt back. It used to stream until the job finished, so a script calling it on a running job hung until it timed out.

Slide text

01 02 03 04 05 hf jobs

huggingface/huggingface_hub · GitHub release v2.2.0 · PR #5090

huggingface_hub 2.2.0

hf jobs stats prints the latest sample and exits

Released October 8, 2026 Listed as breaking

Hugging Face lists this as a breaking change, and scripts that relied on the stream can pass the follow option. The Python call fetch_job_metrics changed the same way.

The snapshot also comes in machine-readable formats. There's a JSON mode, an agent mode and a quiet mode, and they print the raw metrics that the Jobs API returns, instead of the table.

Logs without the follow option used to hang while a job kept printing, and now they return. Stats works when a listed job has finished, and its JSON output is an empty list when zero jobs are running.

Downloads.

In a fresh process, the first parallel downloads could see a placeholder version for hf_xet and treat it as missing, so some files in a snapshot_download skipped Xet. Every download now sees the real version.

snapshot_download, and the hf download command, now print a warning when your allow or ignore patterns match zero files. Until now, a typo in a pattern looked exactly like a successful download.

An offline snapshot_download with a pinned revision now works. Following a resolve redirect keeps the port number. And CommitScheduler had a race that skipped its first scheduled push, which is fixed.

HfFileSystem.

HfFileSystem cached "not found" results. A repo, branch or bucket you created in the same process stayed missing until you cleared the whole cache. Now a refresh on that one path clears it.

Directory listings were appended to the cache, so a listing after a find call could show duplicate entries. A refresh kept deleted files, too. Each new listing now replaces the cached one.

Listing a bucket you had just created, with zero files in it, used to raise FileNotFoundError. In 2.2.0 it returns an empty list.

The hf command.

On Windows, the hf command expanded quoted patterns into local file names. So a download filtered to JSON files, run in a folder that held a JSON file, quietly fetched zero files. Arguments now arrive as typed.

Env and secrets files are now read as UTF-8, with the byte order mark stripped. Upload takes an existing path with glob characters literally. And planning a cache cleanup no longer slows quadratically with the number of files in a revision.

On Windows ARM64 machines, Python reports the processor name in capitals, so the install rule missed it. Installing huggingface_hub there now pulls in hf_xet as well.

Two exceptions change.

Hugging Face lists this as breaking too. load_torch_model used to raise a value error when the checkpoint path didn't exist. Code that catches it is affected. And model_index_to_eval_results now raises a value error when you pass an empty list.

The 2.2.0 release notes on GitHub link every pull request behind these changes, and they point to the Jobs guide for the new snapshot output.