Hugging Face's logo

Spaces:
huggingface-course
/
audio-course-u7-assessment
Runtime error

App Files Community
23

Assessment Space down: "No space left on device" loading facebook/mms-lid-126

#23
by VoicesColeby - opened
Discussion
VoicesColeby

The Unit 7 assessment Space is currently failing to start with a RuntimeError: ... No space left on device (os error 28) while downloading facebook/mms-lid-126, which then cascades into:

OSError: Can't load the model for 'facebook/mms-lid-126'

So the Space sits in RUNTIME_ERROR and submissions cannot be made. The progress tracker Space (MariaK/Check-my-progress-Audio-Course) is also down at the moment.

Looks like the persistent storage / cache for this Space has filled up. A factory reboot (or clearing the HF cache / increasing storage) should fix the mms-lid-126 download.

For reference, my demo is ready and public: VoicesColeby/speech-to-speech-translation (translates speech to Dutch; output is correctly classified as nld by mms-lid-126). I just can'''t submit it while the assessor is down.

cc @sanchit-gandhi — thanks!
VoicesColeby

Update — actual root cause is not the disk error, and I have a working fork.

The current RUNTIME_ERROR is from the Repository(...).git_pull() at the top of app.py:

USERNAMES_DATASET_ID = "huggingface-course/audio-course-u7-hands-on"
usernames_repo = Repository(local_dir="usernames", clone_from=usernames_url, use_auth_token=HF_TOKEN)
usernames_repo.git_pull()

That dataset URL now returns 404 (deleted or perms-revoked), so the Space crashes before demo.launch(). The mms-lid-126 disk error in the original report has likely also been resolved by HF's free-tier disk bump — once the dataset issue is fixed, the LID model loads fine.

I've verified this by duplicating the Space, dropping the dataset clone, and re-pinning a few versions to escape the Python-3.13 / gradio 5.0 / transformers-HEAD churn that's separately breaking a lot of Spaces this week:

Status: Congratulations VoicesColeby! Your demo passed the assessment!
Top language: nld (Dutch) @ 0.99988 confidence

Patch for the official Space

  1. Drop the huggingface-course/audio-course-u7-hands-on clone (or re-create the dataset and re-publish).
  2. README.md YAML: pin python_version: "3.12" and sdk_version: 5.49.1 (gradio 5.0 imports the now-removed huggingface_hub.HfFolder).
  3. requirements.txt: switch git+https://github.com/huggingface/transformers to transformers>=4.45,<4.60 so HEAD ABI changes don't break us; add numpy<2; add gradio_client>=1.4.
  4. app.py: wrap the file input via gradio_client.handle_file("test_short.wav") — modern Pydantic-validated gradio rejects bare path strings.

Full diff + working files: https://huggingface.co/spaces/VoicesColeby/audio-course-u7-assessment-fixed/tree/main

@sanchit-gandhi @MariaK — would appreciate if someone can either restore the usernames dataset + push a small app.py patch, or use this evidence to manually credit Unit 7 completion for users whose submissions have been verified through the fork. Happy to open a PR against the original Space if that's easier.

cc anyone else stuck on this: feel free to verify your submission against the fork. Note: the fork cannot write to the official roster, so passing there does not auto-update MariaK/Check-my-progress-Audio-Course — it is proof-of-completion only.

Sign up or log in to comment