Why this Space is in CONFIG_ERROR: the pinned torch is outside ZeroGPU's supported set — and the cu118 pins must move with it

#13
by nootxlm - opened

This Space never reaches its own code. The platform's config error is explicit: the torch version in requirements.txt is not compatible with ZeroGPU, and it lists the supported versions — 2.13.0, 2.12.1, 2.11.0, 2.10.0, 2.9.1, 2.8.0.

requirements.txt pins torch==2.2.0, with torchaudio==2.2.0 taken from the cu118 wheel index. 2.2.0 is below the supported floor, so the Space is rejected at the config stage on every rebuild.

The fix has to move the coupled pins together, not just the version number:

  • torch==2.8.0 and torchaudio==2.8.0 (the oldest supported pair, so the rest of the stack drifts least), from the matching wheel index — the cu118 index line belongs to the 2.2 pair and should not survive the bump.
  • app.py also installs flash-attn at startup (pip install flash-attn --no-build-isolation), and that build follows whatever torch it finds. Under the current pin it would build against 2.2.0; after the bump it builds against 2.8.0 — worth a line in the same change, because a flash-attn build keyed to the old torch is the next startup failure in line.

Check: after the change, a restart passes the config stage — the error above is the config stage's own verdict, so its absence on the next boot is the proof this was the fault.

Sign up or log in to comment