[ESM/Persimmon/NllbMoe/ESMFold] use safetensors repos to fix Xet FUSE SIGBUS/IOError on bucket CI runners (#49194)
* [ESM/Persimmon/NllbMoe] use safetensors repos to fix Xet FUSE SIGBUS on bucket CI runners
Models facebook/esm2_t36_3B_UR50D, adept/persimmon-8b-chat, and
facebook/nllb-moe-54b only have pytorch .bin files. On bucket CI runners
(/mnt/cache mounted via Xet FUSE), torch.load() triggers mmap page faults
that the Xet FUSE daemon cannot serve for large blobs → SIGBUS crash.
Switch to dedicated hf-internal-testing repos with safetensors weights
which load correctly from Xet FUSE.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* [ESMFold] use safetensors repo to fix Xet FUSE I/O error on multi-gpu bucket runners
facebook/esmfold_v1 only has pytorch_model.bin. On bucket CI runners
(Xet FUSE), torch.load() fails with OSError: [Errno 5] Input/output error
when loading from /mnt/cache. Switch to safetensors repo.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>