transformers
2d4dc67f - Strix Halo: adopt AMD's tuned image + HF cache + token (Ryan's files)

Commit
56 days ago
Strix Halo: adopt AMD's tuned image + HF cache + token (Ryan's files) Replace the Dockerfile and probe workflow with AMD's tested versions: - base ubuntu:24.04 (no ROCm base); torch 2.14.0+rocm7.14 from download.pytorch.org (now ships gfx1151), torchvision 0.29, torchaudio 2.11 (AMD), torchcodec 0.13; trims non-gfx1151 device code (~39GB->14.6GB). - prefetch Hub test fixtures into the image (avoids per-job 429s), via a BuildKit hf_token secret. - add scipy/librosa/soundfile (missing-dep NameErrors). - run workflow mounts persistent HF cache /mnt/scratch/cache/hf -> HF_HOME=/mnt/cache. Wire the real org secret HF_HUB_READ_TOKEN (repo has no HF_TOKEN) into both the run (HF_TOKEN env) and the build (BuildKit hf_token secret for fixture prefetch). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Parents
Loading