$ the-wire · showcase
Transformers drops torch.jit.script and changes Mac serve attention; LeRobot moves to native EMA
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
Two breaking changes landed in transformers and huggingface.js while LeRobot replaced its diffusers-derived EMA with a sharded-training-capable implementation and finally raised its Pillow cap to pick up 13.0.0.
Remove torch.jit.script (#50082) huggingface/transformers
transformers removed torch.jit.script from the codebase entirely. If your serving or export path relied on scripted modules rather than trace or export, verify it against a recent build before pinning forward.
🚨 [serve][breaking] update attn for mac users (#49730) huggingface/transformers
The serve path got a breaking attention update for Mac users, flagged with the release-note marker for breaking. Mac-based inference servers should be re-tested against the new attention implementation rather than assumed unchanged.
feat(train): add LeRobot native EMA support with pluggable decays and sharded training huggingface/lerobot
LeRobot now owns its EMA in lerobot/optim/ema.py, replacing the diffusers EMAModel added earlier: it works under FSDP2 and HSDP, where the previous version ran on the main process only and raised NotImplementedError on sharded runs, and it takes pluggable decay curves via --ema.type including the power-function EMA of EDM2.
chore(dependencies): allow Pillow 13 huggingface/lerobot
LeRobot raises its Pillow cap from <13.0.0 to <14.0.0; Pillow 13.0.0 is the only release carrying fixes for CVE-2026-112395, -112397, -112400, -112404, -112895, -112896, -112897 and -112899 across the TIFF and PDF decoders, and Pillow does not backport to earlier majors.
Add F8_E4M3 GGUF QuantizationType (#2754) huggingface/huggingface.js
huggingface.js registers F8_E4M3 (OCP-FP8 E4M3) as a GGUF quantization type on GGMLFileQuantizationType (42) and GGMLQuantizationType (43), tracking the addition in llama.cpp.
Cloud baseline artifacts carry results only huggingface/transformers-mic
elsewhere: transformers-mic stopped copying contracts and fixtures into the CUDA baseline artifact (it had grown to gigabytes), diffusers fixed compute_embeddings.py for current transformers (load_in_8bit is gone; pass BitsAndBytesConfig, and cast bf16 to float32 before .numpy()), and lerobot-rollout now logs cadence, inference latency and memory every five seconds, tunable with --status_interv...
Action items
- → Audit transformers usage for torch.jit.script, removed in this release huggingface/transformers [plan]
- → Re-test Mac serve attention paths against the breaking update huggingface/transformers [plan]
- → Upgrade lerobot to the release allowing Pillow 13.0.0 to pick up the CVE-2026-1123xx fixes huggingface/lerobot [plan]
