$ the-wire · showcase
Flash and sage varlen prep work under torch.compile dynamic shapes
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
The varlen attention paths in diffusers and the Metal flash SDPA path in kernels-community both landed fixes for failures that only showed up under real execution conditions, and nothing in the day obliges an upgrade.
Fix flash/sage varlen prep under torch.compile with dynamic shapes (#14568) huggingface/diffusers
The no-mask varlen prep path no longer calls `.item()` at all, removing the graph break that Flux hit because it passes no attention mask, so the flash-attn and sage varlen backends are usable under `torch.compile` with dynamic shapes. Zero-length sequences are now handled in the varlen offsets, and a dynamic-compile test was added; masked varlen keeps one `.item()` for the data-dependent key l...
[MPS] Fix metal flash sdpa encoder handoff huggingface/kernels-community
Rather than creating a new compute encoder on the command buffer Torch created, which fails when another encoder is already active, the Metal flash SDPA path reuses Torch's compute encoder directly; the blit path ends compute coalescing before opening its own encoder. That is what made the encoder handoff fail on MPS.
[Bot] Update Inference Providers documentation huggingface/hub-docs
The Inference Providers docs were regenerated by bumping `@huggingface/tasks` and `@huggingface/inference` to latest and re-running the generator, then merged automatically by the documentation workflow. Expect the published provider list to track the current SDK output.
