The Wire · Showcase
TRANSFORMERS FIXES RECURRENT MASKING BUG; MOON-IDE KILLS PAIRING WINDOW
By RepoJournal · Filed · About Hugging Face
10 people shipped this
coyotte508
@coyotte508
3 cited
David Corvoysier
@dacorvo
3 cited
sayakpaul
@sayakpaul
2 cited
Andrei Boldyrev
@abcgco
1 cited
Ferdinand Mom
@3outeille
1 cited
hmellor
@hmellor
1 cited
atharvajoshi10
@atharvajoshi10
1 cited
goyaladitya05
@goyaladitya05
1 cited
akshan-main
@akshan-main
popovaan
@popovaan
A critical padding mask bug in transformers is now fixed for chunked prefill and cache continuation, while moon-ide removes dead startup ceremony and hardens the remote bridge client.
Transformers shipped two critical fixes this cycle. First, `create_recurrent_attention_mask` was returning `None` during multi-token forwards from cache, meaning chunked prefill and cache continuation never zeroed padding in the linear-attention state [1]. Second, distributed training infrastructure landed with DistributedMixin and a new distributed/utils.py module, laying groundwork for FSDP orchestration without breaking the from_pretrained API [2]. Additional model fixes address yarn `mscale_all_dim` across DeepSeek v2 and Mistral 4 [3], and Cosmos 3 Edge patch packing [4]. Over in moon-ide, the startup pairing window is gone: since enrolled IDEs mint codes on demand, the printed QR at serve startup was dead surface [5]. Three fixes landed for the outbound relay client: rustls provider initialization at startup, clean reconnection handling, and silent task kills [6]. The relay hub is now production-deployed on bridge.coyo.dev behind nginx with two new serve flags: --no-idle-exit (to keep the standing relay alive) and --advertise-url (to fix pairing payload routing) [7]. Diffusers deprecated `torch_dtype` in favor of `dtype` following transformers convention [8], and removed the experimental API warning from LoRA modules [9]. The optimum-intel OpenVINO pipeline added image-to-video support for LTX-Video [10]. Funes dropped its automated-first-index guard since cold stores now fill one bounded step per turn from any hook [11], added hermes as a full-pipeline agent [12], and broke opencode support entirely [13].
One email a day. Unsubscribe in one click.
What actually shipped in Hugging Face, written up every day — commits, pull requests, releases, and security advisories.
One email a day. Unsubscribe in one click. Read a past issue →
Action items
- → Upgrade transformers if you use recurrent models with chunked prefill or cache continuation huggingface/transformers [immediate]
- → Review distributed training changes if you plan FSDP usage huggingface/transformers [plan]
- → Update diffusers calls from torch_dtype to dtype huggingface/diffusers [plan]
- → If using funes with opencode, migrate to manual recall now huggingface/funes [immediate]
References
- [1] Fix recurrent-layer padding mask being skipped on continued forwards (chunked prefill, cache continuation) (#47087) huggingface/transformers
- [2] Add distributed runtime utils and DistributedMixin (#47352) huggingface/transformers
- [3] Fix yarn `mscale_all_dim` for DeepSeek v2 and Mistral 4 ↗ huggingface/transformers
- [4] Fix Cosmos 3 Edge Patch packing order ↗ huggingface/transformers
- [5] Remove the startup pairing window; mint phone codes on demand everywhere huggingface/moon-ide
- [6] Fix remote-bridge client: rustls provider panic, reconnect on clean close, startup respawn huggingface/moon-ide
- [7] Add standing-relay serve flags for public nginx-fronted deployment huggingface/moon-ide
- [8] [core] deprecate `torch_dtype` and prefer `dtype` following transformers. ↗ huggingface/diffusers
- [9] [chore] remove experimental api warning from lora modules. (#14248) huggingface/diffusers
- [10] [OpenVINO] Add Image-to-Video support for LTX-Video ↗ huggingface/optimum-intel
- [11] feat(index): drop the automated-first-index guard huggingface/funes
- [12] feat(add): funes add hermes wires the shell-hook and first index huggingface/funes
- [13] feat!: drop opencode support huggingface/funes
