The Wire · Showcase
FSDP DISTRIBUTED TRAINING AND PARAKEET ASR LAND IN TRANSFORMERS
By RepoJournal · Filed · About Hugging Face
Transformers shipped native FSDP and tensor parallelism support with distributed save/load in one release, while kernels team hardened security auditing and operation registration validation across all repos.
The transformers FSDP integration [1] adds fully sharded data parallel training with auto/manual mode detection, shard-on-read loading via DtensorShardOperation, and FSDP-aware flash attention checks. This is the distributed training foundation teams have been waiting for. Parakeet ASR models [2] now support Token-and-Duration Transducer decoding with per-token timestamps and full AutoModel integration, extending beyond CTC-only to match production ASR pipelines. PP-OCRv6 [3] arrives as the new standard for document recognition, adding four new model variants with updated backbones and dedicated image processors. Across kernels and kernels-community, the team locked down security workflows [4] [5] [6] and rolled out operation registration validation [7] to catch misconfigured custom ops at build time. Flash attention variable-length backward pass now works on XPU [8], removing a gap for Intel accelerator users. AsyncGRPOTrainer [9] now handles models with final logits softcapping like Gemma 2, fixing a trainer compatibility issue.
Action items
- → Test FSDP integration in staging if you run distributed training at scale - this replaces DDP for large model training huggingface/transformers [plan]
- → Pin kernels builds and run security audit workflow on your fork to verify the remediation took effect [ref:13] huggingface/kernels-community [immediate]
- → Update any custom ops to follow the naming prefix convention documented in the nix-builder hook [ref:2] huggingface/kernels [plan]
- → Upgrade to latest TRL if you're training with Gemma 2 or other softcapped models using AsyncGRPOTrainer huggingface/trl [monitor]
References
- [1] FSDP + TP & native save/load distributed (#45028) huggingface/transformers
- [2] Parakeet tdt (#44171) huggingface/transformers
- [3] [Model] Add PP-OCRv6 Models Support (#45838) huggingface/transformers
- [4] feat: mention maintainers in the slack security auditing. (#567) huggingface/kernels
- [5] fix(security): remediate workflow vulnerability in .github/workflows/security-audit.yml (#884) huggingface/kernels-community
- [6] feat: mention maintainers in the slack security auditing. (#881) huggingface/kernels-community
- [7] nix-builder: add a hook to detect incorrect op registrations ↗ huggingface/kernels
- [8] flash-attn2: Add flash_attn_varlen_func backward support for XPU ↗ huggingface/kernels-community
- [9] Final logits softcapping support for async GRPO Trainer (#5691) huggingface/trl
FAQ
- What changed in Hugging Face on May 20, 2026?
- Transformers shipped native FSDP and tensor parallelism support with distributed save/load in one release, while kernels team hardened security auditing and operation registration validation across all repos.
- What should Hugging Face teams do about it?
- Test FSDP integration in staging if you run distributed training at scale - this replaces DDP for large model training • Pin kernels builds and run security audit workflow on your fork to verify the remediation took effect [ref:13] • Update any custom ops to follow the naming prefix convention documented in the nix-builder hook [ref:2]
- Which Hugging Face repositories shipped on May 20, 2026?
- huggingface/transformers, huggingface/kernels, huggingface/kernels-community, huggingface/trl