The Wire · Showcase
DIFFUSERS TIGHTENS LORA TRAINING, TRANSFORMERS SHIPS FSDP2 FOUNDATION
By RepoJournal · Filed · About Hugging Face
Hugging Face cut training tests over to pytest while shipping caption dropout and aspect-ratio bucketing for two key LoRA models, and transformers landed its FSDP2 native module ahead of major distributed training rework.
Diffusers merged caption dropout and aspect-ratio bucketing for krea2 and qwen-image [1], two training features that let operators randomly zero captions to improve classifier-free guidance and pack images by ratio instead of padding to fixed sizes. The team also migrated training tests to pytest [2] to standardize infrastructure, with follow-up work to consolidate test mixins and eliminate copy-paste patterns [2]. Transformers shipped FSDP2 native module wrapping and plan verification [3], moving distributed training coordination out of integrations and into core, then cleaned up redundant guards across the codebase [4]. On the multimodal front, transformers updated most commonly used LLM processors in a breaking change that removed small private functions [5]. MLClaw completed OpenAI OAuth integration across five commits, removing the Codex CLI dependency entirely and routing agent models through native OpenClaw harness [6], while hardening credential lifecycle to prevent logout undo and preserve migration intent [7].
One email a day. Unsubscribe in one click.
Keep up with Hugging Face in about 3 minutes a day: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
One email a day. Unsubscribe in one click. Read a past issue →
Action items
- → Review breaking processor changes in transformers [ref:9] before upgrading if you depend on private multimodal functions huggingface/transformers [plan]
- → Test FSDP2 integration if running distributed training at scale huggingface/transformers [monitor]
- → Integrate caption dropout into LoRA training pipelines for better CFG performance huggingface/diffusers [plan]
References
- [1] [lora training] add caption dropout and aspect ratio buckets to krea2 and qwen-image ↗ huggingface/diffusers
- [2] [tests] migrate training tests to pytest. ↗ huggingface/diffusers
- [3] Add FSDP plans to all models (#47165) huggingface/transformers
- [4] Remove redundant guarding for distributed (#47570) huggingface/transformers
- [5] :rotating_light: Processors update the rest ↗ huggingface/transformers
- [6] fix(openai): use Codex agent runtime ↗ huggingface/mlclaw
- [7] fix(credentials): harden Codex credential lifecycle ↗ huggingface/mlclaw