The Wire · Showcase
TRL FIXES MULTIMODAL LOGGING AND QUANTIZATION CRASHES ACROSS GRPO AND QORA PIPELINES
By RepoJournal · Filed · About Hugging Face
TRL shipped four critical fixes overnight for multimodal RL training, including a QLoRA + FSDP1 dtype crash that was breaking mixed-precision setups.
The biggest one: QLoRA + FSDP1 mixed-dtype configurations were crashing in prepare_peft_model, and that's now fixed [1]. Parallel to that, the multimodal logging system got a new control lever with the `log_multimodal` parameter added to both GRPOConfig and RLOOConfig [2], letting you silence image logging when you need lighter checkpoints. GKD also got a vocab_size validation fix for multimodal models [3] that was silently breaking configs that looked valid. On the data side, GRPO and RLOO now handle IterableDataset errors correctly instead of swallowing them [4]. Kernels desk shipped Helion support [5], enabling pure-Python Triton kernels to build as noarch variants, which means one binary for all torch versions. huggingface.js fixed their Deno e2e CI by setting minimumDependencyAge to 0 for local registry testing [6], unblocking the test pipeline.
Action items
- → If you're running QLoRA + FSDP1 with mixed dtypes, pull the latest TRL immediately huggingface/trl [immediate]
- → Review GKD multimodal configs if you're seeing vocab_size validation errors huggingface/trl [plan]
- → Monitor huggingface.js Deno e2e tests; they should be green now huggingface/huggingface.js [monitor]
References
- [1] Fix QLoRA + FSDP1 mixed-dtype crash in prepare_peft_model (#6343) huggingface/trl
- [2] Add `log_multimodal` param to GRPOConfig and RLOOConfig to control image logging (#5408) huggingface/trl
- [3] Fix GKD vocab_size check for multimodal model configs (#6340) huggingface/trl
- [4] Fix error type for IterableDataset in GRPO and RLOO (#6324) huggingface/trl
- [5] Add Helion as a supported Python dependency ↗ huggingface/kernels
- [6] CI fix e2e test (#2284) huggingface/huggingface.js