$ the-wire · showcase
Online DPO fixes Qwen-VL multimodal RoPE, and a CPU parity test for continuous batching
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
13 people shipped this
nrenvo
@nrenvo
1 cited
molbap
@molbap
1 cited
albertvillanova
@albertvillanova
1 cited
clementtisseau
@clementtisseau
1 cited
linoytsaban
@linoytsaban
1 cited
regisss
@regisss
ebezzam
@ebezzam
ydshieh
@ydshieh
github-actions[bot]
@github-actions[bot]
yanchenl-nvidia
@yanchenl-nvidia
tifchen-nvda
@tifchen-nvda
Daniël de Kok
@danieldk
sayakpaul
@sayakpaul
TRL repaired Online DPO for Qwen2-VL and Qwen3-VL after transformers 5.3.0 began requiring mm_token_type_ids for multimodal RoPE, while transformers and diffusers filled in test and CI coverage rather than shipping new surface area.
Pass mm_token_type_ids to the model in Online DPO huggingface/trl
Online DPO training for Qwen-VL models dropped mm_token_type_ids from the processor outputs, so Qwen2-VL and Qwen3-VL failed outright and Qwen2.5-VL trained on wrong positions. The fix keeps the field alongside the other vision inputs and passes it into _forward, extended with zeros for the completion.
Fix several object detection models/loader fix for module tying huggingface/transformers
Models like grounding dino, dab-detr, rt-detr-v2 and d-fine alias weights under two names, such as self.model.decoder.bbox_embed = self.bbox_embed, which the loader did not handle. If you load any of those checkpoints, this is the fix for the object detection path.
test: add parity check to the CPU-only continuous batching test (#49728) huggingface/transformers
The only CPU continuous batching test was @slow and merely confirmed generation ran, so the CPU path had no accuracy coverage. The new test runs greedy generation through both generate_batch and generate on a tiny random Llama in fp32 and asserts the tokens match.
add fibo-2 LoRA training and loading huggingface/diffusers
Bria fibo-2 gets both text-to-image and image-edit DreamBooth LoRA trainers, plus a BriaFibo2LoraLoaderMixin so adapters can load, save and fuse; the model PR shipped PeftAdapterMixin without a loader mixin.
AsyncGRPO: keep filling a micro-batch when another row has free room huggingface/trl
TokenBudgetBatcher._place now seats a full-group sample in the least loaded row it fits without raising the most loaded row, rather than closing the micro-batch when every rank is busy. Elsewhere: Intel XPU CI arrives in transformers, TRL accepts vLLM 0.32.x patch releases, kernels-community moved cv-utils index_select onto torch-stable-abi-shim, and the Hub's download counting picked up EgoHan...
