CI warns with dev dependencies: FutureWarning: `device` is deprecated and will be removed in version 5.18 for `compute_default_rope_parameters`

GRPO VLM colocate: pre-tokenized prompt ids + raw image handoff mis-aligns with vLLM preprocessing (M-RoPE crash / silent prompt corruption)

Optimize `old_per_token_logps` recomputation in GRPOTrainer: per-rollout window check instead of static modulo

CI fails: AssertionError: Param model.visual.blocks.0.norm1.weight is not updated

CI fails with transformers v5.0.0: AttributeError: 'GptOssConfig' object has no attribute 'num_experts'

Add support for models using final_logits_softcapping to AsyncGRPOTrainer

TPOTrainer.evaluate() returns NaN eval_loss while training loss is finite

DPOTrainer does not work for multimodal Gemma 4

Does PPOTrainer use mini_batch_size to update parameters?

CI fails with dev dependencies for gpt-oss models: RuntimeError: You set `ignore_mismatched_sizes` to `False`