Search results

qwen3.5-0.8 grpo trl(most recent version) is not working with vllm 0.17.0

Status: Open.

#5269 In huggingface/trl;

· deter3 opened on Mar 10, 2026

[GOLD][vLLM] On-policy generation decodes prompts with skip_special_tokens=True (chat template stripped) — intentional?

Status: Open.

#5241 In huggingface/trl;

· siyan-zhao opened on Mar 7, 2026

Logging how vllm importance ratios are truncated/masked in GRPOTrainer

Status: Open.

#5231 In huggingface/trl;

· muupan opened on Mar 6, 2026

Integrate GTPO (Group-relative Trajectory-based Policy Optimization) as a new loss_type in GRPOTrainer

Status: Open.

#5226 In huggingface/trl;

· winstonsmith1897 opened on Mar 5, 2026

Re-tokenization bug in GRPO multi-turn tool calling

Status: Open.

#5224 In huggingface/trl;

· qgallouedec opened on Mar 5, 2026

DPOTrainer ref adapter crashes with PEFT target_parameters (MoE models on Transformers 5.x)

Status: Open.

#5222 In huggingface/trl;

· pritmish opened on Mar 5, 2026

Recomputed tensor size does not match when using quantization + LoRA with the DPO Trainer with gradient checkpointing using deepspeed and accelerate

Status: Open.

#5217 In huggingface/trl;

· slhmath opened on Mar 4, 2026

ValueError: The following model_kwargs are not used by the model: ['mm_token_type_ids']

Status: Open.

Bug

#5216 In huggingface/trl;

· albertvillanova opened on Mar 4, 2026

Qwen3.5-9B: Mismatch between tokenized prompt and the start of tokenized prompt+completion

Status: Open.

#5213 In huggingface/trl;

· zouYC2021 opened on Mar 3, 2026

Adaptive Beta-DPO

Status: Open.

#5211 In huggingface/trl;

· giuseppecurci opened on Mar 3, 2026

CI often fails for TestDPOTrainer::test_train_toolcall_data: torch.OutOfMemoryError: CUDA out of memory

Status: Open.

#5207 In huggingface/trl;

· albertvillanova opened on Mar 3, 2026

dpo training with Qwen2.5-0.5B+1000 seq length cost 17GB GPU mem

Status: Open.

#5200 In huggingface/trl;

· 32234 opened on Mar 2, 2026