Search results

CI fails with dev dependencies: ValueError: The following <code>model_kwargs</code> are not used by the model: ['mm_token_type_ids']

Status: Open.

Bug

#5201 In huggingface/trl;

· albertvillanova opened on Mar 2, 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

feat: Variational Sequence-Level Soft Policy Optimization (VESPO)

Status: Open.

#5196 In huggingface/trl;

· casinca opened on Feb 27, 2026

Unify Weight Sync for Rollout Path

Status: Open.

#5195 In huggingface/trl;

· rycerzes opened on Feb 27, 2026

Make openenv/utils.py Backend-Agnostic

Status: Open.

#5194 In huggingface/trl;

· rycerzes opened on Feb 27, 2026

Introduce a Minimal Backend Generation Interface

Status: Open.

#5193 In huggingface/trl;

· rycerzes opened on Feb 27, 2026

Align trl skills CLI DX with HF CLI conventions

Status: Open.

#5191 In huggingface/trl;

· albertvillanova opened on Feb 27, 2026

Does GOLD Trainer support multi-node training similar to GRPO?

Status: Open.

#5167 In huggingface/trl;

· NoviceStone opened on Feb 25, 2026

Support for constrained decoding with JSON Schema in GRPO Trainer + vLLM

Status: Open.

#5154 In huggingface/trl;

· ehofm opened on Feb 23, 2026

Computation of merged probabilities for GOLD

Status: Open.

#5152 In huggingface/trl;

· vman049 opened on Feb 23, 2026

GRPOTrainer tool_mask can become longer than completion_ids after tool-call retokenization

Status: Open.

#5144 In huggingface/trl;

· MichalMraz opened on Feb 21, 2026

Improper buggy SFT training of Llama-3.2-3B-Instruct.

🐛 bugSomething isn't workingSomething isn't working

🚨 Important

Status: Open.

#5138 In huggingface/trl;

· 3rdAT opened on Feb 20, 2026