trl · Issues· 284 open

[Bug] GMPOTrainer with use_liger_kernel=True trains on the GRPO objective

SFTTrainer's QLoRA fp16 downcast also hits QDoRA magnitude vector, causing silent no-op updates

GMPO: use_liger_kernel=True silently bypasses the GMPO loss; MoE aux loss and entropy bonus dropped (gaps not covered by #6095)

GRPO: vLLM importance-sampling ratio is biased when top_p/top_k/min_p truncate sampling

Fused Triton kernels for the loss head

[Tracking] Move off liger-kernel and make the memory-efficient loss the default

router_aux_loss_coef is silently ignored on MoE configs without output_router_logits

Audit the trl-internal-testing repos that have no generation script

Align tiny test model configs with their reference models

[RFC] Criteria for removing code from `trl.experimental`

DPO and KTO skip DDP gradient synchronization with use_liger_kernel=True

get_repetition_penalty_reward accepts non-positive ngram_size

Apple Silicon (MPS): create_model_from_path can segfault while loading bf16 checkpoints

[Bug] get_dataset raises ZeroDivisionError for zero-sum fractions

Support router-based multi-teacher (MOPD) distillation in DistillationTrainer