Issues

• Update the dependency version floors under the agreed criterion

  huggingface/trl#7644 · Filter by author albertvillanova opened on Oct 12, 2026
  

• LoRA on MoE models with shared experts does not train the routed experts with peft main

  huggingface/trl#7637 · Filter by author albertvillanova opened on Oct 12, 2026
  3

  

• DPOTrainer crashes on the first step when use_weighting=True is combined with a loss_type containing "sft"

  huggingface/trl#7630 · Filter by author Sergasgr opened on Oct 11, 2026
  3

• support training decision ("System One") models with SFTTrainer

  huggingface/trl#7623 · Filter by author rycerzes opened on Oct 10, 2026

• RFC: a use_kernels switch, then kernels on by default, then packing on by default

  huggingface/trl#7621 · Filter by author qgallouedec opened on Oct 9, 2026
  4

• SDFT and SDPO build the teacher in fp32 with the default attention when the model is passed as an instance

  huggingface/trl#7618 · Filter by author ningmao-hlyz opened on Oct 9, 2026
  4

• Add ABKD as a loss in DistillationTrainer

  huggingface/trl#7617 · Filter by author ghwang-s opened on Oct 9, 2026

• ORPO log1mexp: finite loss but NaN gradients near zero log-probability

  huggingface/trl#7608 · Filter by author taking-lying-flat opened on Oct 9, 2026
  7

• Proposing to remove SSDTrainer

  huggingface/trl#7596 · Filter by author qgallouedec opened on Oct 8, 2026
  2

• Support Transformer Engine FP8 training through Accelerate in TRL

  huggingface/trl#7592 · Filter by author sbhavani opened on Oct 8, 2026
  2

• CI emits peft RuntimeWarning: The following rank_pattern keys did not match any targeted module and were ignored

  huggingface/trl#7587 · Filter by author albertvillanova opened on Oct 8, 2026
  2

  

• Test real objects instead of hand-built stand-ins

  huggingface/trl#7565 · Filter by author albertvillanova opened on Oct 7, 2026
  

• Some modules fail to import when an optional dependency is missing

  huggingface/trl#7560 · Filter by author albertvillanova opened on Oct 7, 2026
  

• IWOPDTrainer can skip the mandatory vLLM weight sync when resuming from a checkpoint

  huggingface/trl#7544 · Filter by author sahil-mangla opened on Oct 6, 2026
  1

• SFTTrainer and RewardTrainer modify the caller's peft_config when chat_template_path adds tokens

  huggingface/trl#7526 · Filter by author Vedant-Agarwal opened on Oct 5, 2026
  2

• Add score centering to GRPO for vLLM training-inference mismatch

  huggingface/trl#7520 · Filter by author vladbataev opened on Oct 4, 2026
  1

• OnlineDPOTrainer documents processing_class as optional but rejects omission

  huggingface/trl#7513 · Filter by author DaoyuanLi2816 opened on Oct 3, 2026

• Clarify raw held-out dataset evaluation for CPO and ORPO

  huggingface/trl#7512 · Filter by author DaoyuanLi2816 opened on Oct 3, 2026
  1

• BEMACallback fails at train start with ordinary PEFT LoRA

  huggingface/trl#7511 · Filter by author DaoyuanLi2816 opened on Oct 3, 2026
  4

• AsyncGRPOTrainer: support eval_dataset for validation reward

  ✨ enhancement
  huggingface/trl#7509 · Filter by author lewtun opened on Oct 2, 2026
  3

  

• Async GRPO does not train Gemma 4's tool-calling turns

  huggingface/trl#7503 · Filter by author albertvillanova opened on Oct 2, 2026
  

• Make the upkeep of stored chat templates more robust

  huggingface/trl#7501 · Filter by author albertvillanova opened on Oct 2, 2026
  

• Revisit how the tiny models track chat template revisions

  huggingface/trl#7496 · Filter by author albertvillanova opened on Oct 2, 2026
  

• AsyncGRPOTrainer and AsyncDistillationTrainer log the wrong epoch with the epoch-driven stop

  🐛 bug
  huggingface/trl#7482 · Filter by author lewtun opened on Oct 2, 2026
  5

  

• BEMACallback breaks when any parameter is frozen

  huggingface/trl#7480 · Filter by author Shr3yash opened on Oct 1, 2026
  3