Issues

• GOLD: ULD loss and gradients scale with gradient_accumulation_steps

  huggingface/trl#8343 · Filter by author ayushkokande opened on Nov 13, 2026

• Vision collators never append EOS for standard (non-conversational) data in SFT, DPO and KTO

  huggingface/trl#8342 · Filter by author ayushkokande opened on Nov 13, 2026

• SFTTrainer silently ignores formatting_func output for messages/prompt-completion datasets or a custom dataset_text_field

  huggingface/trl#8341 · Filter by author ayushkokande opened on Nov 13, 2026

• LogCompletionsCallback and WeaveCallback don't split prompts across processes, so every row is logged once per rank

  huggingface/trl#8340 · Filter by author ayushkokande opened on Nov 13, 2026

• CI fails intermittently with latest and dev dependencies for continuous batching: torch.AcceleratorError: CUDA error: device-side assert triggered

  huggingface/trl#8339 · Filter by author albertvillanova opened on Nov 13, 2026
  

• [Tracking] Support the transformers cache rename (past_key_values/cache_params -> cache)

  huggingface/trl#8331 · Filter by author albertvillanova opened on Nov 13, 2026
  

• DPOTrainer and RewardTrainer: train_sampling_strategy="group_by_length" always fails (no input_ids column)

  huggingface/trl#8327 · Filter by author himanshushukla12 opened on Nov 13, 2026
  1

• SFT on Gemma3 with Flash Attention trains image tokens with causal attention

  huggingface/trl#8325 · Filter by author albertvillanova opened on Nov 13, 2026
  1

  

• AsyncDistillationTrainer loads students that do not support Flash Attention with a Flash Attention kernel

  huggingface/trl#8324 · Filter by author albertvillanova opened on Nov 13, 2026
  1

  

• CI fails for chat template drift check: HuggingFaceTB/SmolVLM2-500M-Video-Instruct ships a chat template not stored in trl/chat_templates/

  huggingface/trl#8316 · Filter by author albertvillanova opened on Nov 13, 2026
  1

  

• SFT with assistant_only_loss trains the generation prompt's <think>\n on DeepSeek-R1-Distill

  huggingface/trl#8311 · Filter by author albertvillanova opened on Nov 13, 2026
  

• SFT with assistant_only_loss trains the generation prompt's <think>\n on DeepSeek-R1

  huggingface/trl#8310 · Filter by author albertvillanova opened on Nov 13, 2026
  

• DPOTrainer: DPO metrics lose the dataset name when eval_dataset is a dict

  huggingface/trl#8307 · Filter by author AgrawalSujay2006 opened on Nov 13, 2026
  4

• SFT truncation filter retains samples whose only unmasked label is removed by causal shifting

  huggingface/trl#8296 · Filter by author utkarshxdev opened on Nov 13, 2026
  1

• [Tracking] Update the DeepSeek-V3 and DeepSeek-R1-Distill Hub chat templates for transformers 5.23

  huggingface/trl#8280 · Filter by author albertvillanova opened on Nov 12, 2026
  1

  

• SFT with assistant_only_loss trains the generation prompt's \n on Qwen2.5

  huggingface/trl#8278 · Filter by author albertvillanova opened on Nov 12, 2026
  

• CI fails for chat template drift check: meta-models/Muse-Spark-Open-816B-A42B ships a chat template not stored in trl/chat_templates/

  huggingface/trl#8273 · Filter by author albertvillanova opened on Nov 12, 2026
  1

  

• The tiny model revision overrides in tests/conftest.py don't reach the scripts that tests call in-process, which pass revision="main"

  huggingface/trl#8263 · Filter by author albertvillanova opened on Nov 12, 2026
  

• CI fails for latest TRL release with dev dependencies for GPT-OSS and DeepSeek-V3: ValueError: You cannot perform fine-tuning on purely quantized models

  huggingface/trl#8259 · Filter by author albertvillanova opened on Nov 12, 2026
  1

  

• CI fails for latest TRL release with dev dependencies for FLOPs without head_dim: AssertionError: assert 10456320 == 19900416

  huggingface/trl#8258 · Filter by author albertvillanova opened on Nov 12, 2026
  1

  

• GOLDTrainer: no pad_token fallback for plain tokenizers

  huggingface/trl#8255 · Filter by author catherinelee274 opened on Nov 11, 2026

• KTOTrainer: train_sampling_strategy="group_by_length" always fails (no input_ids column)

  huggingface/trl#8254 · Filter by author catherinelee274 opened on Nov 11, 2026

• DistillationTrainer + FSDP2: teacher stays on CPU, step 1 fails with "Expected all tensors to be on the same device"

  huggingface/trl#8253 · Filter by author catherinelee274 opened on Nov 11, 2026
  1

• SFT evaluation fails when compute_loss returns a tensor requiring gradients

  huggingface/trl#8246 · Filter by author PalashSharma20 opened on Nov 10, 2026

• [Tracking] Support more model types in compute_flops_per_token

  huggingface/trl#8237 · Filter by author qgallouedec opened on Nov 10, 2026
  5