All issues

Search results

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

Status: Open.

#7220 In huggingface/trl;

· Saheb opened on Sep 15, 2026

AsyncGRPOTrainer: loss normalization across gradient accumulation and complete-rollout admission

Status: Open.

#7206 In huggingface/trl;

· adithya-s-k opened on Sep 14, 2026

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

Status: Open.

#7202 In huggingface/trl;

· albertvillanova opened on Sep 14, 2026

vLLM encoder cache is not reset after weight sync

Status: Open.

#7198 In huggingface/trl;

· rmaguado opened on Sep 13, 2026

Online DPO: Idefics2 image batches fail in vision tensor duplication and scoring preparation

Status: Open.

#7189 In huggingface/trl;

· DaoyuanLi2816 opened on Sep 12, 2026

[RFC] Criteria for removing code from trl.experimental

Status: Open.

#7182 In huggingface/trl;

· albertvillanova opened on Sep 11, 2026

DPO/KTO training goes NaN with the default SDPA backend on torch 2.13 (cuDNN attention); clean with enable_cudnn_sdp(False) or eager

Status: Open.

Bug

#7181 In huggingface/trl;

· qgallouedec opened on Sep 11, 2026

[GRPO] Pad per minibatch instead of per generation batch

Status: Open.

#7179 In huggingface/trl;

· qgallouedec opened on Sep 10, 2026

Where the GPU runner hours go in the Tests workflow

🔧 maintenanceCI, tooling, tests, or repo upkeepCI, tooling, tests, or repo upkeep

Status: Open.

#7169 In huggingface/trl;

· albertvillanova opened on Sep 10, 2026

Align tiny test model configs with their reference models

Status: Open.

#7137 In huggingface/trl;

· albertvillanova opened on Sep 9, 2026

Agree on an explicit criterion for raising a dependency version floor

Status: Open.

#7116 In huggingface/trl;

· albertvillanova opened on Sep 8, 2026

AsyncGRPO FSDP2 weight sync hangs: interleaved full_tensor() and is_checkpoint_format=True

Status: Open.

#7103 In huggingface/trl;

· sfc-gh-truwase opened on Sep 8, 2026