All issues

Search results

Accounting error in FLOPS calculation

Status: Open.

#6708 In huggingface/trl;

· Harindhar10 opened on Aug 11, 2026

GRPO: a NaN training step is invisible in log_history, so a corrupted run reports a clean loss curve

Status: Open.

#6702 In huggingface/trl;

· behroozazarkhalili opened on Aug 10, 2026

GRPOTrainer silently no-ops (empty optimizer) when continuing from a PeftModel loaded with is_trainable=False and beta != 0

Status: Open.

#6700 In huggingface/trl;

· Meredith2328 opened on Aug 10, 2026

Qwen3.5 VLM Liger GRPO drops multimodal token type IDs

Status: Open.

#6697 In huggingface/trl;

· DimensionSTP opened on Aug 10, 2026

SFT chunked NLL stalls under ZeRO-3 with unequal rank chunk counts

Status: Open.

#6696 In huggingface/trl;

· DimensionSTP opened on Aug 10, 2026

Support status and numerical safety of GRPO + vLLM + bf16 LoRA for long multi-turn tool rollouts in TRL 1.9.2

Status: Open.

#6688 In huggingface/trl;

· yanan1116 opened on Aug 8, 2026

SFTTrainer's default chunked_nll path fails on models whose output has no past_key_values

Status: Open.

#6685 In huggingface/trl;

· Hakureirm opened on Aug 8, 2026

AsyncGRPOTrainer hardcodes dtype=torch.float32, so a 4B OOMs on an 80 GB card

Status: Open.

#6684 In huggingface/trl;

· Mandark-droid opened on Aug 8, 2026

examples/scripts/openenv/opencode.py is broken against current OpenEnv@main - every rollout scores "unscorable"

Status: Open.

#6683 In huggingface/trl;

· Mandark-droid opened on Aug 8, 2026

[Bug] trl vllm-serve fails with tensor_parallel_size > 1: Cannot re-initialize CUDA in forked subprocess

Status: Open.

#6679 In huggingface/trl;

· linfeng-du opened on Aug 8, 2026

pack_dataset with strategy="wrapped" silently duplicates data across map batches

Status: Open.

#6669 In huggingface/trl;

· SuryanshSS1011 opened on Aug 6, 2026

SFTTrainer reports loss 0.0 when every label in the batch is masked

Status: Open.

#6668 In huggingface/trl;

· behroozazarkhalili opened on Aug 5, 2026