Search results

CI fails with minimum versions: AssertionError: gradient mismatch on model.embed_tokens.weight

Status: Open.

#5874 In huggingface/trl;

· albertvillanova opened on May 28, 2026

AsyncGRPO: make NCCL weight sync robust on PCIe-only GPUs (auto fallback from P2P/SHM)

Status: Open.

#5865 In huggingface/trl;

· rycerzes opened on May 27, 2026

[AsyncGRPO] Loading from checkpoint recomputes all vLLM past requests

Status: Open.

#5863 In huggingface/trl;

· ggcr opened on May 27, 2026

[AsyncGRPO] aiohttp limits to 100 reqs when max_inflight_tasks > 100

Status: Open.

#5847 In huggingface/trl;

· ggcr opened on May 26, 2026

Add missing AsyncGRPOTrainer features present in GRPOTrainer

Status: Open.

#5831 In huggingface/trl;

· rycerzes opened on May 25, 2026

GRPOTrainer silently uses near-greedy decoding when temperature=1.0 (transformers >= 4.50 + Qwen2.5)

Status: Open.

#5783 In huggingface/trl;

· timwei0801 opened on May 18, 2026

Optimize old_per_token_logps recomputation in GRPOTrainer: per-rollout window check instead of static modulo

Status: Open.

#5770 In huggingface/trl;

· wengeezhang opened on May 14, 2026

Weighted SFT

Status: Open.

#5761 In huggingface/trl;

· RyanJFriedrich opened on May 13, 2026

[AsyncGRPO] AsyncGRPO add_response_schema not gated on self.tools (diverges from GRPOTrainer)

Status: Open.

#5742 In huggingface/trl;

· aazizyan opened on May 9, 2026

Can gold_trainer support privileged context?

Status: Open.

#5741 In huggingface/trl;

· zouYC2021 opened on May 9, 2026

Expose response_schema= argument on add_response_schema and GRPOTrainer/GRPOConfig

Status: Open.

#5724 In huggingface/trl;

· rycerzes opened on May 7, 2026

🛣️ Path to 30B MoE long-context SFT training

⚡accelerateRelated to accelerateRelated to accelerate

🏋 SFTRelated to SFTRelated to SFT

Status: Open.

#5713 In huggingface/trl;

· AmineDiro opened on May 6, 2026