Search results

Loss calculation of <code>RewardTrainer</code> may be inaccurate when performing gradient accumulation?

⏳ needs more infoAdditional information or clarification is required to proceedAdditional information or clarification is required to proceed

🐛 bugSomething isn't workingSomething isn't working

Status: Open.

#4104 In huggingface/trl;

· jue-jue-zi opened on Sep 18, 2025

DeepSpeed Zero 3 + LoRA 32B model still OOM on 8*H100 (each GPU load one model?)

⚡accelerateRelated to accelerateRelated to accelerate

🏋 GRPORelated to GRPORelated to GRPO

🚀 deepspeedRelated to deepspeedRelated to deepspeed

Status: Open.

#4100 In huggingface/trl;

· Aaronhuang-778 opened on Sep 17, 2025

SFTTrainer fails with DFT loss during evaluation

🏋 SFTRelated to SFTRelated to SFT

🐛 bugSomething isn't workingSomething isn't working

Status: Open.

#4096 In huggingface/trl;

· solenetarride opened on Sep 16, 2025

Logging behaviour when steps_per_generation &gt; gradient_accumulation_steps in GRPOTrainer

🏋 GRPORelated to GRPORelated to GRPO

Status: Open.

#4086 In huggingface/trl;

· vatsal-kr opened on Sep 15, 2025

prompt+completion training vs messages training for VLM

⚡ PEFTRelated to PEFTRelated to PEFT

❓ questionSeeking clarification or more informationSeeking clarification or more information

Status: Open.

#4077 In huggingface/trl;

· weathon opened on Sep 13, 2025

Possible bug in tutorial

❓ questionSeeking clarification or more informationSeeking clarification or more information

🐛 bugSomething isn't workingSomething isn't working

📚 documentationImprovements or additions to documentationImprovements or additions to documentation

Status: Open.

#4071 In huggingface/trl;

· GabrieleGiudic opened on Sep 12, 2025

SFTraine compute_loss after trl 0.22.2 works uncorrect with logits_to_keep

🏋 SFTRelated to SFTRelated to SFT

🐛 bugSomething isn't workingSomething isn't working

Status: Open.

#4066 In huggingface/trl;

· BogdanDidenko opened on Sep 11, 2025

DPOTrainer: pick vision path by processing object type; always pass a tokenizer to chat template

✨ enhancementNew feature or requestNew feature or request

🏋 DPORelated to DPORelated to DPO

Status: Open.

#4062 In huggingface/trl;

· ginkyenglee opened on Sep 11, 2025

GRPO on InternVL3.5 does not support dynamic vision tokens

🏋 GRPORelated to GRPORelated to GRPO

🐛 bugSomething isn't workingSomething isn't working

Status: Open.

#4061 In huggingface/trl;

· sanchit97 opened on Sep 11, 2025

Whether to support GFPO training？

✨ enhancementNew feature or requestNew feature or request

🏋 PPORelated to PPORelated to PPO

Status: Open.

#4044 In huggingface/trl;

· Kangdi123 opened on Sep 9, 2025

why kl = nan when grpo train?

❓ questionSeeking clarification or more informationSeeking clarification or more information

🏋 GRPORelated to GRPORelated to GRPO

Status: Open.

#4040 In huggingface/trl;

· uilstong opened on Sep 9, 2025

SFTTrainer with PEFT model

⚡ PEFTRelated to PEFTRelated to PEFT

🏋 SFTRelated to SFTRelated to SFT

Status: Open.

#4029 In huggingface/trl;

· lylaiyy opened on Sep 8, 2025