[model] add gpt oss

[hiyouga]

Apply LoRA Fine-Tuning on GPT-OSS model in 3 steps

1. Install LlamaFactory and transformers

git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory
pip install -e ".[torch,metrics]" --no-build-isolation
pip install "transformers==4.55.0"

2. Train GPT-OSS on a single GPU (> 44GB) (multi-GPU is also supported)

llamafactory-cli train examples/train_lora/gpt_lora_sft.yaml

3. Merge the LoRA weight into the base model

llamafactory-cli export --model_name_or_path openai/gpt-oss-20b --adapter_name_or_path saves/gpt-20b/lora/sft --export_dir gpt_merged

(Optional) Chat with the fine-tuned model

llamafactory-cli chat --model_name_or_path gpt_merged --template gpt --skip_special_tokens False

Full fine-tuning recipes

See #8837

Use Web UI to fine-tune the model:

[yuimo]

what is the weight precision after finetune? is it still mxfp4 for the moe layer?

[ziheng0924]

When will full-parameter fine-tuning be supported?

[hiyouga]

@ziheng0924 full finetuning is supported

[hiyouga]

@yuimo the lora weights will be fp32 format

[PROoshio]

VLLM推理什么时候支持呢？

[BenjaminBossan]

Hi, I just saw this PR to support gpt-oss. IIUC, with this recipe, only the nn.Linear layers are being targeted. To target the MoE layers, you'd have to use a new PEFT feature, namely LoraConfig(target_parameters=[...]). The OpenAI cookbook has an example of that.

Generally, targeting just the MoE layers may be fine, but the majority of parameters reside in the MoE layers (90% for 20b), so users may want to target those too. On the other hand, this will be much more expensive memory-wise, so there is a trade-off here. If you have questions about the new PEFT feature, LMK.

[hiyouga]

@BenjaminBossan Sure, I agree with you. I'm excited to dive into exploring these new PEFT features. Thank you for pointing it out.

[RalphMao]

VLLM推理什么时候支持呢？

vLLM doesn't support bf16 ckpt yet

vllm-project/vllm#22380

[Pikachu1412]

加载oss 120b模型时CPU内存异常增长，导致oom

[KosmoCHE]

what is the weight precision after finetune? is it still mxfp4 for the moe layer?

it will be bf16, if triton < 3.4.0

[Opdoop]

After finetuning the gpt-oss model, I have a BF16 weight. Is there any tool to create the MXFP4 weight from BF16 weight?

[liuqianchao]

+1, is there any method to convert a trained SFT model in bf16 to mxfp4, or to directly train using native mxfp4?

[bobzhang208]

微调gpt-oss-20b显存占用大概是什么样子呢，在A100上占用大概为50，单是为什么4*3090 24G会oom

[WeiminWu2000]

微调gpt-oss-20b看起来好像不支持 linger-kernel，input context length比较长，内存会爆掉，请问有解决办法来支持linger-kernel吗

[Imbernoulli]

请问gpt-oss-120b全参数微调最少要多少卡？

[Lei-Tin]

Does this allow us to train with the reasoning content within the prompt? Or do we have to process the prompt to allow the gpt template to take in reasoning content as well?