[Feature]: serve gpt-oss BF16 weights

[RalphMao]

🚀 The feature, motivation and pitch

It would be great if vLLM can also support serving GPT-OSS BF16 weights (either upcasted, or finetuned), like this one: https://huggingface.co/unsloth/gpt-oss-20b-BF16/discussions/1

Many people have the needs to finetune or tweak GPT-OSS, but they won't be able to quantize it back to MXFP4 due to the lack of QAT capability

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[Opdoop]

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

[narankhuut]

Is there any updates for serving gpt-oss BF16 weights?

[RayaneKimo]

@Opdoop Check this link : TensorRT-Model-Optimizer
You can use the : convert_oai_mxfp4_weight_only.py script to convert from BF16 into MXFP4 and let me know if it works for you. In the meantime, we still waiting to serve the BF16 weights.

[yiakwy-xpu-ml-framework-team]

@Opdoop this script I added in day in Aug 11 : openai/gpt-oss#108 to create the MXFP4 weight from BF16, much simpler and convient than NV modelopt tools

[geraldstanje]

hi, how to convert mxfp4 format to bf16? so i can run gpt oss 20b on nvidia l40s?