Support mxfp4 for GPT-OSS

[Ying1123]

No description provided.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[zhyncs]

Thanks to the NVIDIA Solution Architect Team for their great work! @zhuofan1123 @liz-badada @xutizhou @linhu-nv

[Byeong-Chan]

@Ying1123

Thank you to code works!

Is it ok about MxFp4Config -> Mxfp4Config?

[EduardDurech]

? I'm using meta-llama/Llama-3.2-3B-Instruct and not quantized

    engine    = sgl.Engine(model_path=a.model, tp_size=int(os.getenv("TP")))
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/sglang/python/sglang/api.py", line 44, in Engine
    from sglang.srt.entrypoints.engine import Engine
  File "/sglang/python/sglang/srt/entrypoints/engine.py", line 42, in <module>
    from sglang.srt.managers.data_parallel_controller import (
  File "/sglang/python/sglang/srt/managers/data_parallel_controller.py", line 38, in <module>
    from sglang.srt.managers.scheduler import run_scheduler_process
  File "/sglang/python/sglang/srt/managers/scheduler.py", line 37, in <module>
    from sglang.srt.configs.model_config import ModelConfig
  File "/sglang/python/sglang/srt/configs/model_config.py", line 31, in <module>
    from sglang.srt.layers.quantization import QUANTIZATION_METHODS
  File "/sglang/python/sglang/srt/layers/quantization/__init__.py", line 69, in <module>
    from sglang.srt.layers.quantization.mxfp4 import Mxfp4Config
  File "/sglang/python/sglang/srt/layers/quantization/mxfp4.py", line 36, in <module>
    from flashinfer import (
ImportError: cannot import name 'mxfp8_quantize' from 'flashinfer' (/usr/local/lib/python3.12/dist-packages/flashinfer/__init__.py). Did you mean: 'fp4_quantize'?

We need to rebuild all our libraries I assume