Support mxfp4 for GPT-OSS (#8843)

Co-authored-by: Co-author fzyzcjy <ch271828n@outlook.com>
Co-authored-by: fzyzcjy <5236035+fzyzcjy@users.noreply.github.com>
Co-authored-by: zhuofan1123 <zhuofanl@nvidia.com>
Co-authored-by: liz-badada <jinyanc@nvidia.com>
Co-authored-by: xutizhou <xutingz@nvidia.com>
Co-authored-by: linhu-nv <linhu@nvidia.com>

python/sglang/srt/server_args.py

@@ -464,6 +464,16 @@ class ServerArgs:
             self.enable_triton_kernel_moe = True
             self.disable_hybrid_swa_memory = True
 
+            quantization_config = getattr(
+                self.get_hf_config(), "quantization_config", None
+            )
+            if (
+                quantization_config is not None
+                and quantization_config.get("quant_method") == "mxfp4"
+            ):
+                # use bf16 for mxfp4 triton kernels
+                self.dtype = "bfloat16"
+
         # Set page size
         if self.page_size is None:
             self.page_size = 1