chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Feat][NVFP4] Enable NVFP4 MoE for Qwen series models (eg. Qwen3-Next) #13761 (#13761)

Browse Source 
Co-authored-by: Kaixi Hou <kaixih@nvidia.com>
This commit is contained in:
Sam
2025-11-27 08:53:45 +08:00
committed by GitHub
parent 231df4b0d4
commit 91e8dc371a
4 changed files with 77 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/server_args.py
View File

@@ -1202,7 +1202,7 @@ class ServerArgs:
if self.quantization is None and quant_method is not None:
self.quantization = quant_method
if (
self.quantization == "fp8"
self.quantization in ("fp8", "modelopt_fp4")
and self.moe_a2a_backend == "none"
and self.moe_runner_backend == "auto"
):
@@ -1229,7 +1229,7 @@ class ServerArgs:
if self.quantization is None and quant_method is not None:
self.quantization = quant_method
if (
self.quantization == "fp8"
(self.quantization == "fp8" or self.quantization == "modelopt_fp4")
and self.moe_a2a_backend == "none"
and self.moe_runner_backend == "auto"
):