EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Fix enable flashinfer mxfp4 moe bf16 check (#8950)

Browse Source
This commit is contained in:
Xiaoyu Zhang
2025-08-08 13:52:09 +08:00
committed by GitHub
parent 39fd178831
commit 76915d68a8
1 changed files with 9 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
python/sglang/srt/server_args.py
View File

@@ -476,8 +476,15 @@ class ServerArgs:
self.attention_backend == "trtllm_mha" self.attention_backend == "trtllm_mha"
or self.attention_backend == "triton" or self.attention_backend == "triton"
) )
quantization_config = getattr(
self.get_hf_config(), "quantization_config", None
)
is_mxfp4_quant_format = (
quantization_config is not None
and quantization_config.get("quant_method") == "mxfp4"
)
if is_sm100_supported(): if is_sm100_supported() and is_mxfp4_quant_format:
self.enable_flashinfer_mxfp4_moe = True self.enable_flashinfer_mxfp4_moe = True
self.enable_triton_kernel_moe = False self.enable_triton_kernel_moe = False
else: else:
@@ -485,13 +492,7 @@ class ServerArgs:
self.disable_hybrid_swa_memory = True self.disable_hybrid_swa_memory = True
quantization_config = getattr( if is_mxfp4_quant_format:
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 # use bf16 for mxfp4 triton kernels
self.dtype = "bfloat16" self.dtype = "bfloat16"