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

ignore the deepgemm check when the model weight with nvfp4 and moe ba… (#12782)

Browse Source
This commit is contained in:
Rain Jiang
2025-11-06 15:27:28 -08:00
committed by GitHub
parent fb314d7bb1
commit a119363f08
1 changed files with 9 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/moe/ep_moe/layer.py
+9 -1
View File
@@ -112,8 +112,16 @@ class DeepEPMoE(FusedMoE):
self.deepep_mode = get_deepep_mode() self.deepep_mode = get_deepep_mode()
if self.deepep_mode.enable_low_latency() and not _is_npu: if (
self.deepep_mode.enable_low_latency()
and not _is_npu
and not (
get_moe_runner_backend().is_flashinfer_cutedsl()
and self.quant_config.get_name() == "modelopt_fp4"
)
):
# NPU supports low_latency deepep without deepgemm # NPU supports low_latency deepep without deepgemm
# FP4 quantization with flashinfer_cutedsl also supports low_latency deepep without deepgemm
assert ( assert (
deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM
), f"DeepEP {self.deepep_mode} mode requires deep_gemm" ), f"DeepEP {self.deepep_mode} mode requires deep_gemm"