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

Support mixing cutedsl and deepgemm backend (#11807)

Browse Source
This commit is contained in:
fzyzcjy
2025-10-21 07:38:35 +08:00
committed by GitHub
parent 184a4df697
commit 0917c5da8c
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/layers/moe/ep_moe/layer.py
View File

@@ -191,11 +191,15 @@ class DeepEPMoE(FusedMoE):
assert deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM and self.use_fp8_w8a8
return self.forward_deepgemm_contiguous(dispatch_output)
elif DispatchOutputChecker.format_is_deepep_ll(dispatch_output):
if get_moe_runner_backend().is_flashinfer_cutedsl():
if (
get_moe_runner_backend().is_flashinfer_cutedsl()
and self.quant_config.get_name() == "modelopt_fp4"
):
return self.forward_flashinfer_cutedsl(
dispatch_output, down_gemm_overlap_args=down_gemm_overlap_args
)
assert deep_gemm_wrapper.ENABLE_JIT_DEEPGEMM and self.use_fp8_w8a8
assert down_gemm_overlap_args is None
return self.forward_deepgemm_masked(dispatch_output)
else:
raise ValueError(