FlashInfer NVFP4 MoE with EP & 2-stream shared expert (#7327)

Co-authored-by: JieXin Liang <Alcanderian@users.noreply.github.com>
Co-authored-by: alcanderian <alcanderian@gmail.com>
This commit is contained in:
Trevor Morris
2025-06-22 13:38:47 -07:00
committed by GitHub
co-authored by JieXin Liang alcanderian
parent edc21cc8ae
commit 5962e70d8d
6 changed files with 182 additions and 20 deletions
@@ -1295,6 +1295,9 @@ class DeepEPMoE(EPMoE):
def get_moe_impl_class():
if global_server_args_dict["enable_deepep_moe"]:
return DeepEPMoE
if global_server_args_dict["enable_flashinfer_moe"]:
# Must come before EPMoE because FusedMoE also supports enable_ep_moe
return FusedMoE
if global_server_args_dict["enable_ep_moe"]:
return EPMoE
return FusedMoE