Support async in DeepEP (#4610)

Co-authored-by: Cheng Wan <cwan39@gatech.edu>
2025-03-23 13:39:56 +08:00
parent c6d549e773
commit ca75741e86
2 changed files with 25 additions and 15 deletions
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -239,6 +239,7 @@ class DeepseekV2MoE(nn.Module):
                num_local_experts=config.n_routed_experts // self.tp_size,
                hidden_size=config.hidden_size,
                params_dtype=config.torch_dtype,
+                async_finish=True,  # TODO
            )

    def forward(