A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bugfix] Fix DeepEP config for DP4TP4 (#23619)
Signed-off-by: Ming Yang <minos.future@gmail.com>
Ming Yang committed
Sep 10, 2025 at 10:37 UTC
40329496301a86044df7b57b8b6f57f5b71e701f
1 file changed
+4
-4
vllm/model_executor/layers/fused_moe/deepep_ht_prepare_finalize.py
+4
-4
@@ -49,14 +49,14 @@ class DeepEPHTPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
49
return torch.int64
50
51
def _get_dispatch_config(self) -> Optional[deep_ep.Config]:
52
- if self.dp_size not in self.available_rank_configs:
52
+ if self.num_dispatchers_ not in self.available_rank_configs:
53
return None
54
- return deep_ep.Buffer.get_dispatch_config(self.dp_size)
54
+ return deep_ep.Buffer.get_dispatch_config(self.num_dispatchers_)
55
56
def _get_combine_config(self) -> Optional[deep_ep.Config]:
57
- if self.dp_size not in self.available_rank_configs:
57
+ if self.num_dispatchers_ not in self.available_rank_configs:
58
return None
59
- return deep_ep.Buffer.get_combine_config(self.dp_size)
59
+ return deep_ep.Buffer.get_combine_config(self.num_dispatchers_)
60
61
def _do_dispatch(
62
self,