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 15, 21:00 UTC
[Misc] Print FusedMoE detail info (#13974)
Jee Jee Li committed
Feb 28, 2025 at 07:53 UTC
67fc426845d17508782a00194c8455b7b64d49f1
1 file changed
+20
vllm/model_executor/layers/fused_moe/layer.py
+20
@@ -737,3 +737,23 @@ class FusedMoE(torch.nn.Module):
737
# If we are in the row parallel case (down_proj)
738
else:
739
param_data[expert_id] = loaded_weight
740
+
741
+ def extra_repr(self) -> str:
742
+
743
+ s = (
744
+ f"global_num_experts={self.global_num_experts}, "
745
+ f"local_num_experts={self.local_num_experts}, "
746
+ f"top_k={self.top_k}, "
747
+ f"intermediate_size_per_partition={self.intermediate_size_per_partition}, " # noqa: E501
748
+ f"tp_size={self.tp_size},\n"
749
+ f"ep_size={self.ep_size}, "
750
+ f"reduce_results={self.reduce_results}, "
751
+ f"renormalize={self.renormalize}, "
752
+ f"use_grouped_topk={self.use_grouped_topk}")
753
+
754
+ if self.use_grouped_topk:
755
+ s += f", num_expert_group={self.num_expert_group}, topk_group={self.topk_group}" # noqa: E501
756
+
757
+ s += f", scoring_func='{self.scoring_func}', activation='{self.activation}'" # noqa: E501
758
+
759
+ return s