xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-03-19 22:17:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

Optimize memory.

Browse Source 
The tensor workspace13 is used only with one shape.

Signed-off-by: Andrey Khalyavin <halyavin@yandex-team.ru>
This commit is contained in:
Andrey Khalyavin 2025-12-16 16:11:56 +03:00
parent 676db55eec
commit a4af2e7b3a
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/model_executor/layers/fused_moe/batched_deep_gemm_moe.py
View File

@ -312,7 +312,7 @@ class BatchedDeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
num_dispatchers = self.num_dispatchers
num_experts = local_num_experts
max_num_tokens = M if self.max_num_tokens is None else self.max_num_tokens
workspace13 = (num_experts, max_num_tokens * num_dispatchers, max(K, N))
workspace13 = (num_experts, max_num_tokens * num_dispatchers, N)
workspace2 = (num_experts, max_num_tokens * num_dispatchers, (N // 2))
output = (num_experts, max_num_tokens * num_dispatchers, K)
return (workspace13, workspace2, output)