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 13, 13:15 UTC
[BugFix] Fix DeepGEMM over-allocating workspace (#28254)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Lucas Wilkinson committed
Nov 10, 2025 at 17:01 UTC
6dec9f61098786690b4ca2140682dbafb849f8d9
1 file changed
+1
-1
vllm/model_executor/layers/fused_moe/deep_gemm_moe.py
+1
-1
@@ -215,7 +215,7 @@ class DeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
215
)
216
assert M_sum % block_m == 0
217
218
- workspace1 = (M_sum, max(N, K))
218
+ workspace1 = (M_sum, N)
219
workspace2 = (M_sum, max(N // 2, K))
220
output = (M, K)
221
return (workspace1, workspace2, output)