[BugFix] Fix DeepGEMM over-allocating workspace (#28254)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
2025-11-10 17:01:17 -05:00
parent bf6a3d0ff5
commit 6dec9f6109
1 changed files with 1 additions and 1 deletions
--- a/vllm/model_executor/layers/fused_moe/deep_gemm_moe.py
+++ b/vllm/model_executor/layers/fused_moe/deep_gemm_moe.py
@@ -215,7 +215,7 @@ class DeepGemmExperts(mk.FusedMoEPermuteExpertsUnpermute):
        )
        assert M_sum % block_m == 0

-        workspace1 = (M_sum, max(N, K))
+        workspace1 = (M_sum, N)
        workspace2 = (M_sum, max(N // 2, K))
        output = (M, K)
        return (workspace1, workspace2, output)