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 11, 07:30 UTC
[BugFix] Workspace allocation during profile run : DeepEPHighThroughput + DeepGEMM (#30899)
Varun Sundar Rabindranath committed
Dec 17, 2025 at 18:00 UTC
e3fc374a9a69dddb16885d810f1e28d3fdd39ebd
1 file changed
+4
-1
vllm/model_executor/layers/fused_moe/modular_kernel.py
+4
-1
@@ -795,7 +795,10 @@ class FusedMoEModularKernel(torch.nn.Module):
795
top_k,
796
global_num_experts,
797
local_num_experts,
798
- expert_tokens_meta,
798
+ # expert_tokens_meta help in allocating optimal/minimal
799
+ # amount of workspace. Mark it None, so we allocate for
800
+ # the worst-case scenario.
801
+ expert_tokens_meta=None,
802
)
803
)
804