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, 05:00 UTC
[moe] Use enable_chunking func (to support disabling chunking) (#29935)
Signed-off-by: Ming Yang <minos.future@gmail.com>
Ming Yang committed
Dec 18, 2025 at 01:02 UTC
8372be2828f16dd339b24d46cb6142c9d0afd004
1 file changed
+2
-2
vllm/model_executor/layers/fused_moe/modular_kernel.py
+2
-2
@@ -743,7 +743,7 @@ class FusedMoEModularKernel(torch.nn.Module):
743
1,
744
(
745
M
746
- if not self.fused_experts.supports_chunking()
746
+ if not self.fused_experts.enable_chunking()
747
else min(M, envs.VLLM_FUSED_MOE_CHUNK_SIZE)
748
),
749
)
@@ -786,7 +786,7 @@ class FusedMoEModularKernel(torch.nn.Module):
786
is_forward_context_available()
787
and get_forward_context().attn_metadata is None
788
)
789
- if is_profile_run and self.fused_experts.supports_chunking() and self.is_dp_ep:
789
+ if is_profile_run and self.fused_experts.enable_chunking() and self.is_dp_ep:
790
max_workspace_13, max_workspace_2, max_fused_out_shape = (
791
self.fused_experts.workspace_shapes(
792
envs.VLLM_FUSED_MOE_CHUNK_SIZE,