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 2, 22:00 UTC
fix: MoE model using shared routed experts crashes on AMD GPUs (#42373)
Signed-off-by: weizhou.lan@daocloud.io <weizhou.lan@daocloud.io>
weizhoublue committed
May 25, 2026 at 12:03 UTC
6cbe448eed751824d608faf9078ef84724d621c1
1 file changed
+2
-2
vllm/model_executor/layers/fused_moe/router/aiter_shared_routed_fused_moe_router.py
+2
-2
@@ -71,7 +71,7 @@ class AiterSharedRoutedFusedMoERouter(BaseRouter):
71
"Number of tokens mismatch"
72
)
73
74
- from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
74
+ from vllm.model_executor.layers.fused_moe.experts.rocm_aiter_moe import (
75
aiter_topK_meta_data,
76
)
77
@@ -125,7 +125,7 @@ class AiterSharedRoutedFusedMoERouter(BaseRouter):
125
)
126
127
if aiter_topK_meta_data is not None:
128
- from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
128
+ from vllm.model_executor.layers.fused_moe.experts.rocm_aiter_moe import (
129
inject_shared_expert_weights,
130
)
131