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 15, 09:45 UTC
[Bugfix] Fix FusedMoEPrepareAndFinalize for cuda-disalike backends (#18178)
Signed-off-by: Mengqing Cao <cmq0113@163.com>
Mengqing Cao committed
May 15, 2025 at 14:16 UTC
70f8b967242633f4cf38a456127f102664d1da3e
1 file changed
+1
vllm/model_executor/layers/fused_moe/layer.py
+1
@@ -43,6 +43,7 @@ if current_platform.is_cuda_alike():
43
from .pplx_prepare_finalize import PplxPrepareAndFinalize
44
else:
45
fused_experts = None # type: ignore
46
+ FusedMoEPrepareAndFinalize = None # type: ignore
47
if is_rocm_aiter_moe_enabled():
48
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import ( # noqa: E501
49
rocm_aiter_biased_group_topk as grouped_topk)