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 14, 05:30 UTC
[Bugfix][CPU] Add placeholder to avoid import errors when using fused_moe ops on platforms without triton (#25137)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Li, Jiang committed
Sep 19, 2025 at 15:41 UTC
825fdb11add30237e7f592f1a132d3913cd632ec
1 file changed
+9
vllm/model_executor/layers/fused_moe/__init__.py
+9
@@ -78,3 +78,12 @@ if HAS_TRITON:
78
"TritonOrDeepGemmExperts",
79
"BatchedTritonOrDeepGemmExperts",
80
]
81
+else:
82
+ # Some model classes directly use the custom ops. Add placeholders
83
+ # to avoid import errors.
84
+ def _raise_exception(method: str):
85
+ raise NotImplementedError(
86
+ f"{method} is not implemented as lack of triton.")
87
+
88
+ fused_topk = lambda *args, **kwargs: _raise_exception("fused_topk")
89
+ fused_experts = lambda *args, **kwargs: _raise_exception("fused_experts")