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 6, 05:45 UTC
[MoE] Move FlashInfer CuteDSL experts into fused_moe/experts/ (#37759)
Signed-off-by: Robert Shaw <robertgshaw2@gmail.com>
Robert Shaw committed
Mar 21, 2026 at 19:15 UTC
6b2fa3a76204a8e1f64db1bc1962f384bd11c005
3 files changed
+2
-2
tests/kernels/moe/test_cutedsl_moe.py
+1
-1
@@ -17,7 +17,7 @@ from flashinfer import fp4_quantize
17
from torch.nn import functional as F
18
19
from vllm.model_executor.layers.activation import SiluAndMul
20
-from vllm.model_executor.layers.fused_moe.flashinfer_cutedsl_moe import (
20
+from vllm.model_executor.layers.fused_moe.experts.flashinfer_cutedsl_moe import (
21
flashinfer_cutedsl_moe_masked,
22
)
23
from vllm.utils.flashinfer import (
vllm/model_executor/layers/fused_moe/experts/flashinfer_cutedsl_moe.py
renamed
vllm/model_executor/layers/fused_moe/oracle/nvfp4.py
+1
-1
@@ -86,7 +86,7 @@ def backend_to_kernel_cls(
86
return [FlashInferExperts]
87
88
elif backend == NvFp4MoeBackend.FLASHINFER_CUTEDSL:
89
- from vllm.model_executor.layers.fused_moe.flashinfer_cutedsl_moe import (
89
+ from vllm.model_executor.layers.fused_moe.experts.flashinfer_cutedsl_moe import ( # noqa: E501
90
FlashInferCuteDSLExperts,
91
)
92