[MoE] Move cutlass moe to fused_moe/experts/ (#40574)

Signed-off-by: Jackmin801 <ongjackm@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Jackmin801
2026-04-24 06:05:49 +00:00
committed by GitHub
co-authored by Claude
parent 9744b699ba
commit 079a4cf399
14 changed files with 20 additions and 18 deletions
@@ -73,15 +73,15 @@ __all__ = [
if HAS_TRITON:
# import to register the custom ops
from vllm.model_executor.layers.fused_moe.cutlass_moe import (
from vllm.model_executor.layers.fused_moe.experts.batched_deep_gemm_moe import (
BatchedDeepGemmExperts,
)
from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import (
CutlassBatchedExpertsFp8,
CutlassExpertsFp8,
CutlassExpertsW4A8Fp8,
cutlass_moe_w4a8_fp8,
)
from vllm.model_executor.layers.fused_moe.experts.batched_deep_gemm_moe import (
BatchedDeepGemmExperts,
)
from vllm.model_executor.layers.fused_moe.experts.deep_gemm_moe import (
DeepGemmExperts,
)
@@ -173,7 +173,7 @@ def backend_to_kernel_cls(
return [TritonOrCutlassExperts]
elif backend == Fp8MoeBackend.BATCHED_VLLM_CUTLASS:
from vllm.model_executor.layers.fused_moe.cutlass_moe import (
from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import (
CutlassBatchedExpertsFp8,
)
@@ -107,7 +107,7 @@ def backend_to_kernel_cls(
return [FlashInferCuteDSLBatchedExperts]
elif backend == NvFp4MoeBackend.VLLM_CUTLASS:
from vllm.model_executor.layers.fused_moe.cutlass_moe import (
from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import (
CutlassExpertsFp4,
)
@@ -10,7 +10,7 @@ from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEQuantConfig,
)
from vllm.model_executor.layers.fused_moe.cutlass_moe import CutlassExpertsFp8
from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import CutlassExpertsFp8
from vllm.model_executor.layers.fused_moe.fallback import FallbackExperts
from vllm.model_executor.layers.fused_moe.fused_moe import TritonExperts
from vllm.platforms import current_platform
@@ -14,7 +14,7 @@ from vllm.model_executor.layers.fused_moe.config import (
FusedMoEQuantConfig,
mxfp4_moe_quant_config,
)
from vllm.model_executor.layers.fused_moe.cutlass_moe import (
from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import (
CutlassExpertsMxfp4,
)
from vllm.model_executor.layers.fused_moe.fused_marlin_moe import (
@@ -149,7 +149,7 @@ class CompressedTensorsW4A4Mxfp4MoEMethod(CompressedTensorsMoEMethod):
if self.use_cutlass_mxfp4:
# Swizzle weight scales from flat checkpoint layout [E, N, K//32]
# to CUTLASS tiled layout [E, numMTiles*numKTiles*512].
from vllm.model_executor.layers.fused_moe.cutlass_moe import (
from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import (
swizzle_mxfp4_scales,
)
@@ -315,7 +315,7 @@ class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):
)
assert self.moe_quant_config is not None
from vllm.model_executor.layers.fused_moe.cutlass_moe import (
from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import (
cutlass_moe_w4a8_fp8,
)