biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Kernels] Split up fused_moe/layer.py, isolate more modular kernel code (#28064)

Browse Source
This commit is contained in:
bnellnm
2025-11-11 09:29:02 -05:00
committed by GitHub
parent fa1970201d
commit a1448b4b69
10 changed files with 1064 additions and 948 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
tests/kernels/moe/modular_kernel_tools/mk_objects.py
View File

@@ -6,6 +6,10 @@ import torch
# Fused experts and PrepareFinalize imports
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
from vllm.model_executor.layers.fused_moe import TritonExperts
from vllm.model_executor.layers.fused_moe.all2all_utils import (
maybe_make_prepare_finalize,
)
from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
BatchedDeepGemmExperts,
)
@@ -21,7 +25,6 @@ from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
BatchedTritonExperts,
NaiveBatchedExperts,
)
from vllm.model_executor.layers.fused_moe.layer import FusedMoEMethodBase, TritonExperts
from vllm.model_executor.layers.fused_moe.prepare_finalize import (
MoEPrepareAndFinalizeNoEP,
)
@@ -399,9 +402,7 @@ def make_prepare_finalize(
quant_config: FusedMoEQuantConfig,
) -> mk.FusedMoEPrepareAndFinalize:
if backend != "naive" and backend is not None:
prepare_finalize = FusedMoEMethodBase._maybe_make_prepare_finalize(
moe, quant_config
)
prepare_finalize = maybe_make_prepare_finalize(moe, quant_config)
assert prepare_finalize is not None
return prepare_finalize
elif prepare_finalize_type == FlashInferCutlassMoEPrepareAndFinalize: