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

[MoE Refactor] Integrate Naive Prepare Finalize into MK (#32567)

Browse Source 
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Signed-off-by: Amir Klein <203507526+amirkl94@users.noreply.github.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: amirkl94 <203507526+amirkl94@users.noreply.github.com>
This commit is contained in:
Robert Shaw
2026-01-26 20:28:02 -05:00
committed by GitHub
parent 6d86fde09c
commit 5a93b9162b
46 changed files with 1018 additions and 876 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
tests/kernels/moe/modular_kernel_tools/mk_objects.py
View File

@@ -7,9 +7,6 @@ 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,
)
@@ -255,13 +252,12 @@ if has_pplx():
)
if has_flashinfer_cutlass_fused_moe() and current_platform.has_device_capability(100):
from vllm.model_executor.layers.fused_moe.flashinfer_a2a_prepare_finalize import ( # noqa: E501
FlashInferCutlassMoEPrepareAndFinalize,
)
from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe import (
FlashInferExperts,
)
from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_prepare_finalize import ( # noqa: E501
FlashInferCutlassMoEPrepareAndFinalize,
create_flashinfer_prepare_finalize,
)
register_prepare_and_finalize(
FlashInferCutlassMoEPrepareAndFinalize,
@@ -429,24 +425,6 @@ if cutlass_fp4_supported() or has_flashinfer_cutlass_fused_moe():
]
def make_prepare_finalize(
prepare_finalize_type: mk.FusedMoEPrepareAndFinalize,
backend: str | None,
moe: FusedMoEConfig,
quant_config: FusedMoEQuantConfig,
) -> mk.FusedMoEPrepareAndFinalize:
if backend != "naive" and backend is not None:
prepare_finalize = maybe_make_prepare_finalize(moe, quant_config)
assert prepare_finalize is not None
return prepare_finalize
elif prepare_finalize_type == FlashInferCutlassMoEPrepareAndFinalize:
return create_flashinfer_prepare_finalize(
use_dp=moe.moe_parallel_config.dp_size > 1
)
else:
return MoEPrepareAndFinalizeNoEP()
def _slice(rank: int, num_local_experts: int, t: torch.Tensor) -> torch.Tensor:
s = rank * num_local_experts
e = s + num_local_experts