[MoE Refactor] Integrate Naive Prepare Finalize into MK (#32567)
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:
co-authored by
Robert Shaw
amirkl94
parent
6d86fde09c
commit
5a93b9162b
@@ -15,9 +15,6 @@ from vllm.model_executor.layers.fused_moe.config import (
|
||||
FusedMoEParallelConfig,
|
||||
RoutingMethodType,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_prepare_finalize import ( # noqa: E501
|
||||
create_flashinfer_prepare_finalize,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
||||
QuantKey,
|
||||
kNvfp4Dynamic,
|
||||
@@ -42,7 +39,6 @@ __all__ = [
|
||||
"is_flashinfer_fp4_cutlass_moe_available",
|
||||
"is_flashinfer_fp4_cutedsl_moe_available",
|
||||
"reorder_w1w3_to_w3w1",
|
||||
"build_flashinfer_fp4_cutlass_moe_prepare_finalize",
|
||||
]
|
||||
|
||||
#
|
||||
@@ -163,17 +159,6 @@ def reorder_w1w3_to_w3w1(
|
||||
)
|
||||
|
||||
|
||||
def build_flashinfer_fp4_cutlass_moe_prepare_finalize(
|
||||
moe: FusedMoEConfig,
|
||||
) -> mk.FusedMoEPrepareAndFinalize:
|
||||
"""Create a FlashInfer CUTLASS fused-MoE prepare finalize kernel"""
|
||||
use_dp = moe.moe_parallel_config.dp_size > 1
|
||||
enable_alltoallv = moe.moe_parallel_config.all2all_backend == "flashinfer_all2allv"
|
||||
return create_flashinfer_prepare_finalize(
|
||||
use_dp=use_dp, use_nvfp4=True, enable_alltoallv=enable_alltoallv
|
||||
)
|
||||
|
||||
|
||||
def prepare_static_weights_for_trtllm_fp4_moe(
|
||||
# args_dequant,
|
||||
# args,
|
||||
|
||||
@@ -4,15 +4,8 @@ from enum import Enum
|
||||
|
||||
import torch
|
||||
|
||||
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
|
||||
from vllm import envs
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe.config import (
|
||||
FusedMoEConfig,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_prepare_finalize import ( # noqa: E501
|
||||
create_flashinfer_prepare_finalize,
|
||||
)
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.utils.math_utils import round_up
|
||||
|
||||
@@ -163,18 +156,6 @@ def make_fp8_moe_alpha_scales_for_fi(
|
||||
return g1_alphas, g2_alphas
|
||||
|
||||
|
||||
def build_flashinfer_fp8_cutlass_moe_prepare_finalize(
|
||||
moe: FusedMoEConfig | None, use_deepseek_fp8_block_scale: bool = False
|
||||
) -> mk.FusedMoEPrepareAndFinalize:
|
||||
"""Create a FlashInfer CUTLASS fused-MoE prepare finalize kernel"""
|
||||
use_dp = moe.moe_parallel_config.dp_size > 1 if moe is not None else False
|
||||
# Propagate block-scale flag so prepare/finalize can skip act quantization
|
||||
# and inform the kernel to consume per-block weight scales.
|
||||
return create_flashinfer_prepare_finalize(
|
||||
use_dp, use_deepseek_fp8_block_scale=use_deepseek_fp8_block_scale
|
||||
)
|
||||
|
||||
|
||||
def get_flashinfer_moe_backend() -> FlashinferMoeBackend:
|
||||
backend_map = {
|
||||
"throughput": FlashinferMoeBackend.CUTLASS,
|
||||
|
||||
Reference in New Issue
Block a user