[NVIDIA] Add SM100 Flashinfer Cutlass MoE fp8 backend (#22357)

Signed-off-by: Amir Klein <203507526+amirkl94@users.noreply.github.com>
This commit is contained in:
amirkl94
2025-08-19 18:01:53 -04:00
committed by GitHub
parent 21dce80ea9
commit a38b8af4c3
6 changed files with 613 additions and 139 deletions
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from enum import Enum
from typing import Any, Callable, Optional, Union
import torch
@@ -27,8 +26,11 @@ from vllm.model_executor.layers.quantization.utils.flashinfer_fp4_moe import (
build_flashinfer_fp4_cutlass_moe_prepare_finalize, reorder_w1w3_to_w3w1,
select_nvfp4_gemm_impl)
from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
apply_flashinfer_per_tensor_scale_fp8, register_moe_scaling_factors,
rotate_flashinfer_fp8_moe_weights, swap_w13_to_w31)
FlashinferMoeBackend, apply_flashinfer_per_tensor_scale_fp8,
build_flashinfer_fp8_cutlass_moe_prepare_finalize,
flashinfer_cutlass_moe_fp8, get_flashinfer_moe_backend,
register_moe_scaling_factors, rotate_flashinfer_fp8_moe_weights,
select_cutlass_fp8_gemm_impl, swap_w13_to_w31)
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import (
apply_fp4_marlin_linear, is_fp4_marlin_supported,
prepare_fp4_layer_for_marlin, prepare_moe_fp4_layer_for_marlin)
@@ -49,11 +51,6 @@ QUANT_ALGOS = ["FP8", "NVFP4"]
KV_CACHE_QUANT_ALGOS = ["FP8"]
class FlashinferMoeBackend(Enum):
TENSORRT_LLM = "TensorRT-LLM"
CUTLASS = "CUTLASS"
class ModelOptFp8Config(QuantizationConfig):
"""Config class for ModelOpt FP8."""
@@ -179,7 +176,7 @@ class ModelOptFp8Config(QuantizationConfig):
elif isinstance(layer, Attention):
return ModelOptFp8KVCacheMethod(self)
elif isinstance(layer, FusedMoE):
return ModelOptFp8MoEMethod(self, layer.moe_config)
return ModelOptFp8MoEMethod(self, layer)
return None
@@ -278,18 +275,49 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
def __init__(
self,
quant_config: ModelOptFp8Config,
moe: FusedMoEConfig,
layer: torch.nn.Module,
) -> None:
super().__init__(moe)
super().__init__(layer.moe_config)
self.layer = layer
self.quant_config = quant_config
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
cutlass_fp8_supported)
self.cutlass_fp8_supported = cutlass_fp8_supported()
self.flashinfer_moe_enabled = False
self.flashinfer_moe_backend: Optional[FlashinferMoeBackend] = None
self.fused_experts: Optional[
mk.FusedMoEModularKernel] = None # type: ignore
if envs.VLLM_USE_FLASHINFER_MOE_FP8 and has_flashinfer_moe():
self.flashinfer_moe_backend = get_flashinfer_moe_backend()
logger.info_once(
"Using FlashInfer MoE FP8 kernels for ModelOptFp8MoEMethod.")
self.flashinfer_moe_enabled = True
f"Using FlashInfer {self.flashinfer_moe_backend.value} kernels"
)
def maybe_make_prepare_finalize(
self,
moe: FusedMoEConfig,
) -> Optional[mk.FusedMoEPrepareAndFinalize]:
if self.fused_experts is not None or \
self.flashinfer_moe_backend != FlashinferMoeBackend.CUTLASS:
return super().maybe_make_prepare_finalize(moe)
prepare_finalize = build_flashinfer_fp8_cutlass_moe_prepare_finalize(
moe,
layer=self.layer,
)
logger.debug_once("%s", prepare_finalize.__class__.__name__)
return prepare_finalize
def select_gemm_impl(
self,
prepare_finalize: mk.FusedMoEPrepareAndFinalize,
moe: FusedMoEConfig,
) -> mk.FusedMoEPermuteExpertsUnpermute:
experts = select_cutlass_fp8_gemm_impl(
moe,
self.layer,
)
logger.debug_once("Using %s", experts.__class__.__name__)
return experts
def create_weights(
self,
@@ -433,11 +461,12 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
layer.w2_input_scale = Parameter(layer.w2_input_scale.max(),
requires_grad=False)
if self.flashinfer_moe_enabled:
if self.flashinfer_moe_backend is not None:
layer.w13_weight.data = swap_w13_to_w31(layer.w13_weight.data)
rotate_flashinfer_fp8_moe_weights(layer.w13_weight,
layer.w2_weight)
register_moe_scaling_factors(layer)
if self.flashinfer_moe_backend == FlashinferMoeBackend.TENSORRT_LLM:
rotate_flashinfer_fp8_moe_weights(layer.w13_weight,
layer.w2_weight)
def apply(
self,
@@ -461,14 +490,13 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
logical_to_physical_map: Optional[torch.Tensor] = None,
logical_replica_count: Optional[torch.Tensor] = None,
) -> torch.Tensor:
assert self.fused_experts is None
if enable_eplb:
raise NotImplementedError(
"EPLB not supported for `ModelOptFp8MoEMethod` yet.")
if self.flashinfer_moe_enabled:
assert activation == 'silu'
if self.flashinfer_moe_backend == FlashinferMoeBackend.TENSORRT_LLM:
assert activation == 'silu', (
f"Expected 'silu' activation but got {activation}")
assert not renormalize
return apply_flashinfer_per_tensor_scale_fp8(
layer=layer,
@@ -495,6 +523,36 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
e_score_correction_bias=e_score_correction_bias,
indices_type=self.topk_indices_dtype,
)
if self.flashinfer_moe_backend == FlashinferMoeBackend.CUTLASS:
assert not renormalize
assert activation == 'silu', (
f"Expected 'silu' activation but got {activation}")
if self.fused_experts is not None:
return self.fused_experts(
x,
layer.w13_weight,
layer.w2_weight,
topk_weights,
topk_ids,
inplace=False,
activation=activation,
global_num_experts=global_num_experts,
expert_map=expert_map,
apply_router_weight_on_input=apply_router_weight_on_input,
)
else:
return flashinfer_cutlass_moe_fp8(
x,
layer,
topk_weights,
topk_ids,
inplace=False,
activation=activation,
global_num_experts=global_num_experts,
expert_map=expert_map,
apply_router_weight_on_input=apply_router_weight_on_input,
)
from vllm.model_executor.layers.fused_moe.fused_moe import (
fused_experts)
return fused_experts(
@@ -951,20 +1009,10 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
self.flashinfer_moe_backend = None
if self.allow_flashinfer:
flashinfer_moe_backend = envs.VLLM_FLASHINFER_MOE_BACKEND
if flashinfer_moe_backend == "throughput":
self.flashinfer_moe_backend = FlashinferMoeBackend.CUTLASS
logger.info_once("Using FlashInfer CUTLASS kernels for "
"ModelOptNvFp4FusedMoE.")
elif flashinfer_moe_backend == "latency":
self.flashinfer_moe_backend = FlashinferMoeBackend.TENSORRT_LLM
logger.info_once("Using FlashInfer TensorRT-LLM kernels for "
"ModelOptNvFp4FusedMoE.")
else:
allowed_backends = ["throughput", "latency"]
raise ValueError(
f"Unknown flashinfer moe backend: {flashinfer_moe_backend}"
f" expected one of {allowed_backends}")
self.flashinfer_moe_backend = get_flashinfer_moe_backend()
logger.info_once(
f"Using FlashInfer {self.flashinfer_moe_backend.value} kernels"
" for ModelOptNvFp4FusedMoE.")
def maybe_make_prepare_finalize(
self,