Skip to content

GitLab

Commit 43fe650e authored by lixh6's avatar lixh6 Committed by zhangzbb
Browse files

[FEATURE] 接入Aiter MoE W8A8 量化模型支持

parent a3776adc
Hide whitespace changes
Inline Side-by-side
Showing with 183 additions and 77 deletions
vllm/envs.py
View file @ 43fe650e
......@@ -167,6 +167,7 @@ if TYPE_CHECKING:
VLLM_MOE_USE_DEEP_GEMM: bool = True
VLLM_USE_DEEP_GEMM_E8M0: bool = True
VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES: bool = True
VLLM_USE_AITER_MOE_W8A8: bool = True
VLLM_DEEP_GEMM_WARMUP: Literal[
"skip",
"full",
......@@ -1290,6 +1291,9 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES": lambda: bool(
int(os.getenv("VLLM_USE_DEEP_GEMM_TMA_ALIGNED_SCALES", "1"))
),
"VLLM_USE_AITER_MOE_W8A8": lambda: bool(
int(os.getenv("VLLM_USE_AITER_MOE_W8A8", "1"))
),
# DeepGemm JITs the kernels on-demand. The warmup attempts to make DeepGemm
# JIT all the required kernels before model execution so there is no
# JIT'ing in the hot-path. However, this warmup increases the engine
......
vllm/model_executor/layers/fused_moe/fused_moe.py
View file @ 43fe650e
......@@ -6,7 +6,9 @@ import functools
import json
import os
import math
import sys
import aiter
from aiter.moe import get_aiter_moe_config, aiter_moe, MoeQuantType
from collections.abc import Callable
from typing import Any, Callable, Dict, List, Optional
......@@ -1858,35 +1860,74 @@ def fused_experts_impl(
cache13 = torch.empty(M * top_k_num * max(N, K if not use_nn_moe else w2.shape[2]), device=hidden_states.device, dtype=hidden_states.dtype)
if use_int8_w8a8 or use_fp8_w8a8:
return fused_experts_impl_int8(hidden_states=hidden_states,
w1=w1,
w2=w2,
topk_weights=topk_weights,
topk_ids=topk_ids,
cache13=cache13,
inplace=inplace,
activation=activation,
apply_router_weight_on_input=apply_router_weight_on_input,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=False,
use_int4_w4a16=False,
per_channel_quant=per_channel_quant,
global_num_experts=global_num_experts,
expert_map=expert_map,
w1_scale=w1_scale,
w2_scale=w2_scale,
w1_zp=w1_zp,
w2_zp=w2_zp,
a1_scale=a1_scale,
a2_scale=a2_scale,
block_shape=block_shape,
use_nn_moe=False,
routed_scaling_factor=routed_scaling_factor,
shared_output=shared_output,
i_q=i_q,
i_s=i_s
)
if envs.VLLM_USE_AITER_MOE_W8A8==True:
K_input = hidden_states.size(1)
actual_N2 = N // 2
quant_type = MoeQuantType.W8A8
status, moe_config = get_aiter_moe_config(
M=num_tokens,
E=global_num_experts,
N1=N,
N2=actual_N2,
K=K_input,
top_k=top_k_num,
block_size=0,
dtype=hidden_states.dtype,
quant_type=quant_type,
)
output = aiter_moe(
hidden_states=hidden_states,
w1=w1,
w2=w2,
topk_weights=topk_weights,
topk_ids=topk_ids,
moe_config=moe_config,
inplace=inplace,
activation=activation,
w1_scale=w1_scale,
w2_scale=w2_scale,
w1_zp=w1_zp,
w2_zp=w2_zp,
a1_scale=a1_scale,
a2_scale=a2_scale,
block_shape=None,
global_num_experts=global_num_experts,
expert_map=expert_map,
routed_scaling_factor=routed_scaling_factor,
)
return output
else:
return fused_experts_impl_int8(hidden_states=hidden_states,
w1=w1,
w2=w2,
topk_weights=topk_weights,
topk_ids=topk_ids,
cache13=cache13,
inplace=inplace,
activation=activation,
apply_router_weight_on_input=apply_router_weight_on_input,
use_fp8_w8a8=use_fp8_w8a8,
use_int8_w8a8=use_int8_w8a8,
use_int8_w8a16=False,
use_int4_w4a16=False,
per_channel_quant=per_channel_quant,
global_num_experts=global_num_experts,
expert_map=expert_map,
w1_scale=w1_scale,
w2_scale=w2_scale,
w1_zp=w1_zp,
w2_zp=w2_zp,
a1_scale=a1_scale,
a2_scale=a2_scale,
block_shape=block_shape,
use_nn_moe=False,
routed_scaling_factor=routed_scaling_factor,
shared_output=shared_output,
i_q=i_q,
i_s=i_s
)
elif use_int4_w4a8 is True:
return fused_experts_impl_w4a8(hidden_states=hidden_states,
w1=w1,
......
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe_marlin.py
View file @ 43fe650e
......@@ -26,6 +26,14 @@ from vllm.model_executor.layers.fused_moe import (
FusedMoEPrepareAndFinalize,
FusedMoeWeightScaleSupported,
)
import aiter
from aiter.test_common import checkAllclose, perftest
from aiter.ops.shuffle import moe_layout_shuffle_gemm1, moe_layout_shuffle_gemm2
from aiter.fused_moe import fused_topk, torch_moe
from aiter import dtypes, ActivationType
from aiter.moe import get_aiter_moe_config, aiter_moe, MoeSolutionType, MoeQuantType
try:
from lmslim.layers.fused_moe.fuse_moe_int8_marlin import fused_experts_impl_int8_marlin
from lmslim.layers.fused_moe.fuse_moe_fp8_marlin import fused_experts_impl_fp8_marlin
......@@ -369,30 +377,44 @@ class CompressedTensorsW8A8Int8MarlinMoEMethod(CompressedTensorsMarlinMoEMethod)
layer.w13_input_scale = None
layer.w2_input_scale = None
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
w1_marlin_list = []
for ii in range(layer.w13_weight.shape[0]):
if not self.use_deepep:
w1_marlin_in = get_w8a8_int8_marlin_weights(layer.w13_weight[ii])
else:
#w1_marlin_in = w8a8_nt_kpack2_marlin_weight(layer.w13_weight[ii])
w1_marlin_in = weight8bit_nt_kpack2_marlin1(layer.w13_weight[ii])
w1_marlin_list.append(w1_marlin_in)
w1_marlin = torch.stack(w1_marlin_list, dim=0)
def shuffle_w8a8_gemm1(self, weight_data):
w_i8 = weight_data.to(torch.int8)
return moe_layout_shuffle_gemm1(w_i8)
del w1_marlin_list
w2_marlin_list = []
for ii in range(layer.w2_weight.shape[0]):
if not self.use_deepep:
w2_marlin_in = get_w8a8_int8_marlin_weights(layer.w2_weight[ii])
else:
#w2_marlin_in = w8a8_nt_kpack2_marlin_weight(layer.w2_weight[ii])
w2_marlin_in = weight8bit_nt_kpack2_marlin1(layer.w2_weight[ii])
w2_marlin_list.append(w2_marlin_in)
w2_marlin = torch.stack(w2_marlin_list, dim=0)
def shuffle_w8a8_gemm2(self, weight_data):
w_i8 = weight_data.to(torch.int8)
return moe_layout_shuffle_gemm2(w_i8)
layer.w13_weight = Parameter(w1_marlin, requires_grad=False)
layer.w2_weight = Parameter(w2_marlin, requires_grad=False)
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
if envs.VLLM_USE_AITER_MOE_W8A8==True:
layer.w13_weight_scale = Parameter(layer.w13_weight_scale.data, requires_grad=False)
layer.w2_weight_scale = Parameter(layer.w2_weight_scale.data, requires_grad=False)
shuffled_w13 = self.shuffle_w8a8_gemm1(layer.w13_weight)
layer.w13_weight = Parameter(shuffled_w13.view(*layer.w13_weight.shape), requires_grad=False)
shuffled_w2 = self.shuffle_w8a8_gemm2(layer.w2_weight)
layer.w2_weight = Parameter(shuffled_w2.view(*layer.w2_weight.shape), requires_grad=False)
else:
w1_marlin_list = []
for ii in range(layer.w13_weight.shape[0]):
if not self.use_deepep:
w1_marlin_in = get_w8a8_int8_marlin_weights(layer.w13_weight[ii])
else:
w1_marlin_in = w8a8_nt_kpack2_marlin_weight(layer.w13_weight[ii])
w1_marlin_list.append(w1_marlin_in)
w1_marlin = torch.stack(w1_marlin_list, dim=0)
del w1_marlin_list
w2_marlin_list = []
for ii in range(layer.w2_weight.shape[0]):
if not self.use_deepep:
w2_marlin_in = get_w8a8_int8_marlin_weights(layer.w2_weight[ii])
else:
w2_marlin_in = w8a8_nt_kpack2_marlin_weight(layer.w2_weight[ii])
w2_marlin_list.append(w2_marlin_in)
w2_marlin = torch.stack(w2_marlin_list, dim=0)
layer.w13_weight = Parameter(w1_marlin, requires_grad=False)
layer.w2_weight = Parameter(w2_marlin, requires_grad=False)
def apply(
self,
......@@ -407,31 +429,70 @@ class CompressedTensorsW8A8Int8MarlinMoEMethod(CompressedTensorsMarlinMoEMethod)
routed_scaling_factor: Optional[float] = 1.0,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
from vllm.model_executor.layers.fused_moe import fused_experts
return fused_experts_impl_int8_marlin(
hidden_states=x,
w1=layer.w13_weight,
w2=layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
activation=layer.activation,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
use_int8_w8a8=True,
per_channel_quant=True,
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,
quant_config=self.moe_quant_config,
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale,
use_nn_moe=False,
i_q=i_q,
i_s=i_s,
shared_output=shared_output,
routed_scaling_factor=routed_scaling_factor,
)
if envs.VLLM_USE_AITER_MOE_W8A8==True:
m_flat = x.view(-1, x.shape[-1])
M = m_flat.shape[0]
E = layer.w13_weight.size(0)
K = x.size(-1)
N1 = layer.w13_weight.size(1)
topk = topk_ids.size(1)
w1_input = layer.w13_weight.view(E, N1, K)
w2_input = layer.w2_weight.view(E, K, N1 // 2)
_, moe_cfg = get_aiter_moe_config(
M=M,
E=E,
N1=N1,
N2=N1 // 2,
K=K,
top_k=topk,
block_size=0,
dtype=x.dtype,
quant_type=MoeQuantType.W8A8,
)
output = aiter_moe(
hidden_states=x,
w1=w1_input,
w2=w2_input,
topk_weights=topk_weights,
topk_ids=topk_ids,
moe_config=moe_cfg,
inplace=False,
activation=getattr(layer, "activation", "silu"),
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
a1_scale=getattr(layer, "w13_input_scale", None),
a2_scale=getattr(layer, "w2_input_scale", None),
global_num_experts=E,
expert_map=getattr(layer, "expert_map", None),
routed_scaling_factor=routed_scaling_factor,
)
return output
else:
return fused_experts_impl_int8_marlin(
hidden_states=x,
w1=layer.w13_weight,
w2=layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
inplace=True,
activation=layer.activation,
apply_router_weight_on_input=layer.apply_router_weight_on_input,
use_int8_w8a8=True,
per_channel_quant=True,
global_num_experts=layer.global_num_experts,
expert_map=layer.expert_map,
quant_config=self.moe_quant_config,
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale,
use_nn_moe=False,
i_q=i_q,
i_s=i_s,
shared_output=shared_output,
routed_scaling_factor=routed_scaling_factor,
)
def select_gemm_impl(
self,
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment