[MoE] Migrate W4A8 CT to oracle kernel setup (#42680)
Signed-off-by: Siddharth Bedekar <bedeksid@gmail.com> Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
co-authored by
OpenAI Codex
parent
8737e4a857
commit
4438b6e7dc
+74
-177
@@ -7,12 +7,7 @@ from compressed_tensors.quantization import (
|
||||
QuantizationArgs,
|
||||
)
|
||||
|
||||
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
|
||||
from vllm import _custom_ops as ops
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
FusedMoEActivationFormat,
|
||||
FusedMoEExpertsModular,
|
||||
FusedMoeWeightScaleSupported,
|
||||
RoutedExperts,
|
||||
SharedExperts,
|
||||
@@ -20,19 +15,18 @@ from vllm.model_executor.layers.fused_moe import (
|
||||
from vllm.model_executor.layers.fused_moe.config import (
|
||||
FusedMoEConfig,
|
||||
FusedMoEQuantConfig,
|
||||
int4_w4afp8_moe_quant_config,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.oracle.w4a8 import (
|
||||
convert_to_w4a8_moe_kernel_format,
|
||||
make_w4a8_moe_kernel,
|
||||
make_w4a8_moe_quant_config,
|
||||
select_w4a8_moe_backend,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe import ( # noqa E501
|
||||
CompressedTensorsMoEMethod,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
||||
convert_bf16_scales_to_fp8,
|
||||
convert_packed_uint4b8_to_signed_int4_inplace,
|
||||
)
|
||||
from vllm.model_executor.utils import replace_parameter, set_weight_attrs
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):
|
||||
def __init__(
|
||||
@@ -56,16 +50,10 @@ class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):
|
||||
assert self.weight_quant.actorder != "group"
|
||||
assert self.group_size == 128, "Only group size 128 supported for W4A8 MoE"
|
||||
|
||||
self.disable_expert_map = False
|
||||
self.layer_name = layer_name
|
||||
|
||||
from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
||||
GroupShape,
|
||||
self.w4a8_backend, self.experts_cls = select_w4a8_moe_backend(
|
||||
config=self.moe,
|
||||
)
|
||||
|
||||
self.quant_fp8 = QuantFP8(static=False, group_shape=GroupShape.PER_TOKEN)
|
||||
|
||||
def create_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
@@ -156,152 +144,74 @@ class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):
|
||||
layer.register_parameter("w13_weight_shape", w13_weight_shape)
|
||||
set_weight_attrs(w13_weight_shape, extra_weight_attrs)
|
||||
|
||||
w13_weight_chan_scale = torch.nn.Parameter(
|
||||
torch.ones(
|
||||
num_experts,
|
||||
2 * intermediate_size_per_partition,
|
||||
dtype=torch.float32,
|
||||
),
|
||||
requires_grad=False,
|
||||
)
|
||||
layer.register_parameter("w13_weight_chan_scale", w13_weight_chan_scale)
|
||||
|
||||
w2_weight_chan_scale = torch.nn.Parameter(
|
||||
torch.ones(num_experts, hidden_size, dtype=torch.float32),
|
||||
requires_grad=False,
|
||||
)
|
||||
layer.register_parameter("w2_weight_chan_scale", w2_weight_chan_scale)
|
||||
|
||||
# don't use input scales
|
||||
layer.w13_input_scale = None
|
||||
layer.w2_input_scale = None
|
||||
|
||||
def process_weights_after_loading(self, layer):
|
||||
device = layer.w13_weight_packed.device
|
||||
|
||||
# STRIDES
|
||||
# A, C
|
||||
self.a_strides1_c_strides2 = torch.full(
|
||||
(layer.local_num_experts,),
|
||||
layer.hidden_size,
|
||||
device=device,
|
||||
dtype=torch.int64,
|
||||
)
|
||||
self.a_strides2 = torch.full(
|
||||
(layer.local_num_experts,),
|
||||
layer.intermediate_size_per_partition,
|
||||
device=device,
|
||||
dtype=torch.int64,
|
||||
)
|
||||
self.c_strides1 = torch.full(
|
||||
(layer.local_num_experts,),
|
||||
2 * layer.intermediate_size_per_partition,
|
||||
device=device,
|
||||
dtype=torch.int64,
|
||||
def process_weights_after_loading(self, layer: RoutedExperts) -> None:
|
||||
(
|
||||
w13_weight_packed,
|
||||
w2_weight_packed,
|
||||
w13_weight_scale,
|
||||
w2_weight_scale,
|
||||
w13_weight_chan_scale,
|
||||
w2_weight_chan_scale,
|
||||
b_strides1,
|
||||
b_strides2,
|
||||
) = convert_to_w4a8_moe_kernel_format(
|
||||
w13_weight_packed=layer.w13_weight_packed,
|
||||
w2_weight_packed=layer.w2_weight_packed,
|
||||
w13_weight_scale=layer.w13_weight_scale,
|
||||
w2_weight_scale=layer.w2_weight_scale,
|
||||
)
|
||||
|
||||
# S (group-wise scales)
|
||||
# sizeof(StrideS) = 16 bytes, so we need to use 2xint64 to encode it
|
||||
self.s_strides1 = torch.zeros(
|
||||
(layer.local_num_experts, 2), device=device, dtype=torch.int64
|
||||
)
|
||||
self.s_strides1[:, 0] = 2 * layer.intermediate_size_per_partition
|
||||
replace_parameter(layer, "w13_weight_packed", w13_weight_packed)
|
||||
replace_parameter(layer, "w2_weight_packed", w2_weight_packed)
|
||||
replace_parameter(layer, "w13_weight_scale", w13_weight_scale)
|
||||
replace_parameter(layer, "w2_weight_scale", w2_weight_scale)
|
||||
replace_parameter(layer, "w13_weight_chan_scale", w13_weight_chan_scale)
|
||||
replace_parameter(layer, "w2_weight_chan_scale", w2_weight_chan_scale)
|
||||
|
||||
self.s_strides2 = torch.zeros(
|
||||
(layer.local_num_experts, 2), device=device, dtype=torch.int64
|
||||
)
|
||||
self.s_strides2[:, 0] = layer.hidden_size
|
||||
self.b_strides1 = b_strides1
|
||||
self.b_strides2 = b_strides2
|
||||
|
||||
# encode and reorder weight tensors, and get the layout to pass to
|
||||
# the grouped gemm kernel. `b_strides1/2` specifies the entire layout
|
||||
convert_packed_uint4b8_to_signed_int4_inplace(layer.w13_weight_packed)
|
||||
# mirror the sync in CutlassW4A8LinearKernel; required for tp>1 correctness
|
||||
torch.accelerator.synchronize()
|
||||
w13_weight_shuffled, self.b_strides1 = (
|
||||
ops.cutlass_encode_and_reorder_int4b_grouped(layer.w13_weight_packed)
|
||||
)
|
||||
replace_parameter(layer, "w13_weight_packed", w13_weight_shuffled)
|
||||
convert_packed_uint4b8_to_signed_int4_inplace(layer.w2_weight_packed)
|
||||
# mirror the sync in CutlassW4A8LinearKernel; required for tp>1 correctness
|
||||
torch.accelerator.synchronize()
|
||||
w2_weight_shuffled, self.b_strides2 = (
|
||||
ops.cutlass_encode_and_reorder_int4b_grouped(layer.w2_weight_packed)
|
||||
)
|
||||
replace_parameter(layer, "w2_weight_packed", w2_weight_shuffled)
|
||||
self.moe_quant_config = self.get_fused_moe_quant_config(layer)
|
||||
if self.moe_quant_config is not None:
|
||||
assert self.experts_cls is not None
|
||||
self.moe_kernel = make_w4a8_moe_kernel(
|
||||
moe_quant_config=self.moe_quant_config,
|
||||
moe_config=self.moe,
|
||||
experts_cls=self.experts_cls,
|
||||
b_strides1=self.b_strides1,
|
||||
b_strides2=self.b_strides2,
|
||||
group_size=self.group_size,
|
||||
routing_tables=layer._expert_routing_tables(),
|
||||
)
|
||||
|
||||
# convert bf16 scales to (fp8_scales, channel_scales)
|
||||
w13_weight_scale, w13_weight_chan_scale = convert_bf16_scales_to_fp8(
|
||||
self.quant_fp8, layer.w13_weight_scale
|
||||
)
|
||||
w2_weight_scale, w2_weight_chan_scale = convert_bf16_scales_to_fp8(
|
||||
self.quant_fp8, layer.w2_weight_scale
|
||||
)
|
||||
|
||||
# register channel scales
|
||||
layer.register_parameter(
|
||||
"w13_weight_chan_scale",
|
||||
torch.nn.Parameter(w13_weight_chan_scale, requires_grad=False),
|
||||
)
|
||||
layer.register_parameter(
|
||||
"w2_weight_chan_scale",
|
||||
torch.nn.Parameter(w2_weight_chan_scale, requires_grad=False),
|
||||
)
|
||||
|
||||
# The scales are stored as (E, N, K // 128) but the kernel expects
|
||||
# (E, K // 128, N) in row-major format, so we need to permute the last 2 dims
|
||||
# and make it contiguous
|
||||
w13_weight_scale_packed = ops.cutlass_pack_scale_fp8(
|
||||
w13_weight_scale.permute(0, 2, 1).contiguous()
|
||||
)
|
||||
replace_parameter(layer, "w13_weight_scale", w13_weight_scale_packed)
|
||||
w2_weight_scale_packed = ops.cutlass_pack_scale_fp8(
|
||||
w2_weight_scale.permute(0, 2, 1).contiguous()
|
||||
)
|
||||
replace_parameter(layer, "w2_weight_scale", w2_weight_scale_packed)
|
||||
|
||||
def maybe_make_prepare_finalize(
|
||||
self,
|
||||
routing_tables: tuple[torch.Tensor, torch.Tensor, torch.Tensor] | None = None,
|
||||
) -> mk.FusedMoEPrepareAndFinalizeModular | None:
|
||||
return super().maybe_make_prepare_finalize(routing_tables)
|
||||
|
||||
def get_fused_moe_quant_config(
|
||||
self, layer: torch.nn.Module
|
||||
) -> FusedMoEQuantConfig | None:
|
||||
# Store quantization scales; both per-group and per-channel
|
||||
# Note we haven't specified the group size here because
|
||||
# the quant config logic assumes group-wise scaling
|
||||
# and channel-wise scaling are exclusive.
|
||||
return int4_w4afp8_moe_quant_config(
|
||||
w1_scale=layer.w13_weight_scale, # group scale
|
||||
w2_scale=layer.w2_weight_scale, # group scale
|
||||
def get_fused_moe_quant_config(self, layer: torch.nn.Module) -> FusedMoEQuantConfig:
|
||||
return make_w4a8_moe_quant_config(
|
||||
w1_scale=layer.w13_weight_scale,
|
||||
w2_scale=layer.w2_weight_scale,
|
||||
g1_alphas=layer.w13_weight_chan_scale,
|
||||
g2_alphas=layer.w2_weight_chan_scale,
|
||||
per_act_token_quant=True, # always use dynamic per-token
|
||||
per_out_ch_quant=True, # always use per-channel
|
||||
)
|
||||
|
||||
def select_gemm_impl(
|
||||
self,
|
||||
prepare_finalize: mk.FusedMoEPrepareAndFinalizeModular,
|
||||
layer: torch.nn.Module,
|
||||
) -> mk.FusedMoEExpertsModular:
|
||||
assert self.moe_quant_config is not None
|
||||
assert (
|
||||
prepare_finalize.activation_format == FusedMoEActivationFormat.Standard
|
||||
), "BatchedExperts not supported"
|
||||
|
||||
from vllm.model_executor.layers.fused_moe import CutlassExpertsW4A8Fp8
|
||||
|
||||
experts: FusedMoEExpertsModular
|
||||
|
||||
logger.debug("CutlassExpertsW4A8Fp8(%s)", self.__class__.__name__)
|
||||
experts = CutlassExpertsW4A8Fp8(
|
||||
out_dtype=self.moe.in_dtype,
|
||||
a_strides1=self.a_strides1_c_strides2,
|
||||
a_strides2=self.a_strides2,
|
||||
b_strides1=self.b_strides1,
|
||||
b_strides2=self.b_strides2,
|
||||
c_strides1=self.c_strides1,
|
||||
c_strides2=self.a_strides1_c_strides2,
|
||||
s_strides1=self.s_strides1,
|
||||
s_strides2=self.s_strides2,
|
||||
moe_config=self.moe,
|
||||
quant_config=self.moe_quant_config,
|
||||
group_size=self.group_size,
|
||||
)
|
||||
|
||||
num_dispatchers = prepare_finalize.num_dispatchers()
|
||||
self.disable_expert_map = (
|
||||
num_dispatchers > 1 or not experts.supports_expert_map()
|
||||
)
|
||||
|
||||
return experts
|
||||
|
||||
def apply(
|
||||
self,
|
||||
layer: RoutedExperts,
|
||||
@@ -311,33 +221,20 @@ class CompressedTensorsW4A8Fp8MoEMethod(CompressedTensorsMoEMethod):
|
||||
shared_experts: SharedExperts | None,
|
||||
shared_experts_input: torch.Tensor | None,
|
||||
) -> torch.Tensor:
|
||||
assert self.moe_quant_config is not None
|
||||
|
||||
from vllm.model_executor.layers.fused_moe.experts.cutlass_moe import (
|
||||
cutlass_moe_w4a8_fp8,
|
||||
)
|
||||
|
||||
return cutlass_moe_w4a8_fp8(
|
||||
x,
|
||||
layer.w13_weight_packed,
|
||||
layer.w2_weight_packed,
|
||||
topk_weights,
|
||||
topk_ids,
|
||||
moe_config=self.moe,
|
||||
quant_config=self.moe_quant_config,
|
||||
assert not self.is_monolithic
|
||||
assert self.moe_kernel is not None
|
||||
return self.moe_kernel.apply(
|
||||
hidden_states=x,
|
||||
w1=layer.w13_weight_packed,
|
||||
w2=layer.w2_weight_packed,
|
||||
topk_weights=topk_weights,
|
||||
topk_ids=topk_ids,
|
||||
activation=layer.activation,
|
||||
global_num_experts=layer.global_num_experts,
|
||||
expert_map=None if self.disable_expert_map else layer.expert_map,
|
||||
a_strides1=self.a_strides1_c_strides2,
|
||||
a_strides2=self.a_strides2,
|
||||
b_strides1=self.b_strides1,
|
||||
b_strides2=self.b_strides2,
|
||||
c_strides1=self.c_strides1,
|
||||
c_strides2=self.a_strides1_c_strides2,
|
||||
s_strides1=self.s_strides1,
|
||||
s_strides2=self.s_strides2,
|
||||
group_size=self.group_size,
|
||||
expert_map=layer.expert_map,
|
||||
apply_router_weight_on_input=layer.apply_router_weight_on_input,
|
||||
shared_experts=shared_experts,
|
||||
shared_experts_input=shared_experts_input,
|
||||
)
|
||||
|
||||
@property
|
||||
|
||||
Reference in New Issue
Block a user