kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-14 06:27:45 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity
77f8e43592
TensorRT-LLMs/tensorrt_llm/quantization/layers.py
Quanfeng Li 839aad4d6e
fix: Add missing parameter for WeightOnlyQuantRowLinear module (#2768) 
Signed-off-by: Quanfeng Li <liquanfeng7@foxmail.com>
2025-03-31 16:20:30 +08:00

3316 lines
135 KiB
Python
Raw Blame History

# SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
from typing import Optional
import numpy as np
import tensorrt as trt
import torch
from .._common import default_net, precision
from .._utils import is_same_dtype, str_dtype_to_torch, trt_dtype_to_torch
from ..functional import (ACT2FN, AllReduceFusionOp, AllReduceParams,
AttentionMaskType, PositionEmbeddingType,
RotaryScalingType, Tensor, allgather, allreduce, cast,
concat, constant, div, embedding, gemm_allreduce,
generate_alibi_slopes, gpt_attention, matmul, mul,
shape, slice, softmax, split, where)
from ..layers import MropeParams, SpecDecodingParams
from ..layers.embedding import Embedding
from ..layers.linear import Linear, RowLinear
from ..module import Module
from ..parameter import Parameter
# isort: off
from .functional import (
block_double_dequantize, dequantize, dynamic_quantize, fp4_gemm,
quantize_to_fp4_tensor, fp8_rowwise_gemm, fp8_rowwise_rms_norm,
postprocess_fp8_rowwise, postprocess_weight_only,
postprocess_weight_only_groupwise, quantize, quantize_fp8_per_token,
quantize_per_token, quantize_tensor, validate_group_size, smooth_quant_gemm,
smooth_quant_layer_norm, smooth_quant_rms_norm,
weight_only_groupwise_quant_matmul, weight_only_quant_matmul,
qserve_gemm_per_group, qserve_gemm_per_channel)
# isort: on
from .mode import GroupwiseQuantAlgo, QuantMode
class Quantize(Module):
"""
Quantize Layer
For per-tensor mode, the scaling factor is a scalar.
For per-channel mode, the scaling factor is a vector.
"""
def __init__(
self,
output_dtype: str = 'int8',
scaling_factor_dtype: str = 'float32',
in_channels: int = -1,
axis=-1,
) -> None:
super().__init__()
self.scaling_factor = Parameter(shape=(in_channels, ) if axis != -1 else
(),
dtype=scaling_factor_dtype)
self.output_dtype = output_dtype
self.axis = axis
def forward(self, x):
return quantize(x, self.scaling_factor.value, self.output_dtype,
self.axis)
class QuantizePerToken(Module):
"""
Quantize Per Token and compute dynamic scales for SmoothQuant
"""
def forward(self, x):
return quantize_per_token(x)
class Dequantize(Module):
"""
Dequantize Layer.
"""
def __init__(self, axis: int = -1) -> None:
super().__init__()
self.scaling_factor = Parameter(shape=(), dtype='float32')
self.axis = axis
def forward(self, input):
return dequantize(input, self.scaling_factor.value, self.axis)
class SmoothQuantLinear(Linear):
def __init__(self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
gather_output=True,
quant_mode=QuantMode(0),
prefer_managed_weight=True):
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=gather_output,
prefer_managed_weight=prefer_managed_weight)
if not quant_mode.has_act_and_weight_quant():
raise ValueError(
"SmoothQuant Linear has to have act+weight quantization mode set"
)
weights_dtype = dtype
if quant_mode.has_act_and_weight_quant():
weights_dtype = "int8"
self.weight = Parameter(shape=(self.out_features, self.in_features),
dtype=weights_dtype,
prefer_managed=self.prefer_managed_weight)
if quant_mode.has_act_and_weight_quant():
scale_shape = (1, self.out_features
) if quant_mode.has_per_channel_scaling() else (1, 1)
self.per_channel_scale = Parameter(shape=scale_shape,
dtype="float32")
if quant_mode.has_act_static_scaling():
self.act_scale = Parameter(shape=(1, 1), dtype="float32")
self.quant_mode = quant_mode
def forward(self, x, lora_runtime_params=None):
assert lora_runtime_params is None, "lora is not supported on SmoothQuantLinear now"
if self.quant_mode.has_act_static_scaling():
per_token_scale = self.act_scale.value
else:
# If we are in SmoothQuant with dynamic activation scaling,
# input x has to be a tuple of int8 tensor and fp32 scaling factors
x, per_token_scale = x
x = smooth_quant_gemm(x, self.weight.value, per_token_scale,
self.per_channel_scale.value,
self.quant_mode.has_per_token_dynamic_scaling(),
self.quant_mode.has_per_channel_scaling(),
self.dtype)
if self.bias is not None:
x = x + self.bias.value
if self.gather_output and self.tp_size > 1 and self.tp_group is not None:
# [dim0, local_dim] -> [dim0 * tp_size, local_dim] --> [dim0, local_dim * tp_size]
x = allgather(x, self.tp_group, gather_dim=1)
return x
SmoothQuantColumnLinear = SmoothQuantLinear
class SmoothQuantRowLinear(RowLinear):
def __init__(
self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
prefer_managed_weight=True,
):
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
prefer_managed_weight=prefer_managed_weight)
if not quant_mode.has_act_and_weight_quant():
raise ValueError(
"SmoothQuant Linear has to have act+weight quantization mode set"
)
weights_dtype = dtype
if quant_mode.has_act_and_weight_quant():
weights_dtype = "int8"
self.weight = Parameter(shape=(self.out_features, self.in_features),
dtype=weights_dtype,
prefer_managed=self.prefer_managed_weight)
self.smoother = Parameter(shape=(1, self.in_features), dtype="float32")
if quant_mode.has_act_and_weight_quant():
scale_shape = (1, self.out_features
) if quant_mode.has_per_channel_scaling() else (1, 1)
self.per_channel_scale = Parameter(shape=scale_shape,
dtype="float32")
if quant_mode.has_act_static_scaling():
self.act_scale = Parameter(shape=(1, 1), dtype="float32")
self.quant_mode = quant_mode
def forward(self, x, lora_runtime_params=None, all_reduce_params=None):
assert lora_runtime_params is None, "lora is not supported on SmoothQuantRowLinear now"
if self.quant_mode.has_act_static_scaling():
per_token_scale = self.act_scale.value
else:
x, per_token_scale = x
x = smooth_quant_gemm(x, self.weight.value, per_token_scale,
self.per_channel_scale.value,
self.quant_mode.has_per_token_dynamic_scaling(),
self.quant_mode.has_per_channel_scaling(),
self.dtype)
if self.tp_size > 1 and self.tp_group is not None:
need_bias = self.bias is not None
fuse_bias_into_all_reduce = need_bias and (
all_reduce_params
is not None) and (all_reduce_params.fusion_op
== AllReduceFusionOp.RESIDUAL_RMS_NORM)
if fuse_bias_into_all_reduce:
all_reduce_params.bias = self.bias.value
x = allreduce(x, self.tp_group, all_reduce_params=all_reduce_params)
if need_bias and not fuse_bias_into_all_reduce:
x = x + self.bias.value
return x
if self.bias is not None:
x = x + self.bias.value
return x
class SmoothQuantLayerNorm(Module):
def __init__(
self,
normalized_shape,
eps=1e-05,
elementwise_affine=True,
bias=True,
dtype=None,
quant_mode=QuantMode(0),
):
super().__init__()
if isinstance(normalized_shape, int):
normalized_shape = (normalized_shape, )
if not quant_mode.has_act_and_weight_quant():
raise ValueError(
"SmoothQuant layer norm has to have some quantization mode set")
self.normalized_shape = tuple(normalized_shape)
self.elementwise_affine = elementwise_affine
if self.elementwise_affine:
self.weight = Parameter(shape=self.normalized_shape, dtype=dtype)
if bias:
self.bias = Parameter(shape=self.normalized_shape, dtype=dtype)
else:
self.register_parameter('bias', None)
else:
self.register_parameter('weight', None)
self.register_parameter('bias', None)
self.eps = eps
self.dtype = dtype
self.quant_mode = quant_mode
if self.quant_mode.has_act_and_weight_quant():
self.scale_to_int = Parameter(shape=(1, ), dtype=dtype)
else:
self.register_parameter('scale_to_int', None)
def forward(self, x):
weight = None if self.weight is None else self.weight.value
bias = None if self.bias is None else self.bias.value
scale = None if self.scale_to_int is None else self.scale_to_int.value
return smooth_quant_layer_norm(
x,
self.normalized_shape,
weight,
bias,
scale,
self.eps,
dynamic_act_scaling=self.quant_mode.has_per_token_dynamic_scaling())
class SmoothQuantRmsNorm(Module):
def __init__(
self,
normalized_shape,
eps=1e-06,
elementwise_affine=True,
dtype=None,
quant_mode=QuantMode(0),
bias=False,
clamp_val=None,
):
super().__init__()
if isinstance(normalized_shape, int):
normalized_shape = (normalized_shape, )
if not quant_mode.has_act_and_weight_quant():
raise ValueError(
"SmoothQuant Rms norm has to have some quantization mode set")
self.normalized_shape = tuple(normalized_shape)
self.elementwise_affine = elementwise_affine
if self.elementwise_affine:
self.weight = Parameter(shape=self.normalized_shape, dtype=dtype)
else:
self.register_parameter('weight', None)
if bias:
self.bias = Parameter(shape=self.normalized_shape, dtype=dtype)
else:
self.register_parameter('bias', None)
if clamp_val:
if not (isinstance(clamp_val, list) and len(clamp_val) == 2):
raise ValueError(f'unsupported clamp_val {clamp_val}')
self.clamp_val = Parameter(np.array(clamp_val, dtype=np.float32),
dtype='float32',
is_buffer=True)
else:
self.register_parameter('clamp_val', None)
self.eps = eps
self.dtype = dtype
self.quant_mode = quant_mode
if self.quant_mode.has_act_and_weight_quant():
self.scale_to_int = Parameter(shape=(1, ), dtype=dtype)
else:
self.register_parameter('scale_to_int', None)
def forward(self, x):
weight = None if self.weight is None else self.weight.value
bias = None if self.bias is None else self.bias.value
scale = None if self.scale_to_int is None else self.scale_to_int.value
clamp_val = None if self.clamp_val is None else self.clamp_val.value
return smooth_quant_rms_norm(
x,
self.normalized_shape,
weight,
bias,
scale,
clamp_val,
self.eps,
dynamic_act_scaling=self.quant_mode.has_per_token_dynamic_scaling())
class QServeW4A8Linear(Linear):
def __init__(self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
gather_output=True,
quant_mode=QuantMode(0)):
assert dtype == "float16" # Currently the kernel only supports float16 output
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=gather_output)
self.quant_mode = quant_mode
assert self.quant_mode.is_qserve_w4a8()
# Only support g128 now.
if self.quant_mode.has_per_group_scaling():
self.group_size = 128
else:
self.group_size = -1
self.weight = Parameter(shape=(self.out_features,
self.in_features // 2),
dtype="int8")
self.s1_scales = Parameter(shape=(self.out_features, ), dtype="float16")
if self.group_size == -1:
self.s1_szeros = Parameter(shape=(self.out_features, ),
dtype="float16")
else:
self.s2_scales = Parameter(
shape=(self.in_features // self.group_size, self.out_features),
dtype="int8")
self.s2_szeros = Parameter(
shape=(self.in_features // self.group_size, self.out_features),
dtype="int8")
def forward(self, x):
if self.group_size == -1:
x, per_token_scale, per_token_sum = x
x = qserve_gemm_per_channel(x, per_token_scale, per_token_sum,
self.weight.value, self.s1_scales.value,
self.s1_szeros.value)
else:
x, per_token_scale = x
x = qserve_gemm_per_group(x, per_token_scale, self.weight.value,
self.s1_scales.value,
self.s2_scales.value,
self.s2_szeros.value, self.group_size)
if self.bias is not None:
x = x + self.bias.value
if self.gather_output and self.tp_size > 1 and self.tp_group is not None:
# [dim0, local_dim] -> [dim0 * tp_size, local_dim] --> [dim0, local_dim * tp_size]
x = allgather(x, self.tp_group, gather_dim=1)
return x
QServeW4A8ColumnLinear = QServeW4A8Linear
class QServeW4A8RowLinear(RowLinear):
def __init__(
self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
):
assert dtype == "float16" # Currently the kernel only supports float16 output
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size)
self.quant_mode = quant_mode
assert self.quant_mode.is_qserve_w4a8()
# Only supports 128g now.
if self.quant_mode.has_per_group_scaling():
self.group_size = 128
else:
self.group_size = -1
self.weight = Parameter(shape=(self.out_features,
self.in_features // 2),
dtype="int8")
self.s1_scales = Parameter(shape=(self.out_features, ), dtype="float16")
if self.group_size == -1:
self.s1_szeros = Parameter(shape=(self.out_features, ),
dtype="float16")
else:
self.s2_scales = Parameter(
shape=(self.in_features // self.group_size, self.out_features),
dtype="int8")
self.s2_szeros = Parameter(
shape=(self.in_features // self.group_size, self.out_features),
dtype="int8")
def forward(self, x, all_reduce_params=None):
if self.group_size == -1:
x, per_token_scale, per_token_sum = x
x = qserve_gemm_per_channel(x, per_token_scale, per_token_sum,
self.weight.value, self.s1_scales.value,
self.s1_szeros.value)
else:
x, per_token_scale = x
x = qserve_gemm_per_group(x, per_token_scale, self.weight.value,
self.s1_scales.value,
self.s2_scales.value,
self.s2_szeros.value, self.group_size)
if self.tp_size > 1 and self.tp_group is not None:
need_bias = self.bias is not None
fuse_bias_into_all_reduce = need_bias and (
all_reduce_params
is not None) and (all_reduce_params.fusion_op
== AllReduceFusionOp.RESIDUAL_RMS_NORM)
if fuse_bias_into_all_reduce:
all_reduce_params.bias = self.bias.value
x = allreduce(x, self.tp_group, all_reduce_params=all_reduce_params)
if need_bias and not fuse_bias_into_all_reduce:
x = x + self.bias.value
return x
if self.bias is not None:
x = x + self.bias.value
return x
class Fp8RowwiseRmsNorm(Module):
def __init__(
self,
normalized_shape,
eps=1e-06,
elementwise_affine=True,
dtype=None,
quant_mode=QuantMode(0),
bias=False,
clamp_val=None,
):
super().__init__()
if isinstance(normalized_shape, int):
normalized_shape = (normalized_shape, )
if not quant_mode.has_fp8_rowwise():
raise ValueError(
"Fp8 Rowwise Rms norm has to have some quantization mode set")
self.normalized_shape = tuple(normalized_shape)
self.elementwise_affine = elementwise_affine
if self.elementwise_affine:
self.weight = Parameter(shape=self.normalized_shape, dtype=dtype)
else:
self.register_parameter('weight', None)
if bias:
self.bias = Parameter(shape=self.normalized_shape, dtype=dtype)
else:
self.register_parameter('bias', None)
if clamp_val:
if not (isinstance(clamp_val, list) and len(clamp_val) == 2):
raise ValueError(f'unsupported clamp_val {clamp_val}')
self.clamp_val = Parameter(np.array(clamp_val, dtype=np.float32),
dtype='float32',
is_buffer=True)
else:
self.register_parameter('clamp_val', None)
self.eps = eps
self.dtype = dtype
self.quant_mode = quant_mode
def forward(self, x):
weight = None if self.weight is None else self.weight.value
bias = None if self.bias is None else self.bias.value
scale = None
clamp_val = None if self.clamp_val is None else self.clamp_val.value
return fp8_rowwise_rms_norm(
x,
self.normalized_shape,
weight,
bias,
scale,
clamp_val,
self.eps,
dynamic_act_scaling=self.quant_mode.has_fp8_rowwise())
class Fp8RowwiseLinear(Linear):
def __init__(self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
gather_output=True,
quant_mode=QuantMode(0)):
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=gather_output)
if not quant_mode.has_fp8_rowwise():
raise ValueError(
"Fp8 Rowwise Linear has to have act+weight quantization mode set"
)
weights_dtype = dtype
if quant_mode.has_fp8_rowwise():
weights_dtype = "fp8"
self.weight = Parameter(shape=(self.out_features, self.in_features),
dtype=weights_dtype)
if quant_mode.has_fp8_rowwise():
self.per_channel_scale = Parameter(shape=(self.out_features, ),
dtype="float32")
self.quant_mode = quant_mode
self.tllm_to_externel_key_dict = {"weight": ["weight", "weight_scale"]}
def forward(self, x, lora_runtime_params=None):
assert lora_runtime_params is None, "lora is not supported on SmoothQuantLinear now"
x, per_token_scale = x
x = fp8_rowwise_gemm(x, self.weight.value, per_token_scale,
self.per_channel_scale.value,
self.quant_mode.has_per_token_dynamic_scaling(),
self.quant_mode.has_per_channel_scaling())
if self.bias is not None:
x = x + self.bias.value
if self.gather_output and self.tp_size > 1 and self.tp_group is not None:
# [dim0, local_dim] -> [dim0 * tp_size, local_dim] --> [dim0, local_dim * tp_size]
x = allgather(x, self.tp_group, gather_dim=1)
return x
def postprocess(self, tllm_key, weights, **kwargs):
if "per_channel_scale" in tllm_key:
return {}
return postprocess_fp8_rowwise(tllm_key, weights, **kwargs)
Fp8RowwiseColumnLinear = Fp8RowwiseLinear
class Fp8RowwiseRowLinear(RowLinear):
def __init__(
self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
):
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size)
if not quant_mode.has_fp8_rowwise():
raise ValueError(
"Fp8 Rowwise Linear has to have act+weight quantization mode set"
)
weights_dtype = dtype
if quant_mode.has_fp8_rowwise():
weights_dtype = "fp8"
self.weight = Parameter(shape=(self.out_features, self.in_features),
dtype=weights_dtype)
if quant_mode.has_fp8_rowwise():
self.per_channel_scale = Parameter(shape=(self.out_features, ),
dtype="float32")
self.quant_mode = quant_mode
self.tllm_to_externel_key_dict = {"weight": ["weight", "weight_scale"]}
def forward(self, x, lora_runtime_params=None, all_reduce_params=None):
assert lora_runtime_params is None, "lora is not supported on SmoothQuantRowLinear now"
x, per_token_scale = x
x = fp8_rowwise_gemm(x, self.weight.value, per_token_scale,
self.per_channel_scale.value,
self.quant_mode.has_fp8_rowwise(),
self.quant_mode.has_fp8_rowwise())
if self.tp_size > 1 and self.tp_group is not None:
need_bias = self.bias is not None
fuse_bias_into_all_reduce = need_bias and (
all_reduce_params
is not None) and (all_reduce_params.fusion_op
== AllReduceFusionOp.RESIDUAL_RMS_NORM)
if fuse_bias_into_all_reduce:
all_reduce_params.bias = self.bias.value
x = allreduce(x, self.tp_group, all_reduce_params=all_reduce_params)
if need_bias and not fuse_bias_into_all_reduce:
x = x + self.bias.value
return x
if self.bias is not None:
x = x + self.bias.value
return x
def postprocess(self, tllm_key, weights, **kwargs):
if "per_channel_scale" in tllm_key:
return {}
return postprocess_fp8_rowwise(tllm_key, weights, **kwargs)
class WeightOnlyQuantLinear(Linear):
def __init__(
self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
tp_rank=0,
gather_output=True,
quant_mode=QuantMode.use_weight_only(),
transa=False,
transb=False,
is_qkv=False,
prefer_managed_weight=True,
):
multiple = 64 * tp_size
self.is_padded = False
if in_features % multiple > 0:
in_features = math.ceil(in_features / multiple) * multiple
self.is_padded = True
if out_features % multiple > 0:
out_features = math.ceil(out_features / multiple) * multiple
self.is_padded = True
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=gather_output,
is_qkv=is_qkv,
prefer_managed_weight=prefer_managed_weight)
if quant_mode.is_int8_weight_only():
self.weight_only_quant_mode = 1
quant_type_size_in_bits = 8
elif quant_mode.is_int4_weight_only():
self.weight_only_quant_mode = 2
quant_type_size_in_bits = 4
# we use a fake tensor with data_type = int8
self.weight = Parameter(shape=(self.in_features,
int(self.out_features *
quant_type_size_in_bits / 8)),
dtype="int8",
prefer_managed=self.prefer_managed_weight)
scale_shape = (self.out_features, )
self.per_channel_scale = Parameter(shape=scale_shape, dtype=dtype)
self.transa = transa
self.transb = transb
self.tp_rank = tp_rank
if self.is_padded:
self.tp_dim = -1
self.quant_mode = quant_mode
def forward(self, x, lora_runtime_params=None):
# ootb has not supported int4 yet.
if self.weight_only_quant_mode == 2 and not default_net(
).plugin_config.weight_only_quant_matmul_plugin:
raise TypeError(
"Int4 Weight-only Quant MatMul is only supported with plugin")
hidden_state = x
x = weight_only_quant_matmul(x, self.weight.value,
self.per_channel_scale.value,
self.weight_only_quant_mode, self.dtype,
self.transa, self.transb)
if default_net(
).plugin_config.lora_plugin and lora_runtime_params is not None:
x = x + self.lora(hidden_state,
lora_runtime_params=lora_runtime_params)
if self.bias is not None:
x = x + self.bias.value
if self.gather_output and self.tp_size > 1 and self.tp_group is not None:
# [dim0, local_dim] -> [dim0 * tp_size, local_dim] --> [dim0, local_dim * tp_size]
x = allgather(x, self.tp_group, gather_dim=1)
return x
def postprocess(self, tllm_key, weights, **kwargs):
if "per_channel_scale" in tllm_key:
return {}
weights = super().postprocess(tllm_key, weights, **kwargs)[tllm_key]
weights = weights.to(str_dtype_to_torch(self.dtype))
return postprocess_weight_only(
tllm_key, weights,
torch.int8 if self.weight_only_quant_mode == 1 else torch.quint4x2,
self)
WeightOnlyQuantColumnLinear = WeightOnlyQuantLinear
class WeightOnlyQuantRowLinear(RowLinear):
def __init__(
self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
tp_rank=0,
quant_mode=QuantMode.use_weight_only(),
prefer_managed_weight=True,
is_expert=False,
):
multiple = 64 * tp_size
self.is_padded = False
if in_features % multiple > 0:
in_features = math.ceil(in_features / multiple) * multiple
self.is_padded = True
if out_features % multiple > 0:
out_features = math.ceil(out_features / multiple) * multiple
self.is_padded = True
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
prefer_managed_weight=prefer_managed_weight,
is_expert=is_expert)
if quant_mode.is_int8_weight_only():
self.weight_only_quant_mode = 1
elif quant_mode.is_int4_weight_only():
self.weight_only_quant_mode = 2
#we use a fake tensor with data_type = int8
self.weight = Parameter(shape=(self.in_features,
int(self.out_features /
self.weight_only_quant_mode)),
dtype="int8",
prefer_managed=prefer_managed_weight)
self.per_channel_scale = Parameter(shape=(self.out_features, ),
dtype=dtype)
self.tp_rank = tp_rank
if self.is_padded:
self.tp_dim = -1
self.quant_mode = quant_mode
def forward(self, x, lora_runtime_params=None, all_reduce_params=None):
hidden_state = x
x = weight_only_quant_matmul(x, self.weight.value,
self.per_channel_scale.value,
self.weight_only_quant_mode, self.dtype)
if default_net(
).plugin_config.lora_plugin and lora_runtime_params is not None:
x = x + self.lora(hidden_state,
lora_runtime_params=lora_runtime_params)
if self.tp_size > 1 and self.tp_group is not None:
need_bias = self.bias is not None
fuse_bias_into_all_reduce = need_bias and (
all_reduce_params
is not None) and (all_reduce_params.fusion_op
== AllReduceFusionOp.RESIDUAL_RMS_NORM)
if fuse_bias_into_all_reduce:
all_reduce_params.bias = self.bias.value
if not self.is_expert:
x = allreduce(x,
self.tp_group,
all_reduce_params=all_reduce_params)
if need_bias and not fuse_bias_into_all_reduce:
bias = cast(self.bias.value, x.dtype)
x = x + bias
else:
if need_bias and not fuse_bias_into_all_reduce:
bias = cast(self.bias.value, x.dtype)
x = x + bias / self.tp_size
return x
if self.bias is not None:
x = x + self.bias.value
return x
def postprocess(self, tllm_key, weights, **kwargs):
if "per_channel_scale" in tllm_key:
return {}
weights = weights.to(str_dtype_to_torch(self.dtype))
return postprocess_weight_only(
tllm_key, weights,
torch.int8 if self.weight_only_quant_mode == 1 else torch.quint4x2,
self)
class WeightOnlyQuantEmbedding(Embedding):
def __init__(
self,
num_embeddings: int,
embedding_dim: int,
dtype: Optional[str] = None,
tp_size: int = 1,
tp_group: Optional[list] = None,
sharding_dim: int = 0,
tp_rank: Optional[int] = None,
quant_mode=QuantMode.use_weight_only(),
):
super().__init__(
num_embeddings,
embedding_dim,
dtype, # dtype,
tp_size,
tp_group,
sharding_dim,
tp_rank)
# only support int8 wo now
# TODO support int4 wo
self.quant_mode = quant_mode
self.per_token_scale = Parameter(shape=(self.num_embeddings, ),
dtype=dtype)
if sharding_dim == 1:
self.weight = Parameter(shape=(self.num_embeddings,
self.embedding_dim // self.tp_size),
dtype="int8")
elif sharding_dim == 0:
self.weight = Parameter(shape=(math.ceil(
self.num_embeddings / self.tp_size), self.embedding_dim),
dtype="int8")
def forward(self, x):
result = embedding(x,
self.weight.value,
tp_size=self.tp_size,
tp_group=self.tp_group,
sharding_dim=self.sharding_dim,
tp_rank=self.tp_rank,
per_token_scale=self.per_token_scale.value)
return result
def unpack_int32_into_int8(w_packed):
# Unpack inputs packed in int32/float32 into uint4 and store them in int8 format
w_packed_int4x2 = w_packed.contiguous().view(torch.uint8)
w_unpacked = torch.zeros(w_packed_int4x2.shape[0],
w_packed_int4x2.shape[1] * 2,
dtype=torch.int8)
w_unpacked[:, ::2] = w_packed_int4x2 % 16
w_unpacked[:, 1::2] = w_packed_int4x2 // 16
w_unpacked = w_unpacked.view(-1, 8)[:, [0, 4, 1, 5, 2, 6, 3, 7]].view(
w_unpacked.shape)
return w_unpacked.contiguous()
class WeightOnlyGroupwiseQuantLinear(Linear):
def __init__(
self,
in_features,
out_features,
group_size=128,
pre_quant_scale=False,
zero=False,
bias=False,
dtype=None,
tp_group=None,
tp_size=1,
tp_rank=0,
gather_output=True,
use_w4a8_awq=False,
use_int8_weight=False,
is_qkv=False,
prefer_managed_weight=True,
):
multiple = max((128 if use_w4a8_awq else 64), group_size) * tp_size
self.is_padded = False
if in_features % multiple > 0:
in_features = math.ceil(in_features / multiple) * multiple
self.is_padded = True
if out_features % multiple > 0:
out_features = math.ceil(out_features / multiple) * multiple
self.is_padded = True
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=gather_output,
is_qkv=is_qkv,
prefer_managed_weight=prefer_managed_weight)
self.quant_algo = (
use_int8_weight * GroupwiseQuantAlgo.INT8_WEIGHT +
use_w4a8_awq * GroupwiseQuantAlgo.W4A8_ALPHA +
pre_quant_scale * GroupwiseQuantAlgo.PRE_QUANT_SCALE +
zero * GroupwiseQuantAlgo.ZERO + bias * GroupwiseQuantAlgo.BIAS)
self.group_size = group_size
# packed in FP16 format (INT4*4 -> FP16, INT8*2 -> FP16)
pack_ratio = 2 if use_int8_weight else 4
self.weight = Parameter(shape=(self.in_features,
self.out_features // pack_ratio),
dtype=dtype,
prefer_managed=self.prefer_managed_weight)
scale_shape = (self.in_features // group_size, self.out_features)
self.weights_scaling_factor = Parameter(shape=scale_shape, dtype=dtype)
self.tp_rank = tp_rank
if self.is_padded:
self.tp_dim = -1
self.pre_quant_scale = pre_quant_scale
self.use_w4a8_awq = use_w4a8_awq
if pre_quant_scale:
self.prequant_scaling_factor = Parameter(shape=(1,
self.in_features),
dtype=dtype)
else:
self.register_parameter('prequant_scaling_factor', None)
if zero:
self.zero = Parameter(shape=scale_shape, dtype=dtype)
else:
self.register_parameter('zero', None)
if use_w4a8_awq:
self.alpha = Parameter(shape=(1, ), dtype="float32")
else:
self.register_parameter('alpha', None)
validate_group_size(self)
if pre_quant_scale:
self.tllm_to_externel_key_dict = {
"weights_scaling_factor": "weight_scale",
"prequant_scaling_factor": "input_quantizer._pre_quant_scale",
} # AWQ
else:
self.tllm_to_externel_key_dict = {
"weight": ["qweight", "scales", "qzeros"]
} # GPTQ
def forward(self, x, lora_runtime_params=None):
pre_quant_scale = self.prequant_scaling_factor.value if self.prequant_scaling_factor else None
zero = self.zero.value if self.zero else None
bias = self.bias.value if self.bias else None
alpha = self.alpha if self.alpha else None
hidden_state = x
x = weight_only_groupwise_quant_matmul(
x, pre_quant_scale, self.weight.value,
self.weights_scaling_factor.value, zero, bias, alpha,
self.quant_algo, self.group_size, self.dtype)
if default_net(
).plugin_config.lora_plugin and lora_runtime_params is not None:
x = x + self.lora(hidden_state,
lora_runtime_params=lora_runtime_params)
if self.gather_output and self.tp_size > 1 and self.tp_group is not None:
# [dim0, local_dim] -> [dim0 * tp_size, local_dim] --> [dim0, local_dim * tp_size]
x = allgather(x, self.tp_group, gather_dim=1)
return x
def postprocess(self, tllm_key, weights, **kwargs):
if tllm_key.endswith("zero") or (
self.prequant_scaling_factor is None
and tllm_key.endswith("weights_scaling_factor")):
return {}
torch_dtype = str_dtype_to_torch(self.dtype)
return postprocess_weight_only_groupwise(tllm_key, weights, torch_dtype,
self, **kwargs)
WeightOnlyGroupwiseQuantColumnLinear = WeightOnlyGroupwiseQuantLinear
class WeightOnlyGroupwiseQuantRowLinear(RowLinear):
def __init__(self,
in_features,
out_features,
group_size=128,
pre_quant_scale=False,
zero=False,
bias=False,
dtype=None,
tp_group=None,
tp_size=1,
tp_rank=0,
use_w4a8_awq=False,
use_int8_weight=False,
prefer_managed_weight=True):
multiple = max((128 if use_w4a8_awq else 64), group_size) * tp_size
self.is_padded = False
if in_features % multiple > 0:
in_features = math.ceil(in_features / multiple) * multiple
self.is_padded = True
if out_features % multiple > 0:
out_features = math.ceil(out_features / multiple) * multiple
self.is_padded = True
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
prefer_managed_weight=prefer_managed_weight)
self.quant_algo = (
use_int8_weight * GroupwiseQuantAlgo.INT8_WEIGHT +
use_w4a8_awq * GroupwiseQuantAlgo.W4A8_ALPHA +
pre_quant_scale * GroupwiseQuantAlgo.PRE_QUANT_SCALE +
zero * GroupwiseQuantAlgo.ZERO + bias * GroupwiseQuantAlgo.BIAS)
self.group_size = group_size
# packed in FP16 format (INT4*4 -> FP16, INT8*2 -> FP16)
pack_ratio = 2 if use_int8_weight else 4
self.weight = Parameter(shape=(self.in_features,
self.out_features // pack_ratio),
dtype=dtype,
prefer_managed=self.prefer_managed_weight)
scale_shape = (self.in_features // group_size, self.out_features)
self.weights_scaling_factor = Parameter(shape=scale_shape, dtype=dtype)
self.tp_rank = tp_rank
if self.is_padded:
self.tp_dim = -1
self.pre_quant_scale = pre_quant_scale
self.use_w4a8_awq = use_w4a8_awq
if pre_quant_scale:
self.prequant_scaling_factor = Parameter(shape=(1,
self.in_features),
dtype=dtype)
else:
self.register_parameter('prequant_scaling_factor', None)
if zero:
self.zero = Parameter(shape=scale_shape, dtype=dtype)
else:
self.register_parameter('zero', None)
if use_w4a8_awq:
self.alpha = Parameter(shape=(1, ), dtype="float32")
self.activation_scaling_factor = Parameter(shape=(1, ),
dtype=trt.float32)
else:
self.register_parameter('alpha', None)
self.register_parameter('activation_scaling_factor', None)
validate_group_size(self)
if pre_quant_scale:
self.tllm_to_externel_key_dict = {
"weights_scaling_factor": "weight_scale",
"prequant_scaling_factor": "input_quantizer._pre_quant_scale",
} # AWQ
else:
self.tllm_to_externel_key_dict = {
"weight": ["qweight", "scales", "qzeros"]
} # GPTQ
def forward(self, x, lora_runtime_params=None, all_reduce_params=None):
pre_quant_scale = self.prequant_scaling_factor.value if self.prequant_scaling_factor else None
zero = self.zero.value if self.zero else None
bias = self.bias.value if self.bias else None
alpha = self.alpha if self.alpha else None
activation_scaling_factor = self.activation_scaling_factor.value if self.activation_scaling_factor else None
if self.alpha and x.dtype == trt.fp8:
x = dequantize(x, activation_scaling_factor, -1, self.dtype)
hidden_state = x
x = weight_only_groupwise_quant_matmul(
x, pre_quant_scale, self.weight.value,
self.weights_scaling_factor.value, zero, bias, alpha,
self.quant_algo, self.group_size, self.dtype)
if default_net(
).plugin_config.lora_plugin and lora_runtime_params is not None:
x = x + self.lora(hidden_state,
lora_runtime_params=lora_runtime_params)
if self.tp_size > 1 and self.tp_group is not None:
x = allreduce(x, self.tp_group, all_reduce_params=all_reduce_params)
return x
def postprocess(self, tllm_key, weights, **kwargs):
if tllm_key.endswith("zero") or (
self.prequant_scaling_factor is None
and tllm_key.endswith("weights_scaling_factor")):
return {}
torch_dtype = str_dtype_to_torch(self.dtype)
return postprocess_weight_only_groupwise(tllm_key, weights, torch_dtype,
self, **kwargs)
class SmoothQuantMLP(Module):
def __init__(
self,
hidden_size,
ffn_hidden_size,
hidden_act,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
):
super().__init__()
if hidden_act not in ACT2FN:
raise ValueError(
'unsupported activation function: {}'.format(hidden_act))
fc_output_size = 2 * ffn_hidden_size if hidden_act == 'swiglu' else ffn_hidden_size
self.fc = SmoothQuantColumnLinear(hidden_size,
fc_output_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False,
quant_mode=quant_mode)
self.proj = SmoothQuantRowLinear(ffn_hidden_size,
hidden_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
quant_mode=quant_mode)
self.hidden_act = hidden_act
self.quant_mode = quant_mode
self.dtype = dtype
if self.quant_mode.has_act_static_scaling():
self.quantization_scaling_factor = Parameter(shape=(1, ),
dtype='float32')
else:
self.register_parameter('quantization_scaling_factor', None)
def forward(self, hidden_states, lora_layer_params=None):
inter = self.fc(hidden_states)
inter = ACT2FN[self.hidden_act](inter)
value = cast(self.proj.smoother.value, inter.dtype)
inter = inter / value
if self.quant_mode.has_act_and_weight_quant():
if self.quant_mode.has_act_static_scaling():
# Avoid quantization layers as it breaks int8 plugins
inter = quantize_tensor(inter,
self.quantization_scaling_factor.value)
else:
# Quantize per token outputs tuple:
# quantized tensor and scaling factors per token
inter = quantize_per_token(inter)
output = self.proj(inter)
return output
class Int8SmoothQuantRowLinear(RowLinear):
def __init__(self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
prefer_managed_weight=True):
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
prefer_managed_weight=prefer_managed_weight)
self.activation_scaling_factor = Parameter(shape=(1, ),
dtype=trt.float32)
self.weights_scaling_factor = Parameter(shape=(self.out_features, ),
dtype=trt.float32)
self.prequant_scaling_factor = Parameter(shape=(self.in_features, ),
dtype=dtype)
self.weight = Parameter(shape=(self.out_features, self.in_features),
dtype=trt.int8,
prefer_managed=self.prefer_managed_weight)
def forward(self, x, lora_runtime_params=None, all_reduce_params=None):
lora_hidden_state = x if lora_runtime_params is not None else None
if default_net().strongly_typed:
assert is_same_dtype(
x.dtype,
self.dtype), f"Got input type {x.dtype}, expecting {self.dtype}"
x = mul(x, self.prequant_scaling_factor.value)
x = cast(x, self.activation_scaling_factor.value.dtype)
quantized_out = quantize(x, self.activation_scaling_factor.value,
'int8')
dequantized_out = dequantize(quantized_out,
self.activation_scaling_factor.value, -1,
self.activation_scaling_factor.value.dtype)
dequantized_out = cast(dequantized_out, self.dtype)
w_deq_out = dequantize(self.weight.value,
self.weights_scaling_factor.value, 0,
self.weights_scaling_factor.value.dtype)
w_deq_out = cast(w_deq_out, self.dtype)
return self.multiply_collect(dequantized_out,
w_deq_out,
gemm_plugin=None,
all_reduce_params=all_reduce_params,
lora_runtime_params=lora_runtime_params,
lora_hidden_state=lora_hidden_state)
class Int8SmoothQuantLinear(Linear):
def __init__(
self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
gather_output=True,
prefer_managed_weight=True,
):
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=gather_output,
prefer_managed_weight=prefer_managed_weight)
self.activation_scaling_factor = Parameter(shape=(1, ),
dtype=trt.float32)
self.weights_scaling_factor = Parameter(shape=(self.out_features, ),
dtype=trt.float32)
self.prequant_scaling_factor = Parameter(shape=(self.in_features, ),
dtype=dtype)
self.weight = Parameter(shape=(self.out_features, self.in_features),
dtype=trt.int8,
prefer_managed=self.prefer_managed_weight)
def forward(self, x, lora_runtime_params=None):
lora_hidden_state = x if lora_runtime_params is not None else None
if default_net().strongly_typed:
assert is_same_dtype(
x.dtype,
self.dtype), f"Got input type {x.dtype}, expecting {self.dtype}"
x = mul(x, self.prequant_scaling_factor.value)
x = cast(x, self.activation_scaling_factor.value.dtype)
quantized_out = quantize(x, self.activation_scaling_factor.value,
'int8')
dequantized_out = dequantize(quantized_out,
self.activation_scaling_factor.value, -1,
self.activation_scaling_factor.value.dtype)
dequantized_out = cast(dequantized_out, self.dtype)
w_deq_out = dequantize(self.weight.value,
self.weights_scaling_factor.value, 0,
self.weights_scaling_factor.value.dtype)
w_deq_out = cast(w_deq_out, self.dtype)
return self.multiply_collect(dequantized_out,
w_deq_out,
gemm_plugin=None,
lora_runtime_params=lora_runtime_params,
lora_hidden_state=lora_hidden_state)
class FP8Linear(Linear):
def __init__(self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
gather_output=True,
prefer_managed_weight=True,
is_qkv=False):
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=gather_output,
prefer_managed_weight=prefer_managed_weight,
is_qkv=is_qkv)
self.weight = Parameter(shape=(self.out_features, self.in_features),
dtype='fp8',
prefer_managed=self.prefer_managed_weight)
self.activation_scaling_factor = Parameter(shape=(1, ),
dtype=trt.float32)
self.weights_scaling_factor = Parameter(shape=(1, ), dtype=trt.float32)
if self.is_qkv:
self.tllm_to_externel_key_dict = {
"activation_scaling_factor": "input_scale",
"weight": ["weight", "weight_scale"],
"weights_scaling_factor": "",
}
else:
self.tllm_to_externel_key_dict = {
"activation_scaling_factor": "input_scale",
"weights_scaling_factor": "weight_scale",
}
def forward(self, x, lora_runtime_params=None):
assert lora_runtime_params is None or default_net(
).plugin_config.lora_plugin == self.dtype
if default_net().strongly_typed:
assert default_net().plugin_config.user_buffer or is_same_dtype(
x.dtype,
self.dtype), f"Got input type {x.dtype}, expecting {self.dtype}"
alpha = self.weights_scaling_factor.raw_value * self.activation_scaling_factor.raw_value
activation_scaling_factor = constant(
self.activation_scaling_factor.raw_value)
activation_scaling_factor = cast(activation_scaling_factor, self.dtype)
if x.dtype != trt.fp8:
quantized_out = quantize(x, activation_scaling_factor, 'fp8')
lora_hidden_state = x if lora_runtime_params is not None else None
else:
quantized_out = x
# TODO: add fp8 LoRA support
lora_hidden_state = dequantize(
x, activation_scaling_factor, -1,
self.dtype) if lora_runtime_params is not None else None
weights_scaling_factor = cast(self.weights_scaling_factor.value,
self.dtype)
if self.weight.value.dtype != trt.fp8:
w_quant_out = quantize(self.weight.value, weights_scaling_factor,
'fp8')
else:
w_quant_out = self.weight.value
gemm_plugin = default_net().plugin_config.gemm_plugin
low_latency_gemm_plugin = default_net(
).plugin_config.low_latency_gemm_plugin
if (low_latency_gemm_plugin == "fp8"):
return self.multiply_collect(
quantized_out,
w_quant_out,
gemm_plugin=None,
low_latency_gemm_plugin=low_latency_gemm_plugin,
use_fp8=True,
alpha=alpha,
lora_runtime_params=lora_runtime_params,
lora_hidden_state=lora_hidden_state)
elif gemm_plugin == 'fp8':
return self.multiply_collect(
quantized_out,
w_quant_out,
gemm_plugin=gemm_plugin,
use_fp8=True,
alpha=alpha,
lora_runtime_params=lora_runtime_params,
lora_hidden_state=lora_hidden_state)
else:
dequantized_out = dequantize(quantized_out,
activation_scaling_factor, -1,
self.dtype)
w_deq_out = dequantize(w_quant_out, weights_scaling_factor, -1,
self.dtype)
return self.multiply_collect(
dequantized_out,
w_deq_out,
gemm_plugin=None,
use_fp8=True,
lora_runtime_params=lora_runtime_params,
lora_hidden_state=lora_hidden_state)
def postprocess(self, tllm_key, weights, **kwargs):
if self.is_qkv:
if tllm_key.endswith("activation_scaling_factor"):
return max(weights).reshape(1, ).to(torch.float32)
elif tllm_key.endswith("weights_scaling_factor"):
return {}
elif tllm_key.endswith("weight"):
assert len(weights) == 6
weight_scaling_factors = weights[1::2]
new_amax = max(weight_scaling_factors).reshape(1, ).to(
torch.float32)
for qkv_idx in range(3):
idx = qkv_idx * 2
weights[idx] = weights[idx].view(torch.float8_e4m3fn).to(
torch.float32)
weights[idx] *= weights[idx + 1]
weights[idx] /= new_amax
weights[idx] = weights[idx].to(torch.float8_e4m3fn)
weights = torch.cat(weights[::2])
scales = new_amax
return {
tllm_key: weights,
tllm_key.replace("weight", "weights_scaling_factor"): scales
}
else:
return super().postprocess(tllm_key, weights, **kwargs)
if tllm_key.endswith("scaling_factor"):
return weights.reshape(1, ).to(torch.float32)
elif tllm_key.endswith("weight"):
return weights.view(torch.float8_e4m3fn)
elif tllm_key.endswith("bias"):
return weights.to(trt_dtype_to_torch(self.bias.dtype))
class FP8RowLinear(RowLinear):
def __init__(self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
prefer_managed_weight=True,
is_expert=False):
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
prefer_managed_weight=prefer_managed_weight,
is_expert=is_expert)
self.weight = Parameter(
shape=(self.out_features, self.in_features),
dtype="fp8",
prefer_managed=self.prefer_managed_weight,
)
self.activation_scaling_factor = Parameter(shape=(1, ),
dtype=trt.float32)
self.weights_scaling_factor = Parameter(shape=(1, ), dtype=trt.float32)
self.tllm_to_externel_key_dict = {
"activation_scaling_factor": "input_scale",
"weights_scaling_factor": "weight_scale",
}
def forward(self, x, lora_runtime_params=None, all_reduce_params=None):
assert lora_runtime_params is None or default_net(
).plugin_config.lora_plugin == self.dtype
alpha = self.weights_scaling_factor.raw_value * self.activation_scaling_factor.raw_value
activation_scaling_factor = cast(self.activation_scaling_factor.value,
self.dtype)
if x.dtype != trt.fp8:
quantized_out = quantize(x, activation_scaling_factor, 'fp8')
lora_hidden_state = x if lora_runtime_params is not None else None
else:
quantized_out = x
# TODO: add fp8 LoRA support
lora_hidden_state = dequantize(
x, activation_scaling_factor, -1,
self.dtype) if lora_runtime_params is not None else None
weights_scaling_factor = cast(self.weights_scaling_factor.value,
self.dtype)
if self.weight.value.dtype != trt.fp8:
w_quant_out = quantize(self.weight.value, weights_scaling_factor,
'fp8')
else:
w_quant_out = self.weight.value
gemm_plugin = default_net().plugin_config.gemm_plugin
low_latency_gemm_plugin = default_net(
).plugin_config.low_latency_gemm_plugin
gemm_allreduce_plugin = default_net(
).plugin_config.gemm_allreduce_plugin
if gemm_allreduce_plugin:
ret = self.multiply_collect(quantized_out,
w_quant_out,
gemm_plugin=None,
use_fp8=True,
alpha=alpha,
lora_runtime_params=lora_runtime_params,
lora_hidden_state=lora_hidden_state,
all_reduce_params=all_reduce_params)
elif (low_latency_gemm_plugin == "fp8"):
ret = self.multiply_collect(
quantized_out,
w_quant_out,
gemm_plugin=None,
low_latency_gemm_plugin=low_latency_gemm_plugin,
use_fp8=True,
alpha=alpha,
lora_runtime_params=lora_runtime_params,
lora_hidden_state=lora_hidden_state,
all_reduce_params=all_reduce_params)
elif gemm_plugin == 'fp8':
ret = self.multiply_collect(quantized_out,
w_quant_out,
gemm_plugin=gemm_plugin,
use_fp8=True,
alpha=alpha,
lora_runtime_params=lora_runtime_params,
lora_hidden_state=lora_hidden_state,
all_reduce_params=all_reduce_params)
else:
dequantized_out = dequantize(quantized_out,
activation_scaling_factor, -1,
self.dtype)
w_deq_out = dequantize(w_quant_out, weights_scaling_factor, -1,
self.dtype)
ret = self.multiply_collect(dequantized_out,
w_deq_out,
gemm_plugin=None,
use_fp8=True,
lora_runtime_params=lora_runtime_params,
lora_hidden_state=lora_hidden_state,
all_reduce_params=all_reduce_params)
return ret
def postprocess(self, tllm_key, weights, **kwargs):
if tllm_key.endswith("scaling_factor"):
return weights.reshape(1, ).to(torch.float32)
elif tllm_key.endswith("weight"):
return weights.view(torch.float8_e4m3fn)
elif tllm_key.endswith("bias"):
return weights.to(str_dtype_to_torch(self.bias.dtype))
class Fp8RowwiseMLP(Module):
def __init__(
self,
hidden_size,
ffn_hidden_size,
hidden_act,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
clamp_val=None,
):
super().__init__()
if hidden_act not in ACT2FN:
raise ValueError(
'unsupported activation function: {}'.format(hidden_act))
fc_output_size = 2 * ffn_hidden_size if hidden_act == 'swiglu' else ffn_hidden_size
self.fc = Fp8RowwiseColumnLinear(hidden_size,
fc_output_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False,
quant_mode=quant_mode)
self.proj = Fp8RowwiseRowLinear(ffn_hidden_size,
hidden_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
quant_mode=quant_mode)
self.hidden_size = hidden_size
self.ffn_hidden_size = ffn_hidden_size
self.hidden_act = hidden_act
self.bias = bias
self.dtype = dtype
self.tp_group = tp_group
self.tp_size = tp_size
self.quant_mode = quant_mode
if clamp_val:
if not (isinstance(clamp_val, list) and len(clamp_val) == 2):
raise ValueError(f'unsupported clamp_val {clamp_val}')
self.clamp_val = Parameter(np.array(clamp_val, dtype=np.float32),
dtype='float32',
is_buffer=True)
else:
self.register_parameter('clamp_val', None)
def forward(self, hidden_states, lora_layer_params=None):
assert lora_layer_params is None, f"lora is not supported on {self.__class__.__name__} now"
inter = self.fc(hidden_states)
inter = ACT2FN[self.hidden_act](inter)
if self.quant_mode.has_fp8_rowwise():
# Quantize per token outputs tuple:
# quantized tensor and scaling factors per token
clamp_val = None if self.clamp_val is None else self.clamp_val.value
inter = quantize_fp8_per_token(inter, clamp_val)
output = self.proj(inter)
return output
class Fp8RowwiseGatedMLP(Fp8RowwiseMLP):
def __init__(
self,
hidden_size,
ffn_hidden_size,
hidden_act,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
clamp_val=None,
):
super().__init__(hidden_size,
ffn_hidden_size,
hidden_act,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
quant_mode=quant_mode,
clamp_val=clamp_val)
if hidden_act not in ACT2FN:
raise ValueError(
'unsupported activation function: {}'.format(hidden_act))
self.gate = Fp8RowwiseColumnLinear(hidden_size,
ffn_hidden_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False,
quant_mode=quant_mode)
def forward(self, hidden_states, lora_layer_params=None):
assert lora_layer_params is None, f"lora is not supported on {self.__class__.__name__} now"
inter = self.fc(hidden_states)
inter = ACT2FN[self.hidden_act](inter)
gate = self.gate(hidden_states)
inter_x_gate = inter * gate
if self.quant_mode.has_fp8_rowwise():
# Quantize per token outputs tuple:
# quantized tensor and scaling factors per token
clamp_val = None if self.clamp_val is None else self.clamp_val.value
inter_x_gate = quantize_fp8_per_token(inter_x_gate, clamp_val)
output = self.proj(inter_x_gate)
return output
class Fp8RowwiseFusedGatedMLP(Module):
def __init__(
self,
hidden_size,
ffn_hidden_size,
hidden_act,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
clamp_val=None,
):
super().__init__()
self.hidden_size = hidden_size
self.ffn_hidden_size = ffn_hidden_size
self.hidden_act = hidden_act
self.bias = bias
self.dtype = dtype
self.tp_group = tp_group
self.tp_size = tp_size
self.quant_mode = quant_mode
self.fused_fc = Fp8RowwiseColumnLinear(hidden_size,
ffn_hidden_size * 2,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False,
quant_mode=quant_mode)
self.proj = Fp8RowwiseRowLinear(ffn_hidden_size,
hidden_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
quant_mode=quant_mode)
if clamp_val:
if not (isinstance(clamp_val, list) and len(clamp_val) == 2):
raise ValueError(f'unsupported clamp_val {clamp_val}')
self.clamp_val = Parameter(np.array(clamp_val, dtype=np.float32),
dtype='float32',
is_buffer=True)
else:
self.register_parameter('clamp_val', None)
def forward(self, hidden_states, lora_layer_params=None):
assert lora_layer_params is None, f"lora is not supported on {self.__class__.__name__} now"
inter = self.fused_fc(hidden_states)
if self.hidden_act == 'silu':
inter = ACT2FN['swiglu'](inter)
elif self.hidden_act == 'gelu':
inter = ACT2FN['geglu'](inter)
else:
raise NotImplementedError(
f"Activation {self.hidden_act} not yet implemented for {self.__class__.__name__}."
)
if self.quant_mode.has_fp8_rowwise():
# Quantize per token outputs tuple:
# quantized tensor and scaling factors per token
clamp_val = None if self.clamp_val is None else self.clamp_val.value
inter = quantize_fp8_per_token(inter, clamp_val)
output = self.proj(inter)
return output
class Fp8RowwiseAttention(Module):
def __init__(self,
*,
local_layer_idx,
hidden_size,
num_attention_heads,
num_kv_heads=None,
max_position_embeddings=1024,
num_layers=1,
apply_query_key_layer_scaling=False,
attention_head_size=None,
attention_mask_type=AttentionMaskType.padding,
bias=True,
dense_bias=None,
dtype=None,
position_embedding_type=PositionEmbeddingType.learned_absolute,
rotary_embedding_base=10000.0,
rotary_embedding_scaling=None,
rotary_embedding_percentage=1.0,
tp_group=None,
tp_size=1,
tp_rank=0,
scale_alibi_bias=False,
paged_kv_cache=False,
quant_mode=QuantMode(0),
clamp_val=None):
super().__init__()
self.local_layer_idx = local_layer_idx
self.attention_mask_type = attention_mask_type
self.attention_head_size = hidden_size // num_attention_heads if attention_head_size is None else attention_head_size
self.num_attention_heads = num_attention_heads // tp_size
self.num_kv_heads = num_kv_heads
self.num_attention_kv_heads = (
num_kv_heads + tp_size - 1
) // tp_size if num_kv_heads is not None else self.num_attention_heads
self.hidden_size = hidden_size // tp_size
self.max_position_embeddings = 0 if max_position_embeddings is None else max_position_embeddings
self.tp_size = tp_size
self.tp_rank = tp_rank
self.dense_bias = dense_bias
if dense_bias is None:
self.dense_bias = bias
self.num_layers = num_layers
self.apply_query_key_layer_scaling = apply_query_key_layer_scaling
self.norm_factor = math.sqrt(self.attention_head_size)
self.q_scaling = 1
if self.apply_query_key_layer_scaling:
self.norm_factor *= self.num_layers
self.q_scaling *= self.num_layers
# Whether to scale ALiBi bias. Mathematically, it's equivalent to
# normalizing QK after adding bias.
# - False, inv_sqrt_Dh * Q*K^T + alibi_bias
# - True, inv_sqrt_Dh * Q*K^T + inv_sqrt_Dh * alibi_bias
self.scale_alibi_bias = scale_alibi_bias
self.position_embedding_type = position_embedding_type
self.paged_kv_cache = paged_kv_cache
self.rotary_embedding_base = rotary_embedding_base
self.rotary_embedding_scale_type = RotaryScalingType.none
self.rotary_embedding_scale = 1.0
self.rotary_embedding_dim = 0
if rotary_embedding_scaling is not None:
rotary_scaling_type = rotary_embedding_scaling.get(
"type", rotary_embedding_scaling.get("rope_type"))
self.rotary_embedding_scale_type = RotaryScalingType.from_string(
rotary_scaling_type)
self.rotary_embedding_scale = rotary_embedding_scaling.get(
"factor", 1.0)
if self.position_embedding_type.is_rope():
self.rotary_embedding_dim = int(self.attention_head_size *
rotary_embedding_percentage)
elif self.position_embedding_type.is_alibi():
alibi_scale = 1. / self.norm_factor if self.scale_alibi_bias else 1.
alibi_slopes = generate_alibi_slopes(self.num_attention_heads *
self.tp_size,
tp_size=self.tp_size,
tp_rank=self.tp_rank,
alibi_scale=alibi_scale)
self.register_parameter(
'alibi_slopes',
Parameter(alibi_slopes, dtype='float32', is_buffer=True))
self.quant_mode = quant_mode
self.dtype = dtype
self.register_parameter('kv_cache_scaling_factor', None)
self.qkv = Fp8RowwiseColumnLinear(
hidden_size,
tp_size * self.num_attention_heads * self.attention_head_size +
(2 * tp_size * self.num_attention_kv_heads *
self.attention_head_size),
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False,
quant_mode=quant_mode)
self.dense = Fp8RowwiseRowLinear(tp_size * self.num_attention_heads *
self.attention_head_size,
hidden_size,
bias=self.dense_bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
quant_mode=quant_mode)
if clamp_val:
if not (isinstance(clamp_val, list) and len(clamp_val) == 2):
raise ValueError(f'unsupported clamp_val {clamp_val}')
self.clamp_val = Parameter(np.array(clamp_val, dtype=np.float32),
dtype='float32',
is_buffer=True)
else:
self.register_parameter('clamp_val', None)
self.use_lora = False
def forward(
self,
hidden_states: Tensor,
attention_mask=None,
use_cache=False,
kv_cache_params=None,
attention_params=None,
spec_decoding_params=None,
encoder_output=None,
position_embedding=None,
norm_before_bmm1=False,
lora_layer_params=None,
all_reduce_params: Optional[AllReduceParams] = None,
):
assert lora_layer_params is None, f"lora is not supported on {self.__class__.__name__} now"
qkv = self.qkv(hidden_states)
alibi_slopes = None
if self.position_embedding_type == PositionEmbeddingType.alibi:
alibi_slopes = self.alibi_slopes.value
dtype = trt.float32
if default_net().plugin_config.gpt_attention_plugin or default_net(
).plugin_config.inflight_batching_gpt_attention_plugin:
dtype = hidden_states.dtype if self.quant_mode.has_act_static_scaling(
) else hidden_states[0].dtype
if dtype == trt.int8:
dtype = trt.float16
alibi_slopes = cast(alibi_slopes, dtype)
if spec_decoding_params is None:
spec_decoding_params = SpecDecodingParams()
assert default_net().plugin_config.gpt_attention_plugin
assert attention_params.is_valid(
default_net().plugin_config.gpt_attention_plugin,
default_net().plugin_config.remove_input_padding, use_cache)
if use_cache:
assert kv_cache_params.is_valid(
default_net().plugin_config.gpt_attention_plugin)
assert self.attention_mask_type == AttentionMaskType.causal, \
'Plugin only support masked MHA.'
if self.kv_cache_scaling_factor is not None:
kv_orig_quant_scale = self.kv_cache_rcp_scaling_factor.value
kv_quant_orig_scale = self.kv_cache_scaling_factor.value
else:
kv_orig_quant_scale = None
kv_quant_orig_scale = None
if self.position_embedding_type.is_rope():
rotary_inv_freq = attention_params.rotary_inv_freq
rotary_cos_sin = attention_params.embed_positions_for_gpt_attention
else:
rotary_inv_freq = None
rotary_cos_sin = None
context, past_key_value = gpt_attention(
qkv=qkv,
past_key_value=kv_cache_params.get_first_past_key_value(),
sequence_length=attention_params.sequence_length,
host_past_key_value_lengths=kv_cache_params.
host_past_key_value_lengths,
host_max_attention_window_sizes=kv_cache_params.
host_max_attention_window_sizes,
host_sink_token_length=kv_cache_params.host_sink_token_length,
context_lengths=attention_params.context_lengths,
cache_indirection=kv_cache_params.cache_indirection,
host_request_types=attention_params.host_request_types,
layer_idx=self.local_layer_idx,
num_heads=self.num_attention_heads,
num_kv_heads=self.num_attention_kv_heads,
num_kv_heads_origin=self.num_kv_heads,
hidden_size_per_head=self.attention_head_size,
q_scaling=self.q_scaling,
rotary_embedding_dim=self.rotary_embedding_dim,
rotary_embedding_base=self.rotary_embedding_base,
rotary_embedding_scale_type=self.rotary_embedding_scale_type,
rotary_embedding_scale=self.rotary_embedding_scale,
rotary_embedding_max_positions=self.max_position_embeddings,
position_embedding_type=self.position_embedding_type,
rotary_inv_freq=rotary_inv_freq,
rotary_cos_sin=rotary_cos_sin,
kv_orig_quant_scale=kv_orig_quant_scale,
kv_quant_orig_scale=kv_quant_orig_scale,
kv_cache_quant_mode=self.quant_mode,
max_context_length=attention_params.max_context_length,
alibi_slopes=alibi_slopes,
tp_size=self.tp_size,
tp_rank=self.tp_rank,
kv_cache_block_offsets=kv_cache_params.kv_cache_block_offsets,
host_kv_cache_block_offsets=kv_cache_params.
host_kv_cache_block_offsets,
host_kv_cache_pool_pointers=kv_cache_params.
host_kv_cache_pool_pointers,
host_kv_cache_pool_mapping=kv_cache_params.
host_kv_cache_pool_mapping,
host_context_lengths=attention_params.host_context_lengths,
use_cache=use_cache,
spec_decoding_generation_lengths=spec_decoding_params.
spec_decoding_generation_lengths,
spec_decoding_position_offsets=spec_decoding_params.
spec_decoding_position_offsets,
spec_decoding_packed_mask=spec_decoding_params.
spec_decoding_packed_mask,
host_runtime_perf_knobs=attention_params.host_runtime_perf_knobs,
host_context_progress=attention_params.host_context_progress)
if self.quant_mode.has_fp8_rowwise():
# Quantize per token outputs tuple:
# quantized tensor and scaling factors per token
clamp_val = None if self.clamp_val is None else self.clamp_val.value
context = quantize_fp8_per_token(context, clamp_val)
context = self.dense(
context,
all_reduce_params=all_reduce_params,
)
if use_cache:
return (context, past_key_value)
return context
def postprocess(self, tllm_key, weights, **kwargs):
if tllm_key.endswith("kv_cache_scaling_factor") and weights is None:
return {tllm_key: torch.ones(1, )}
else:
return {tllm_key: weights}
class FP4Linear(Linear):
def __init__(
self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
gather_output=True,
prefer_managed_weight=True,
is_qkv=False,
):
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=gather_output,
prefer_managed_weight=prefer_managed_weight,
is_qkv=is_qkv)
self.scaling_vector_size = 16
assert self.in_features % self.scaling_vector_size == 0, \
"Input features must be a multiple of 16 for FP4 GEMM"
self.weight = Parameter(shape=(self.out_features, self.in_features),
dtype=trt.fp4)
self.weights_block_scaling_factor = Parameter(
shape=(self.out_features,
self.in_features // self.scaling_vector_size),
dtype=trt.fp8)
self.weights_block_scaling_factor_interleaved = Parameter(
shape=(self.out_features,
self.in_features // self.scaling_vector_size),
dtype=trt.fp8)
self.weights_global_scaling_factor = Parameter(shape=(1, ),
dtype=trt.float32)
self.activation_global_scaling_factor = Parameter(shape=(1, ),
dtype=trt.float32)
# alpha = 1.0 / (weight_global_scale * act_global_scale)
self.alpha = Parameter(shape=(1, ), dtype=trt.float32)
if self.is_qkv:
self.tllm_to_externel_key_dict = {
"weight":
["weight", "weight_scale", "weight_scale_2", "input_scale"],
"weights_block_scaling_factor":
"",
"weights_block_scaling_factor_interleaved":
"",
"weights_global_scaling_factor":
"",
"activation_global_scaling_factor":
"",
"alpha":
"",
}
else:
self.tllm_to_externel_key_dict = {
"weight": ["weight"],
"weights_block_scaling_factor": "weight_scale",
"weights_block_scaling_factor_interleaved": "weight_scale",
"weights_global_scaling_factor": "weight_scale_2",
"activation_global_scaling_factor": "input_scale",
"alpha": ["weight_scale_2", "input_scale"],
}
def forward(self, x, lora_runtime_params=None):
assert lora_runtime_params is None, "lora is not supported on FP4Linear now"
if isinstance(x, (tuple, list)):
fp4_x, act_per_block_scale = x
else:
if default_net().plugin_config.gemm_plugin == 'nvfp4':
fp4_x, act_per_block_scale = quantize_to_fp4_tensor(
x, div(1, self.activation_global_scaling_factor.value))
else:
fp4_x, act_per_block_scale = dynamic_quantize(
x, self.activation_global_scaling_factor.value)
if default_net().plugin_config.gemm_plugin == 'nvfp4':
x = fp4_gemm(fp4_x, act_per_block_scale, self.weight.value,
self.weights_block_scaling_factor_interleaved.value,
self.alpha.value, self.dtype)
else:
quant_w = self.weight.value
scale_w = self.weights_block_scaling_factor.value
dequant_w = block_double_dequantize(
quant_w,
scale_w,
self.weights_global_scaling_factor.value,
dtype=trt.float16)
dequant_x = block_double_dequantize(
fp4_x,
act_per_block_scale,
self.activation_global_scaling_factor.value,
dtype=trt.float16)
x = matmul(dequant_x, dequant_w, transb=True).cast(self.dtype)
if self.bias is not None:
x = x + self.bias.value
if self.gather_output and self.tp_size > 1 and self.tp_group is not None:
# [dim0, local_dim] -> [dim0 * tp_size, local_dim] --> [dim0, local_dim * tp_size]
x = allgather(x, self.tp_group, gather_dim=1)
return x
def postprocess(self, tllm_key, weights, **kwargs):
if not any([
tllm_key.endswith(suffix)
for suffix in self.tllm_to_externel_key_dict
]):
return super().postprocess(tllm_key, weights, **kwargs)
if self.is_qkv:
if tllm_key.endswith("weight"):
assert len(weights) == 12
qkv_weight = weights[0::4]
qkv_block_scale = weights[1::4]
qkv_global_scale = weights[2::4]
qkv_input_scale = weights[3::4]
# Ckpt uses qkv max is guaranteed. So no need to re-quantize.
qkv_input_scale = max(qkv_input_scale).reshape(1, ).to(
torch.float32)
qkv_global_scale = max(qkv_global_scale).reshape(1, ).to(
torch.float32)
qkv_weight = torch.cat(qkv_weight)
qkv_block_scale = torch.cat(qkv_block_scale)
return {
tllm_key:
qkv_weight,
tllm_key.replace("weight", "weights_block_scaling_factor"):
qkv_block_scale,
tllm_key.replace(
'weight', "weights_block_scaling_factor_interleaved"):
torch.ops.tensorrt_llm.nvfp4_block_scale_interleave(
qkv_block_scale.view(
torch.uint8).cpu().contiguous()).reshape(
qkv_block_scale.shape).view(
torch.float8_e4m3fn),
tllm_key.replace("weight", "weights_global_scaling_factor"):
qkv_global_scale,
tllm_key.replace("weight", "activation_global_scaling_factor"):
qkv_input_scale,
tllm_key.replace("weight", "alpha"):
qkv_input_scale * qkv_global_scale,
}
else:
return {}
else:
if tllm_key.endswith("weight"):
return weights
elif tllm_key.endswith("weights_block_scaling_factor"):
return weights
elif tllm_key.endswith("weights_block_scaling_factor_interleaved"):
return torch.ops.tensorrt_llm.nvfp4_block_scale_interleave(
weights.view(torch.uint8).cpu().contiguous()).reshape(
weights.shape).view(torch.float8_e4m3fn)
elif tllm_key.endswith("weights_global_scaling_factor"):
return weights.float()
elif tllm_key.endswith('activation_global_scaling_factor'):
return weights.float()
elif tllm_key.endswith('alpha'):
weight_global_sf = weights[0].float()
act_global_sf = weights[1].float()
return act_global_sf * weight_global_sf
class FP4RowLinear(RowLinear):
def __init__(
self,
in_features,
out_features,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
):
super().__init__(in_features,
out_features,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size)
self.scaling_vector_size = 16
assert self.in_features % self.scaling_vector_size == 0, \
"Input features must be a multiple of 16 for FP4 GEMM"
self.weight = Parameter(shape=(self.out_features, self.in_features),
dtype=trt.fp4)
self.weights_block_scaling_factor = Parameter(
shape=(self.out_features,
self.in_features // self.scaling_vector_size),
dtype=trt.fp8)
self.weights_block_scaling_factor_interleaved = Parameter(
shape=(self.out_features,
self.in_features // self.scaling_vector_size),
dtype=trt.fp8)
self.weights_global_scaling_factor = Parameter(shape=(1, ),
dtype=trt.float32)
self.activation_global_scaling_factor = Parameter(shape=(1, ),
dtype=trt.float32)
# alpha = 1.0 / (weight_global_scale * act_global_scale)
self.alpha = Parameter(shape=(1, ), dtype=trt.float32)
self.tllm_to_externel_key_dict = {
"weight": ["weight"],
"weights_block_scaling_factor": "weight_scale",
"weights_block_scaling_factor_interleaved": "weight_scale",
"weights_global_scaling_factor": "weight_scale_2",
"activation_global_scaling_factor": "input_scale",
"alpha": ["weight_scale_2", "input_scale"],
}
def forward(self, x, lora_runtime_params=None, all_reduce_params=None):
assert lora_runtime_params is None, "lora is not supported on FP4Linear now"
if isinstance(x, (tuple, list)):
fp4_x, act_per_block_scale = x
else:
if default_net().plugin_config.gemm_plugin == "nvfp4":
fp4_x, act_per_block_scale = quantize_to_fp4_tensor(
x, div(1.0, self.activation_global_scaling_factor.value))
else:
# WAR for FP8 output attention
if x.dtype == trt.fp8:
# Since the scale is NVFP4 scale, we need to make it back to fp8 scale
new_scale_factor = self.activation_global_scaling_factor.raw_value
new_scale_factor = constant(new_scale_factor * 6)
x = dequantize(x, new_scale_factor, 0,
new_scale_factor.dtype)
fp4_x, act_per_block_scale = dynamic_quantize(
x, self.activation_global_scaling_factor.value)
if default_net().plugin_config.gemm_allreduce_plugin:
x = gemm_allreduce(
a=fp4_x,
b=self.weight.value,
a_sf=act_per_block_scale,
b_sf=self.weights_block_scaling_factor_interleaved.value,
transa=False, # row-major
transb=True, # col-major
alpha=self.alpha.value,
group=self.tp_group, # ranks participating
fp8_inputs_override=False)
else:
if default_net().plugin_config.gemm_plugin == "nvfp4":
x = fp4_gemm(
fp4_x, act_per_block_scale, self.weight.value,
self.weights_block_scaling_factor_interleaved.value,
self.alpha.value, self.dtype)
else:
quant_w = self.weight.value
scale_w = self.weights_block_scaling_factor.value
dequant_x = block_double_dequantize(
fp4_x, act_per_block_scale,
self.activation_global_scaling_factor.value, trt.float16)
dequant_w = block_double_dequantize(
quant_w, scale_w, self.weights_global_scaling_factor.value,
trt.float16)
x = matmul(dequant_x, dequant_w, transb=True).cast(self.dtype)
if self.tp_size > 1 and self.tp_group is not None:
need_bias = self.bias is not None
fuse_bias_into_all_reduce = need_bias and (
all_reduce_params
is not None) and (all_reduce_params.fusion_op
== AllReduceFusionOp.RESIDUAL_RMS_NORM)
if fuse_bias_into_all_reduce:
all_reduce_params.bias = self.bias.value
x = allreduce(x,
self.tp_group,
all_reduce_params=all_reduce_params)
if need_bias and not fuse_bias_into_all_reduce:
x = x + self.bias.value
return x
if self.bias is not None:
x = x + self.bias.value
return x
def postprocess(self, tllm_key, weights, **kwargs):
if not any([
tllm_key.endswith(suffix)
for suffix in self.tllm_to_externel_key_dict
]):
return super().postprocess(tllm_key, weights, **kwargs)
if tllm_key.endswith("weight"):
return weights
elif tllm_key.endswith("weights_block_scaling_factor"):
return weights
elif tllm_key.endswith("weights_block_scaling_factor_interleaved"):
return torch.ops.tensorrt_llm.nvfp4_block_scale_interleave(
weights.view(torch.uint8).cpu().contiguous()).reshape(
weights.shape).view(torch.float8_e4m3fn)
elif tllm_key.endswith("weights_global_scaling_factor"):
return weights.float()
elif tllm_key.endswith('activation_global_scaling_factor'):
return weights.float()
elif tllm_key.endswith('alpha'):
weight_global_sf = weights[0].float()
act_global_sf = weights[1].float()
return act_global_sf * weight_global_sf
class SmoothQuantGatedMLP(SmoothQuantMLP):
def __init__(
self,
hidden_size,
ffn_hidden_size,
hidden_act,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
):
super().__init__(hidden_size,
ffn_hidden_size,
hidden_act,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
quant_mode=quant_mode)
if hidden_act not in ACT2FN:
raise ValueError(
'unsupported activation function: {}'.format(hidden_act))
self.gate = SmoothQuantColumnLinear(hidden_size,
ffn_hidden_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False,
quant_mode=quant_mode)
if self.quant_mode.has_act_static_scaling():
self.quantization_scaling_factor = Parameter(shape=(1, ),
dtype='float32')
else:
self.register_parameter('quantization_scaling_factor', None)
def forward(self, hidden_states, lora_layer_params=None):
assert lora_layer_params is None, f"lora is not supported on {self.__class__.__name__} now"
inter = self.fc(hidden_states)
inter = ACT2FN[self.hidden_act](inter)
gate = self.gate(hidden_states)
inter_x_gate = inter * gate
smoother = cast(self.proj.smoother.value, self.dtype)
inter_x_gate = inter_x_gate / smoother
if self.quant_mode.has_act_and_weight_quant():
if self.quant_mode.has_act_static_scaling():
# Avoid quantization layers as it breaks int8 plugins
inter_x_gate = quantize_tensor(
inter_x_gate, self.quantization_scaling_factor.value)
else:
# Quantize per token outputs tuple:
# quantized tensor and scaling factors per token
inter_x_gate = quantize_per_token(inter_x_gate)
output = self.proj(inter_x_gate)
return output
class SmoothQuantAttention(Module):
def __init__(self,
*,
local_layer_idx,
hidden_size,
num_attention_heads,
num_kv_heads=None,
max_position_embeddings=1024,
num_layers=1,
apply_query_key_layer_scaling=False,
attention_head_size=None,
attention_mask_type=AttentionMaskType.padding,
bias=True,
dense_bias=None,
dtype=None,
position_embedding_type=PositionEmbeddingType.learned_absolute,
rotary_embedding_base=10000.0,
rotary_embedding_scaling=None,
rotary_embedding_percentage=1.0,
tp_group=None,
tp_size=1,
tp_rank=0,
scale_alibi_bias=False,
paged_kv_cache=False,
quant_mode=QuantMode(0)):
super().__init__()
self.local_layer_idx = local_layer_idx
self.attention_mask_type = attention_mask_type
self.attention_head_size = hidden_size // num_attention_heads if attention_head_size is None else attention_head_size
self.num_attention_heads = num_attention_heads // tp_size
self.num_kv_heads = num_kv_heads
self.num_attention_kv_heads = (
num_kv_heads + tp_size - 1
) // tp_size if num_kv_heads is not None else self.num_attention_heads
self.hidden_size = hidden_size // tp_size
self.max_position_embeddings = 0 if max_position_embeddings is None else max_position_embeddings
self.tp_size = tp_size
self.tp_rank = tp_rank
self.dense_bias = dense_bias
if dense_bias is None:
self.dense_bias = bias
self.num_layers = num_layers
self.apply_query_key_layer_scaling = apply_query_key_layer_scaling
self.norm_factor = math.sqrt(self.attention_head_size)
self.q_scaling = 1
if self.apply_query_key_layer_scaling:
self.norm_factor *= self.num_layers
self.q_scaling *= self.num_layers
# Whether to scale ALiBi bias. Mathematically, it's equivalent to
# normalizing QK after adding bias.
# - False, inv_sqrt_Dh * Q*K^T + alibi_bias
# - True, inv_sqrt_Dh * Q*K^T + inv_sqrt_Dh * alibi_bias
self.scale_alibi_bias = scale_alibi_bias
self.position_embedding_type = position_embedding_type
self.paged_kv_cache = paged_kv_cache
self.rotary_embedding_base = rotary_embedding_base
self.rotary_embedding_scale_type = RotaryScalingType.none
self.rotary_embedding_scale = 1.0
self.rotary_embedding_dim = 0
if rotary_embedding_scaling is not None:
rotary_scaling_type = rotary_embedding_scaling.get(
"type", rotary_embedding_scaling.get("rope_type"))
self.rotary_embedding_scale_type = RotaryScalingType.from_string(
rotary_scaling_type)
self.rotary_embedding_scale = rotary_embedding_scaling.get(
"factor", 1.0)
if self.position_embedding_type.is_rope():
self.rotary_embedding_dim = int(self.attention_head_size *
rotary_embedding_percentage)
elif self.position_embedding_type.is_alibi():
alibi_scale = 1. / self.norm_factor if self.scale_alibi_bias else 1.
alibi_slopes = generate_alibi_slopes(self.num_attention_heads *
self.tp_size,
tp_size=self.tp_size,
tp_rank=self.tp_rank,
alibi_scale=alibi_scale)
self.register_parameter(
'alibi_slopes',
Parameter(alibi_slopes, dtype='float32', is_buffer=True))
self.quant_mode = quant_mode
self.dtype = dtype
if self.quant_mode.has_act_static_scaling():
self.quantization_scaling_factor = Parameter(shape=(1, ),
dtype='float32')
else:
self.register_parameter('quantization_scaling_factor', None)
qkv_quant_mode = quant_mode
if self.quant_mode.has_act_and_weight_quant():
# We need to hijack quant_mode for QKV because QKV always uses per channel scaling
qkv_quant_mode = QuantMode.from_description(
True, True, quant_mode.has_per_token_dynamic_scaling(), True)
self.register_parameter('kv_cache_scaling_factor', None)
self.qkv = SmoothQuantColumnLinear(
hidden_size,
tp_size * self.num_attention_heads * self.attention_head_size +
(2 * tp_size * self.num_attention_kv_heads *
self.attention_head_size),
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False,
quant_mode=qkv_quant_mode)
self.dense = SmoothQuantRowLinear(tp_size * self.num_attention_heads *
self.attention_head_size,
hidden_size,
bias=self.dense_bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
quant_mode=quant_mode)
self.use_lora = False
def forward(
self,
hidden_states: Tensor,
attention_mask=None,
use_cache=False,
kv_cache_params=None,
attention_params=None,
spec_decoding_params=None,
mrope_params=None,
encoder_output=None,
position_embedding=None,
norm_before_bmm1=False,
lora_layer_params=None,
all_reduce_params: Optional[AllReduceParams] = None,
):
assert lora_layer_params is None, f"lora is not supported on {self.__class__.__name__} now"
qkv = self.qkv(hidden_states)
alibi_slopes = None
if self.position_embedding_type == PositionEmbeddingType.alibi:
alibi_slopes = self.alibi_slopes.value
dtype = trt.float32
if default_net().plugin_config.gpt_attention_plugin or default_net(
).plugin_config.inflight_batching_gpt_attention_plugin:
dtype = hidden_states.dtype if self.quant_mode.has_act_static_scaling(
) else hidden_states[0].dtype
if dtype == trt.int8:
dtype = trt.float16
alibi_slopes = cast(alibi_slopes, dtype)
if spec_decoding_params is None:
spec_decoding_params = SpecDecodingParams()
if mrope_params is None:
mrope_params = MropeParams()
if default_net().plugin_config.gpt_attention_plugin:
assert attention_params.is_valid(
default_net().plugin_config.gpt_attention_plugin,
default_net().plugin_config.remove_input_padding, use_cache)
if use_cache:
assert kv_cache_params.is_valid(
default_net().plugin_config.gpt_attention_plugin)
assert self.attention_mask_type == AttentionMaskType.causal, \
'Plugin only support masked MHA.'
if self.kv_cache_scaling_factor is not None:
kv_orig_quant_scale = self.kv_cache_rcp_scaling_factor.value
kv_quant_orig_scale = self.kv_cache_scaling_factor.value
else:
kv_orig_quant_scale = None
kv_quant_orig_scale = None
if self.position_embedding_type.is_rope():
rotary_inv_freq = attention_params.rotary_inv_freq
rotary_cos_sin = attention_params.embed_positions_for_gpt_attention
else:
rotary_inv_freq = None
rotary_cos_sin = None
context, past_key_value = gpt_attention(
qkv=qkv,
past_key_value=kv_cache_params.get_first_past_key_value(),
sequence_length=attention_params.sequence_length,
host_past_key_value_lengths=kv_cache_params.
host_past_key_value_lengths,
host_max_attention_window_sizes=kv_cache_params.
host_max_attention_window_sizes,
host_sink_token_length=kv_cache_params.host_sink_token_length,
context_lengths=attention_params.context_lengths,
cache_indirection=kv_cache_params.cache_indirection,
host_request_types=attention_params.host_request_types,
layer_idx=self.local_layer_idx,
num_heads=self.num_attention_heads,
num_kv_heads=self.num_attention_kv_heads,
num_kv_heads_origin=self.num_kv_heads,
hidden_size_per_head=self.attention_head_size,
q_scaling=self.q_scaling,
rotary_embedding_dim=self.rotary_embedding_dim,
rotary_embedding_base=self.rotary_embedding_base,
rotary_embedding_scale_type=self.rotary_embedding_scale_type,
rotary_embedding_scale=self.rotary_embedding_scale,
rotary_embedding_max_positions=self.max_position_embeddings,
position_embedding_type=self.position_embedding_type,
rotary_inv_freq=rotary_inv_freq,
rotary_cos_sin=rotary_cos_sin,
kv_orig_quant_scale=kv_orig_quant_scale,
kv_quant_orig_scale=kv_quant_orig_scale,
kv_cache_quant_mode=self.quant_mode,
max_context_length=attention_params.max_context_length,
alibi_slopes=alibi_slopes,
tp_size=self.tp_size,
tp_rank=self.tp_rank,
kv_cache_block_offsets=kv_cache_params.kv_cache_block_offsets,
host_kv_cache_block_offsets=kv_cache_params.
host_kv_cache_block_offsets,
host_kv_cache_pool_pointers=kv_cache_params.
host_kv_cache_pool_pointers,
host_kv_cache_pool_mapping=kv_cache_params.
host_kv_cache_pool_mapping,
host_context_lengths=attention_params.host_context_lengths,
use_cache=use_cache,
spec_decoding_generation_lengths=spec_decoding_params.
spec_decoding_generation_lengths,
spec_decoding_position_offsets=spec_decoding_params.
spec_decoding_position_offsets,
spec_decoding_packed_mask=spec_decoding_params.
spec_decoding_packed_mask,
mrope_rotary_cos_sin=mrope_params.mrope_rotary_cos_sin,
mrope_position_deltas=mrope_params.mrope_position_deltas,
host_runtime_perf_knobs=attention_params.
host_runtime_perf_knobs,
host_context_progress=attention_params.host_context_progress,
)
else:
assert self.paged_kv_cache == False
def transpose_for_scores(x):
new_x_shape = concat([
shape(x, 0),
shape(x, 1), self.num_attention_heads,
self.attention_head_size
])
return x.view(new_x_shape).permute([0, 2, 1, 3])
query, key, value = split(qkv, self.hidden_size, dim=2)
query = transpose_for_scores(query)
key = transpose_for_scores(key)
value = transpose_for_scores(value)
past_key_value = kv_cache_params.get_first_past_key_value()
if past_key_value is not None:
def dequantize_tensor(x, scale):
# Cast from int8 to dtype
casted_x = cast(x, self.dtype)
return casted_x * scale
if self.quant_mode.has_int8_kv_cache():
past_key_value = dequantize_tensor(
past_key_value, self.kv_dequantization_scale.value)
# past_key_value [bs, 2, num_heads, max_seq_len, head_dim]
past_key, past_value = split(past_key_value, 1, dim=1)
key_shape = concat([
shape(past_key, 0),
shape(past_key, 2),
shape(past_key, 3),
shape(past_key, 4)
])
past_key = past_key.view(key_shape, zero_is_placeholder=False)
past_value = past_value.view(key_shape,
zero_is_placeholder=False)
key = concat([past_key, key], dim=2)
value = concat([past_value, value], dim=2)
def merge_caches():
key_inflated_shape = concat([
shape(key, 0), 1,
shape(key, 1),
shape(key, 2),
shape(key, 3)
])
inflated_key = key.view(key_inflated_shape,
zero_is_placeholder=False)
inflated_value = value.view(key_inflated_shape,
zero_is_placeholder=False)
past_key_value = concat([inflated_key, inflated_value], dim=1)
return past_key_value
if self.attention_mask_type == AttentionMaskType.causal:
query_length = shape(query, 2)
key_length = shape(key, 2)
starts = concat([0, 0, key_length - query_length, 0])
sizes = concat([1, 1, query_length, key_length])
buffer = constant(
np.expand_dims(
np.tril(
np.ones(
(self.max_position_embeddings,
self.max_position_embeddings))).astype(bool),
(0, 1)))
causal_mask = slice(buffer, starts, sizes)
key = key.permute([0, 1, 3, 2])
with precision("float32"):
attention_scores = matmul(query, key)
if self.attention_mask_type == AttentionMaskType.causal:
attention_scores = where(causal_mask, attention_scores,
-10000.0)
attention_scores = attention_scores / self.norm_factor
attention_probs = softmax(attention_scores, dim=-1)
context = matmul(attention_probs, value,
use_fp32_acc=False).permute([0, 2, 1, 3])
context = context.view(
concat([shape(context, 0),
shape(context, 1), self.hidden_size]))
past_key_value = merge_caches()
if use_cache and self.quant_mode.has_int8_kv_cache():
past_key_value = quantize_tensor(
past_key_value, self.kv_quantization_scale.value)
value = cast(self.dense.smoother.value, context.dtype)
context = context / value
if self.quant_mode.has_act_and_weight_quant():
if self.quant_mode.has_act_static_scaling():
# Avoid quantization layers as it breaks int8 plugins
context = quantize_tensor(
context, self.quantization_scaling_factor.value)
else:
# Quantize per token outputs tuple:
# quantized tensor and scaling factors per token
context = quantize_per_token(context)
context = self.dense(
context,
all_reduce_params=all_reduce_params,
)
if use_cache:
return (context, past_key_value)
return context
# TODO: Duplicates SmoothQuantRmsNorm
class QServeRmsNorm(Module):
def __init__(self,
normalized_shape,
eps=1e-06,
elementwise_affine=False,
dtype=None,
quant_mode=QuantMode(0),
bias=False,
clamp_val=None):
super().__init__()
if isinstance(normalized_shape, int):
normalized_shape = (normalized_shape, )
assert quant_mode.is_qserve_w4a8()
self.normalized_shape = tuple(normalized_shape)
self.elementwise_affine = elementwise_affine
if self.elementwise_affine:
self.weight = Parameter(shape=self.normalized_shape, dtype=dtype)
else:
self.register_parameter('weight', None)
if bias:
self.bias = Parameter(shape=self.normalized_shape, dtype=dtype)
else:
self.register_parameter('bias', None)
if clamp_val:
if not (isinstance(clamp_val, list) and len(clamp_val) == 2):
raise ValueError(f'unsupported clamp_val {clamp_val}')
self.clamp_val = Parameter(np.array(clamp_val, dtype=np.float32),
dtype='float32',
is_buffer=True)
else:
self.register_parameter('clamp_val', None)
self.eps = eps
self.dtype = dtype
self.quant_mode = quant_mode
if self.quant_mode.has_act_and_weight_quant():
self.scale_to_int = Parameter(shape=(1, ), dtype=dtype)
else:
self.register_parameter('scale_to_int', None)
def forward(self, x):
weight = None if self.weight is None else self.weight.value
bias = None if self.bias is None else self.bias.value
scale = None if self.scale_to_int is None else self.scale_to_int.value
clamp_val = None if self.clamp_val is None else self.clamp_val.value
return smooth_quant_rms_norm(
x,
self.normalized_shape,
weight,
bias,
scale,
clamp_val,
self.eps,
dynamic_act_scaling=True,
scale_dtype='float16',
sum_per_token=not self.quant_mode.has_per_group_scaling(),
sum_dtype='float16')
# TODO: Mostly duplicates SmoothQuantMLP.
# TODO: MLP could represent GatedMLP if hidden_act=='swiglu'.
class QServeMLP(Module):
def __init__(
self,
hidden_size,
ffn_hidden_size,
hidden_act,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
):
super().__init__()
if hidden_act not in ACT2FN:
raise ValueError(
'unsupported activation function: {}'.format(hidden_act))
fc_output_size = 2 * ffn_hidden_size if hidden_act == 'swiglu' else ffn_hidden_size
self.fc = QServeW4A8ColumnLinear(hidden_size,
fc_output_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False,
quant_mode=quant_mode)
self.proj = QServeW4A8RowLinear(ffn_hidden_size,
hidden_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
quant_mode=quant_mode)
self.hidden_act = hidden_act
self.quant_mode = quant_mode
self.dtype = dtype
def forward(self, hidden_states):
inter = self.fc(hidden_states)
inter = ACT2FN[self.hidden_act](inter)
inter = quantize_per_token(
inter,
scale_dtype='float16',
sum_per_token=not self.quant_mode.has_per_group_scaling(),
sum_dtype='float16')
output = self.proj(inter)
return output
# TODO: Mostly duplicates SmoothQuantGatedMLP.
class QServeGatedMLP(QServeMLP):
def __init__(
self,
hidden_size,
ffn_hidden_size,
hidden_act,
bias=True,
dtype=None,
tp_group=None,
tp_size=1,
quant_mode=QuantMode(0),
):
super().__init__(hidden_size,
ffn_hidden_size,
hidden_act,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
quant_mode=quant_mode)
if hidden_act not in ACT2FN:
raise ValueError(
'unsupported activation function: {}'.format(hidden_act))
self.gate = QServeW4A8Linear(hidden_size,
ffn_hidden_size,
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False,
quant_mode=quant_mode)
def forward(self, hidden_states, lora_layer_params=None):
assert lora_layer_params is None, "lora_layer_params not supported"
inter = self.fc(hidden_states)
inter = ACT2FN[self.hidden_act](inter)
gate = self.gate(hidden_states)
inter_x_gate = inter * gate
inter_x_gate = quantize_per_token(
inter_x_gate,
scale_dtype='float16',
sum_per_token=not self.quant_mode.has_per_group_scaling(),
sum_dtype='float16')
output = self.proj(inter_x_gate)
return output
# TODO: Duplicates SmoothQuantAttention.
class QServeAttention(Module):
def __init__(self,
*,
local_layer_idx,
hidden_size,
num_attention_heads,
num_kv_heads=None,
max_position_embeddings=1024,
num_layers=1,
apply_query_key_layer_scaling=False,
attention_head_size=None,
attention_mask_type=AttentionMaskType.padding,
bias=True,
dense_bias=None,
dtype=None,
position_embedding_type=PositionEmbeddingType.learned_absolute,
rotary_embedding_base=10000.0,
rotary_embedding_scaling=None,
rotary_embedding_percentage=1.0,
tp_group=None,
tp_size=1,
tp_rank=0,
scale_alibi_bias=False,
paged_kv_cache=False,
quant_mode=QuantMode(0)):
super().__init__()
self.local_layer_idx = local_layer_idx
self.attention_mask_type = attention_mask_type
self.attention_head_size = hidden_size // num_attention_heads if attention_head_size is None else attention_head_size
self.num_attention_heads = num_attention_heads // tp_size
self.num_kv_heads = num_kv_heads
self.num_attention_kv_heads = (
num_kv_heads + tp_size - 1
) // tp_size if num_kv_heads is not None else self.num_attention_heads
self.hidden_size = hidden_size // tp_size
self.max_position_embeddings = 0 if max_position_embeddings is None else max_position_embeddings
self.tp_size = tp_size
self.tp_rank = tp_rank
self.dense_bias = dense_bias
if dense_bias is None:
self.dense_bias = bias
self.num_layers = num_layers
self.apply_query_key_layer_scaling = apply_query_key_layer_scaling
self.norm_factor = math.sqrt(self.attention_head_size)
self.q_scaling = 1
if self.apply_query_key_layer_scaling:
self.norm_factor *= self.num_layers
self.q_scaling *= self.num_layers
# Whether to scale ALiBi bias. Mathematically, it's equivalent to
# normalizing QK after adding bias.
# - False, inv_sqrt_Dh * Q*K^T + alibi_bias
# - True, inv_sqrt_Dh * Q*K^T + inv_sqrt_Dh * alibi_bias
self.scale_alibi_bias = scale_alibi_bias
self.position_embedding_type = position_embedding_type
self.paged_kv_cache = paged_kv_cache
self.rotary_embedding_base = rotary_embedding_base
self.rotary_embedding_scale_type = RotaryScalingType.none
self.rotary_embedding_scale = 1.0
self.rotary_embedding_dim = 0
if rotary_embedding_scaling is not None:
rotary_scaling_type = rotary_embedding_scaling.get(
"type", rotary_embedding_scaling.get("rope_type"))
self.rotary_embedding_scale_type = RotaryScalingType.from_string(
rotary_scaling_type)
self.rotary_embedding_scale = rotary_embedding_scaling.get(
"factor", 1.0)
if self.position_embedding_type.is_rope():
self.rotary_embedding_dim = int(self.attention_head_size *
rotary_embedding_percentage)
elif self.position_embedding_type.is_alibi():
alibi_scale = 1. / self.norm_factor if self.scale_alibi_bias else 1.
alibi_slopes = generate_alibi_slopes(self.num_attention_heads *
self.tp_size,
tp_size=self.tp_size,
tp_rank=self.tp_rank,
alibi_scale=alibi_scale)
self.register_parameter(
'alibi_slopes',
Parameter(alibi_slopes, dtype='float32', is_buffer=True))
self.quant_mode = quant_mode
self.dtype = dtype
# QServe does not use act static scaling
# if self.quant_mode.has_act_static_scaling():
# self.quantization_scaling_factor = Parameter(shape=(1, ),
# dtype='float32')
# else:
# self.register_parameter('quantization_scaling_factor', None)
qkv_quant_mode = quant_mode
if self.quant_mode.has_act_and_weight_quant():
# We need to hijack quant_mode for QKV because QKV always uses per channel scaling
qkv_quant_mode = QuantMode.from_description(
True, True, quant_mode.has_per_token_dynamic_scaling(), True)
self.register_parameter('kv_cache_scaling_factor', None)
self.qkv = QServeW4A8ColumnLinear(
hidden_size,
tp_size * self.num_attention_heads * self.attention_head_size +
(2 * tp_size * self.num_attention_kv_heads *
self.attention_head_size),
bias=bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
gather_output=False,
quant_mode=qkv_quant_mode)
self.dense = QServeW4A8RowLinear(tp_size * self.num_attention_heads *
self.attention_head_size,
hidden_size,
bias=self.dense_bias,
dtype=dtype,
tp_group=tp_group,
tp_size=tp_size,
quant_mode=quant_mode)
self.use_lora = False
def forward(
self,
hidden_states: Tensor,
attention_mask=None,
use_cache=False,
kv_cache_params=None,
attention_params=None,
spec_decoding_params=None,
encoder_output=None,
position_embedding=None,
norm_before_bmm1=False,
lora_layer_params=None,
all_reduce_params: Optional[AllReduceParams] = None,
):
assert lora_layer_params is None, "lora is not supported on SmoothQuantAttention now"
if default_net().plugin_config.qserve_gemm_plugin:
qkv = self.qkv(hidden_states)
else:
raise ValueError("qserve_gemm_plugin is not set")
alibi_slopes = None
if self.position_embedding_type == PositionEmbeddingType.alibi:
alibi_slopes = self.alibi_slopes.value
dtype = trt.float32
if default_net().plugin_config.gpt_attention_plugin or default_net(
).plugin_config.inflight_batching_gpt_attention_plugin:
dtype = hidden_states.dtype if self.quant_mode.has_act_static_scaling(
) else hidden_states[0].dtype
if dtype == trt.int8:
dtype = trt.float16
alibi_slopes = cast(alibi_slopes, dtype)
if spec_decoding_params is None:
spec_decoding_params = SpecDecodingParams()
if default_net().plugin_config.gpt_attention_plugin:
assert attention_params.is_valid(
default_net().plugin_config.gpt_attention_plugin,
default_net().plugin_config.remove_input_padding, use_cache)
if use_cache:
assert kv_cache_params.is_valid(
default_net().plugin_config.gpt_attention_plugin)
assert self.attention_mask_type == AttentionMaskType.causal, \
'Plugin only support masked MHA.'
if self.kv_cache_scaling_factor is not None:
kv_orig_quant_scale = self.kv_cache_rcp_scaling_factor.value
kv_quant_orig_scale = self.kv_cache_scaling_factor.value
else:
kv_orig_quant_scale = None
kv_quant_orig_scale = None
if self.position_embedding_type.is_rope():
rotary_inv_freq = attention_params.rotary_inv_freq
rotary_cos_sin = attention_params.embed_positions_for_gpt_attention
else:
rotary_inv_freq = None
rotary_cos_sin = None
context, past_key_value = gpt_attention(
qkv=qkv,
past_key_value=kv_cache_params.get_first_past_key_value(),
sequence_length=attention_params.sequence_length,
host_past_key_value_lengths=kv_cache_params.
host_past_key_value_lengths,
host_max_attention_window_sizes=kv_cache_params.
host_max_attention_window_sizes,
host_sink_token_length=kv_cache_params.host_sink_token_length,
context_lengths=attention_params.context_lengths,
cache_indirection=kv_cache_params.cache_indirection,
host_request_types=attention_params.host_request_types,
layer_idx=self.local_layer_idx,
num_heads=self.num_attention_heads,
num_kv_heads=self.num_attention_kv_heads,
num_kv_heads_origin=self.num_kv_heads,
hidden_size_per_head=self.attention_head_size,
q_scaling=self.q_scaling,
rotary_embedding_dim=self.rotary_embedding_dim,
rotary_embedding_base=self.rotary_embedding_base,
rotary_embedding_scale_type=self.rotary_embedding_scale_type,
rotary_embedding_scale=self.rotary_embedding_scale,
rotary_embedding_max_positions=self.max_position_embeddings,
position_embedding_type=self.position_embedding_type,
rotary_inv_freq=rotary_inv_freq,
rotary_cos_sin=rotary_cos_sin,
kv_orig_quant_scale=kv_orig_quant_scale,
kv_quant_orig_scale=kv_quant_orig_scale,
kv_cache_quant_mode=self.quant_mode,
max_context_length=attention_params.max_context_length,
alibi_slopes=alibi_slopes,
tp_size=self.tp_size,
tp_rank=self.tp_rank,
kv_cache_block_offsets=kv_cache_params.kv_cache_block_offsets,
host_kv_cache_block_offsets=kv_cache_params.
host_kv_cache_block_offsets,
host_kv_cache_pool_pointers=kv_cache_params.
host_kv_cache_pool_pointers,
host_kv_cache_pool_mapping=kv_cache_params.
host_kv_cache_pool_mapping,
host_context_lengths=attention_params.host_context_lengths,
use_cache=use_cache,
spec_decoding_generation_lengths=spec_decoding_params.
spec_decoding_generation_lengths,
spec_decoding_position_offsets=spec_decoding_params.
spec_decoding_position_offsets,
spec_decoding_packed_mask=spec_decoding_params.
spec_decoding_packed_mask,
host_runtime_perf_knobs=attention_params.
host_runtime_perf_knobs,
host_context_progress=attention_params.host_context_progress)
else:
assert self.paged_kv_cache == False
def transpose_for_scores(x):
new_x_shape = concat([
shape(x, 0),
shape(x, 1), self.num_attention_heads,
self.attention_head_size
])
return x.view(new_x_shape).permute([0, 2, 1, 3])
query, key, value = split(qkv, self.hidden_size, dim=2)
query = transpose_for_scores(query)
key = transpose_for_scores(key)
value = transpose_for_scores(value)
past_key_value = kv_cache_params.get_first_past_key_value()
if past_key_value is not None:
def dequantize_tensor(x, scale):
# Cast from int8 to dtype
casted_x = cast(x, self.dtype)
return casted_x * scale
if self.quant_mode.has_int8_kv_cache():
past_key_value = dequantize_tensor(
past_key_value, self.kv_dequantization_scale.value)
# past_key_value [bs, 2, num_heads, max_seq_len, head_dim]
past_key, past_value = split(past_key_value, 1, dim=1)
key_shape = concat([
shape(past_key, 0),
shape(past_key, 2),
shape(past_key, 3),
shape(past_key, 4)
])
past_key = past_key.view(key_shape, zero_is_placeholder=False)
past_value = past_value.view(key_shape,
zero_is_placeholder=False)
key = concat([past_key, key], dim=2)
value = concat([past_value, value], dim=2)
def merge_caches():
key_inflated_shape = concat([
shape(key, 0), 1,
shape(key, 1),
shape(key, 2),
shape(key, 3)
])
inflated_key = key.view(key_inflated_shape,
zero_is_placeholder=False)
inflated_value = value.view(key_inflated_shape,
zero_is_placeholder=False)
past_key_value = concat([inflated_key, inflated_value], dim=1)
return past_key_value
if self.attention_mask_type == AttentionMaskType.causal:
query_length = shape(query, 2)
key_length = shape(key, 2)
starts = concat([0, 0, key_length - query_length, 0])
sizes = concat([1, 1, query_length, key_length])
buffer = constant(
np.expand_dims(
np.tril(
np.ones(
(self.max_position_embeddings,
self.max_position_embeddings))).astype(bool),
(0, 1)))
causal_mask = slice(buffer, starts, sizes)
key = key.permute([0, 1, 3, 2])
with precision("float32"):
attention_scores = matmul(query, key)
if self.attention_mask_type == AttentionMaskType.causal:
attention_scores = where(causal_mask, attention_scores,
-10000.0)
attention_scores = attention_scores / self.norm_factor
attention_probs = softmax(attention_scores, dim=-1)
context = matmul(attention_probs, value,
use_fp32_acc=False).permute([0, 2, 1, 3])
context = context.view(
concat([shape(context, 0),
shape(context, 1), self.hidden_size]))
past_key_value = merge_caches()
if use_cache and self.quant_mode.has_int8_kv_cache():
past_key_value = quantize_tensor(
past_key_value, self.kv_quantization_scale.value)
# Quantize per token outputs tuple:
# quantized tensor and scaling factors per token
context = quantize_per_token(
context,
scale_dtype='float16',
sum_per_token=not self.quant_mode.has_per_group_scaling(),
sum_dtype='float16')
context = self.dense(context, all_reduce_params=all_reduce_params)
if use_cache:
return (context, past_key_value)
return context
Reference in New Issue View Git Blame Copy Permalink