File size: 12,961 Bytes
13a5289 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 | # Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
#
# 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.
from functools import wraps
from math import ceil
from typing import Optional
import torch
from compressed_tensors.quantization.quant_args import (
QuantizationArgs,
QuantizationStrategy,
round_to_quantized_type,
)
from compressed_tensors.quantization.quant_config import QuantizationStatus
from compressed_tensors.quantization.quant_scheme import QuantizationScheme
from compressed_tensors.quantization.utils import (
calculate_range,
compute_dynamic_scales_and_zp,
)
from compressed_tensors.utils import safe_permute
from torch.nn import Module
__all__ = [
"quantize",
"dequantize",
"fake_quantize",
"wrap_module_forward_quantized",
"forward_quantize",
]
@torch.no_grad()
def quantize(
x: torch.Tensor,
scale: torch.Tensor,
zero_point: torch.Tensor,
args: QuantizationArgs,
dtype: Optional[torch.dtype] = None,
g_idx: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
Quantize the input tensor x using the QuantizationStrategy specified in args.
Quantization can be done per tensor, channel, token or group. For group
quantization, the group_size must be divisible by the column size. The input scale
and zero_points are reshaped to support vectorization (Assumes 1 is the
channel dimension)
:param x: Input tensor
:param scale: scale tensor
:param zero_point: zero point tensor
:param args: quantization args dictating how to quantize x
:param dtype: optional dtype to cast the quantized output to
:param g_idx: optional mapping from column index to group index
:return: fake quantized tensor
"""
return _process_quantization(
x=x,
scale=scale,
zero_point=zero_point,
args=args,
dtype=dtype,
do_quantize=True,
do_dequantize=False,
g_idx=g_idx,
)
@torch.no_grad()
def dequantize(
x_q: torch.Tensor,
scale: torch.Tensor,
zero_point: Optional[torch.Tensor] = None,
args: Optional[QuantizationArgs] = None,
dtype: Optional[torch.dtype] = None,
g_idx: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
Dequantize a quantized input tensor x_q based on the strategy specified in args. If
args is not provided, the strategy will be inferred.
:param x: quantized input tensor
:param scale: scale tensor
:param zero_point: zero point tensor
:param args: quantization args used to quantize x_q
:param dtype: optional dtype to cast the dequantized output to
:param g_idx: optional mapping from column index to group index
:return: dequantized float tensor
"""
if args is None:
if scale.ndim == 0 or scale.ndim == 1:
args = QuantizationArgs(strategy=QuantizationStrategy.TENSOR)
elif scale.ndim == 2:
if scale.shape[1] == 1:
args = QuantizationArgs(strategy=QuantizationStrategy.CHANNEL)
else:
group_size = int(x_q.shape[1] / scale.shape[1])
args = QuantizationArgs(
strategy=QuantizationStrategy.GROUP, group_size=group_size
)
else:
raise ValueError(
f"Could not infer a quantization strategy from scale with {scale.ndim} "
"dimmensions. Expected 0 or 2 dimmensions."
)
if dtype is None:
dtype = scale.dtype
return _process_quantization(
x=x_q,
scale=scale,
zero_point=zero_point,
args=args,
do_quantize=False,
do_dequantize=True,
dtype=dtype,
g_idx=g_idx,
)
@torch.no_grad()
def fake_quantize(
x: torch.Tensor,
scale: torch.Tensor,
zero_point: torch.Tensor,
args: QuantizationArgs,
g_idx: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
Fake quantize the input tensor x by quantizing then dequantizing with
the QuantizationStrategy specified in args. Quantization can be done per tensor,
channel, token or group. For group quantization, the group_size must be divisible
by the column size. The input scale and zero_points are reshaped to support
vectorization (Assumes 1 is the channel dimension)
:param x: Input tensor
:param scale: scale tensor
:param zero_point: zero point tensor
:param args: quantization args dictating how to quantize x
:param g_idx: optional mapping from column index to group index
:return: fake quantized tensor
"""
return _process_quantization(
x=x,
scale=scale,
zero_point=zero_point,
args=args,
do_quantize=True,
do_dequantize=True,
g_idx=g_idx,
)
@torch.no_grad()
def _process_quantization(
x: torch.Tensor,
scale: torch.Tensor,
zero_point: torch.Tensor,
args: QuantizationArgs,
g_idx: Optional[torch.Tensor] = None,
dtype: Optional[torch.dtype] = None,
do_quantize: bool = True,
do_dequantize: bool = True,
) -> torch.Tensor:
q_min, q_max = calculate_range(args, x.device)
group_size = args.group_size
if args.strategy == QuantizationStrategy.GROUP:
output_dtype = dtype if dtype is not None else x.dtype
output = torch.zeros_like(x).to(output_dtype)
columns = output.shape[1]
# TODO: make validation step for inputs
while scale.ndim < 2:
# pad scale and zero point dims for slicing
scale = scale.unsqueeze(1)
zero_point = zero_point.unsqueeze(1) if zero_point is not None else None
if columns >= group_size:
if columns % group_size != 0:
raise ValueError(
"tensor column shape must be divisble "
f"by the given group_size {group_size}"
)
# support column-order (default) quantization as well as other orderings
# such as activation ordering. Below checks if g_idx has been initialized
is_column_order = g_idx is None or -1 in g_idx
if is_column_order:
num_groups = int(ceil(columns / group_size))
group_sizes = torch.full((num_groups,), group_size, dtype=torch.int)
else:
group_indices, group_sizes = torch.unique(g_idx, return_counts=True)
group_sizes = group_sizes[torch.argsort(group_indices)]
perm = torch.argsort(g_idx)
x = safe_permute(x, perm, dim=1)
# TODO: experiment with vectorizing for loop for performance
end = 0
for index, group_count in enumerate(group_sizes):
sc = scale[:, index].view(-1, 1)
zp = zero_point[:, index].view(-1, 1) if zero_point is not None else None
start = end
end = start + group_count
if do_quantize:
output[:, start:end] = _quantize(
x[:, start:end],
sc,
zp,
q_min,
q_max,
args,
dtype=dtype,
)
if do_dequantize:
input = output[:, start:end] if do_quantize else x[:, start:end]
output[:, start:end] = _dequantize(input, sc, zp)
if not is_column_order:
output = safe_permute(output, torch.argsort(perm), dim=1)
else: # covers channel, token and tensor strategies
if do_quantize:
output = _quantize(
x,
scale,
zero_point,
q_min,
q_max,
args,
dtype=dtype,
)
if do_dequantize:
output = _dequantize(output if do_quantize else x, scale, zero_point)
return output
def wrap_module_forward_quantized(module: Module, scheme: QuantizationScheme):
# expects a module already initialized and injected with the parameters in
# initialize_module_for_quantization
if hasattr(module.forward, "__func__"):
forward_func_orig = module.forward.__func__
else:
forward_func_orig = module.forward.func
@wraps(forward_func_orig) # ensures docstring, names, etc are propagated
def wrapped_forward(self, *args, **kwargs):
if not getattr(module, "quantization_enabled", True):
# quantization is disabled on forward passes, return baseline
# forward call
return forward_func_orig.__get__(module, module.__class__)(*args, **kwargs)
input_ = args[0]
compressed = module.quantization_status == QuantizationStatus.COMPRESSED
if scheme.input_activations is not None:
# prehook should calibrate activations before forward call
input_ = forward_quantize(module, input_, "input", scheme.input_activations)
if scheme.weights is not None and not compressed:
# calibrate and (fake) quantize weights when applicable
unquantized_weight = self.weight.data.clone()
self.weight.data = forward_quantize(
module, self.weight, "weight", scheme.weights
)
# perform wrapped forward call
output = forward_func_orig.__get__(module, module.__class__)(
input_, *args[1:], **kwargs
)
# restore back to unquantized_value
if scheme.weights is not None and not compressed:
self.weight.data = unquantized_weight
if scheme.output_activations is not None:
# forward-hook should calibrate/forward_quantize
if (
module.quantization_status == QuantizationStatus.CALIBRATION
and not scheme.output_activations.dynamic
):
return output
output = forward_quantize(
module, output, "output", scheme.output_activations
)
return output
# bind wrapped forward to module class so reference to `self` is correct
bound_wrapped_forward = wrapped_forward.__get__(module, module.__class__)
# set forward to wrapped forward
setattr(module, "forward", bound_wrapped_forward)
def forward_quantize(
module: Module, value: torch.Tensor, base_name: str, args: "QuantizationArgs"
) -> torch.Tensor:
# in compressed mode, the weight is already compressed and quantized so we don't
# need to run fake quantization
if (
module.quantization_status == QuantizationStatus.COMPRESSED
and base_name == "weight"
):
return value
if value.numel() == 0:
# if the tensor is empty,
# skip quantization
return value
g_idx = getattr(module, "weight_g_idx", None)
if args.dynamic:
# dynamic quantization - determine the scale/zp on the fly
scale, zero_point = compute_dynamic_scales_and_zp(value=value, args=args)
else:
# static quantization - get scale and zero point from layer
scale = getattr(module, f"{base_name}_scale")
zero_point = getattr(module, f"{base_name}_zero_point", None)
return fake_quantize(
x=value,
scale=scale,
zero_point=zero_point,
args=args,
g_idx=g_idx,
)
@torch.no_grad()
def _quantize(
x: torch.Tensor,
scale: torch.Tensor,
zero_point: torch.Tensor,
q_min: torch.Tensor,
q_max: torch.Tensor,
args: QuantizationArgs,
dtype: Optional[torch.dtype] = None,
) -> torch.Tensor:
scaled = x / scale
if zero_point is not None:
scaled += zero_point.to(x.dtype)
# clamp first because cast isn't guaranteed to be saturated (ie for fp8)
clamped_value = torch.clamp(
scaled,
q_min,
q_max,
)
quantized_value = round_to_quantized_type(clamped_value, args)
if dtype is not None:
quantized_value = quantized_value.to(dtype)
return quantized_value
@torch.no_grad()
def _dequantize(
x_q: torch.Tensor,
scale: torch.Tensor,
zero_point: torch.Tensor = None,
dtype: Optional[torch.dtype] = None,
) -> torch.Tensor:
dequant_value = x_q.to(scale.dtype)
if zero_point is not None:
dequant_value = dequant_value - zero_point.to(scale.dtype)
dequant_value = dequant_value * scale
if dtype is not None:
dequant_value = dequant_value.to(dtype)
return dequant_value
|