File size: 16,232 Bytes
76f9669 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 | # 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.
import logging
import math
from typing import Generator, Optional, Tuple
import torch
from compressed_tensors.quantization.quant_args import (
FP4_E2M1_DATA,
FP8_E4M3_DATA,
FloatArgs,
QuantizationArgs,
QuantizationStrategy,
QuantizationType,
round_to_quantized_type_dtype,
)
from compressed_tensors.quantization.quant_scheme import QuantizationScheme
from compressed_tensors.quantization.utils.mxfp4_utils import (
generate_mxfp4_scales,
maybe_convert_from_mxfp4_exp,
should_generatre_mxfp4_scales,
)
from compressed_tensors.utils import deprecated
from loguru import logger
from torch import FloatTensor, IntTensor, Tensor
from torch.nn import Module
__all__ = [
"is_module_quantized",
"is_model_quantized",
"module_type",
"get_torch_bit_depth",
"can_quantize",
"KV_CACHE_TARGETS",
"is_kv_cache_quant_scheme",
"iter_named_leaf_modules",
"iter_named_quantizable_modules",
"compute_dynamic_scales_and_zp",
"calculate_range",
"calculate_qparams",
"generate_gparam",
"strategy_cdiv",
]
# target the self_attn layer
# QuantizedKVParameterCache is responsible for obtaining the k_scale and v_scale
KV_CACHE_TARGETS = ["re:.*self_attn$"]
_LOGGER: logging.Logger = logging.getLogger(__name__)
def calculate_qparams(
min_vals: Tensor,
max_vals: Tensor,
quantization_args: QuantizationArgs,
global_scale: Optional[Tensor] = None,
) -> Tuple[FloatTensor, IntTensor]:
"""
:param min_vals: tensor of min value(s) to calculate scale(s) and zero point(s)
from
:param max_vals: tensor of max value(s) to calculate scale(s) and zero point(s)
from
:param quantization_args: settings to quantization
:param global_scale: additional global scale to scale the locally generated scale
currently only applied/supported for Fp4
:return: tuple of the calculated scale(s) and zero point(s). For FP4, the calculated
scale is of dtype FP8
"""
# based on the implementations for consuming quantized values,
# 0.0 must always be representable within the quantized range
min_vals = torch.min(min_vals, torch.zeros_like(min_vals))
max_vals = torch.max(max_vals, torch.zeros_like(max_vals))
device = min_vals.device
bit_min, bit_max = calculate_range(quantization_args, device)
bit_range = bit_max - bit_min
# 1. Generate scale and zero-point
if quantization_args.symmetric:
max_val_pos = torch.max(torch.abs(min_vals), torch.abs(max_vals))
if should_generatre_mxfp4_scales(args=quantization_args):
scales = generate_mxfp4_scales(x=max_val_pos)
else:
scales = max_val_pos / (float(bit_range) / 2)
zero_points = torch.zeros(scales.shape, device=device, dtype=min_vals.dtype)
else:
if (
quantization_args.num_bits == 4
and quantization_args.type == QuantizationType.FLOAT
):
raise NotImplementedError(
"Asymmetric Quantization is not supported for FP4"
)
scales = (max_vals - min_vals) / float(bit_range)
zero_points = bit_min - (min_vals / scales)
zero_points = torch.clamp(zero_points, bit_min, bit_max)
# 2. Conditionally scale the generated local scale by a global_scale
if global_scale is not None:
scales = global_scale * scales
# 3. Conditionally round the scale to the quantized dtype, if scale_dtype is set
if quantization_args.scale_dtype is not None:
scales = round_to_quantized_type_dtype(
scales, dtype=quantization_args.scale_dtype
)
# 4. Optionally remove exponent
scales = maybe_convert_from_mxfp4_exp(quantization_args, scales)
# 5. Update any 0s with small values to
# prevent div by 0
eps = _get_dtype_eps(
dtype=quantization_args.scale_dtype
if quantization_args.scale_dtype is not None
else scales.dtype
)
scales = torch.where(
scales == 0,
torch.tensor(eps, dtype=scales.dtype, device=device),
scales,
)
# 6. Round the zp to zp_dtype
zero_points = round_to_quantized_type_dtype(
zero_points, dtype=quantization_args.zp_dtype, cast_to_original_dtype=False
)
if scales.ndim == 0:
scales = scales.reshape(1)
zero_points = zero_points.reshape(1)
return scales, zero_points
def compute_dynamic_scales_and_zp(
value: Tensor,
args: QuantizationArgs,
module: torch.nn.Module,
global_scale: Optional[Tensor] = None,
):
"""
Returns the computed scales and zero points for dynamic activation
quantization.
:param value: tensor to calculate quantization parameters for
:param args: quantization args
:param reduce_dims: optional tuple of dimensions to reduce along,
returned scale and zero point will be shaped (1,) along the
reduced dimensions
:return: tuple of scale and zero point derived from the observed tensor
"""
keep_dims = True
if args.strategy == QuantizationStrategy.TOKEN:
dim = {0, 1}
reduce_dims = tuple(idx for idx in range(value.ndim) if idx not in dim)
elif args.strategy == QuantizationStrategy.TENSOR:
reduce_dims = None
elif args.strategy in (
QuantizationStrategy.TENSOR_GROUP,
QuantizationStrategy.GROUP,
):
reduce_dims = -1
keep_dims = False
reshaped_dims = (
math.ceil(value.shape[-1] / args.group_size),
args.group_size,
)
value = value.unflatten(-1, reshaped_dims)
else:
supported_strategies = (
QuantizationStrategy.TOKEN,
QuantizationStrategy.TENSOR,
QuantizationStrategy.TENSOR_GROUP,
QuantizationStrategy.GROUP,
)
raise ValueError(
"Dynamic quantization is only supported for ",
f"{supported_strategies}",
)
if not reduce_dims:
min_val, max_val = torch.aminmax(value)
else:
min_val = torch.amin(value, dim=reduce_dims, keepdims=keep_dims)
max_val = torch.amax(value, dim=reduce_dims, keepdims=keep_dims)
return calculate_qparams(min_val, max_val, args, global_scale=global_scale)
def calculate_range(quantization_args: QuantizationArgs, device: str) -> Tuple:
"""
Calculated the effective quantization range for the given Quantization Args
:param quantization_args: quantization args to get range of
:param device: device to store the range to
:return: tuple endpoints for the given quantization range
"""
if quantization_args.type == QuantizationType.INT:
bit_range = 2**quantization_args.num_bits
q_max = torch.tensor(bit_range / 2 - 1, device=device)
q_min = torch.tensor(-bit_range / 2, device=device)
elif quantization_args.type == QuantizationType.FLOAT:
if quantization_args.num_bits == 8:
q_max = torch.tensor(FP8_E4M3_DATA.max, device=device)
q_min = torch.tensor(FP8_E4M3_DATA.min, device=device)
elif quantization_args.num_bits == 4:
q_max = torch.tensor(FP4_E2M1_DATA.max, device=device)
q_min = torch.tensor(FP4_E2M1_DATA.min, device=device)
else:
raise NotImplementedError(
"Range calculation only supported for 4 and 8 bits"
)
else:
raise ValueError(f"Invalid quantization type {quantization_args.type}")
return q_min, q_max
def is_module_quantized(module: Module) -> bool:
"""
Check if a module is quantized, based on the existence of a non-empty quantization
scheme
:param module: pytorch module to check
:return: True if module is quantized, False otherwise
"""
if not hasattr(module, "quantization_scheme"):
return False
if module.quantization_scheme.weights is not None:
return True
if module.quantization_scheme.input_activations is not None:
return True
if module.quantization_scheme.output_activations is not None:
return True
return False
def is_model_quantized(model: Module) -> bool:
"""
Check if any modules in a model are quantized, based on the existence of a non-empty
quantization scheme in at least one module
:param model: pytorch model
:return: True if model is quantized, False otherwise
"""
return any(is_module_quantized(submodule) for submodule in model.modules())
def module_type(module: Module) -> str:
"""
Gets a string representation of a module type
:module: pytorch module to get type of
:return: module type as a string
"""
return type(module).__name__
@deprecated(
message="This function will be removed in a future release. "
"Please use `model.named_modules()` and filter by "
"compressed_tensors.InternalModule if neceessary"
)
def iter_named_leaf_modules(model: Module) -> Generator[Tuple[str, Module], None, None]:
"""
Yields modules that do not have any submodules except observers. The observers
themselves are not yielded
:param model: model to get leaf modules of
:returns: generator tuple of (name, leaf_submodule)
"""
for name, submodule in model.named_modules():
children = list(submodule.children())
# TODO: verify if an observer would ever be attached in this case/remove check
if len(children) == 0 and "observer" in name:
yield name, submodule
else:
if len(children) > 0:
named_children, children = zip(*list(submodule.named_children()))
has_non_observer_children = False
for i in range(len(children)):
child_name = named_children[i]
if "observer" not in child_name:
has_non_observer_children = True
if not has_non_observer_children:
yield name, submodule
@deprecated(
message="This function will be removed in a future release. "
"Please use `model.named_modules()` and filter by "
"compressed_tensors.InternalModule if neceessary"
)
def iter_named_quantizable_modules(
model: Module,
include_children: bool = True,
include_attn: bool = False,
include_mlp: bool = False,
) -> Generator[Tuple[str, Module], None, None]:
"""
Yield name and submodule of
- leaf modules, set by include_children
- attention modyles, set by include_attn
:param model: model to get leaf modules of
:param include_children: flag to get the leaf modules
:param inlcude_attn: flag to get the attention modules
:returns: generator tuple of (name, submodule)
"""
for name, submodule in model.named_modules():
# TODO: verify if an observer would ever be attached in this case/remove check
if include_children:
children = list(submodule.children())
if len(children) == 0 and "observer" not in name:
yield name, submodule
else:
if len(children) > 0:
named_children, children = zip(*list(submodule.named_children()))
has_non_observer_children = False
for i in range(len(children)):
child_name = named_children[i]
if "observer" not in child_name:
has_non_observer_children = True
if not has_non_observer_children:
yield name, submodule
if include_attn:
if name.endswith("self_attn"):
yield name, submodule
if include_mlp:
if name.endswith("mlp"):
yield name, submodule
def get_torch_bit_depth(value: torch.Tensor) -> int:
"""
Determine the number of bits used to represent the dtype of a tensor
:param value: tensor to check bit depth of
:return: bit depth of each element in the value tensor
"""
try:
bit_depth = torch.finfo(value.dtype).bits
except TypeError:
bit_depth = torch.iinfo(value.dtype).bits
return bit_depth
def can_quantize(value: torch.Tensor, quant_args: "QuantizationArgs") -> bool: # noqa
"""
Checks if value can be quantized by quant_args.
:param value: tensor to check for quantization
:param quant_args: QuantizationArgs to use for quantization
:return: False if value is already quantized to quant_args or value is incompatible
with quant_args, True if value can be quantized with quant_args
"""
bit_depth = get_torch_bit_depth(value)
requested_depth = quant_args.num_bits
if bit_depth < quant_args.num_bits:
_LOGGER.warn(
f"Can't quantize tensor with bit depth {bit_depth} to {requested_depth}."
"The QuantizationArgs provided are not compatible with the input tensor."
)
return bit_depth > quant_args.num_bits
@deprecated()
def is_kv_cache_quant_scheme(scheme: QuantizationScheme) -> bool:
"""
Check whether the QuantizationScheme targets the kv cache.
It does if all the following criteria are met:
- the scheme targets either exactly match the KV_CACHE_TARGETS
or the match KV_CACHE_TARGETS regex pattern
- the scheme quantizes output_activations (we want to quantize the
outputs from the KV_CACHE_TARGETS, as their correspond to the
keys and values that are to be saved in the cache)
:param scheme: The QuantizationScheme to investigate
:return: boolean flag
"""
for target in scheme.targets:
if target in KV_CACHE_TARGETS:
return True
return False
def generate_gparam(
updated_min_val: torch.Tensor,
updated_max_val: torch.Tensor,
scale_data: Optional[FloatArgs] = FP8_E4M3_DATA,
quant_data: Optional[FloatArgs] = FP4_E2M1_DATA,
dtype: Optional[torch.dtype] = torch.float32,
):
"""
Generate a global scale for an entire tensor (input_tensor).
Goal of the scale is to ensure that the quantization (local) scale
falls into the approproiate dtype range.
E.g. for NVFP4, group (local) scales are in dtype FP8. The global_scale
attempts to use the entire FP8 dtype range while mapping a per-group max
to the FP4 max.
"""
min_vals = torch.min(updated_min_val, torch.zeros_like(updated_min_val))
max_vals = torch.max(updated_max_val, torch.zeros_like(updated_max_val))
max_val_pos = torch.max(torch.abs(min_vals), torch.abs(max_vals))
global_scale = scale_data.max * quant_data.max / max_val_pos
return global_scale.to(dtype).reshape([1])
def strategy_cdiv(
value: int,
divisor: int,
strategy: Optional[QuantizationStrategy],
strict: bool = False,
) -> int:
dividend = math.ceil(value / divisor)
if dividend * divisor != value:
message = (
f"{strategy} quantization strategy requires strict division of "
f"weight/activation size {value} and group/block size {divisor}. "
"consider reducing the group/block size or ignoring modules with "
f"weights not divisible by {divisor}"
)
if strict:
raise ValueError(message)
else:
logger.bind(log_once=True).warning(message)
return dividend
def _get_dtype_eps(dtype: torch.dtype) -> float:
if dtype == FP8_E4M3_DATA.dtype:
return 0.125
elif dtype == FP4_E2M1_DATA.dtype:
return 0.25
elif torch.is_floating_point(torch.tensor([], dtype=dtype)):
return torch.finfo(dtype).eps
else:
return 1
|