forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathquantization_config.py
More file actions
676 lines (562 loc) 路 30 KB
/
Copy pathquantization_config.py
File metadata and controls
676 lines (562 loc) 路 30 KB
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
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
#!/usr/bin/env python
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. 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.
"""
Adapted from
https://github.com/huggingface/transformers/blob/52cb4034ada381fe1ffe8d428a1076e5411a8026/src/transformers/utils/quantization_config.py
"""
import copy
import importlib.metadata
import inspect
import json
import os
from dataclasses import dataclass
from enum import Enum
from functools import partial
from typing import Any, Dict, List, Optional, Union
from packaging import version
from ..utils import is_torch_available, is_torchao_available, logging
if is_torch_available():
import torch
logger = logging.get_logger(__name__)
class QuantizationMethod(str, Enum):
BITS_AND_BYTES = "bitsandbytes"
GGUF = "gguf"
TORCHAO = "torchao"
@dataclass
class QuantizationConfigMixin:
"""
Mixin class for quantization config
"""
quant_method: QuantizationMethod
_exclude_attributes_at_init = []
@classmethod
def from_dict(cls, config_dict, return_unused_kwargs=False, **kwargs):
"""
Instantiates a [`QuantizationConfigMixin`] from a Python dictionary of parameters.
Args:
config_dict (`Dict[str, Any]`):
Dictionary that will be used to instantiate the configuration object.
return_unused_kwargs (`bool`,*optional*, defaults to `False`):
Whether or not to return a list of unused keyword arguments. Used for `from_pretrained` method in
`PreTrainedModel`.
kwargs (`Dict[str, Any]`):
Additional parameters from which to initialize the configuration object.
Returns:
[`QuantizationConfigMixin`]: The configuration object instantiated from those parameters.
"""
config = cls(**config_dict)
to_remove = []
for key, value in kwargs.items():
if hasattr(config, key):
setattr(config, key, value)
to_remove.append(key)
for key in to_remove:
kwargs.pop(key, None)
if return_unused_kwargs:
return config, kwargs
else:
return config
def to_json_file(self, json_file_path: Union[str, os.PathLike]):
"""
Save this instance to a JSON file.
Args:
json_file_path (`str` or `os.PathLike`):
Path to the JSON file in which this configuration instance's parameters will be saved.
use_diff (`bool`, *optional*, defaults to `True`):
If set to `True`, only the difference between the config instance and the default
`QuantizationConfig()` is serialized to JSON file.
"""
with open(json_file_path, "w", encoding="utf-8") as writer:
config_dict = self.to_dict()
json_string = json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
writer.write(json_string)
def to_dict(self) -> Dict[str, Any]:
"""
Serializes this instance to a Python dictionary. Returns:
`Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance.
"""
return copy.deepcopy(self.__dict__)
def __iter__(self):
"""allows `dict(obj)` for situations where obj may be a dict or QuantizationConfigMixin"""
for attr, value in copy.deepcopy(self.__dict__).items():
yield attr, value
def __repr__(self):
return f"{self.__class__.__name__} {self.to_json_string()}"
def to_json_string(self, use_diff: bool = True) -> str:
"""
Serializes this instance to a JSON string.
Args:
use_diff (`bool`, *optional*, defaults to `True`):
If set to `True`, only the difference between the config instance and the default `PretrainedConfig()`
is serialized to JSON string.
Returns:
`str`: String containing all the attributes that make up this configuration instance in JSON format.
"""
if use_diff is True:
config_dict = self.to_diff_dict()
else:
config_dict = self.to_dict()
return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
def update(self, **kwargs):
"""
Updates attributes of this class instance with attributes from `kwargs` if they match existing attributes,
returning all the unused kwargs.
Args:
kwargs (`Dict[str, Any]`):
Dictionary of attributes to tentatively update this class.
Returns:
`Dict[str, Any]`: Dictionary containing all the key-value pairs that were not used to update the instance.
"""
to_remove = []
for key, value in kwargs.items():
if hasattr(self, key):
setattr(self, key, value)
to_remove.append(key)
# Remove all the attributes that were updated, without modifying the input dict
unused_kwargs = {key: value for key, value in kwargs.items() if key not in to_remove}
return unused_kwargs
@dataclass
class BitsAndBytesConfig(QuantizationConfigMixin):
"""
This is a wrapper class about all possible attributes and features that you can play with a model that has been
loaded using `bitsandbytes`.
This replaces `load_in_8bit` or `load_in_4bit`therefore both options are mutually exclusive.
Currently only supports `LLM.int8()`, `FP4`, and `NF4` quantization. If more methods are added to `bitsandbytes`,
then more arguments will be added to this class.
Args:
load_in_8bit (`bool`, *optional*, defaults to `False`):
This flag is used to enable 8-bit quantization with LLM.int8().
load_in_4bit (`bool`, *optional*, defaults to `False`):
This flag is used to enable 4-bit quantization by replacing the Linear layers with FP4/NF4 layers from
`bitsandbytes`.
llm_int8_threshold (`float`, *optional*, defaults to 6.0):
This corresponds to the outlier threshold for outlier detection as described in `LLM.int8() : 8-bit Matrix
Multiplication for Transformers at Scale` paper: https://arxiv.org/abs/2208.07339 Any hidden states value
that is above this threshold will be considered an outlier and the operation on those values will be done
in fp16. Values are usually normally distributed, that is, most values are in the range [-3.5, 3.5], but
there are some exceptional systematic outliers that are very differently distributed for large models.
These outliers are often in the interval [-60, -6] or [6, 60]. Int8 quantization works well for values of
magnitude ~5, but beyond that, there is a significant performance penalty. A good default threshold is 6,
but a lower threshold might be needed for more unstable models (small models, fine-tuning).
llm_int8_skip_modules (`List[str]`, *optional*):
An explicit list of the modules that we do not want to convert in 8-bit. This is useful for models such as
Jukebox that has several heads in different places and not necessarily at the last position. For example
for `CausalLM` models, the last `lm_head` is typically kept in its original `dtype`.
llm_int8_enable_fp32_cpu_offload (`bool`, *optional*, defaults to `False`):
This flag is used for advanced use cases and users that are aware of this feature. If you want to split
your model in different parts and run some parts in int8 on GPU and some parts in fp32 on CPU, you can use
this flag. This is useful for offloading large models such as `google/flan-t5-xxl`. Note that the int8
operations will not be run on CPU.
llm_int8_has_fp16_weight (`bool`, *optional*, defaults to `False`):
This flag runs LLM.int8() with 16-bit main weights. This is useful for fine-tuning as the weights do not
have to be converted back and forth for the backward pass.
bnb_4bit_compute_dtype (`torch.dtype` or str, *optional*, defaults to `torch.float32`):
This sets the computational type which might be different than the input type. For example, inputs might be
fp32, but computation can be set to bf16 for speedups.
bnb_4bit_quant_type (`str`, *optional*, defaults to `"fp4"`):
This sets the quantization data type in the bnb.nn.Linear4Bit layers. Options are FP4 and NF4 data types
which are specified by `fp4` or `nf4`.
bnb_4bit_use_double_quant (`bool`, *optional*, defaults to `False`):
This flag is used for nested quantization where the quantization constants from the first quantization are
quantized again.
bnb_4bit_quant_storage (`torch.dtype` or str, *optional*, defaults to `torch.uint8`):
This sets the storage type to pack the quanitzed 4-bit prarams.
kwargs (`Dict[str, Any]`, *optional*):
Additional parameters from which to initialize the configuration object.
"""
_exclude_attributes_at_init = ["_load_in_4bit", "_load_in_8bit", "quant_method"]
def __init__(
self,
load_in_8bit=False,
load_in_4bit=False,
llm_int8_threshold=6.0,
llm_int8_skip_modules=None,
llm_int8_enable_fp32_cpu_offload=False,
llm_int8_has_fp16_weight=False,
bnb_4bit_compute_dtype=None,
bnb_4bit_quant_type="fp4",
bnb_4bit_use_double_quant=False,
bnb_4bit_quant_storage=None,
**kwargs,
):
self.quant_method = QuantizationMethod.BITS_AND_BYTES
if load_in_4bit and load_in_8bit:
raise ValueError("load_in_4bit and load_in_8bit are both True, but only one can be used at the same time")
self._load_in_8bit = load_in_8bit
self._load_in_4bit = load_in_4bit
self.llm_int8_threshold = llm_int8_threshold
self.llm_int8_skip_modules = llm_int8_skip_modules
self.llm_int8_enable_fp32_cpu_offload = llm_int8_enable_fp32_cpu_offload
self.llm_int8_has_fp16_weight = llm_int8_has_fp16_weight
self.bnb_4bit_quant_type = bnb_4bit_quant_type
self.bnb_4bit_use_double_quant = bnb_4bit_use_double_quant
if bnb_4bit_compute_dtype is None:
self.bnb_4bit_compute_dtype = torch.float32
elif isinstance(bnb_4bit_compute_dtype, str):
self.bnb_4bit_compute_dtype = getattr(torch, bnb_4bit_compute_dtype)
elif isinstance(bnb_4bit_compute_dtype, torch.dtype):
self.bnb_4bit_compute_dtype = bnb_4bit_compute_dtype
else:
raise ValueError("bnb_4bit_compute_dtype must be a string or a torch.dtype")
if bnb_4bit_quant_storage is None:
self.bnb_4bit_quant_storage = torch.uint8
elif isinstance(bnb_4bit_quant_storage, str):
if bnb_4bit_quant_storage not in ["float16", "float32", "int8", "uint8", "float64", "bfloat16"]:
raise ValueError(
"`bnb_4bit_quant_storage` must be a valid string (one of 'float16', 'float32', 'int8', 'uint8', 'float64', 'bfloat16') "
)
self.bnb_4bit_quant_storage = getattr(torch, bnb_4bit_quant_storage)
elif isinstance(bnb_4bit_quant_storage, torch.dtype):
self.bnb_4bit_quant_storage = bnb_4bit_quant_storage
else:
raise ValueError("bnb_4bit_quant_storage must be a string or a torch.dtype")
if kwargs and not all(k in self._exclude_attributes_at_init for k in kwargs):
logger.warning(f"Unused kwargs: {list(kwargs.keys())}. These kwargs are not used in {self.__class__}.")
self.post_init()
@property
def load_in_4bit(self):
return self._load_in_4bit
@load_in_4bit.setter
def load_in_4bit(self, value: bool):
if not isinstance(value, bool):
raise TypeError("load_in_4bit must be a boolean")
if self.load_in_8bit and value:
raise ValueError("load_in_4bit and load_in_8bit are both True, but only one can be used at the same time")
self._load_in_4bit = value
@property
def load_in_8bit(self):
return self._load_in_8bit
@load_in_8bit.setter
def load_in_8bit(self, value: bool):
if not isinstance(value, bool):
raise TypeError("load_in_8bit must be a boolean")
if self.load_in_4bit and value:
raise ValueError("load_in_4bit and load_in_8bit are both True, but only one can be used at the same time")
self._load_in_8bit = value
def post_init(self):
r"""
Safety checker that arguments are correct - also replaces some NoneType arguments with their default values.
"""
if not isinstance(self.load_in_4bit, bool):
raise TypeError("load_in_4bit must be a boolean")
if not isinstance(self.load_in_8bit, bool):
raise TypeError("load_in_8bit must be a boolean")
if not isinstance(self.llm_int8_threshold, float):
raise TypeError("llm_int8_threshold must be a float")
if self.llm_int8_skip_modules is not None and not isinstance(self.llm_int8_skip_modules, list):
raise TypeError("llm_int8_skip_modules must be a list of strings")
if not isinstance(self.llm_int8_enable_fp32_cpu_offload, bool):
raise TypeError("llm_int8_enable_fp32_cpu_offload must be a boolean")
if not isinstance(self.llm_int8_has_fp16_weight, bool):
raise TypeError("llm_int8_has_fp16_weight must be a boolean")
if self.bnb_4bit_compute_dtype is not None and not isinstance(self.bnb_4bit_compute_dtype, torch.dtype):
raise TypeError("bnb_4bit_compute_dtype must be torch.dtype")
if not isinstance(self.bnb_4bit_quant_type, str):
raise TypeError("bnb_4bit_quant_type must be a string")
if not isinstance(self.bnb_4bit_use_double_quant, bool):
raise TypeError("bnb_4bit_use_double_quant must be a boolean")
if self.load_in_4bit and not version.parse(importlib.metadata.version("bitsandbytes")) >= version.parse(
"0.39.0"
):
raise ValueError(
"4 bit quantization requires bitsandbytes>=0.39.0 - please upgrade your bitsandbytes version"
)
def is_quantizable(self):
r"""
Returns `True` if the model is quantizable, `False` otherwise.
"""
return self.load_in_8bit or self.load_in_4bit
def quantization_method(self):
r"""
This method returns the quantization method used for the model. If the model is not quantizable, it returns
`None`.
"""
if self.load_in_8bit:
return "llm_int8"
elif self.load_in_4bit and self.bnb_4bit_quant_type == "fp4":
return "fp4"
elif self.load_in_4bit and self.bnb_4bit_quant_type == "nf4":
return "nf4"
else:
return None
def to_dict(self) -> Dict[str, Any]:
"""
Serializes this instance to a Python dictionary. Returns:
`Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance.
"""
output = copy.deepcopy(self.__dict__)
output["bnb_4bit_compute_dtype"] = str(output["bnb_4bit_compute_dtype"]).split(".")[1]
output["bnb_4bit_quant_storage"] = str(output["bnb_4bit_quant_storage"]).split(".")[1]
output["load_in_4bit"] = self.load_in_4bit
output["load_in_8bit"] = self.load_in_8bit
return output
def __repr__(self):
config_dict = self.to_dict()
return f"{self.__class__.__name__} {json.dumps(config_dict, indent=2, sort_keys=True)}\n"
def to_diff_dict(self) -> Dict[str, Any]:
"""
Removes all attributes from config which correspond to the default config attributes for better readability and
serializes to a Python dictionary.
Returns:
`Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance,
"""
config_dict = self.to_dict()
# get the default config dict
default_config_dict = BitsAndBytesConfig().to_dict()
serializable_config_dict = {}
# only serialize values that differ from the default config
for key, value in config_dict.items():
if value != default_config_dict[key]:
serializable_config_dict[key] = value
return serializable_config_dict
@dataclass
class GGUFQuantizationConfig(QuantizationConfigMixin):
"""This is a config class for GGUF Quantization techniques.
Args:
compute_dtype: (`torch.dtype`, defaults to `torch.float32`):
This sets the computational type which might be different than the input type. For example, inputs might be
fp32, but computation can be set to bf16 for speedups.
"""
def __init__(self, compute_dtype: Optional["torch.dtype"] = None):
self.quant_method = QuantizationMethod.GGUF
self.compute_dtype = compute_dtype
self.pre_quantized = True
# TODO: (Dhruv) Add this as an init argument when we can support loading unquantized checkpoints.
self.modules_to_not_convert = None
if self.compute_dtype is None:
self.compute_dtype = torch.float32
@dataclass
class TorchAoConfig(QuantizationConfigMixin):
"""This is a config class for torchao quantization/sparsity techniques.
Args:
quant_type (`str`):
The type of quantization we want to use, currently supporting:
- **Integer quantization:**
- Full function names: `int4_weight_only`, `int8_dynamic_activation_int4_weight`,
`int8_weight_only`, `int8_dynamic_activation_int8_weight`
- Shorthands: `int4wo`, `int4dq`, `int8wo`, `int8dq`
- **Floating point 8-bit quantization:**
- Full function names: `float8_weight_only`, `float8_dynamic_activation_float8_weight`,
`float8_static_activation_float8_weight`
- Shorthands: `float8wo`, `float8wo_e5m2`, `float8wo_e4m3`, `float8dq`, `float8dq_e4m3`,
`float8_e4m3_tensor`, `float8_e4m3_row`,
- **Floating point X-bit quantization:**
- Full function names: `fpx_weight_only`
- Shorthands: `fpX_eAwB`, where `X` is the number of bits (between `1` to `7`), `A` is the number
of exponent bits and `B` is the number of mantissa bits. The constraint of `X == A + B + 1` must
be satisfied for a given shorthand notation.
- **Unsigned Integer quantization:**
- Full function names: `uintx_weight_only`
- Shorthands: `uint1wo`, `uint2wo`, `uint3wo`, `uint4wo`, `uint5wo`, `uint6wo`, `uint7wo`
modules_to_not_convert (`List[str]`, *optional*, default to `None`):
The list of modules to not quantize, useful for quantizing models that explicitly require to have some
modules left in their original precision.
kwargs (`Dict[str, Any]`, *optional*):
The keyword arguments for the chosen type of quantization, for example, int4_weight_only quantization
supports two keyword arguments `group_size` and `inner_k_tiles` currently. More API examples and
documentation of arguments can be found in
https://github.com/pytorch/ao/tree/main/torchao/quantization#other-available-quantization-techniques
Example:
```python
from diffusers import FluxTransformer2DModel, TorchAoConfig
quantization_config = TorchAoConfig("int8wo")
transformer = FluxTransformer2DModel.from_pretrained(
"black-forest-labs/Flux.1-Dev",
subfolder="transformer",
quantization_config=quantization_config,
torch_dtype=torch.bfloat16,
)
```
"""
def __init__(self, quant_type: str, modules_to_not_convert: Optional[List[str]] = None, **kwargs) -> None:
self.quant_method = QuantizationMethod.TORCHAO
self.quant_type = quant_type
self.modules_to_not_convert = modules_to_not_convert
# When we load from serialized config, "quant_type_kwargs" will be the key
if "quant_type_kwargs" in kwargs:
self.quant_type_kwargs = kwargs["quant_type_kwargs"]
else:
self.quant_type_kwargs = kwargs
TORCHAO_QUANT_TYPE_METHODS = self._get_torchao_quant_type_to_method()
if self.quant_type not in TORCHAO_QUANT_TYPE_METHODS.keys():
is_floating_quant_type = self.quant_type.startswith("float") or self.quant_type.startswith("fp")
if is_floating_quant_type and not self._is_cuda_capability_atleast_8_9():
raise ValueError(
f"Requested quantization type: {self.quant_type} is not supported on GPUs with CUDA capability <= 8.9. You "
f"can check the CUDA capability of your GPU using `torch.cuda.get_device_capability()`."
)
raise ValueError(
f"Requested quantization type: {self.quant_type} is not supported or is an incorrect `quant_type` name. If you think the "
f"provided quantization type should be supported, please open an issue at https://github.com/huggingface/diffusers/issues."
)
method = TORCHAO_QUANT_TYPE_METHODS[self.quant_type]
signature = inspect.signature(method)
all_kwargs = {
param.name
for param in signature.parameters.values()
if param.kind in [inspect.Parameter.KEYWORD_ONLY, inspect.Parameter.POSITIONAL_OR_KEYWORD]
}
unsupported_kwargs = list(self.quant_type_kwargs.keys() - all_kwargs)
if len(unsupported_kwargs) > 0:
raise ValueError(
f'The quantization method "{quant_type}" does not support the following keyword arguments: '
f"{unsupported_kwargs}. The following keywords arguments are supported: {all_kwargs}."
)
@classmethod
def _get_torchao_quant_type_to_method(cls):
r"""
Returns supported torchao quantization types with all commonly used notations.
"""
if is_torchao_available():
# TODO(aryan): Support autoquant and sparsify
from torchao.quantization import (
float8_dynamic_activation_float8_weight,
float8_static_activation_float8_weight,
float8_weight_only,
fpx_weight_only,
int4_weight_only,
int8_dynamic_activation_int4_weight,
int8_dynamic_activation_int8_weight,
int8_weight_only,
uintx_weight_only,
)
# TODO(aryan): Add a note on how to use PerAxis and PerGroup observers
from torchao.quantization.observer import PerRow, PerTensor
def generate_float8dq_types(dtype: torch.dtype):
name = "e5m2" if dtype == torch.float8_e5m2 else "e4m3"
types = {}
for granularity_cls in [PerTensor, PerRow]:
# Note: Activation and Weights cannot have different granularities
granularity_name = "tensor" if granularity_cls is PerTensor else "row"
types[f"float8dq_{name}_{granularity_name}"] = partial(
float8_dynamic_activation_float8_weight,
activation_dtype=dtype,
weight_dtype=dtype,
granularity=(granularity_cls(), granularity_cls()),
)
return types
def generate_fpx_quantization_types(bits: int):
types = {}
for ebits in range(1, bits):
mbits = bits - ebits - 1
types[f"fp{bits}_e{ebits}m{mbits}"] = partial(fpx_weight_only, ebits=ebits, mbits=mbits)
non_sign_bits = bits - 1
default_ebits = (non_sign_bits + 1) // 2
default_mbits = non_sign_bits - default_ebits
types[f"fp{bits}"] = partial(fpx_weight_only, ebits=default_ebits, mbits=default_mbits)
return types
INT4_QUANTIZATION_TYPES = {
# int4 weight + bfloat16/float16 activation
"int4wo": int4_weight_only,
"int4_weight_only": int4_weight_only,
# int4 weight + int8 activation
"int4dq": int8_dynamic_activation_int4_weight,
"int8_dynamic_activation_int4_weight": int8_dynamic_activation_int4_weight,
}
INT8_QUANTIZATION_TYPES = {
# int8 weight + bfloat16/float16 activation
"int8wo": int8_weight_only,
"int8_weight_only": int8_weight_only,
# int8 weight + int8 activation
"int8dq": int8_dynamic_activation_int8_weight,
"int8_dynamic_activation_int8_weight": int8_dynamic_activation_int8_weight,
}
# TODO(aryan): handle torch 2.2/2.3
FLOATX_QUANTIZATION_TYPES = {
# float8_e5m2 weight + bfloat16/float16 activation
"float8wo": partial(float8_weight_only, weight_dtype=torch.float8_e5m2),
"float8_weight_only": float8_weight_only,
"float8wo_e5m2": partial(float8_weight_only, weight_dtype=torch.float8_e5m2),
# float8_e4m3 weight + bfloat16/float16 activation
"float8wo_e4m3": partial(float8_weight_only, weight_dtype=torch.float8_e4m3fn),
# float8_e5m2 weight + float8 activation (dynamic)
"float8dq": float8_dynamic_activation_float8_weight,
"float8_dynamic_activation_float8_weight": float8_dynamic_activation_float8_weight,
# ===== Matrix multiplication is not supported in float8_e5m2 so the following errors out.
# However, changing activation_dtype=torch.float8_e4m3 might work here =====
# "float8dq_e5m2": partial(
# float8_dynamic_activation_float8_weight,
# activation_dtype=torch.float8_e5m2,
# weight_dtype=torch.float8_e5m2,
# ),
# **generate_float8dq_types(torch.float8_e5m2),
# ===== =====
# float8_e4m3 weight + float8 activation (dynamic)
"float8dq_e4m3": partial(
float8_dynamic_activation_float8_weight,
activation_dtype=torch.float8_e4m3fn,
weight_dtype=torch.float8_e4m3fn,
),
**generate_float8dq_types(torch.float8_e4m3fn),
# float8 weight + float8 activation (static)
"float8_static_activation_float8_weight": float8_static_activation_float8_weight,
# For fpx, only x <= 8 is supported by default. Other dtypes can be explored by users directly
# fpx weight + bfloat16/float16 activation
**generate_fpx_quantization_types(3),
**generate_fpx_quantization_types(4),
**generate_fpx_quantization_types(5),
**generate_fpx_quantization_types(6),
**generate_fpx_quantization_types(7),
}
UINTX_QUANTIZATION_DTYPES = {
"uintx_weight_only": uintx_weight_only,
"uint1wo": partial(uintx_weight_only, dtype=torch.uint1),
"uint2wo": partial(uintx_weight_only, dtype=torch.uint2),
"uint3wo": partial(uintx_weight_only, dtype=torch.uint3),
"uint4wo": partial(uintx_weight_only, dtype=torch.uint4),
"uint5wo": partial(uintx_weight_only, dtype=torch.uint5),
"uint6wo": partial(uintx_weight_only, dtype=torch.uint6),
"uint7wo": partial(uintx_weight_only, dtype=torch.uint7),
# "uint8wo": partial(uintx_weight_only, dtype=torch.uint8), # uint8 quantization is not supported
}
QUANTIZATION_TYPES = {}
QUANTIZATION_TYPES.update(INT4_QUANTIZATION_TYPES)
QUANTIZATION_TYPES.update(INT8_QUANTIZATION_TYPES)
QUANTIZATION_TYPES.update(UINTX_QUANTIZATION_DTYPES)
if cls._is_cuda_capability_atleast_8_9():
QUANTIZATION_TYPES.update(FLOATX_QUANTIZATION_TYPES)
return QUANTIZATION_TYPES
else:
raise ValueError(
"TorchAoConfig requires torchao to be installed, please install with `pip install torchao`"
)
@staticmethod
def _is_cuda_capability_atleast_8_9() -> bool:
if not torch.cuda.is_available():
raise RuntimeError("TorchAO requires a CUDA compatible GPU and installation of PyTorch.")
major, minor = torch.cuda.get_device_capability()
if major == 8:
return minor >= 9
return major >= 9
def get_apply_tensor_subclass(self):
TORCHAO_QUANT_TYPE_METHODS = self._get_torchao_quant_type_to_method()
return TORCHAO_QUANT_TYPE_METHODS[self.quant_type](**self.quant_type_kwargs)
def __repr__(self):
r"""
Example of how this looks for `TorchAoConfig("uint4wo", group_size=32)`:
```
TorchAoConfig {
"modules_to_not_convert": null,
"quant_method": "torchao",
"quant_type": "uint4wo",
"quant_type_kwargs": {
"group_size": 32
}
}
```
"""
config_dict = self.to_dict()
return f"{self.__class__.__name__} {json.dumps(config_dict, indent=2, sort_keys=True)}\n"