forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathutils.py
More file actions
60 lines (44 loc) 路 2.39 KB
/
Copy pathutils.py
File metadata and controls
60 lines (44 loc) 路 2.39 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
import torch.nn as nn
from ...utils import is_accelerate_available, logging
logger = logging.get_logger(__name__)
if is_accelerate_available():
from accelerate import init_empty_weights
def _replace_with_quanto_layers(model, quantization_config, modules_to_not_convert: list, pre_quantized=False):
# Quanto imports diffusers internally. These are placed here to avoid circular imports
from optimum.quanto import QLinear, freeze, qfloat8, qint2, qint4, qint8
def _get_weight_type(dtype: str):
return {"float8": qfloat8, "int8": qint8, "int4": qint4, "int2": qint2}[dtype]
def _replace_layers(model, quantization_config, modules_to_not_convert):
has_children = list(model.children())
if not has_children:
return model
for name, module in model.named_children():
_replace_layers(module, quantization_config, modules_to_not_convert)
if name in modules_to_not_convert:
continue
if isinstance(module, nn.Linear):
with init_empty_weights():
qlinear = QLinear(
in_features=module.in_features,
out_features=module.out_features,
bias=module.bias is not None,
dtype=module.weight.dtype,
weights=_get_weight_type(quantization_config.weights_dtype),
)
model._modules[name] = qlinear
model._modules[name].source_cls = type(module)
model._modules[name].requires_grad_(False)
return model
model = _replace_layers(model, quantization_config, modules_to_not_convert)
has_been_replaced = any(isinstance(replaced_module, QLinear) for _, replaced_module in model.named_modules())
if not has_been_replaced:
logger.warning(
f"{model.__class__.__name__} does not appear to have any `nn.Linear` modules. Quantization will not be applied."
" Please check your model architecture, or submit an issue on Github if you think this is a bug."
" https://github.com/huggingface/diffusers/issues/new"
)
# We need to freeze the pre_quantized model in order for the loaded state_dict and model state dict
# to match when trying to load weights with load_model_dict_into_meta
if pre_quantized:
freeze(model)
return model