Upload modeling.py with huggingface_hub
Browse files- modeling.py +107 -0
modeling.py
ADDED
@@ -0,0 +1,107 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
os.environ["PATH"] = "/usr/local/cuda/bin:" + os.environ["PATH"]
|
4 |
+
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
|
5 |
+
|
6 |
+
import bitblas
|
7 |
+
import torch
|
8 |
+
import torch.nn as nn
|
9 |
+
from transformers import BertConfig, BertModel, PreTrainedModel, PretrainedConfig,AutoModel,AutoConfig,BertPreTrainedModel
|
10 |
+
|
11 |
+
class bitlinear(bitblas.Linear):
|
12 |
+
def __init__(
|
13 |
+
self,
|
14 |
+
in_features: int,
|
15 |
+
out_features: int,
|
16 |
+
bias: bool = False,
|
17 |
+
A_dtype: str = "float16",
|
18 |
+
W_dtype: str = "int2",
|
19 |
+
accum_dtype: str = "float16",
|
20 |
+
out_dtype: str = "float16",
|
21 |
+
group_size: int = -1,
|
22 |
+
with_scaling: bool = False,
|
23 |
+
with_zeros: bool = False,
|
24 |
+
zeros_mode: str = None,
|
25 |
+
opt_M: list = [1, 16, 32, 64, 128, 256, 512],
|
26 |
+
fast_decoding: bool = True,
|
27 |
+
alpha: torch.dtype = torch.float16,
|
28 |
+
b:torch.Tensor=None
|
29 |
+
):
|
30 |
+
super().__init__(
|
31 |
+
in_features=in_features,
|
32 |
+
out_features=out_features,
|
33 |
+
bias=bias,
|
34 |
+
A_dtype=A_dtype,
|
35 |
+
W_dtype=W_dtype,
|
36 |
+
accum_dtype=accum_dtype,
|
37 |
+
out_dtype=out_dtype,
|
38 |
+
group_size=group_size,
|
39 |
+
with_scaling=with_scaling,
|
40 |
+
with_zeros=with_zeros,
|
41 |
+
zeros_mode=zeros_mode,
|
42 |
+
opt_M=opt_M,
|
43 |
+
fast_decoding=fast_decoding,
|
44 |
+
)
|
45 |
+
self.alpha = nn.Parameter(alpha,requires_grad=False)
|
46 |
+
self.b = nn.Parameter(b,requires_grad=False)
|
47 |
+
|
48 |
+
def forward(self, A: torch.Tensor, out: torch.Tensor = None) -> torch.Tensor:
|
49 |
+
out = super().forward(A, out)
|
50 |
+
out *= self.alpha
|
51 |
+
if self.b is not None:
|
52 |
+
out += self.b.view(1, -1).expand_as(out)
|
53 |
+
return out.to(torch.float32)
|
54 |
+
|
55 |
+
|
56 |
+
class TernaryBertConfig(BertConfig):
|
57 |
+
model_type = "ternarybert"
|
58 |
+
def __init__(self, **kwargs):
|
59 |
+
super().__init__(**kwargs)
|
60 |
+
|
61 |
+
|
62 |
+
class TernaryBert(PreTrainedModel):
|
63 |
+
#config_class = TernaryBertConfig
|
64 |
+
config_class = BertConfig
|
65 |
+
|
66 |
+
def __init__(self, config):
|
67 |
+
super().__init__(config)
|
68 |
+
self.bert = BertModel(config)
|
69 |
+
self.replace_linear2bitblas(self.bert)
|
70 |
+
|
71 |
+
#def forward(self, input_ids, attention_mask=None,token_type_ids=None):
|
72 |
+
# return self.bert(input_ids, attention_mask=attention_mask,token_type_ids=token_type_ids)
|
73 |
+
def forward(self, **kwargs):
|
74 |
+
return self.bert(**kwargs)
|
75 |
+
|
76 |
+
def convert_to_bitlinear(self,layer):
|
77 |
+
bitlayer = bitlinear(
|
78 |
+
in_features=layer.in_features,
|
79 |
+
out_features=layer.out_features,
|
80 |
+
bias=False,
|
81 |
+
A_dtype="float16", # activation A dtype
|
82 |
+
W_dtype="int2", # weight W dtype
|
83 |
+
accum_dtype="float16", # accumulation dtype
|
84 |
+
out_dtype="float16", # output dtype
|
85 |
+
# configs for weight only quantization
|
86 |
+
group_size=-1, # setting for grouped quantization
|
87 |
+
with_scaling=False, # setting for scaling factor
|
88 |
+
with_zeros=False, # setting for zeros
|
89 |
+
zeros_mode=None, # setting for how to calculating zeros
|
90 |
+
# Target optimization var for dynamic symbolic.
|
91 |
+
# For detailed information please checkout docs/PythonAPI.md
|
92 |
+
# By default, the optimization var is [1, 16, 32, 64, 128, 256, 512]
|
93 |
+
opt_M=[1, 16, 32, 64, 128, 256, 512],
|
94 |
+
fast_decoding=True,
|
95 |
+
alpha=torch.tensor(1.).to(torch.float16),
|
96 |
+
b = layer.bias.data.to(torch.float16)
|
97 |
+
)
|
98 |
+
return bitlayer
|
99 |
+
|
100 |
+
def replace_linear2bitblas(self,model):
|
101 |
+
for name, module in model.named_children():
|
102 |
+
if isinstance(module, nn.Linear):
|
103 |
+
new_layer = self.convert_to_bitlinear(module)
|
104 |
+
setattr(model, name, new_layer)
|
105 |
+
elif len(list(module.children())) > 0:
|
106 |
+
self.replace_linear2bitblas(module)
|
107 |
+
|