-
Notifications
You must be signed in to change notification settings - Fork 8
/
irqlora.py
182 lines (154 loc) · 8.85 KB
/
irqlora.py
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
from tqdm import tqdm
import peft
import torch
import operator
import numpy as np
import bitsandbytes as bnb
from peft.tuners.lora import LoraLayer
from functools import reduce # Required in Python 3
import bitsandbytes.functional as bnb_F
from torch import Tensor
from scipy.stats import norm
from bitsandbytes.functional import create_fp8_map, create_dynamic_map
cache_folder_path = ''
module_num = 0
sigma = 1 / norm.ppf(torch.linspace(0.9677083, 0.5, 9)[:-1]).tolist()[0]
def replace_to_qlora_model(model, model_fp, blocksize2=256, tau_range=0.1, tau_n=100):
model.model = _replace_with_ours_lora_4bit_linear(model.model, model_fp=model_fp, blocksize2=blocksize2, tau_range=tau_range, tau_n=tau_n)
return model
def prod(iterable):
return reduce(operator.mul, iterable, 1)
normal_map_fp8 = create_dynamic_map()
def quantize_tensor(X, L, idx=False):
L = L.to(X.device)
X_shape = X.shape
X_expanded = X.reshape(-1, 1)
L_reshaped = L.reshape(1, -1)
abs_diff = torch.abs(X_expanded - L_reshaped)
min_index = torch.argmin(abs_diff, dim=-1)
min_index = torch.tensor(min_index, dtype=torch.uint8).to(L.device).reshape(X_shape)
return min_index
def dequantize_tensor(X, L):
L = L.to(X.device)
return torch.index_select(L, dim=0, index=torch.as_tensor(X, dtype=torch.int32).reshape(-1)).reshape(X.shape)
@torch.no_grad()
def nf4_quant(weight, weight_shape, tau, compress_statistics, quant_type, device):
weight = weight.reshape(-1, 256, 64).to(device)
tau = tau.reshape(-1, 256, 1).to(device)
_weight = (weight - tau).reshape(weight_shape)
nf4_weight = bnb.nn.Params4bit(_weight, requires_grad=False, compress_statistics=compress_statistics, quant_type=quant_type).cuda(0)
tau2 = tau.abs().max(dim=1, keepdim=True)[0]
tau1 = quantize_tensor(tau / tau2, normal_map_fp8)
return nf4_weight, tau1.reshape(-1, 256), tau2.reshape(-1, 1)
@torch.no_grad()
def evaluate_entropy(weight_int8, blocksize):
device = weight_int8.device
_weight_int8 = weight_int8.reshape(-1, 1)
weight_nf4 = torch.cat((_weight_int8//16, _weight_int8%16), 1).reshape(1, -1, blocksize)
weight_nf4_repeat = weight_nf4.repeat(16, 1, 1).to(device)
values = torch.tensor(range(16)).reshape(16, 1, 1).to(device)
freqs = (weight_nf4_repeat==values).sum(dim=-1, keepdim=True) / blocksize
entropy = -freqs * torch.log2(freqs)
entropy = torch.where(torch.isnan(entropy), 0, entropy)
entropy = entropy.sum(dim=0)
return entropy
@torch.no_grad()
def search(fp_weight: Tensor, fp_weight_shape, compress_statistics, quant_type, device, tau_range=0.1, tau_n=51, blocksize=64, blocksize2=256):
fp_weight = fp_weight.reshape(-1, blocksize2, blocksize).to(device)
tau0 = fp_weight.median(2, keepdim=True)[0] # [-1, 256, 1]
absmax = (fp_weight - tau0).abs().max(2, keepdim=True)[0]
entropy_max, factor_best = None, None
for factor in tqdm(np.linspace(-tau_range*sigma, tau_range*sigma, tau_n*2+1)):
tau = factor * absmax + tau0
nf4_weight, _, _ = nf4_quant(fp_weight, fp_weight_shape, tau, compress_statistics, quant_type, device)
entropy = evaluate_entropy(nf4_weight, blocksize)
if entropy_max is None:
entropy_max = entropy
factor_best = torch.full_like(entropy, factor)
else:
factor_best = torch.where(entropy > entropy_max, factor, factor_best)
entropy_max = torch.max(entropy_max, entropy)
tau = factor_best.reshape(-1, 256, 1) * absmax + tau0
nf4_weight, tau1, tau2 = nf4_quant(fp_weight, fp_weight_shape, tau, compress_statistics, quant_type, device)
return nf4_weight, tau1, tau2
class IRQLoraLinear4bit(bnb.nn.Linear4bit, LoraLayer):
def __init__(
self, old_model, model_fp=None, blocksize2=256, tau_range=0.1, tau_n=51
):
for key, value in old_model.__dict__.items():
setattr(self, key, value)
fp_weight = model_fp.weight.data.contiguous().to('cpu')
fp_weight_shape = fp_weight.shape
compress_statistics, quant_type, device = self.base_layer.weight.compress_statistics, self.base_layer.weight.quant_type, self.base_layer.weight.device
del self.base_layer.weight, model_fp
torch.cuda.empty_cache()
self.base_layer.weight, self.base_layer.tau_quant, self.base_layer.tau_absmax = search(
fp_weight=fp_weight,
fp_weight_shape=fp_weight_shape,
compress_statistics=compress_statistics,
quant_type=quant_type,
device=device,
tau_range=tau_range, tau_n=tau_n,
blocksize2=blocksize2
)
self.base_layer.tau_quant = self.base_layer.tau_quant.to(device)
self.base_layer.tau_absmax = self.base_layer.tau_absmax.to(device)
del fp_weight
torch.cuda.empty_cache()
self.lora_default_A_scale = torch.nn.Parameter(torch.zeros([1], dtype=self.lora_A.default.weight.dtype).to(self.base_layer.weight.device), requires_grad=True)
self.lora_default_B_scale = torch.nn.Parameter(torch.zeros([1], dtype=self.lora_A.default.weight.dtype).to(self.base_layer.weight.device), requires_grad=True)
def forward(self, x: torch.Tensor):
if self.base_layer.bias is not None and self.base_layer.bias.dtype != x.dtype:
self.base_layer.bias.data = self.base_layer.bias.data.to(x.dtype)
if getattr(self.base_layer.weight, 'quant_state', None) is None:
print('FP4 quantization state not initialized. Please call .cuda() or .to(device) on the LinearFP4 layer first.')
inp_dtype = x.dtype
if self.base_layer.compute_dtype is not None:
x = x.to(self.base_layer.compute_dtype)
bias = None if self.base_layer.bias is None else self.base_layer.bias.to(self.base_layer.compute_dtype)
with torch.no_grad():
fp_B = bnb_F.dequantize_fp4(self.base_layer.weight, self.base_layer.weight.quant_state).to(x.dtype)
tau = (dequantize_tensor(self.base_layer.tau_quant, normal_map_fp8).reshape(-1, 256, 1) * self.base_layer.tau_absmax.reshape(-1, 1, 1)).to(fp_B.device)
blocksize = torch.prod(torch.tensor(fp_B.shape)) / torch.prod(torch.tensor(tau.shape))
fp_B = (fp_B.reshape(-1, blocksize.int().item()) + tau.reshape(-1, 1)).reshape(fp_B.shape).to(x.dtype)
out = torch.nn.functional.linear(x, fp_B, bias)
out = out.to(inp_dtype)
result = out
if self.disable_adapters or self.active_adapter[0] not in self.lora_A.keys():
return result
elif self.r[self.active_adapter[0]] > 0:
result = result.clone()
if not torch.is_autocast_enabled():
expected_dtype = result.dtype
x = x.to(self.lora_A[self.active_adapter[0]].weight.dtype)
x = self.lora_A[self.active_adapter[0]](self.lora_dropout[self.active_adapter[0]](x)) + self.lora_default_A_scale * x.reshape([_ for _ in x.shape[:-1]] + [self.lora_A[self.active_adapter[0]].out_features] + [-1]).mean(dim=-1)
x = (self.lora_B[self.active_adapter[0]](x).reshape([_ for _ in x.shape] + [-1]) + self.lora_default_B_scale * x.unsqueeze(-1)).reshape([_ for _ in x.shape[:-1]] + [-1])
output = x.to(expected_dtype) * self.scaling[self.active_adapter[0]]
else:
x = self.lora_A[self.active_adapter[0]](self.lora_dropout[self.active_adapter[0]](x)) + self.lora_default_A_scale * x.reshape([_ for _ in x.shape[:-1]] + [self.lora_A[self.active_adapter[0]].out_features] + [-1]).mean(dim=-1)
x = (self.lora_B[self.active_adapter[0]](x).reshape([_ for _ in x.shape] + [-1]) + self.lora_default_B_scale * x.unsqueeze(-1)).reshape([_ for _ in x.shape[:-1]] + [-1])
output = x * self.scaling[self.active_adapter[0]]
result += output
return result
def _replace_with_ours_lora_4bit_linear(
model, current_key_name=None, model_fp=None, blocksize2=256, tau_range=0.5, tau_n=51
):
for name, module in model.named_children():
if current_key_name is None:
current_key_name = []
current_key_name.append(name)
if isinstance(module, peft.tuners.lora.Linear4bit):
model._modules[name] = IRQLoraLinear4bit(model._modules[name], model_fp=model_fp._modules[name], blocksize2=blocksize2, tau_range=tau_range, tau_n=tau_n)
if len(list(module.children())) > 0:
if name in model_fp._modules:
_ = _replace_with_ours_lora_4bit_linear(
module,
current_key_name, model_fp._modules[name], blocksize2, tau_range, tau_n
)
else:
_ = _replace_with_ours_lora_4bit_linear(
module,
current_key_name, None, blocksize2, tau_range, tau_n
)
current_key_name.pop(-1)
return model