Create model.py
Browse files- diff_lora/model.py +83 -0
diff_lora/model.py
ADDED
@@ -0,0 +1,83 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
import math
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import torch.nn.functional as F
|
6 |
+
|
7 |
+
class DiffLoRALinear(nn.Module):
|
8 |
+
"""
|
9 |
+
Fused DiffLoRALinear implements a differential low-rank adapter:
|
10 |
+
Δy = (α/r) * [A_pos @ B_pos - τ * (A_neg @ B_neg)]
|
11 |
+
The fused version computes:
|
12 |
+
update = x_dropped @ concat(A_pos, A_neg) @ concat(B_pos, -τ * B_neg)
|
13 |
+
This version explicitly moves τ to the same device as the input.
|
14 |
+
"""
|
15 |
+
def __init__(self, in_features: int, out_features: int, r: int = 8,
|
16 |
+
lora_alpha: float = 16.0, dropout: float = 0.0,
|
17 |
+
merge_weights: bool = False, init_method: str = "kaiming"):
|
18 |
+
super().__init__()
|
19 |
+
# Base linear layer (frozen)
|
20 |
+
self.linear = nn.Linear(in_features, out_features, bias=False)
|
21 |
+
self.linear.weight.requires_grad = False
|
22 |
+
self.in_features = in_features
|
23 |
+
self.out_features = out_features
|
24 |
+
self.r = r
|
25 |
+
self.scaling = lora_alpha / r
|
26 |
+
self.merge_weights = merge_weights
|
27 |
+
self.merged = False
|
28 |
+
self.lora_dropout = nn.Dropout(dropout) if dropout > 0.0 else nn.Identity()
|
29 |
+
# Low-rank parameters for positive and negative components
|
30 |
+
self.A_pos = nn.Parameter(torch.zeros(in_features, r))
|
31 |
+
self.B_pos = nn.Parameter(torch.zeros(r, out_features))
|
32 |
+
self.A_neg = nn.Parameter(torch.zeros(in_features, r))
|
33 |
+
self.B_neg = nn.Parameter(torch.zeros(r, out_features))
|
34 |
+
self.tau = nn.Parameter(torch.tensor(1.0)) # Scalar parameter
|
35 |
+
self.reset_parameters(init_method)
|
36 |
+
|
37 |
+
def reset_parameters(self, init_method: str = "kaiming"):
|
38 |
+
if init_method == "kaiming":
|
39 |
+
nn.init.kaiming_uniform_(self.A_pos, a=math.sqrt(5))
|
40 |
+
nn.init.zeros_(self.B_pos)
|
41 |
+
nn.init.kaiming_uniform_(self.A_neg, a=math.sqrt(5))
|
42 |
+
nn.init.zeros_(self.B_neg)
|
43 |
+
elif init_method == "xavier":
|
44 |
+
nn.init.xavier_uniform_(self.A_pos)
|
45 |
+
nn.init.zeros_(self.B_pos)
|
46 |
+
nn.init.xavier_uniform_(self.A_neg)
|
47 |
+
nn.init.zeros_(self.B_neg)
|
48 |
+
else:
|
49 |
+
raise ValueError(f"Unknown init_method: {init_method}")
|
50 |
+
|
51 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
52 |
+
if self.merge_weights and self.merged:
|
53 |
+
return self.linear(x)
|
54 |
+
base_out = self.linear(x)
|
55 |
+
x_dropped = self.lora_dropout(x)
|
56 |
+
# Ensure tau is on the same device as x
|
57 |
+
tau = self.tau.to(x.device)
|
58 |
+
# Concatenate positive and negative parameters along the rank dimension
|
59 |
+
combined_A = torch.cat([self.A_pos, self.A_neg], dim=1) # (in_features, 2*r)
|
60 |
+
combined_B = torch.cat([self.B_pos, -tau * self.B_neg], dim=0) # (2*r, out_features)
|
61 |
+
update = x_dropped @ combined_A @ combined_B
|
62 |
+
delta = self.scaling * update
|
63 |
+
return base_out + delta
|
64 |
+
|
65 |
+
def replace_linear_with_diff_lora(module: nn.Module, target_regex: str, r: int):
|
66 |
+
"""
|
67 |
+
Recursively replace nn.Linear modules whose names match target_regex
|
68 |
+
with DiffLoRALinear modules using rank r.
|
69 |
+
"""
|
70 |
+
for name, child in module.named_children():
|
71 |
+
if isinstance(child, nn.Linear) and re.search(target_regex, name, re.IGNORECASE):
|
72 |
+
new_layer = DiffLoRALinear(
|
73 |
+
in_features=child.in_features,
|
74 |
+
out_features=child.out_features,
|
75 |
+
r=r,
|
76 |
+
lora_alpha=16.0,
|
77 |
+
dropout=0.1,
|
78 |
+
merge_weights=False,
|
79 |
+
)
|
80 |
+
new_layer.linear.weight.data.copy_(child.weight.data)
|
81 |
+
setattr(module, name, new_layer)
|
82 |
+
else:
|
83 |
+
replace_linear_with_diff_lora(child, target_regex, r)
|