Create utils.py
Browse files
utils.py
ADDED
@@ -0,0 +1,127 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn.functional as F
|
3 |
+
from torch import nn
|
4 |
+
from torch.cuda.amp import custom_fwd, custom_bwd
|
5 |
+
|
6 |
+
from bitsandbytes.functional import quantize_blockwise, dequantize_blockwise
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
class FrozenBNBLinear(nn.Module):
|
11 |
+
def __init__(self, weight, absmax, code, bias=None):
|
12 |
+
assert isinstance(bias, nn.Parameter) or bias is None
|
13 |
+
super().__init__()
|
14 |
+
self.out_features, self.in_features = weight.shape
|
15 |
+
self.register_buffer("weight", weight.requires_grad_(False))
|
16 |
+
self.register_buffer("absmax", absmax.requires_grad_(False))
|
17 |
+
self.register_buffer("code", code.requires_grad_(False))
|
18 |
+
self.adapter = None
|
19 |
+
self.bias = bias
|
20 |
+
|
21 |
+
def forward(self, input):
|
22 |
+
output = DequantizeAndLinear.apply(input, self.weight, self.absmax, self.code, self.bias)
|
23 |
+
if self.adapter:
|
24 |
+
output += self.adapter(input)
|
25 |
+
return output
|
26 |
+
|
27 |
+
@classmethod
|
28 |
+
def from_linear(cls, linear: nn.Linear) -> "FrozenBNBLinear":
|
29 |
+
weights_int8, state = quantize_blockise_lowmemory(linear.weight)
|
30 |
+
return cls(weights_int8, *state, linear.bias)
|
31 |
+
|
32 |
+
def __repr__(self):
|
33 |
+
return f"{self.__class__.__name__}({self.in_features}, {self.out_features})"
|
34 |
+
|
35 |
+
|
36 |
+
class DequantizeAndLinear(torch.autograd.Function):
|
37 |
+
@staticmethod
|
38 |
+
@custom_fwd
|
39 |
+
def forward(ctx, input: torch.Tensor, weights_quantized: torch.ByteTensor,
|
40 |
+
absmax: torch.FloatTensor, code: torch.FloatTensor, bias: torch.FloatTensor):
|
41 |
+
weights_deq = dequantize_blockwise(weights_quantized, absmax=absmax, code=code)
|
42 |
+
ctx.save_for_backward(input, weights_quantized, absmax, code)
|
43 |
+
ctx._has_bias = bias is not None
|
44 |
+
return F.linear(input, weights_deq, bias)
|
45 |
+
|
46 |
+
@staticmethod
|
47 |
+
@custom_bwd
|
48 |
+
def backward(ctx, grad_output: torch.Tensor):
|
49 |
+
assert not ctx.needs_input_grad[1] and not ctx.needs_input_grad[2] and not ctx.needs_input_grad[3]
|
50 |
+
input, weights_quantized, absmax, code = ctx.saved_tensors
|
51 |
+
# grad_output: [*batch, out_features]
|
52 |
+
weights_deq = dequantize_blockwise(weights_quantized, absmax=absmax, code=code)
|
53 |
+
grad_input = grad_output @ weights_deq
|
54 |
+
grad_bias = grad_output.flatten(0, -2).sum(dim=0) if ctx._has_bias else None
|
55 |
+
return grad_input, None, None, None, grad_bias
|
56 |
+
|
57 |
+
|
58 |
+
class FrozenBNBEmbedding(nn.Module):
|
59 |
+
def __init__(self, weight, absmax, code):
|
60 |
+
super().__init__()
|
61 |
+
self.num_embeddings, self.embedding_dim = weight.shape
|
62 |
+
self.register_buffer("weight", weight.requires_grad_(False))
|
63 |
+
self.register_buffer("absmax", absmax.requires_grad_(False))
|
64 |
+
self.register_buffer("code", code.requires_grad_(False))
|
65 |
+
self.adapter = None
|
66 |
+
|
67 |
+
def forward(self, input, **kwargs):
|
68 |
+
with torch.no_grad():
|
69 |
+
# note: both quantuized weights and input indices are *not* differentiable
|
70 |
+
weight_deq = dequantize_blockwise(self.weight, absmax=self.absmax, code=self.code)
|
71 |
+
output = F.embedding(input, weight_deq, **kwargs)
|
72 |
+
if self.adapter:
|
73 |
+
output += self.adapter(input)
|
74 |
+
return output
|
75 |
+
|
76 |
+
@classmethod
|
77 |
+
def from_embedding(cls, embedding: nn.Embedding) -> "FrozenBNBEmbedding":
|
78 |
+
weights_int8, state = quantize_blockise_lowmemory(embedding.weight)
|
79 |
+
return cls(weights_int8, *state)
|
80 |
+
|
81 |
+
def __repr__(self):
|
82 |
+
return f"{self.__class__.__name__}({self.num_embeddings}, {self.embedding_dim})"
|
83 |
+
|
84 |
+
|
85 |
+
def quantize_blockise_lowmemory(matrix: torch.Tensor, chunk_size: int = 2 ** 20):
|
86 |
+
assert chunk_size % 4096 == 0
|
87 |
+
code = None
|
88 |
+
chunks = []
|
89 |
+
absmaxes = []
|
90 |
+
flat_tensor = matrix.view(-1)
|
91 |
+
for i in range((matrix.numel() - 1) // chunk_size + 1):
|
92 |
+
input_chunk = flat_tensor[i * chunk_size: (i + 1) * chunk_size].clone()
|
93 |
+
quantized_chunk, (absmax_chunk, code) = quantize_blockwise(input_chunk, code=code)
|
94 |
+
chunks.append(quantized_chunk)
|
95 |
+
absmaxes.append(absmax_chunk)
|
96 |
+
|
97 |
+
matrix_i8 = torch.cat(chunks).reshape_as(matrix)
|
98 |
+
absmax = torch.cat(absmaxes)
|
99 |
+
return matrix_i8, (absmax, code)
|
100 |
+
|
101 |
+
|
102 |
+
def convert_to_int8(model):
|
103 |
+
"""Convert linear and embedding modules to 8-bit with optional adapters"""
|
104 |
+
for module in list(model.modules()):
|
105 |
+
for name, child in module.named_children():
|
106 |
+
if isinstance(child, nn.Linear):
|
107 |
+
print(name, child)
|
108 |
+
setattr(
|
109 |
+
module,
|
110 |
+
name,
|
111 |
+
FrozenBNBLinear(
|
112 |
+
weight=torch.zeros(child.out_features, child.in_features, dtype=torch.uint8),
|
113 |
+
absmax=torch.zeros((child.weight.numel() - 1) // 4096 + 1),
|
114 |
+
code=torch.zeros(256),
|
115 |
+
bias=child.bias,
|
116 |
+
),
|
117 |
+
)
|
118 |
+
elif isinstance(child, nn.Embedding):
|
119 |
+
setattr(
|
120 |
+
module,
|
121 |
+
name,
|
122 |
+
FrozenBNBEmbedding(
|
123 |
+
weight=torch.zeros(child.num_embeddings, child.embedding_dim, dtype=torch.uint8),
|
124 |
+
absmax=torch.zeros((child.weight.numel() - 1) // 4096 + 1),
|
125 |
+
code=torch.zeros(256),
|
126 |
+
)
|
127 |
+
)
|