-
Notifications
You must be signed in to change notification settings - Fork 4
/
generate_test_data.py
135 lines (109 loc) · 3.84 KB
/
generate_test_data.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
import math
import os
import torch
import torch.nn.functional as F
from torch import nn
name_to_tensor = {}
# Generate Linear.
inputs = torch.randn(3, 768)
linear = nn.Linear(in_features=768, out_features=4 * 768)
outputs = linear(inputs)
linear_no_bias = nn.Linear(in_features=768, out_features=4 * 768, bias=False)
linear_no_bias.weight.data = linear.weight.data
outputs_no_bias = linear_no_bias(inputs)
name_to_tensor.update(
{
"linear_weight": linear.weight,
"linear_bias": linear.bias,
"linear_inputs": inputs,
"linear_outputs": outputs,
"linear_outputs_no_bias": outputs_no_bias,
}
)
# Generate GELU.
def gelu(x):
"""Gaussian Error Linear Unit (GELU) activation function.
Copied from the nanogpt repo and identical to OpenAI GPT2 implementation. Paper:
https://arxiv.org/abs/1606.08415
"""
# fmt: off
return 0.5 * x * (1.0 + torch.tanh(math.sqrt(2.0 / math.pi) * (x + 0.044715 * torch.pow(x, 3.0))))
# fmt: on
inputs = torch.randn(3, 768)
outputs = gelu(inputs)
name_to_tensor.update({"gelu_inputs": inputs, "gelu_outputs": outputs})
# Generate softmax.
inputs = torch.randn(3, 768)
outputs = F.softmax(inputs, dim=-1)
name_to_tensor.update({"softmax_inputs": inputs, "softmax_outputs": outputs})
# Generate Embedding.
embedding = nn.Embedding(10, 768)
inputs = torch.randint(0, 10, (3,))
outputs = embedding(inputs)
name_to_tensor.update(
{
"embedding_weight": embedding.weight,
"embedding_inputs": inputs,
"embedding_outputs": outputs,
}
)
# Generate LayerNorm.
layer_norm = nn.LayerNorm(768)
inputs = torch.randn(3, 768)
outputs = layer_norm(inputs)
name_to_tensor.update(
{
"layer_norm_weight": layer_norm.weight,
"layer_norm_bias": layer_norm.bias,
"layer_norm_inputs": inputs,
"layer_norm_outputs": outputs,
}
)
# Generate causal self attention.
batch_size, seq_len, n_head, head_dim = 1, 5, 12, 64
n_embed = n_head * head_dim
# Generate transpose intermediaries.
inputs = torch.randn(batch_size, seq_len, n_head, head_dim)
outputs = inputs.transpose(1, 2)
name_to_tensor.update({"transpose_inputs": inputs, "transpose_outputs": outputs})
# Generate split intermediaries.
inputs = torch.randn(batch_size, seq_len, 3 * n_embed)
q, k, v = inputs.split(n_embed, dim=2)
name_to_tensor.update(
{"split_inputs": inputs, "split_q": q, "split_k": k, "split_v": v}
)
inputs = torch.randn(batch_size, seq_len, n_embed)
c_attn = nn.Linear(in_features=n_embed, out_features=3 * n_embed)
outputs = c_attn(inputs)
name_to_tensor.update(
{
"attn_inputs": inputs,
"attn_c_attn_weight": c_attn.weight,
"attn_c_attn_bias": c_attn.bias,
}
)
# Generate intermediaries from scaled dot product attention.
q, k, v = outputs.split(n_embed, dim=2)
q = q.view(batch_size, seq_len, n_head, n_embed // n_head).transpose(1, 2)
k = k.view(batch_size, seq_len, n_head, n_embed // n_head).transpose(1, 2)
v = v.view(batch_size, seq_len, n_head, n_embed // n_head).transpose(1, 2)
mask = torch.tril(torch.ones(seq_len, seq_len).view(1, 1, seq_len, seq_len))
attn = q @ k.transpose(-2, -1) / math.sqrt(k.size(-1))
attn = attn.masked_fill(mask[:, :, :seq_len, :seq_len] == 0, float("-inf"))
attn = F.softmax(attn, dim=-1)
outputs = attn @ v
name_to_tensor.update({"sdpa_q": q, "sdpa_k": k, "sdpa_v": v, "sdpa_outputs": outputs})
inputs = outputs.transpose(1, 2).contiguous().view(batch_size, seq_len, n_embed)
c_proj = nn.Linear(n_embed, n_embed)
outputs = c_proj(inputs)
name_to_tensor.update(
{
"attn_c_proj_weight": c_proj.weight,
"attn_c_proj_bias": c_proj.bias,
"attn_outputs": outputs,
}
)
for name, tensor in name_to_tensor.items():
if not os.path.exists(f"models/test/{name}"):
with open(f"models/test/{name}", "wb") as file_:
file_.write(tensor.reshape(-1).detach().numpy().tobytes())