-
Notifications
You must be signed in to change notification settings - Fork 433
/
gen_test_data.py
162 lines (129 loc) · 4.85 KB
/
gen_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
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
import os
import sys
import math
import torch
from torch.nn import functional as F
## Symlink this file into minGPT directory to import, and run it from this directory
## (Python sux)
# create a GPT instance
from mingpt.model import GPT
from mingpt.utils import set_seed
set_seed(3407)
model_config = GPT.get_default_config()
model_config.model_type = 'gpt-nano'
model_config.vocab_size = 3
model_config.block_size = 11
model = GPT(model_config)
model.eval()
model.load_state_dict(torch.load('mingpt/model.pt'))
def tensor_to_json(tensor):
import base64
data = base64.b64encode(tensor.detach().numpy().tobytes()).decode()
return {"shape": list(tensor.shape), "dtype": str(tensor.dtype), "data": data}
def save_tensor_dict_to_json(dict, filename, extra=None):
import json
items = { k: tensor_to_json(v) for k, v in dict.items() }
if extra is not None:
items = { **extra, **items }
with open(filename, 'w') as f:
json.dump(items, f, indent=4)
extra = {'config': model_config.to_dict()}
save_tensor_dict_to_json(model.state_dict(), 'public/gpt-nano-sort-model.json', extra)
t0 = model.get_submodule('transformer.h.0')
t0Attn = t0.get_submodule('attn')
n_head = model_config.n_head
n_embd = model_config.n_embd
B = 3
T = model_config.block_size
C = n_embd
torch.random.manual_seed(34)
# transformer_input = torch.randn(B, T, C, requires_grad=False)
# print(transformer_input.flatten().tolist()[:10])
def mlp_forward_with_capture(tModule, x):
fc = tModule.c_fc(x)
gelu = tModule.act(fc)
res = tModule.c_proj(gelu)
return res, { 'fc': fc, 'gelu': gelu }
def block_forward_with_capture(tModule, x):
ln1 = tModule.ln_1(x)
attn, attn_partials = transformer_forward_with_capture(tModule.attn, ln1)
attnResid = x + attn
ln2 = tModule.ln_2(attnResid)
mlp, mlp_partials = mlp_forward_with_capture(tModule.mlp, ln2)
mlpResid = attnResid + mlp
return mlpResid, {
'ln1': ln1,
**attn_partials,
'attnResid': attnResid,
'ln2': ln2,
**mlp_partials,
'mlp': mlp,
'mlpResid': mlpResid,
}
def transformer_forward_with_capture(tModule, x):
B, T, C = x.shape
qkv = tModule.c_attn(x)
q, k, v = qkv.split(n_embd, dim=2)
k = k.view(B, T, n_head, C // n_head).transpose(1, 2) # (B, nh, T, hs)
q = q.view(B, T, n_head, C // n_head).transpose(1, 2) # (B, nh, T, hs)
v = v.view(B, T, n_head, C // n_head).transpose(1, 2) # (B, nh, T, hs)
# causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
att = att.masked_fill(tModule.bias[:,:,:T,:T] == 0, float('-inf'))
attSm = F.softmax(att, dim=-1)
# att = self.attn_dropout(att)
y = attSm @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
# output projection
yProj = tModule.c_proj(y)
partials = {
'q': q, 'k': k, 'v': v, # projected vectors (B, nh, T, hs)
'qkv': qkv,
'att': att, 'attSm': attSm, # attention (B, nh, T, T)
'y': y, 'yProj': yProj, # output (B, T, C)
}
return yProj, partials
def gpt_forward_with_capture(model, idx):
b, t = idx.size()
assert t == T, f"For testing, only block size {T} is supported"
pos = torch.arange(0, t, dtype=torch.long).unsqueeze(0) # shape (1, t)
# forward the GPT model itself
tok_emb = model.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
pos_emb = model.transformer.wpe(pos) # position embeddings of shape (1, t, n_embd)
x = tok_emb + pos_emb
partials = {
'idx': idx.type(torch.float32),
'tok_emb': tok_emb,
'pos_emb': pos_emb,
'x': x,
}
return x, partials
idx = torch.tensor([[0, 0, 2, 1, 0, 1, 0, 0, 0, 0, 0]], dtype=torch.long)
extraIdx = torch.cat([
torch.randint(0, 3, (B - 1, 6), dtype=torch.long),
torch.zeros((B - 1, 5), dtype=torch.long),
], dim=1)
if B > 1:
extraIdx[1, 0] = 1
idx = torch.cat([idx, extraIdx], dim=0)
print(idx)
transformer_input, partials0 = gpt_forward_with_capture(model, idx)
res, partials = block_forward_with_capture(t0, transformer_input)
partials = { **partials0, **partials }
resActual = t0(transformer_input)
if not torch.equal(res, resActual):
print('ERROR: test block output does not match model output')
x = transformer_input
for i, block in enumerate(model.transformer.h):
x = block(x)
partials[f'block{i}'] = x
x = model.transformer.ln_f(x)
partials['ln_f'] = x
x = model.lm_head(x)
partials['lm_head'] = x
probs = F.softmax(x, dim=-1)
partials['probs'] = probs
print(model_config.to_dict())
extra = {'config': { **model_config.to_dict(), 'B': B }}
save_tensor_dict_to_json(partials, 'public/gpt-nano-sort-t0-partials.json', extra)
print({ k: v.shape for k, v in partials.items() })