rome-replication/gpt_tests.py at main · NixGD/rome-replication · GitHub

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
from dataclasses import dataclass
import einops
import math
import numpy as np
import time
import torch
from torch import nn
import torch.nn.functional as F
import transformers
from typing import Optional

from _gpt_sol import _UnidirectionalAttention, _GPT2Block, _GPT2


def _check_equal(tensor1, tensor2):
    assert torch.allclose(tensor1, tensor2, atol=1e-4, rtol=1e-4)
    print("Congrats! You've passed the test!")


def test_unidirectional_attn(Attention):
    kwargs = dict(hidden_size=24, num_heads=4)
    encodings = torch.randn(1, 5, kwargs['hidden_size'])

    torch.manual_seed(545)
    _attn = _UnidirectionalAttention(**kwargs)
    _out = _attn(encodings)

    torch.manual_seed(545)
    attn = Attention(**kwargs)
    out = attn(encodings)

    _check_equal(out, _out)


def test_attn_cache(Attention):
    kwargs = dict(hidden_size=24, num_heads=4)
    head_size = kwargs['hidden_size'] // kwargs['num_heads']
    encodings = torch.randn(1, 1, kwargs['hidden_size'])
    past_key_values = torch.randn(kwargs['num_heads'], 3, 2 * head_size)

    torch.manual_seed(945)
    _attn = _UnidirectionalAttention(**kwargs)
    _out = _attn(encodings, past_key_values=past_key_values, return_key_values=True)

    torch.manual_seed(945)
    attn = Attention(**kwargs)
    out = attn(encodings, past_key_values=past_key_values, return_key_values=True)

    print('Checking encoding:')
    _check_equal(out[0], _out[0])
    print('Checking new key and value:')
    _check_equal(out[1], _out[1])


def test_gpt_block(GPT2Block):
    kwargs = dict(hidden_size=48, layer_norm_epsilon=1e-4, dropout=0.0, num_heads=4)
    x = torch.randn(1, 5, 48)

    torch.manual_seed(710)
    _block = _GPT2Block(**kwargs)
    _out = _block(x)

    torch.manual_seed(710)
    block = GPT2Block(**kwargs)
    out = block(x)

    _check_equal(out, _out)


def test_gpt(GPT2):
    config = dict(num_layers=2, num_heads=4, vocab_size=100, hidden_size=64,
                  max_position_embeddings=32, dropout=0.0, layer_norm_epsilon=1e-4)
    x = torch.randint(0, config['vocab_size'], (1, 5))

    torch.manual_seed(1010)
    _gpt = _GPT2(**config)
    _output = _gpt(x)

    torch.manual_seed(1010)
    gpt = GPT2(**config)
    output = gpt(x)

    print('Checking logits:')
    _check_equal(_output.logits, output.logits)
    print('Checking final encodings:')
    _check_equal(_output.final_encoding, output.final_encoding)


def test_gpt_cache(GPT2):
    config = dict(num_layers=2, num_heads=4, vocab_size=100, hidden_size=64,
                  max_position_embeddings=500, dropout=0.0, layer_norm_epsilon=1e-4)
    x = torch.randint(0, config['vocab_size'], (1, 500))

    torch.manual_seed(1010)
    gpt = GPT2(**config)
    t = time.time()
    for i in range(1, x.shape[1]+1):
        output_nocache = gpt(x[:, :i])
    t1 = time.time() - t

    torch.manual_seed(1010)
    gpt = GPT2(**config, use_cache=True)
    t = time.time()
    for i in range(1, x.shape[1]+1):
        output_cache = gpt(x[:, :i])
    t2 = time.time() - t

    if torch.allclose(output_cache.logits, output_nocache.logits, rtol=1e-4, atol=1e-4):
        print('Congrats! Your GPT returns the same results with and without cache.')
        print(f'It took {t1:.3f}s to generate a 500-token sentence without cache and '
              f'{t2:.3f}s with cache.')
    else:
        print('Your GPT returns different results when using cache.')


def _copy_weight_bias(mine, theirs, transpose=False):
    if transpose:
        mine.weight.copy_(theirs.weight.T)
    else:
        mine.weight.copy_(theirs.weight)
    if mine.bias is not None:
        mine.bias.copy_(theirs.bias)


def get_pretrained_gpt():
    pretrained_gpt = transformers.AutoModelForCausalLM.from_pretrained("gpt2")
    tokenizer = transformers.AutoTokenizer.from_pretrained("gpt2")
    config = dict(num_layers=12, num_heads=12, vocab_size=50257, hidden_size=768,
                  max_position_embeddings=1024, dropout=0.1, layer_norm_epsilon=1e-5)
    my_gpt = _GPT2(**config, tokenizer=tokenizer)
    for p in my_gpt.parameters():
        p.requires_grad = False

    my_gpt.token_embedding.weight.copy_(pretrained_gpt.transformer.wte.weight)
    my_gpt.pos_embedding.weight.copy_(pretrained_gpt.transformer.wpe.weight)
    _copy_weight_bias(my_gpt.ln, pretrained_gpt.transformer.ln_f)

    for my_block, hf_block in zip(my_gpt.blocks, pretrained_gpt.transformer.h):
        _copy_weight_bias(my_block.ln1, hf_block.ln_1)
        _copy_weight_bias(my_block.attn.attn_lin, hf_block.attn.c_attn, transpose=True)
        _copy_weight_bias(my_block.attn.out_lin, hf_block.attn.c_proj, transpose=True)
        _copy_weight_bias(my_block.ln2, hf_block.ln_2)
        _copy_weight_bias(my_block.linear1, hf_block.mlp.c_fc, transpose=True)
        _copy_weight_bias(my_block.linear2, hf_block.mlp.c_proj, transpose=True)
    return my_gpt