CocktailSGD/convert_llama_to_hf.py at master · DS3Lab/CocktailSGD · 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
import os
import argparse
import torch

import torch
import torch.nn as nn

from modules.llama_modules import LlamaForCausalLM
from modules.llama_modules import LlamaConfig, LlamaTokenizer

from transformers.modeling_utils import no_init_weights
import os


def create_emtpy_llama(config):

    import torch
    import torch.nn as nn

    _reset_parameters_linear = nn.Linear.reset_parameters
    def dummy(*args, **kargs):
        pass
    nn.Linear.reset_parameters = dummy

    # 1. disable init for faster initialization
    # 2. avoid tie token embeddings with lm_head, as we train them separately.
    with no_init_weights(_enable=True):
        model = LlamaForCausalLM(config).eval()

    nn.Linear.reset_parameters = _reset_parameters_linear

    return model

def load_decentralized_checkpoint(model, checkpoint_path, n_stages=2, n_layer_per_stage=16, ):
    input_path = checkpoint_path

    n_layers = len(model.model.layers)
    assert n_stages * n_layer_per_stage >= len(model.model.layers)
    # assert model.lm_head.weight.data is not model.transformer.wte.weight.data

    for i in range(n_stages):

        print(f'loading stage {i}')

        checkpoint = torch.load(os.path.join(input_path, f'prank_{i}_checkpoint.pt'), map_location=torch.device("cpu"))

        if i == 0:
            _tmp = {k[len(f"{0}."):]:v for k,v in checkpoint.items() if k.startswith(f"0.")}
            # torch.save(_tmp, os.path.join(output_path, f'pytorch_embs.pt'))
            model.model.embed_tokens.weight.data[:] = _tmp['embed_tokens.weight']

            for j in range(n_layer_per_stage):
                _tmp = {k[len(f"{j+1}."):]:v for k,v in checkpoint.items() if k.startswith(f"{j+1}.")}
                if len(_tmp) == 0:
                    break
                # torch.save(_tmp, os.path.join(output_path, f'pytorch_{j}.pt'))
                model.model.layers[j].load_state_dict(_tmp)

        elif i == n_stages - 1:
            for j in range(n_layer_per_stage):
                if i*n_layer_per_stage + j == n_layers:
                    break
                _tmp = {k[len(f"{j}."):]:v for k,v in checkpoint.items() if k.startswith(f"{j}.")}
                if len(_tmp) == 0:
                    break
                # torch.save(_tmp, os.path.join(output_path, f'pytorch_{i*n_layer_per_stage + j}.pt'))
                model.model.layers[i*n_layer_per_stage + j].load_state_dict(_tmp)
            else:
                j += 1

            _tmp = {k[len(f"{j}."):]:v for k,v in checkpoint.items() if k.startswith(f"{j}.")}
            if len(_tmp) == 0:
                break
            # torch.save(_tmp, os.path.join(output_path, f'pytorch_lm_head.pt'))
            model.model.norm.weight.data[:] = _tmp['norm.weight']
            if 'norm.bias' in _tmp:
                model.model.norm.bias.data[:] = _tmp['norm.bias']
            model.lm_head.weight.data[:] = _tmp['lm_head.weight']
            if 'lm_head.bias' in _tmp:
                model.lm_head.bias.data[:] = _tmp['lm_head.bias']

        else:
            for j in range(n_layer_per_stage):
                _tmp = {k[len(f"{j}."):]:v for k,v in checkpoint.items() if k.startswith(f"{j}.")}
                if len(_tmp) == 0:
                    break
                # torch.save(_tmp, os.path.join(output_path, f'pytorch_{i*n_layer_per_stage + j}.pt'))
                model.model.layers[i*n_layer_per_stage + j].load_state_dict(_tmp)

    return model


if __name__ == '__main__':

    parser = argparse.ArgumentParser(description='Convert HF checkpoints')
    parser.add_argument('--config-name', type=str, default='EleutherAI/gpt-neox-20b',
                        help='config-name')
    parser.add_argument('--ckpt-path', type=str, default=None,
                        help='ckpt-path')
    parser.add_argument('--save-path', type=str, default=None,
                        help='save-path')
    parser.add_argument('--n-stages', type=int, default=8,
                        help='pipeline group size')
    parser.add_argument('--n-layer-per-stage', type=int, default=6,
                        help='n layers per GPU device')
    parser.add_argument('--fp16', default=False, action='store_true')
    args = parser.parse_args()

    assert args.ckpt_path is not None
    assert args.save_path is not None

    if not os.path.exists(args.save_path):
        os.mkdir(args.save_path)

    # LlamaForCausalLM LlamaConfig LlamaTokenizer
    print('loading config...')
    config = LlamaConfig.from_pretrained(args.config_name)
    print('loaded config.')
    print('loading tokenizer...')
    tokenizer = LlamaTokenizer.from_pretrained(args.config_name)
    print('loaded tokenizer.')
    print('creating empty model...')
    model = create_emtpy_llama(config)
    if args.fp16:
        model = model.half()
    print('created empty model.')
    print('loading model ckpt...')
    load_decentralized_checkpoint(
        model, args.ckpt_path, n_stages=args.n_stages, n_layer_per_stage=args.n_layer_per_stage,
    )
    print('loaded model ckpt.')

    print('saving HF model...')
    model.save_pretrained(args.save_path)
    print(f'saved HF model to `{args.save_path}`')
    config.save_pretrained(args.save_path)
    tokenizer.save_pretrained(args.save_path)