-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathinfer.py
163 lines (143 loc) · 7.14 KB
/
infer.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
163
import os
import time
import readline
import argparse
import torch
from tokenizer import Tokenizer
from model import GPT
class InferenceGPT:
def __init__(
self,
model_path=os.path.join(os.path.dirname(__file__), "checkpoint/checkpoint.pt"),
lora_path=None,
device="cuda",
is_instruct=True,
max_seq_length=None,
temperature=1.0,
top_k=5,
repetition_penalty=1.2,
profile=False
):
self.model_path = model_path
self.lora_path = lora_path
self.device = device
self.model = None
self.tokenizer = None
self.encode = None
self.decode = None
self.is_instruct = is_instruct
self.max_seq_length = max_seq_length
self.temperature = temperature
self.top_k = top_k
self.repetition_penalty = repetition_penalty
# 读取模型检查点和训练配置
print(f"Loading model from `{self.model_path}`...")
checkpoint = torch.load(self.model_path, map_location=device)
train_config = checkpoint['train_config']
model_config = checkpoint['model_config']
tokenizer_config = checkpoint['tokenizer_config']
self.max_seq_length = model_config.block_size if self.max_seq_length is None or self.max_seq_length > model_config.block_size else self.max_seq_length
# 设置随机种子与训练设置一致
torch.manual_seed(train_config.random_seed)
torch.cuda.manual_seed(train_config.random_seed)
# 加载模型状态
self.model = GPT(model_config)
self.model.load_state_dict(checkpoint['model'], strict=False)
# 加载LoRA模块（若有）
if self.lora_path is not None:
print(f"Loading LoRA module from `{self.lora_path}`...")
lora_ckpt = torch.load(self.lora_path, map_location=device)
lora_train_config = lora_ckpt['train_config']
self.model.to_lora(
lora_rank=lora_train_config.lora_rank,
lora_alpha=lora_train_config.lora_alpha,
lora_dropout=lora_train_config.lora_dropout)
self.model.load_lora_state_dict(lora_ckpt["lora"])
# 设置为推理状态
self.model.eval()
self.model.to(device)
print(f"╭───────────┬───────────┬────────┬──────┬─────────┬──────────┬────────┬──────────────────╮")
print(f"│ \x1b[1mBlockSize │ VocabSize │ Layers │ Embd │ Q_Heads │ KV_Heads │ Hidden │ Parameters\x1b[0m │")
print(f"├───────────┼───────────┼────────┼──────┼─────────┼──────────┼────────┼──────────────────┤")
print(f"│{'{:^11d}'.format(model_config.block_size, end='')}│{'{:^11d}'.format(model_config.vocab_size, end='')}│{'{:^8d}'.format(model_config.n_layer, end='')}│{'{:^6d}'.format(model_config.n_embd, end='')}│{'{:^9d}'.format(model_config.n_head, end='')}│{'{:^10d}'.format(model_config.n_kv_head, end='')}│{'{:^8d}'.format(model_config.n_hidden, end='')}│{'{:^18,d}'.format(self.model.get_num_params(), end='')}│")
print(f"╰───────────┴───────────┴────────┴──────┴─────────┴──────────┴────────┴──────────────────╯")
# 读取分词器
self.tokenizer = Tokenizer()
self.tokenizer.load_from_config_dict(tokenizer_config)
self.encode = lambda s: self.tokenizer.encode(s)
self.decode = lambda l: self.tokenizer.decode(l)
# 性能计时器
self.profile = profile
self.token_count = 0
self.times = []
self.tps_record = []
def measure(self, interval):
t = time.time_ns()
self.times.append(t)
if len(self.times) > 4:
self.times = self.times[1:]
t_avg = (self.times[3] - self.times[0]) / 3
return interval / (t_avg / 1e9)
else:
return False
def typewriter(self, token_tensor):
token_list = token_tensor[0].tolist()
chars = self.decode(token_list)
self.token_count = self.token_count + 1
if self.profile and self.token_count % 4 == 0:
tps = self.measure(4)
if tps:
self.tps_record.append(tps)
if "<|eos|>" in chars:
print(chars.split("<|eos|>")[0], end="", flush=True)
return False
elif "<|padding|>" in chars:
print(chars.replace("<|padding|>", ""), end="", flush=True)
return False
else:
print(chars, end="", flush=True)
return True
def run(self):
with torch.no_grad():
while True:
try:
prompt = input("\x1b[32;1mHomo:\x1b[0m ")
except EOFError:
break
if self.is_instruct:
prompt = f"<|instruct_mark|>{prompt}<|response_mark|>"
x = torch.tensor(self.encode(prompt), dtype=torch.long, device=self.device)[None, ...]
print("\x1b[34;1mNano:\x1b[0m ", end="", flush=True)
y = self.model.auto_regressive_generate(x, self.max_seq_length, self.temperature, self.top_k, self.repetition_penalty, callback=self.typewriter)
print("\n")
if self.profile:
print(f"TPS = {[round(tps) for tps in self.tps_record]}\n")
self.times = []
self.tps_record = []
self.token_count = 0
def main():
print(f"\n\x1b[36;1mNano Language Model\x1b[0m - https://github.com/bd4sur/Nano")
parser = argparse.ArgumentParser(description="Sample (to inference) from Nano model for text generation and question answering.")
parser.add_argument("-m", "--model", type=str, default="checkpoint/ckpt.pt")
parser.add_argument("-l", "--lora", type=str, default=None)
parser.add_argument("-i", "--instruct", action="store_true")
parser.add_argument("-s", "--max_seq_length", type=int, default=None)
parser.add_argument("-t", "--temperature", type=float, default=1.0)
parser.add_argument("-k", "--top_k", type=int, default=5)
parser.add_argument("-r", "--repetition_penalty", type=float, default=1.2)
parser.add_argument("-p", "--profile", action='store_true')
args = parser.parse_args()
infer = InferenceGPT(
args.model,
args.lora,
device="cuda" if torch.cuda.is_available() else "cpu",
is_instruct=args.instruct,
max_seq_length=args.max_seq_length,
temperature=args.temperature,
top_k=args.top_k,
repetition_penalty=args.repetition_penalty,
profile=args.profile
)
infer.run()
if __name__ == "__main__":
main()