low_bit_llama/llama_2b_inference.py at main · GreenBitAI/low_bit_llama · 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
import argparse
import sys

import torch

from model import load_llama_model

if not torch.cuda.is_available():
    print("CUDA is needed to run the model.")
    sys.exit(0)

parser = argparse.ArgumentParser("Run inference with low-bit LLaMA models.")
parser.add_argument("-s", "--model-size", choices=["1.1b", "1.1B", "3b", "3B", "7b", "7B", "13b", "13B", "30b", "30B", "70b", "70B", "70b-chat", "70B-Chat"], required=False, default="7B", type=str, help="Which model size to use.")
parser.add_argument("-v", "--llama-version", choices=[1, 2], required=False, default=1, type=int, help="which version to evaluate")
parser.add_argument("-g", "--groupsize", choices=[8, 16, 32], required=False, default=32, type=int, help="Specify quantization groups")

args = parser.parse_args()
args.model_size = args.model_size.upper()

if args.llama_version == 1:
    model_uri = f'GreenBitAI/LLaMA-{args.model_size}-2bit'

    if args.model_size in ["3b", "3B", "30b", "30B"]:
        model_uri = model_uri + f'-groupsize{args.groupsize}'

else:
    model_uri = f'GreenBitAI/LLaMA-2-{args.model_size}-2bit'

    if args.model_size in ["1.1B", "7B", "70B", "70B-CHAT"]:
        model_uri = model_uri + f'-groupsize{args.groupsize}'
    else:
        raise NotImplemented

if args.groupsize == 32 and args.model_size not in ["1.1b", "1.1B"]:
    asym = True
else:
    asym = False

bits = 2

if bits == 2:
    if asym:
        double_groupsize = -1
    else:
        if args.groupsize == 32:
            double_groupsize=32
        else:
            if args.llama_version == 1:
                double_groupsize=64
            else:
                double_groupsize=32
else:
    if args.model_size in ["3b", "3B"]:
        double_groupsize=64
    elif args.model_size in ["7b", "7B"]:
        double_groupsize=256

v1 = (args.llama_version==1) and args.model_size in ["7b", "7B"]

cache_dir = './cache'

model, tokenizer = load_llama_model(model_uri, cache_dir=cache_dir, groupsize=args.groupsize, double_groupsize=double_groupsize, v1=v1, bits=2, half=True, asym=asym)
model.eval()

if "CHAT" in args.model_size:
    prompt = '''You are a helpful assistant, please tell me the meaning of life.'''
else:
    prompt = '''The meaning of life is'''

batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)

batch = {k: v.cuda() for k, v in batch.items()}
model.cuda()

for i in range(10):
    with torch.no_grad():
        generated = model.generate(
            inputs=batch["input_ids"],
            do_sample=True,
            use_cache=True,
            repetition_penalty=1.5,
            max_new_tokens=256,
            temperature=0.9,
            top_p=0.95,
            top_k=20,
            return_dict_in_generate=True,
            output_attentions=False,
            output_hidden_states=False,
            output_scores=False
        )
    result_text = tokenizer.decode(generated['sequences'].cpu().tolist()[0])
    print(result_text + "\n")