File size: 4,718 Bytes
07423df
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os

from llm_studio.src.utils.config_utils import load_config_yaml

os.environ["TOKENIZERS_PARALLELISM"] = "false"
os.environ["OMP_NUM_THREADS"] = "1"
os.environ["MKL_NUM_THREADS"] = "1"
os.environ["OPENBLAS_NUM_THREADS"] = "1"
os.environ["VECLIB_MAXIMUM_THREADS"] = "1"
os.environ["NUMEXPR_NUM_THREADS"] = "1"
os.environ["TOKENIZERS_PARALLELISM"] = "false"


import argparse

import numpy as np
import torch

from llm_studio.src.datasets.text_utils import get_tokenizer
from llm_studio.src.utils.modeling_utils import load_checkpoint, set_generation_config


def parse_param(cfg, prompt):
    prompt = prompt.replace("--", "")
    parts = prompt.split(" ")
    args = [" ".join(parts[i : i + 2]) for i in range(0, len(parts), 2)]
    for arg in args:
        splitted_arg = arg.split(" ")
        setattr(
            cfg.prediction,
            splitted_arg[0],
            type(getattr(cfg.prediction, splitted_arg[0]))(splitted_arg[1]),
        )
        print(
            f"Permanently changed {splitted_arg[0]} to",
            getattr(cfg.prediction, splitted_arg[0]),
        )
    return cfg


if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Sample prompting.")
    parser.add_argument(
        "-e",
        "--experiment",
        type=str,
        required=True,
        help="Name of the experiment output folder",
    )
    parser.add_argument(
        "-d", "--device", type=str, required=False, default="cuda:0", help="Device"
    )

    args, unknown = parser.parse_known_args()
    DEVICE = args.device

    cfg = load_config_yaml(os.path.join(args.experiment, "cfg.yaml"))

    cfg.training.epochs = 0

    cfg.environment._device = DEVICE
    cfg.environment._local_rank = 0

    cfg.tokenizer.padding_quantile = 0

    cfg.environment.mixed_precision = True
    cfg.architecture.gradient_checkpointing = False
    cfg.architecture.pretrained = False

    cfg.prediction.max_length_inference = 256

    if cfg.dataset.text_prompt_start == "":
        cfg.dataset.text_prompt_start = "\n"

    # cfg.prediction.min_length_inference = 2
    # cfg.prediction.max_length_inference = 256
    # cfg.prediction.repetition_penalty = 1.5
    # cfg.prediction.temperature = 0.3
    # cfg.prediction.num_beams = 2
    # cfg.prediction.do_sample = False
    # cfg.prediction.top_p = 0.9
    # cfg.prediction.top_k = 40

    tokenizer = get_tokenizer(cfg)

    print("Loading model weights...")

    with torch.device(DEVICE):
        model = cfg.architecture.model_class(cfg)
        cfg.architecture.pretrained_weights = os.path.join(
            args.experiment, "checkpoint.pth"
        )
        load_checkpoint(cfg, model, strict=True)

    model = model.to(DEVICE).eval()
    model.backbone.use_cache = True
    model.backbone = set_generation_config(model.backbone, cfg.prediction)

    print()
    print("=============")
    print(
        "You can change inference parameters on the fly by typing --param value, "
        "such as --num_beams 4. You can also chain them such as --num_beams 4 "
        "--top_k 30."
    )
    print()

    while True:
        prompt = input("Please enter some prompt (type 'exit' to stop): ")

        try:
            if prompt.lower() == "exit":
                break

            if prompt.lower().startswith("--"):
                cfg = parse_param(cfg, prompt)
                model.backbone = set_generation_config(model.backbone, cfg.prediction)
                continue

            prompt = cfg.dataset.dataset_class.parse_prompt(cfg, prompt)

            print(prompt)

            inputs = cfg.dataset.dataset_class.encode(
                tokenizer, prompt, cfg.tokenizer.max_length_prompt, "left"
            )
            inputs["prompt_input_ids"] = inputs.pop("input_ids").unsqueeze(0).to(DEVICE)
            inputs["prompt_attention_mask"] = (
                inputs.pop("attention_mask").unsqueeze(0).to(DEVICE)
            )

            output = {}
            with torch.no_grad():
                with torch.cuda.amp.autocast():
                    output["predicted_answer_ids"] = (
                        model.generate(inputs, cfg).detach().cpu()
                    )

            predicted_text = [
                tokenizer.decode(ids, skip_special_tokens=True)
                for ids in output["predicted_answer_ids"]
            ]
            output["predicted_text"] = np.array(predicted_text)

            output = cfg.dataset.dataset_class.clean_output(output, cfg)

            output = output["predicted_text"][0]

            print(output)
            print()
        except Exception as e:
            print("Error: {}".format(e))
            print("Something went wrong, please try again.")