elelems/main.py at main · Scr44gr/elelems · 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
import argparse
import sys

from elelems.core.data.prepare import prepare_dataset
from elelems.core.generate import generate_text, generate_text_stream
from elelems.core.train import train_model

PUNCTUATION_TOKENS = {".", ",", ";", ":", "!", "?"}


def build_parser() -> argparse.ArgumentParser:
    parser = argparse.ArgumentParser(description="Dumb LLM trainer and generator")
    subparsers = parser.add_subparsers(dest="command", required=True)

    prepare_parser = subparsers.add_parser("prepare", help="Clean and encode raw corpus")
    prepare_parser.add_argument("--raw-path", default="data/raws/animals.txt")
    prepare_parser.add_argument("--processed-dir", default="data/processed")
    prepare_parser.add_argument("--min-freq", type=int, default=1)
    prepare_parser.add_argument("--val-ratio", type=float, default=0.1)
    prepare_parser.add_argument("--seed", type=int, default=42)

    train_parser = subparsers.add_parser("train", help="Train mini transformer")
    train_parser.add_argument("--processed-dir", default="data/processed")
    train_parser.add_argument("--checkpoints-dir", default="data/checkpoints")
    train_parser.add_argument("--epochs", type=int, default=30)
    train_parser.add_argument("--batch-size", type=int, default=32)
    train_parser.add_argument("--lr", type=float, default=3e-4)
    train_parser.add_argument("--d-model", type=int, default=64)
    train_parser.add_argument("--n-heads", type=int, default=2)
    train_parser.add_argument("--n-layers", type=int, default=2)
    train_parser.add_argument("--dropout", type=float, default=0.1)
    train_parser.add_argument("--device", default=None)

    generate_parser = subparsers.add_parser("generate", help="Generate text from checkpoint")
    generate_parser.add_argument("--checkpoint", default="data/checkpoints/mini_llm.pt")
    generate_parser.add_argument("--prompt", required=True)
    generate_parser.add_argument("--max-new-tokens", type=int, default=24)
    generate_parser.add_argument("--min-new-tokens", type=int, default=4)
    generate_parser.add_argument("--temperature", type=float, default=1.0)
    generate_parser.add_argument("--top-k", type=int, default=0)
    generate_parser.add_argument("--top-p", type=float, default=1.0)
    generate_parser.add_argument("--repetition-penalty", type=float, default=1.05)
    generate_parser.add_argument("--stream", action="store_true")
    generate_parser.add_argument("--device", default=None)

    return parser


def main() -> None:
    parser = build_parser()
    args = parser.parse_args()

    if args.command == "prepare":
        prepare_dataset(
            raw_path=args.raw_path,
            processed_dir=args.processed_dir,
            min_freq=args.min_freq,
            val_ratio=args.val_ratio,
            seed=args.seed,
        )
        print(f"Dataset ready in: {args.processed_dir}")
        return

    if args.command == "train":
        checkpoint = train_model(
            processed_dir=args.processed_dir,
            checkpoints_dir=args.checkpoints_dir,
            epochs=args.epochs,
            batch_size=args.batch_size,
            lr=args.lr,
            d_model=args.d_model,
            n_heads=args.n_heads,
            n_layers=args.n_layers,
            dropout=args.dropout,
            device=args.device,
        )
        print(f"Model saved in: {checkpoint}")
        return

    if args.command == "generate":
        if args.stream:
            print(args.prompt, end="", flush=True)
            for token in generate_text_stream(
                checkpoint_path=args.checkpoint,
                prompt=args.prompt,
                max_new_tokens=args.max_new_tokens,
                min_new_tokens=args.min_new_tokens,
                temperature=args.temperature,
                top_k=args.top_k,
                top_p=args.top_p,
                repetition_penalty=args.repetition_penalty,
                device=args.device,
            ):
                if token in PUNCTUATION_TOKENS:
                    sys.stdout.write(token)
                else:
                    sys.stdout.write(f" {token}")
                sys.stdout.flush()
            print()
            return

        output = generate_text(
            checkpoint_path=args.checkpoint,
            prompt=args.prompt,
            max_new_tokens=args.max_new_tokens,
            min_new_tokens=args.min_new_tokens,
            temperature=args.temperature,
            top_k=args.top_k,
            top_p=args.top_p,
            repetition_penalty=args.repetition_penalty,
            device=args.device,
        )
        print(output)
        return


if __name__ == "__main__":
    main()