File size: 6,270 Bytes
6d5c79f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
"""Medium-sized Zenex MoE training using a real BPE tokenizer.

Loads SentencePiece tokenizer, encodes corpus once (cached as .npy),
trains a larger MoE config for many more steps. Still CPU-friendly.
"""

from __future__ import annotations

import argparse
import logging
import math
import time
from pathlib import Path

import numpy as np
import sentencepiece as spm
import torch

from architecture.config import ModelConfig
from architecture.model import ZenexMoEDecoder

LOGGER = logging.getLogger(__name__)


def build_medium_config(vocab_size: int) -> ModelConfig:
    """Bigger MoE that still trains on a laptop CPU in a few hours."""

    return ModelConfig(
        num_layers=6,
        hidden_size=384,
        num_attention_heads_q=8,
        num_attention_heads_kv=8,
        head_dim=48,
        vocab_size=vocab_size,
        ffn_intermediate_size=1024,
        num_routed_experts=8,
        num_shared_experts=1,
        top_k_experts=2,
        mla_kv_latent_dim=192,
        mla_q_latent_dim=192,
        mtp_predict_steps=1,
    )


def encode_corpus(sp: spm.SentencePieceProcessor, corpus: Path, cache: Path) -> np.ndarray:
    """Tokenize corpus once and cache as .npy on disk."""

    if cache.exists():
        LOGGER.info("Loading cached token array: %s", cache)
        return np.load(cache)

    LOGGER.info("Encoding corpus with BPE: %s", corpus)
    tokens: list[int] = []
    eos = sp.eos_id() if sp.eos_id() >= 0 else 3
    chunk_size = 0
    with corpus.open("r", encoding="utf-8") as f:
        for line in f:
            line = line.strip()
            if not line:
                continue
            ids = sp.encode(line, out_type=int)
            tokens.extend(ids)
            tokens.append(eos)
            chunk_size += 1
            if chunk_size % 50_000 == 0:
                LOGGER.info("Encoded %d lines, %d tokens", chunk_size, len(tokens))

    arr = np.asarray(tokens, dtype=np.int32)
    cache.parent.mkdir(parents=True, exist_ok=True)
    np.save(cache, arr)
    LOGGER.info("Saved %d tokens to %s", len(arr), cache)
    return arr


def get_batch(data: np.ndarray, seq_len: int, batch_size: int, device: torch.device):
    """Sample random contiguous chunks."""

    ix = np.random.randint(0, len(data) - seq_len - 1, size=batch_size)
    x = np.stack([data[i : i + seq_len].astype(np.int64) for i in ix])
    y = np.stack([data[i + 1 : i + seq_len + 1].astype(np.int64) for i in ix])
    return torch.from_numpy(x).to(device), torch.from_numpy(y).to(device)


def cosine_lr(step: int, total: int, warmup: int, peak: float, min_lr: float) -> float:
    if step < warmup:
        return peak * (step + 1) / max(1, warmup)
    progress = (step - warmup) / max(1, total - warmup)
    return min_lr + 0.5 * (peak - min_lr) * (1 + math.cos(math.pi * progress))


def pick_device() -> torch.device:
    """Pick best available device: CUDA > Apple MPS > CPU."""

    if torch.cuda.is_available():
        return torch.device("cuda")
    if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
        return torch.device("mps")
    return torch.device("cpu")


def train(args: argparse.Namespace) -> None:
    """Main training loop."""

    device = pick_device()
    LOGGER.info("Training on device: %s", device)

    sp = spm.SentencePieceProcessor(model_file=str(args.tokenizer))
    LOGGER.info("Tokenizer loaded: vocab=%d", sp.vocab_size())

    data = encode_corpus(sp, args.corpus, args.cache)
    LOGGER.info("Corpus tokens: %d", len(data))

    cfg = build_medium_config(vocab_size=sp.vocab_size())
    model = ZenexMoEDecoder(cfg).to(device)
    n_params = sum(p.numel() for p in model.parameters())
    LOGGER.info("Model size: %.2fM params", n_params / 1e6)

    optimizer = torch.optim.AdamW(
        model.parameters(), lr=args.peak_lr, betas=(0.9, 0.95),
        eps=1e-8, weight_decay=0.1,
    )
    model.train()

    args.output.parent.mkdir(parents=True, exist_ok=True)
    start = time.time()
    for step in range(1, args.steps + 1):
        lr = cosine_lr(step, args.steps, warmup=max(50, args.steps // 50),
                       peak=args.peak_lr, min_lr=args.peak_lr * 0.1)
        for g in optimizer.param_groups:
            g["lr"] = lr

        x, y = get_batch(data, args.seq_len, args.batch_size, device)
        optimizer.zero_grad(set_to_none=True)
        out = model(x)
        logits = out.logits
        loss = torch.nn.functional.cross_entropy(
            logits.reshape(-1, logits.size(-1)), y.reshape(-1)
        )
        loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
        optimizer.step()

        if step % args.log_every == 0 or step == 1:
            elapsed = time.time() - start
            tps = step * args.batch_size * args.seq_len / max(1e-6, elapsed)
            LOGGER.info(
                "step=%d loss=%.4f lr=%.2e tok/s=%.0f elapsed=%.1fs",
                step, float(loss.item()), lr, tps, elapsed,
            )

        if step % args.save_every == 0 or step == args.steps:
            torch.save(model.state_dict(), args.output)
            LOGGER.info("Checkpoint saved -> %s", args.output)


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Medium MoE training with BPE tokenizer")
    parser.add_argument("--corpus", type=Path, default=Path("/mnt/nas/data/corpus.txt"))
    parser.add_argument("--cache", type=Path, default=Path("/mnt/nas/data/corpus.tokens.npy"))
    parser.add_argument("--tokenizer", type=Path, default=Path("/mnt/nas/tokenizer/tokenizer.model"))
    parser.add_argument("--output", type=Path, default=Path("/mnt/nas/checkpoints/medium.pt"))
    parser.add_argument("--steps", type=int, default=3000)
    parser.add_argument("--seq-len", type=int, default=256)
    parser.add_argument("--batch-size", type=int, default=8)
    parser.add_argument("--peak-lr", type=float, default=3e-4)
    parser.add_argument("--log-every", type=int, default=20)
    parser.add_argument("--save-every", type=int, default=500)
    return parser.parse_args()


def main() -> None:
    logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
    args = parse_args()
    train(args)


if __name__ == "__main__":
    main()