File size: 4,780 Bytes
0523608
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Shared utilities for nanochat sweep scripts.

Extracted from lr_sweep.py / warmup_sweep.py / research_compare.py /
actual_lr_research_sweep.py to eliminate copy-paste.
"""
from __future__ import annotations

import os
import shutil
import subprocess
import sys

import torch


# ---------------------------------------------------------------------------
# DDP runner
# ---------------------------------------------------------------------------

def resolve_runner() -> list[str]:
    """Return torchrun command prefix, capped to available GPUs.

    Reads NPROC_PER_NODE (set by shell scripts) or NANOCHAT_NPROC as the
    requested worker count and caps it to the number of visible CUDA devices.
    """
    nproc_requested = int(os.environ.get("NPROC_PER_NODE", os.environ.get("NANOCHAT_NPROC", 8)))
    gpu_count = max(torch.cuda.device_count(), 1)
    nproc = min(nproc_requested, gpu_count)
    if nproc < nproc_requested:
        print(
            f"[sweep] Requested {nproc_requested} DDP workers but only "
            f"{gpu_count} GPU(s) available — using {nproc}."
        )
    torchrun = shutil.which("torchrun")
    if torchrun:
        return [torchrun, "--standalone", f"--nproc_per_node={nproc}"]
    return [
        sys.executable, "-m", "torch.distributed.run",
        "--standalone", f"--nproc_per_node={nproc}",
    ]


# ---------------------------------------------------------------------------
# Architecture sizing
# ---------------------------------------------------------------------------

def model_dims(depth: int, aspect_ratio: int = 0) -> tuple[int, int, int, int]:
    """Return (aspect_ratio, head_dim, model_dim, research_dim) for a given depth.

    research_dim is ~1/8th of model_dim, rounded up to the nearest head_dim
    multiple, capped at model_dim.
    """
    if aspect_ratio <= 0:
        aspect_ratio = 57 if depth == 9 else 64
    head_dim = 128
    base_dim = depth * aspect_ratio
    model_dim = ((base_dim + head_dim - 1) // head_dim) * head_dim
    raw_research_dim = max(model_dim // 8, 1)
    research_dim = ((raw_research_dim + head_dim - 1) // head_dim) * head_dim
    research_dim = min(research_dim, model_dim)
    return aspect_ratio, head_dim, model_dim, research_dim


def estimate_tokens_from_base(
    depth: int,
    target_ratio: float = 10.5,
    tokenizer_dir: str | None = None,
) -> int:
    """Chinchilla-style optimal token count: target_ratio × scaling_params.

    Mirrors base_train.py exactly (transformer_matrices + lm_head).
    """
    from nanochat.gpt import GPT, GPTConfig

    vocab_size = 32768
    try:
        from nanochat.tokenizer import get_tokenizer
        tokenizer = get_tokenizer(tokenizer_dir=tokenizer_dir)
        vocab_size = tokenizer.get_vocab_size()
    except Exception:
        pass

    _, head_dim, model_dim, _ = model_dims(depth)
    num_heads = model_dim // head_dim
    config = GPTConfig(
        sequence_len=2048, vocab_size=vocab_size,
        n_layer=depth, n_head=num_heads, n_kv_head=num_heads, n_embd=model_dim,
    )
    with torch.device("meta"):
        model = GPT(config)
    counts = model.num_scaling_params()
    scaling_params = counts["transformer_matrices"] + counts["lm_head"]
    return int(scaling_params * target_ratio)


# ---------------------------------------------------------------------------
# Environment helpers
# ---------------------------------------------------------------------------

def check_and_prepare_env(args, label: str = "sweep") -> None:
    """Ensure data shards and tokenizer exist, downloading/training if needed."""
    from nanochat.common import get_base_dir
    from nanochat.dataset import resolve_data_dir, list_parquet_files

    data_dir = getattr(args, "data_dir", None) or resolve_data_dir()
    tokenizer_dir = getattr(args, "tokenizer_dir", None) or os.path.join(get_base_dir(), "tokenizer")
    os.makedirs(data_dir, exist_ok=True)
    os.makedirs(tokenizer_dir, exist_ok=True)

    shards = list_parquet_files(data_dir=data_dir)
    if not shards:
        max_shards = getattr(args, "max_shards", -1)
        num_files = max_shards if max_shards and max_shards > 0 else 2
        cmd = [sys.executable, "-m", "nanochat.dataset", "-n", str(num_files), "--data-dir", data_dir]
        print(f"[{label}] Downloading data: {' '.join(cmd)}")
        subprocess.run(cmd, check=True)

    tokenizer_pkl = os.path.join(tokenizer_dir, "tokenizer.pkl")
    if not os.path.exists(tokenizer_pkl):
        cmd = [
            sys.executable, "-m", "scripts.tok_train",
            "--max-chars", "10000000",
            "--data-dir", data_dir,
            "--tokenizer-dir", tokenizer_dir,
        ]
        print(f"[{label}] Training tokenizer: {' '.join(cmd)}")
        subprocess.run(cmd, check=True)