Blame for code/train.py

"""Tiny GPT (nanoGPT-style) on one dataset size, with its own BPE.

python train.py 1M [--vocab 4096 --layers 6 --heads 6 --dim 192 --epochs 10]

Reports validation loss per genre as bits per byte, which is comparable across
datasets even though each one has its own tokenizer.
"""
import argparse
import json
import math
import shutil
import time
from pathlib import Path

import torch
import torch.nn as nn
import torch.nn.functional as F
from tokenizers import Tokenizer, models, pre_tokenizers, decoders, trainers

ROOT = Path(__file__).resolve().parent.parent
EOT = "<|endoftext|>"


def get_tokenizer(data, vocab, path):
    if path.exists():
        return Tokenizer.from_file(str(path))
    tok = Tokenizer(models.BPE())
    tok.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)
    tok.decoder = decoders.ByteLevel()
    tr = trainers.BpeTrainer(vocab_size=vocab, special_tokens=[EOT],
                             initial_alphabet=pre_tokenizers.ByteLevel.alphabet())
    tok.train_from_iterator((d for d in data.split(EOT) if d.strip()), tr)
    tok.save(str(path))
    return tok


def encode(tok, text):
    eot = tok.token_to_id(EOT)
    docs = [d.strip("\n") for d in text.split(EOT) if d.strip()]
    ids = []
    for i in range(0, len(docs), 20000):  # batched: much faster than one call per document
        for e in tok.encode_batch(docs[i:i + 20000]):
            ids += e.ids + [eot]
    return torch.tensor(ids, dtype=torch.long)


class Block(nn.Module):
    def __init__(self, d, h, p=0.0):
        super().__init__()
        self.h, self.p, self.drop = h, p, nn.Dropout(p)
        self.ln1, self.ln2 = nn.LayerNorm(d), nn.LayerNorm(d)
        self.qkv, self.proj = nn.Linear(d, 3 * d, bias=False), nn.Linear(d, d, bias=False)
        self.fc, self.out = nn.Linear(d, 4 * d, bias=False), nn.Linear(4 * d, d, bias=False)

    def forward(self, x):
        B, T, C = x.shape
        q, k, v = self.qkv(self.ln1(x)).view(B, T, 3, self.h, C // self.h).permute(2, 0, 3, 1, 4)
        att = F.scaled_dot_product_attention(q, k, v, is_causal=True, dropout_p=self.p if self.training else 0.0)
        x = x + self.drop(self.proj(att.transpose(1, 2).reshape(B, T, C)))
        return x + self.drop(self.out(F.gelu(self.fc(self.ln2(x)))))


class GPT(nn.Module):
    def __init__(self, vocab, block, layers, heads, dim, p=0.0):
        super().__init__()
        self.block = block
        self.tok, self.pos = nn.Embedding(vocab, dim), nn.Embedding(block, dim)
        self.drop = nn.Dropout(p)
        self.blocks = nn.ModuleList(Block(dim, heads, p) for _ in range(layers))
        self.ln = nn.LayerNorm(dim)
        self.head = nn.Linear(dim, vocab, bias=False)
        self.head.weight = self.tok.weight  # tied embeddings
        self.apply(self._init)

    @staticmethod
    def _init(m):
        if isinstance(m, (nn.Linear, nn.Embedding)):
            nn.init.normal_(m.weight, std=0.02)

    def forward(self, idx, targets=None):
        x = self.drop(self.tok(idx) + self.pos(torch.arange(idx.size(1), device=idx.device)))
        for b in self.blocks:
            x = b(x)
        logits = self.head(self.ln(x))
        loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.reshape(-1)) if targets is not None else None
        return logits, loss

    @torch.no_grad()
    def generate(self, idx, n, temp=0.8, top_k=40, top_p=1.0, rep=1.0, no_repeat=0, stop=None):
        """Batch size 1. rep: CTRL-style penalty on the last 128 tokens; no_repeat: ban any token
        that would repeat an n-gram already in the text; stop: token id that ends generation."""
        for _ in range(n):
            logits, _ = self(idx[:, -self.block:])
            logits = logits[:, -1] / temp
            if rep != 1.0:
                recent = idx[0, -128:].unique()
                l = logits[0, recent]
                logits[0, recent] = torch.where(l > 0, l / rep, l * rep)
            if no_repeat and idx.size(1) >= no_repeat:
                seq, k = idx[0].tolist(), no_repeat - 1
                prefix = tuple(seq[-k:])
                banned = {seq[i + k] for i in range(len(seq) - k) if tuple(seq[i:i + k]) == prefix}
                if banned:
                    logits[0, list(banned)] = -float("inf")
            v, _ = torch.topk(logits, top_k)
            logits[logits < v[:, [-1]]] = -float("inf")
            if top_p < 1.0:
                probs, order = F.softmax(logits, -1).sort(descending=True)
                drop = probs.cumsum(-1) - probs > top_p  # keep tokens until the mass passes top_p
                logits[0, order[0][drop[0]]] = -float("inf")
            nxt = torch.multinomial(F.softmax(logits, -1), 1)
            if stop is not None and nxt.item() == stop:
                break
            idx = torch.cat([idx, nxt], 1)
        return idx


@torch.no_grad()
def eval_bpb(model, ids, nbytes, block, dev, amp):
    """Bits per byte over an encoded validation file (non-overlapping windows)."""
    model.eval()
    total, count = 0.0, 0
    for i in range(0, len(ids) - 1, block):
        x = ids[i:i + block].to(dev)[None]
        y = ids[i + 1:i + block + 1].to(dev)[None]
        x = x[:, :y.size(1)]
        with torch.autocast("cuda", dtype=torch.bfloat16, enabled=amp):
            _, loss = model(x, y)
        total += loss.item() * y.numel()
        count += y.numel()
    model.train()
    # summed nats over all tokens, divided by bytes: independent of the tokenizer
    return total / math.log(2) / nbytes


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("size")
    ap.add_argument("--vocab", type=int, default=4096)
    ap.add_argument("--layers", type=int, default=6)
    ap.add_argument("--heads", type=int, default=6)
    ap.add_argument("--dim", type=int, default=192)
    ap.add_argument("--block", type=int, default=256)
    ap.add_argument("--batch", type=int, default=64)
    ap.add_argument("--epochs", type=float, default=10)
    ap.add_argument("--steps", type=int, default=0, help="fixed step count (overrides --epochs); used to match compute across datasets")
    ap.add_argument("--dropout", type=float, default=0.0)
    ap.add_argument("--eval-every", type=int, default=0, help="print validation bits per byte every N steps")
    ap.add_argument("--lr", type=float, default=1e-3)
    ap.add_argument("--tag", default="")
    ap.add_argument("--seed", type=int, default=0)
    ap.add_argument("--tokenizer", default=None, help="reuse this tokenizer.json instead of training one (same vocabulary across mixes)")
    ap.add_argument("--compile", action="store_true", help="torch.compile the model (faster on CUDA)")
    ap.add_argument("--data", default=None, help="folder with train.txt (default out/<size>)")
    ap.add_argument("--valid", default=None, help="folder with valid_*.txt (default: --data); "
                    "baselines are scored on OUR validation sets")
    a = ap.parse_args()

    torch.manual_seed(a.seed)
    dev = "cuda" if torch.cuda.is_available() else "cpu"
    amp = dev == "cuda" and torch.cuda.is_bf16_supported()
    data_dir = Path(a.data) if a.data else ROOT / "out" / a.size
    run = ROOT / "runs" / f"{a.size}{a.tag}"
    run.mkdir(parents=True, exist_ok=True)

    train_text = (data_dir / "train.txt").read_text(encoding="utf-8")
    if a.tokenizer:
        shutil.copy(a.tokenizer, run / "tokenizer.json")
    tok = get_tokenizer(train_text, a.vocab, run / "tokenizer.json")
    train = encode(tok, train_text)
    valid_dir = Path(a.valid) if a.valid else data_dir
    valids = {p.stem.replace("valid_", ""): p.read_text(encoding="utf-8") for p in sorted(valid_dir.glob("valid_*.txt"))}
    valids = {g: (encode(tok, t), len(t.replace(EOT, "").encode("utf-8"))) for g, t in valids.items()}
    model = GPT(tok.get_vocab_size(), a.block, a.layers, a.heads, a.dim, a.dropout).to(dev)
    n_params = sum(p.numel() for p in model.parameters())
    steps = a.steps or int(a.epochs * len(train) / (a.batch * a.block))
    print(f"{a.size}: {len(train)/1e6:.2f}M tokens, {n_params/1e6:.2f}M params, {steps} steps, device {dev}, bf16 {amp}", flush=True)

    train = train.to(dev)
    offs = torch.arange(a.block + 1, device=dev)
    fwd = torch.compile(model) if a.compile and dev == "cuda" else model
    opt = torch.optim.AdamW(model.parameters(), lr=a.lr, betas=(0.9, 0.95), weight_decay=0.1)
    sched = lambda s: a.lr * min(1, (s + 1) / 100) * (0.1 + 0.9 * 0.5 * (1 + math.cos(math.pi * s / steps)))
    t0 = time.time()
    for s in range(steps):
        for g in opt.param_groups:
            g["lr"] = sched(s)
        ix = torch.randint(len(train) - a.block - 1, (a.batch,), device=dev)
        win = train[ix[:, None] + offs]
        x, y = win[:, :-1], win[:, 1:]
        with torch.autocast("cuda", dtype=torch.bfloat16, enabled=amp):
            _, loss = fwd(x, y)
        opt.zero_grad(set_to_none=True)
        loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
        opt.step()
        if s % 100 == 0 or s == steps - 1:
            print(f"step {s}/{steps} loss {loss.item():.3f} {time.time()-t0:.0f}s", flush=True)
        if a.eval_every and s and s % a.eval_every == 0:
            bpb = {g: round(eval_bpb(model, ids, nb, a.block, dev, amp), 3) for g, (ids, nb) in valids.items()}
            print(f"  eval step {s}: {bpb}", flush=True)

    res = {"size": a.size, "params_M": round(n_params / 1e6, 2), "tokens_M": round(len(train) / 1e6, 2),
           "steps": steps, "seconds": round(time.time() - t0),
           "tokens_per_s": round(steps * a.batch * a.block / (time.time() - t0)),
           "config": {k: v for k, v in vars(a).items() if k not in ("data", "valid")},
           "valid_bits_per_byte": {g: round(eval_bpb(model, ids, nb, a.block, dev, amp), 4) for g, (ids, nb) in valids.items()}}
    prompt = torch.tensor([[tok.token_to_id(EOT)]], device=dev)
    res["sample"] = tok.decode(model.generate(prompt, 150)[0].tolist())
    (run / "results.json").write_text(json.dumps(res, indent=1), encoding="utf-8")
    torch.save(model.state_dict(), run / "model.pt")
    print(json.dumps(res, indent=1))


if __name__ == "__main__":
    main()