main.py

import torch
import torch.nn as nn
from torch.nn import functional as F
import time
import sys
import os
from pathlib import Path
import tiktoken

SCRIPT_DIR = Path(__file__).resolve().parent

print("[llm]")
print()
print(f"Device: {'CUDA' if torch.cuda.is_available() else 'CPU'}")
print(f"PyTorch: {torch.__version__}")

start = time.time()
cleaned_path = Path(os.environ.get("data", SCRIPT_DIR / "input.txt"))
train_split = 0.9
seed = 1337
batch_size = 2
block_size = 20
max_iters = 10000
eval_interval = 1
learning_rate = 3e-4
device = 'cuda' if torch.cuda.is_available() else 'cpu'
eval_iters = 1
n_embd = 6
n_head = 4
n_layer = 4
dropout = 0.0

torch.manual_seed(seed)


def get_miniq_tokenizer(encoding_name="o200k_base"):
    miniq_tokenizer = tiktoken.get_encoding(encoding_name)
    miniq_vocab_size = miniq_tokenizer.n_vocab
    return miniq_tokenizer, miniq_vocab_size


def miniq_encode(text, tokenizer): return tokenizer.encode(text)
def miniq_decode(tokens, tokenizer): return tokenizer.decode(tokens)


with open(cleaned_path, 'r', encoding='utf-8') as f:
    text = f.read()

miniq_tokenizer, vocab_size = get_miniq_tokenizer("o200k_base")
encoded_data = miniq_encode(text, miniq_tokenizer)
data = torch.tensor(encoded_data, dtype=torch.long)
n = int(train_split * len(data))
train_data = data[:n]
val_data = data[n:]


def get_batch(split):
    data_split = train_data if split == 'train' else val_data
    ix = torch.randint(len(data_split) - block_size, (batch_size,))
    x = torch.stack([data_split[i:i + block_size] for i in ix])
    y = torch.stack([data_split[i + 1:i + block_size + 1] for i in ix])
    x, y = x.to(device), y.to(device)
    return x, y


@torch.no_grad()
def estimate_loss():
    out = {}
    miniq_model.eval()
    for split in ['train', 'val']:
        losses = torch.zeros(eval_iters)
        for k in range(eval_iters):
            X, Y = get_batch(split)
            _, loss = miniq_model(X, Y)
            losses[k] = loss.item()
        out[split] = losses.mean()
    miniq_model.train()
    return out


class MiniQuadtrixHead(nn.Module):
    def __init__(self, head_size):
        super().__init__()
        self.key = nn.Linear(n_embd, head_size, bias=False)
        self.query = nn.Linear(n_embd, head_size, bias=False)
        self.value = nn.Linear(n_embd, head_size, bias=False)
        self.register_buffer('tril', torch.tril(
            torch.ones(block_size, block_size)))
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        B, T, C = x.shape
        k = self.key(x)
        q = self.query(x)
        wei = q @ k.transpose(-2, -1) * k.shape[-1] ** -0.5
        wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf'))
        wei = F.softmax(wei, dim=-1)
        wei = self.dropout(wei)
        return wei @ self.value(x)


class MiniQuadtrixMHA(nn.Module):
    def __init__(self, num_heads, head_size):
        super().__init__()
        self.heads = nn.ModuleList(
            [MiniQuadtrixHead(head_size) for _ in range(num_heads)])
        self.proj = nn.Linear(head_size * num_heads, n_embd)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        out = torch.cat([h(x) for h in self.heads], dim=-1)
        return self.dropout(self.proj(out))


class MiniQuadtrixFFN(nn.Module):
    def __init__(self, n_embd):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(n_embd, 4 * n_embd),
            nn.ReLU(),
            nn.Linear(4 * n_embd, n_embd),
            nn.Dropout(dropout),
        )

    def forward(self, x):
        return self.net(x)


class MiniQuadtrixBlock(nn.Module):
    def __init__(self, n_embd, n_head):
        super().__init__()
        head_size = n_embd // n_head
        self.sa = MiniQuadtrixMHA(n_head, head_size)
        self.ffwd = MiniQuadtrixFFN(n_embd)
        self.ln1 = nn.LayerNorm(n_embd)
        self.ln2 = nn.LayerNorm(n_embd)

    def forward(self, x):
        x = x + self.sa(self.ln1(x))
        x = x + self.ffwd(self.ln2(x))
        return x


class MiniQuadtrix(nn.Module):
    def __init__(self):
        super().__init__()
        self.token_embedding_table = nn.Embedding(vocab_size, n_embd)
        self.position_embedding_table = nn.Embedding(block_size, n_embd)
        self.blocks = nn.Sequential(
            *[MiniQuadtrixBlock(n_embd, n_head=n_head) for _ in range(n_layer)])
        self.ln_f = nn.LayerNorm(n_embd)
        self.lm_head = nn.Linear(n_embd, vocab_size)
        self.apply(self._init_weights)

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
            if module.bias is not None:
                torch.nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)

    def forward(self, idx, targets=None):
        B, T = idx.shape
        tok_emb = self.token_embedding_table(idx)
        pos_emb = self.position_embedding_table(torch.arange(T, device=device))
        x = tok_emb + pos_emb
        x = self.blocks(x)
        x = self.ln_f(x)
        logits = self.lm_head(x)

        if targets is None:
            loss = None
        else:
            B, T, C = logits.shape
            logits = logits.view(B * T, C)
            targets = targets.view(B * T)
            loss = F.cross_entropy(logits, targets)
        return logits, loss

    def generate(self, idx, max_new_tokens):
        for _ in range(max_new_tokens):
            idx_cond = idx[:, -block_size:]
            logits, _ = self(idx_cond)
            logits = logits[:, -1, :]
            probs = F.softmax(logits, dim=-1)
            idx_next = torch.multinomial(probs, num_samples=1)
            idx = torch.cat((idx, idx_next), dim=1)
        return idx


miniq_model = MiniQuadtrix().to(device)
miniq_n_params = sum(p.numel() for p in miniq_model.parameters())
miniq_optimizer = torch.optim.AdamW(miniq_model.parameters(), lr=learning_rate)

print("CONFIG")
print(f"Seed: {seed}")
print(f"Batch size: {batch_size}")
print(f"Block size: {block_size}")
print(f"Learning rate: {learning_rate}")
print(f"Layers: {n_layer}")
print(f"Heads: {n_head}")
print(f"Embedding dim: {n_embd}")
print(f"Dropout: {dropout}")
print(f"Parameters: {miniq_n_params:,}")
print(f"Train tokens: {len(train_data):,}")
print(f"Val tokens: {len(val_data):,}")
print(f"Data file: {str(cleaned_path)}")
print()

best_val_loss = float('inf')
train_start = time.time()
prev_loss = None

for iter in range(max_iters):
    if iter % eval_interval == 0 or iter == max_iters - 1:
        losses = estimate_loss()
        elapsed = time.time() - train_start
        total_norm = 0
        for p in miniq_model.parameters():
            if p.grad is not None:
                param_norm = p.grad.detach().data.norm(2)
                total_norm += param_norm.item() ** 2
        total_norm = total_norm ** 0.5

        tokens_per_sec = (iter + 1) * batch_size * \
            block_size / elapsed if elapsed > 0 else 0

        is_best = losses['val'] < best_val_loss
        if is_best:
            best_val_loss = losses['val']
            torch.save(miniq_model.state_dict(), 'llm.pt')

        print(
            f"step {iter} | loss: {losses['train']:.6f} | lr {learning_rate:.4e} | norm: {total_norm:.4f} | dt: {elapsed*1000:.2f}ms | tok/sec: {tokens_per_sec:.2f}")
        sys.stdout.flush()

    xb, yb = get_batch('train')
    logits, loss = miniq_model(xb, yb)
    miniq_optimizer.zero_grad(set_to_none=True)
    loss.backward()
    miniq_optimizer.step()

total_time = time.time() - train_start

print()
print(f"Duration: {int(total_time // 60)}m {int(total_time % 60):02d}s")
print(f"Best val loss: {best_val_loss:.4f}")
print()

miniq_model.load_state_dict(torch.load(
    'mini-quadtrix.pt', map_location=device, weights_only=True))
miniq_model.eval()

print("INFERENCE")
print()

try:
    while True:
        prompt = input("user > ").strip()
        if prompt.lower() in ("quit", "exit", "q"):
            print()
            print("Session ended.")
            break
        if not prompt:
            continue

        encoded_prompt = miniq_encode(prompt, miniq_tokenizer)
        context = torch.tensor(
            [encoded_prompt], dtype=torch.long, device=device)

        with torch.no_grad():
            output_ids = miniq_model.generate(context, max_new_tokens=200)

        new_tokens = output_ids[0][len(encoded_prompt):].tolist()
        response = miniq_decode(new_tokens, miniq_tokenizer).strip()

        print()
        print(f"Model > {response}")
        print()

except KeyboardInterrupt:
    print()
    print("Interrupted.")

wall_clock = time.time() - start

print()
print(f"Training: {int(total_time // 60)}m {int(total_time % 60):02d}s")
print(f"Total: {int(wall_clock // 60)}m {int(wall_clock % 60):02d}s")
print()