train.py

"""
CodeGPT Training Script.

Supports:
  - Single GPU and multi-GPU (DDP) training
  - Fill-in-the-Middle (FIM) data augmentation
  - Mixed precision (float16/bfloat16)
  - Gradient accumulation
  - Cosine learning rate schedule with warmup
  - Checkpoint save/resume
  - wandb logging (optional)

Usage:
    # Single GPU
    python train.py config/train_codegpt.py

    # Multi-GPU (DDP)
    torchrun --standalone --nproc_per_node=4 train.py config/train_codegpt.py

    # Override config
    python train.py config/train_codegpt_small.py --batch_size=32 --max_iters=5000
"""

# Memory allocator optimization (from autoresearch): reduces fragmentation on long runs
import os
os.environ["PYTORCH_ALLOC_CONF"] = "expandable_segments:True"

import gc
import time
import math
import pickle
import random
from contextlib import nullcontext

import numpy as np
import torch
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.distributed import init_process_group, destroy_process_group

from model import CodeGPT, CodeGPTConfig
from tokenizer import apply_fim_transform, SPECIAL_TOKENS

# ---------- default config ----------
# I/O
out_dir = 'out-codegpt'
eval_interval = 1000
log_interval = 10
eval_iters = 200
eval_only = False
always_save_checkpoint = True
init_from = 'scratch'  # 'scratch', 'resume', 'gpt2', 'gpt2-medium', etc.

# wandb
wandb_log = False
wandb_project = 'codegpt'
wandb_run_name = 'run' + str(time.time())

# data
dataset = 'python_code'
gradient_accumulation_steps = 8
batch_size = 16
block_size = 1024

# model
n_layer = 12
n_head = 12
n_embd = 768
dropout = 0.1
bias = False

# code-specific
fim_enabled = True
fim_rate = 0.5
fim_spm_rate = 0.5

# adamw optimizer
learning_rate = 3e-4
max_iters = 100000
weight_decay = 0.1
beta1 = 0.9
beta2 = 0.95
grad_clip = 1.0

# learning rate schedule
decay_lr = True
warmup_iters = 1000
lr_decay_iters = 100000
min_lr = 3e-5

# DDP
backend = 'nccl'

# system
device = 'cuda' if torch.cuda.is_available() else ('mps' if torch.backends.mps.is_available() else 'cpu')
dtype = 'bfloat16' if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else 'float16'
compile = True if torch.cuda.is_available() else False

# ---------- end default config ----------

# apply config file and CLI overrides
from configurator import configure
configure()

# ---------- DDP setup ----------
ddp = int(os.environ.get('RANK', -1)) != -1
if ddp:
    init_process_group(backend=backend)
    ddp_rank = int(os.environ['RANK'])
    ddp_local_rank = int(os.environ['LOCAL_RANK'])
    ddp_world_size = int(os.environ['WORLD_SIZE'])
    device = f'cuda:{ddp_local_rank}'
    torch.cuda.set_device(device)
    master_process = ddp_rank == 0
    seed_offset = ddp_rank
    assert gradient_accumulation_steps % ddp_world_size == 0
    gradient_accumulation_steps //= ddp_world_size
else:
    master_process = True
    seed_offset = 0
    ddp_world_size = 1

tokens_per_iter = gradient_accumulation_steps * ddp_world_size * batch_size * block_size
if master_process:
    print(f"tokens per iteration: {tokens_per_iter:,}")
    os.makedirs(out_dir, exist_ok=True)

torch.manual_seed(1337 + seed_offset)
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
torch.set_float32_matmul_precision("high")
device_type = 'cuda' if 'cuda' in device else ('mps' if 'mps' in device else 'cpu')
ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)

# ---------- data loading ----------
data_dir = os.path.join('data', dataset)


def get_batch(split):
    data = np.memmap(os.path.join(data_dir, f'{split}.bin'), dtype=np.uint16, mode='r')
    ix = torch.randint(len(data) - block_size, (batch_size,))
    x = torch.stack([torch.from_numpy((data[i:i + block_size]).astype(np.int64)) for i in ix])
    y = torch.stack([torch.from_numpy((data[i + 1:i + 1 + block_size]).astype(np.int64)) for i in ix])

    # apply FIM transformation for code completion training
    if fim_enabled and split == 'train':
        x_fim = []
        y_fim = []
        for b in range(batch_size):
            tokens = x[b].tolist()
            tokens_transformed = apply_fim_transform(tokens, fim_rate=fim_rate, fim_spm_rate=fim_spm_rate)
            # pad or truncate to block_size
            if len(tokens_transformed) > block_size:
                tokens_transformed = tokens_transformed[:block_size]
            elif len(tokens_transformed) < block_size:
                tokens_transformed = tokens_transformed + [SPECIAL_TOKENS["<|fim_pad|>"]] * (block_size - len(tokens_transformed))
            x_fim.append(torch.tensor(tokens_transformed[:-1], dtype=torch.long))

            # target: shifted by 1 (same length as x), with padding tokens masked
            target = tokens_transformed[1:]
            target_tensor = torch.tensor(target, dtype=torch.long)
            # mask padding tokens in loss
            target_tensor[target_tensor == SPECIAL_TOKENS["<|fim_pad|>"]] = -1
            y_fim.append(target_tensor)

        x = torch.stack(x_fim)
        y = torch.stack(y_fim)

    if device_type == 'cuda':
        x, y = x.pin_memory().to(device, non_blocking=True), y.pin_memory().to(device, non_blocking=True)
    else:
        x, y = x.to(device), y.to(device)
    return x, y


# ---------- model init ----------
meta_path = os.path.join(data_dir, 'meta.pkl')
meta_vocab_size = None
if os.path.exists(meta_path):
    with open(meta_path, 'rb') as f:
        meta = pickle.load(f)
    meta_vocab_size = meta.get('vocab_size', None)
    if master_process:
        print(f"found vocab_size = {meta_vocab_size} (from {meta_path})")

model_args = dict(
    n_layer=n_layer, n_head=n_head, n_embd=n_embd,
    block_size=block_size, bias=bias, dropout=dropout,
    fim_enabled=fim_enabled,
)

iter_num = 0
best_val_loss = 1e9

if init_from == 'scratch':
    if master_process:
        print("Initializing CodeGPT model from scratch")
    if meta_vocab_size is not None:
        model_args['vocab_size'] = meta_vocab_size
    config = CodeGPTConfig(**model_args)
    model = CodeGPT(config)

elif init_from == 'resume':
    if master_process:
        print(f"Resuming training from {out_dir}")
    ckpt_path = os.path.join(out_dir, 'ckpt.pt')
    checkpoint = torch.load(ckpt_path, map_location=device, weights_only=False)
    checkpoint_model_args = checkpoint['model_args']
    for k in ['n_layer', 'n_head', 'n_embd', 'block_size', 'bias', 'vocab_size']:
        if k in checkpoint_model_args:
            model_args[k] = checkpoint_model_args[k]
    config = CodeGPTConfig(**model_args)
    model = CodeGPT(config)
    state_dict = checkpoint['model']
    # fix key prefixes from DDP
    unwanted_prefix = '_orig_mod.'
    for k, v in list(state_dict.items()):
        if k.startswith(unwanted_prefix):
            state_dict[k[len(unwanted_prefix):]] = state_dict.pop(k)
    model.load_state_dict(state_dict)
    iter_num = checkpoint['iter_num']
    best_val_loss = checkpoint['best_val_loss']

elif init_from.startswith('gpt2'):
    if master_process:
        print(f"Initializing from pretrained {init_from}")
    model = CodeGPT.from_pretrained(init_from, override_args={'dropout': dropout})
    # expand vocab for code special tokens
    model.expand_vocab(CodeGPTConfig.vocab_size)
    if block_size < model.config.block_size:
        model.crop_block_size(block_size)
    model_args['vocab_size'] = model.config.vocab_size

model.to(device)

# ---------- GradScaler for mixed precision ----------
scaler = torch.amp.GradScaler(device_type, enabled=(dtype == 'float16'))

# ---------- optimizer ----------
optimizer = model.configure_optimizers(weight_decay, learning_rate, (beta1, beta2), device_type)
if init_from == 'resume':
    optimizer.load_state_dict(checkpoint['optimizer'])
checkpoint = None  # free memory

# ---------- compile ----------
if compile:
    if master_process:
        print("compiling the model... (takes a ~minute)")
    model = torch.compile(model)

# ---------- DDP wrapper ----------
if ddp:
    model = DDP(model, device_ids=[ddp_local_rank])


# ---------- loss estimation ----------
@torch.no_grad()
def estimate_loss():
    out = {}
    model.eval()
    for split in ['train', 'val']:
        losses = torch.zeros(eval_iters)
        for k in range(eval_iters):
            X, Y = get_batch(split)
            with ctx:
                logits, loss = model(X, Y)
            losses[k] = loss.item()
        out[split] = losses.mean()
    model.train()
    return out


# ---------- learning rate schedule ----------
def get_lr(it):
    # linear warmup
    if it < warmup_iters:
        return learning_rate * it / warmup_iters
    # cosine decay
    if it > lr_decay_iters:
        return min_lr
    decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)
    coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))
    return min_lr + coeff * (learning_rate - min_lr)


# ---------- logging ----------
if wandb_log and master_process:
    import wandb
    wandb.init(project=wandb_project, name=wandb_run_name, config={
        'batch_size': batch_size, 'block_size': block_size,
        'n_layer': n_layer, 'n_head': n_head, 'n_embd': n_embd,
        'learning_rate': learning_rate, 'max_iters': max_iters,
        'fim_enabled': fim_enabled, 'fim_rate': fim_rate,
        'dataset': dataset,
    })

# ---------- training loop ----------
if master_process:
    print(f"Starting CodeGPT training")
    print(f"  dataset: {dataset}")
    print(f"  FIM enabled: {fim_enabled} (rate={fim_rate})")
    print(f"  device: {device}, dtype: {dtype}")

X, Y = get_batch('train')
t0 = time.time()
local_iter_num = 0
raw_model = model.module if ddp else model
running_mfu = -1.0

while True:
    # set learning rate
    lr = get_lr(iter_num) if decay_lr else learning_rate
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

    # evaluate and checkpoint
    if iter_num % eval_interval == 0 and master_process:
        losses = estimate_loss()
        print(f"step {iter_num}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")
        if wandb_log:
            wandb.log({
                "iter": iter_num,
                "train/loss": losses['train'],
                "val/loss": losses['val'],
                "lr": lr,
                "mfu": running_mfu * 100,
            })
        if losses['val'] < best_val_loss or always_save_checkpoint:
            best_val_loss = min(best_val_loss, losses['val'])
            if iter_num > 0:
                checkpoint = {
                    'model': raw_model.state_dict(),
                    'optimizer': optimizer.state_dict(),
                    'model_args': model_args,
                    'iter_num': iter_num,
                    'best_val_loss': best_val_loss,
                    'config': raw_model.config,
                }
                print(f"saving checkpoint to {out_dir}")
                torch.save(checkpoint, os.path.join(out_dir, 'ckpt.pt'))

    if eval_only:
        break

    # forward/backward with gradient accumulation
    for micro_step in range(gradient_accumulation_steps):
        if ddp:
            model.require_backward_grad_sync = (micro_step == gradient_accumulation_steps - 1)
        with ctx:
            logits, loss = model(X, Y)
            loss = loss / gradient_accumulation_steps
        X, Y = get_batch('train')
        scaler.scale(loss).backward()

    # gradient clipping
    if grad_clip != 0.0:
        scaler.unscale_(optimizer)
        torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)

    scaler.step(optimizer)
    scaler.update()
    optimizer.zero_grad(set_to_none=True)

    # logging
    t1 = time.time()
    dt = t1 - t0
    t0 = t1
    if iter_num % log_interval == 0 and master_process:
        lossf = loss.item() * gradient_accumulation_steps
        if local_iter_num >= 5:
            mfu = raw_model.estimate_mfu(batch_size * gradient_accumulation_steps, dt)
            running_mfu = mfu if running_mfu == -1.0 else 0.9 * running_mfu + 0.1 * mfu
        print(f"iter {iter_num}: loss {lossf:.4f}, time {dt*1000:.2f}ms, mfu {running_mfu*100:.2f}%")

    iter_num += 1
    local_iter_num += 1

    # GC management (from autoresearch): freeze after first iter to prevent ~500ms GC stalls
    if local_iter_num == 1:
        gc.collect()
        gc.freeze()
        gc.disable()

    if iter_num > max_iters:
        break

if ddp:
    destroy_process_group()

print("Training complete!")