standalone serverless simple character level transformer

train_modal_standalone.py

import os
import sys
import time
import math
import pickle
from contextlib import nullcontext
from pathlib import Path
import subprocess
from dataclasses import dataclass
import inspect

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

# Modal imports
import modal

# ============================================================================
# CONFIGURATION - All settings embedded here, no CLI args needed
# ============================================================================

# Modal configuration
N_GPUS = 4  # Number of GPUs to use
GPU_TYPE = "A100"  # GPU type: "A100", "H200", "A10G", etc.

# Training configuration for Shakespeare character-level model
CONFIG = {
    # I/O
    "out_dir": "/data/checkpoints/shakespeare",
    "eval_interval": 250,  # Will be auto-adjusted based on epochs
    "log_interval": 10,  # Will be auto-adjusted based on epochs
    "eval_iters": 200,
    "eval_only": False,
    "always_save_checkpoint": True,
    "init_from": "scratch",
    
    # wandb logging
    "wandb_log": False,
    "wandb_project": "nanogpt-shakespeare",
    "wandb_run_name": "shakespeare-char",
    
    # data
    "dataset": "shakespeare_char",
    "gradient_accumulation_steps": 4,  # Must be divisible by N_GPUS
    "batch_size": 64,
    "block_size": 256,
    
    # model
    "n_layer": 6,
    "n_head": 6,
    "n_embd": 384,
    "dropout": 0.2,
    "bias": False,
    
    # training epochs (max_iters will be calculated automatically)
    "num_epochs": 21.0,  # Set the number of epochs you want
    
    # adamw optimizer
    "learning_rate": 1e-3,
    "max_iters": None,  # Will be calculated based on num_epochs
    "weight_decay": 1e-1,
    "beta1": 0.9,
    "beta2": 0.95,
    "grad_clip": 1.0,
    
    # learning rate decay settings
    "decay_lr": True,  # Will be auto-adjusted based on epochs
    "warmup_iters": None,  # Will be calculated as percentage of max_iters
    "lr_decay_iters": None,  # Will be set to max_iters
    "min_lr": 1e-4,
    
    # DDP settings
    "backend": "nccl",
    
    # system
    "device": "cuda",
    "dtype": "bfloat16",
    "compile": True,
}

# ============================================================================
# MODEL DEFINITION - Embedded from model.py
# ============================================================================

class LayerNorm(nn.Module):
    """ LayerNorm but with an optional bias. PyTorch doesn't support simply bias=False """

    def __init__(self, ndim, bias):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(ndim))
        self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None

    def forward(self, input):
        return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)

class CausalSelfAttention(nn.Module):

    def __init__(self, config):
        super().__init__()
        assert config.n_embd % config.n_head == 0
        # key, query, value projections for all heads, but in a batch
        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
        # output projection
        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
        # regularization
        self.attn_dropout = nn.Dropout(config.dropout)
        self.resid_dropout = nn.Dropout(config.dropout)
        self.n_head = config.n_head
        self.n_embd = config.n_embd
        self.dropout = config.dropout
        # flash attention make GPU go brrrrr but support is only in PyTorch >= 2.0
        self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
        if not self.flash:
            print("WARNING: using slow attention. Flash Attention requires PyTorch >= 2.0")
            # causal mask to ensure that attention is only applied to the left in the input sequence
            self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size))
                                        .view(1, 1, config.block_size, config.block_size))

    def forward(self, x):
        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)

        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
        q, k, v  = self.c_attn(x).split(self.n_embd, dim=2)
        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)

        # causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
        if self.flash:
            # efficient attention using Flash Attention CUDA kernels
            y = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=self.dropout if self.training else 0, is_causal=True)
        else:
            # manual implementation of attention
            att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
            att = att.masked_fill(self.bias[:,:,:T,:T] == 0, float('-inf'))
            att = F.softmax(att, dim=-1)
            att = self.attn_dropout(att)
            y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side

        # output projection
        y = self.resid_dropout(self.c_proj(y))
        return y

class MLP(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.c_fc    = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)
        self.gelu    = nn.GELU()
        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)
        self.dropout = nn.Dropout(config.dropout)

    def forward(self, x):
        x = self.c_fc(x)
        x = self.gelu(x)
        x = self.c_proj(x)
        x = self.dropout(x)
        return x

class Block(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.ln_1 = LayerNorm(config.n_embd, bias=config.bias)
        self.attn = CausalSelfAttention(config)
        self.ln_2 = LayerNorm(config.n_embd, bias=config.bias)
        self.mlp = MLP(config)

    def forward(self, x):
        x = x + self.attn(self.ln_1(x))
        x = x + self.mlp(self.ln_2(x))
        return x

@dataclass
class GPTConfig:
    block_size: int = 1024
    vocab_size: int = 50304 # GPT-2 vocab_size of 50257, padded up to nearest multiple of 64 for efficiency
    n_layer: int = 12
    n_head: int = 12
    n_embd: int = 768
    dropout: float = 0.0
    bias: bool = True # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster

class GPT(nn.Module):

    def __init__(self, config):
        super().__init__()
        assert config.vocab_size is not None
        assert config.block_size is not None
        self.config = config

        self.transformer = nn.ModuleDict(dict(
            wte = nn.Embedding(config.vocab_size, config.n_embd),
            wpe = nn.Embedding(config.block_size, config.n_embd),
            drop = nn.Dropout(config.dropout),
            h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
            ln_f = LayerNorm(config.n_embd, bias=config.bias),
        ))
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        # with weight tying when using torch.compile() some warnings get generated:
        # "UserWarning: functional_call was passed multiple values for tied weights.
        # This behavior is deprecated and will be an error in future versions"
        # not 100% sure what this is, so far seems to be harmless. TODO investigate
        self.transformer.wte.weight = self.lm_head.weight # https://paperswithcode.com/method/weight-tying

        # init all weights
        self.apply(self._init_weights)
        # apply special scaled init to the residual projections, per GPT-2 paper
        for pn, p in self.named_parameters():
            if pn.endswith('c_proj.weight'):
                torch.nn.init.normal_(p, mean=0.0, std=0.02/math.sqrt(2 * config.n_layer))

        # report number of parameters
        print("number of parameters: %.2fM" % (self.get_num_params()/1e6,))

    def get_num_params(self, non_embedding=True):
        """
        Return the number of parameters in the model.
        For non-embedding count (default), the position embeddings get subtracted.
        The token embeddings would too, except due to the parameter sharing these
        params are actually used as weights in the final layer, so we include them.
        """
        n_params = sum(p.numel() for p in self.parameters())
        if non_embedding:
            n_params -= self.transformer.wpe.weight.numel()
        return n_params

    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):
        device = idx.device
        b, t = idx.size()
        assert t <= self.config.block_size, f"Cannot forward sequence of length {t}, block size is only {self.config.block_size}"
        pos = torch.arange(0, t, dtype=torch.long, device=device) # shape (t)

        # forward the GPT model itself
        tok_emb = self.transformer.wte(idx) # token embeddings of shape (b, t, n_embd)
        pos_emb = self.transformer.wpe(pos) # position embeddings of shape (t, n_embd)
        x = self.transformer.drop(tok_emb + pos_emb)
        for block in self.transformer.h:
            x = block(x)
        x = self.transformer.ln_f(x)

        if targets is not None:
            # if we are given some desired targets also calculate the loss
            logits = self.lm_head(x)
            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1)
        else:
            # inference-time mini-optimization: only forward the lm_head on the very last position
            logits = self.lm_head(x[:, [-1], :]) # note: using list [-1] to preserve the time dim
            loss = None

        return logits, loss

    def crop_block_size(self, block_size):
        # model surgery to decrease the block size if necessary
        # e.g. we may load the GPT2 pretrained model checkpoint (block size 1024)
        # but want to use a smaller block size for some smaller, simpler model
        assert block_size <= self.config.block_size
        self.config.block_size = block_size
        self.transformer.wpe.weight = nn.Parameter(self.transformer.wpe.weight[:block_size])
        for block in self.transformer.h:
            if hasattr(block.attn, 'bias'):
                block.attn.bias = block.attn.bias[:,:,:block_size,:block_size]

    def configure_optimizers(self, weight_decay, learning_rate, betas, device_type):
        # start with all of the candidate parameters
        param_dict = {pn: p for pn, p in self.named_parameters()}
        # filter out those that do not require grad
        param_dict = {pn: p for pn, p in param_dict.items() if p.requires_grad}
        # create optim groups. Any parameters that is 2D will be weight decayed, otherwise no.
        # i.e. all weight tensors in matmuls + embeddings decay, all biases and layernorms don't.
        decay_params = [p for n, p in param_dict.items() if p.dim() >= 2]
        nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2]
        optim_groups = [
            {'params': decay_params, 'weight_decay': weight_decay},
            {'params': nodecay_params, 'weight_decay': 0.0}
        ]
        num_decay_params = sum(p.numel() for p in decay_params)
        num_nodecay_params = sum(p.numel() for p in nodecay_params)
        print(f"num decayed parameter tensors: {len(decay_params)}, with {num_decay_params:,} parameters")
        print(f"num non-decayed parameter tensors: {len(nodecay_params)}, with {num_nodecay_params:,} parameters")
        # Create AdamW optimizer and use the fused version if it is available
        fused_available = 'fused' in inspect.signature(torch.optim.AdamW).parameters
        use_fused = fused_available and device_type == 'cuda'
        extra_args = dict(fused=True) if use_fused else dict()
        optimizer = torch.optim.AdamW(optim_groups, lr=learning_rate, betas=betas, **extra_args)
        print(f"using fused AdamW: {use_fused}")

        return optimizer

    def estimate_mfu(self, fwdbwd_per_iter, dt):
        """ estimate model flops utilization (MFU) in units of A100 bfloat16 peak FLOPS """
        # first estimate the number of flops we do per iteration.
        # see PaLM paper Appendix B as ref: https://arxiv.org/abs/2204.02311
        N = self.get_num_params()
        cfg = self.config
        L, H, Q, T = cfg.n_layer, cfg.n_head, cfg.n_embd//cfg.n_head, cfg.block_size
        flops_per_token = 6*N + 12*L*H*Q*T
        flops_per_fwdbwd = flops_per_token * T
        flops_per_iter = flops_per_fwdbwd * fwdbwd_per_iter
        # express our flops throughput as ratio of A100 bfloat16 peak flops
        flops_achieved = flops_per_iter * (1.0/dt) # per second
        flops_promised = 312e12 # A100 GPU bfloat16 peak flops is 312 TFLOPS
        mfu = flops_achieved / flops_promised
        return mfu

    @torch.no_grad()
    def generate(self, idx, max_new_tokens, temperature=1.0, top_k=None):
        """
        Take a conditioning sequence of indices idx (LongTensor of shape (b,t)) and complete
        the sequence max_new_tokens times, feeding the predictions back into the model each time.
        Most likely you'll want to make sure to be in model.eval() mode of operation for this.
        """
        for _ in range(max_new_tokens):
            # if the sequence context is growing too long we must crop it at block_size
            idx_cond = idx if idx.size(1) <= self.config.block_size else idx[:, -self.config.block_size:]
            # forward the model to get the logits for the index in the sequence
            logits, _ = self(idx_cond)
            # pluck the logits at the final step and scale by desired temperature
            logits = logits[:, -1, :] / temperature
            # optionally crop the logits to only the top k options
            if top_k is not None:
                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
                logits[logits < v[:, [-1]]] = -float('Inf')
            # apply softmax to convert logits to (normalized) probabilities
            probs = F.softmax(logits, dim=-1)
            # sample from the distribution
            idx_next = torch.multinomial(probs, num_samples=1)
            # append sampled index to the running sequence and continue
            idx = torch.cat((idx, idx_next), dim=1)

        return idx

# ============================================================================
# DATA PREPARATION
# ============================================================================

def ensure_shakespeare_data(data_root="/data"):
    """Download and prepare Shakespeare dataset if not exists"""
    import requests
    
    data_dir = os.path.join(data_root, "shakespeare_char")
    
    # Check if prepared data already exists
    train_path = os.path.join(data_dir, "train.bin")
    val_path = os.path.join(data_dir, "val.bin")
    meta_path = os.path.join(data_dir, "meta.pkl")
    
    if os.path.exists(train_path) and os.path.exists(val_path) and os.path.exists(meta_path):
        print(f"Shakespeare data already prepared in {data_dir}")
        return
    
    # Create directory
    os.makedirs(data_dir, exist_ok=True)
    
    # Download the tiny shakespeare dataset
    input_file_path = os.path.join(data_dir, 'input.txt')
    if not os.path.exists(input_file_path):
        print("Downloading Shakespeare dataset...")
        data_url = 'https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt'
        with open(input_file_path, 'w') as f:
            f.write(requests.get(data_url).text)
    
    with open(input_file_path, 'r') as f:
        data = f.read()
    print(f"length of dataset in characters: {len(data):,}")
    
    # get all the unique characters that occur in this text
    chars = sorted(list(set(data)))
    vocab_size = len(chars)
    print("all the unique characters:", ''.join(chars))
    print(f"vocab size: {vocab_size:,}")
    
    # create a mapping from characters to integers
    stoi = { ch:i for i,ch in enumerate(chars) }
    itos = { i:ch for i,ch in enumerate(chars) }
    def encode(s):
        return [stoi[c] for c in s] # encoder: take a string, output a list of integers
    def decode(l):
        return ''.join([itos[i] for i in l]) # decoder: take a list of integers, output a string
    
    # create the train and test splits
    n = len(data)
    train_data = data[:int(n*0.9)]
    val_data = data[int(n*0.9):]
    
    # encode both to integers
    train_ids = encode(train_data)
    val_ids = encode(val_data)
    print(f"train has {len(train_ids):,} tokens")
    print(f"val has {len(val_ids):,} tokens")
    
    # export to bin files
    train_ids = np.array(train_ids, dtype=np.uint16)
    val_ids = np.array(val_ids, dtype=np.uint16)
    train_ids.tofile(train_path)
    val_ids.tofile(val_path)
    
    # save the meta information as well, to help us encode/decode later
    meta = {
        'vocab_size': vocab_size,
        'itos': itos,
        'stoi': stoi,
    }
    with open(meta_path, 'wb') as f:
        pickle.dump(meta, f)
    
    print("Data preparation complete!")

# ============================================================================
# TRAINING SCRIPT
# ============================================================================

def train():
    """Main training function that runs under torchrun"""
    
    # Load configuration
    cfg = CONFIG
    
    # Setup DDP
    ddp = int(os.environ.get('RANK', -1)) != -1
    if ddp:
        init_process_group(backend=cfg['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 cfg['gradient_accumulation_steps'] % ddp_world_size == 0
        gradient_accumulation_steps = cfg['gradient_accumulation_steps'] // ddp_world_size
    else:
        # single gpu
        master_process = True
        seed_offset = 0
        ddp_world_size = 1
        device = cfg['device']
        gradient_accumulation_steps = cfg['gradient_accumulation_steps']
    
    tokens_per_iter = gradient_accumulation_steps * ddp_world_size * cfg['batch_size'] * cfg['block_size']
    print(f"tokens per iteration will be: {tokens_per_iter:,}")
    
    if master_process:
        os.makedirs(cfg['out_dir'], exist_ok=True)
    
    torch.manual_seed(1337 + seed_offset)
    torch.backends.cuda.matmul.allow_tf32 = True
    torch.backends.cudnn.allow_tf32 = True
    device_type = 'cuda' if 'cuda' in device else 'cpu'
    
    # Data setup
    data_dir = os.path.join("/data" if os.path.exists("/data") else "data", cfg['dataset'])
    
    # Calculate dataset size and iterations needed for requested epochs
    train_data_path = os.path.join(data_dir, 'train.bin')
    if os.path.exists(train_data_path):
        train_data = np.memmap(train_data_path, dtype=np.uint16, mode='r')
        dataset_tokens = len(train_data)
        print(f"Training dataset has {dataset_tokens:,} tokens")
        
        # Calculate iterations needed for the requested number of epochs
        if cfg['num_epochs'] is not None:
            iterations_per_epoch = dataset_tokens / tokens_per_iter
            cfg['max_iters'] = int(math.ceil(cfg['num_epochs'] * iterations_per_epoch))
            print(f"For {cfg['num_epochs']} epochs, need {cfg['max_iters']} iterations")
            print(f"Each epoch is ~{iterations_per_epoch:.1f} iterations")
            
            # Auto-adjust other parameters based on total iterations
            if cfg['warmup_iters'] is None:
                # Default 2% warmup
                cfg['warmup_iters'] = max(1, int(0.02 * cfg['max_iters']))
            
            if cfg['lr_decay_iters'] is None:
                cfg['lr_decay_iters'] = cfg['max_iters']
            
            # Adjust eval/log intervals for short runs
            if cfg['max_iters'] < 20:
                cfg['eval_interval'] = max(1, cfg['max_iters'] // 4)
                cfg['log_interval'] = 1
                cfg['eval_iters'] = min(50, cfg['eval_iters'])
                print(f"Adjusted for short run: eval_interval={cfg['eval_interval']}, log_interval={cfg['log_interval']}")
            
            # Disable learning rate decay for very short runs
            if cfg['max_iters'] < 10:
                cfg['decay_lr'] = False
                cfg['warmup_iters'] = 0
                print("Disabled learning rate decay for very short run")
        del train_data  # Free memory
    else:
        if cfg['max_iters'] is None:
            raise ValueError("Cannot calculate max_iters: training data not found and max_iters not specified")
    
    def get_batch(split):
        # We recreate np.memmap every batch to avoid a memory leak
        if split == 'train':
            data = np.memmap(os.path.join(data_dir, 'train.bin'), dtype=np.uint16, mode='r')
        else:
            data = np.memmap(os.path.join(data_dir, 'val.bin'), dtype=np.uint16, mode='r')
        ix = torch.randint(len(data) - cfg['block_size'], (cfg['batch_size'],))
        x = torch.stack([torch.from_numpy((data[i:i+cfg['block_size']]).astype(np.int64)) for i in ix])
        y = torch.stack([torch.from_numpy((data[i+1:i+1+cfg['block_size']]).astype(np.int64)) for i in ix])
        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
    
    # Init these up here
    iter_num = 0
    best_val_loss = 1e9
    
    # 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['vocab_size']
        print(f"found vocab_size = {meta_vocab_size} (inside {meta_path})")
    
    # Model configuration
    model_args = dict(
        n_layer=cfg['n_layer'],
        n_head=cfg['n_head'],
        n_embd=cfg['n_embd'],
        block_size=cfg['block_size'],
        bias=cfg['bias'],
        vocab_size=meta_vocab_size if meta_vocab_size is not None else 50304,
        dropout=cfg['dropout']
    )
    
    if cfg['init_from'] == 'scratch':
        print("Initializing a new model from scratch")
        gptconf = GPTConfig(**model_args)
        model = GPT(gptconf)
    elif cfg['init_from'] == 'resume':
        print(f"Resuming training from {cfg['out_dir']}")
        ckpt_path = os.path.join(cfg['out_dir'], 'ckpt.pt')
        checkpoint = torch.load(ckpt_path, map_location=device)
        checkpoint_model_args = checkpoint['model_args']
        for k in ['n_layer', 'n_head', 'n_embd', 'block_size', 'bias', 'vocab_size']:
            model_args[k] = checkpoint_model_args[k]
        gptconf = GPTConfig(**model_args)
        model = GPT(gptconf)
        state_dict = checkpoint['model']
        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']
    
    # Move model to device
    model.to(device)
    
    # Initialize a GradScaler
    scaler = torch.cuda.amp.GradScaler(enabled=(cfg['dtype'] == 'float16'))
    
    # Optimizer
    optimizer = model.configure_optimizers(cfg['weight_decay'], cfg['learning_rate'], 
                                         (cfg['beta1'], cfg['beta2']), device_type)
    if cfg['init_from'] == 'resume' and 'optimizer' in checkpoint:
        optimizer.load_state_dict(checkpoint['optimizer'])
    checkpoint = None # free up memory
    
    # Compile the model
    if cfg['compile']:
        print("compiling the model... (takes a ~minute)")
        unoptimized_model = model
        model = torch.compile(model)
    
    # Wrap model into DDP container
    if ddp:
        model = DDP(model, device_ids=[ddp_local_rank])
    
    # Training helpers
    ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[cfg['dtype']]
    ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)
    
    @torch.no_grad()
    def estimate_loss():
        out = {}
        model.eval()
        for split in ['train', 'val']:
            losses = torch.zeros(cfg['eval_iters'])
            for k in range(cfg['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 decay scheduler (cosine with warmup)
    def get_lr(it):
        # Linear warmup
        if it < cfg['warmup_iters']:
            return cfg['learning_rate'] * (it + 1) / (cfg['warmup_iters'] + 1)
        # If it > lr_decay_iters, return min learning rate
        if it > cfg['lr_decay_iters']:
            return cfg['min_lr']
        # In between, use cosine decay
        decay_ratio = (it - cfg['warmup_iters']) / (cfg['lr_decay_iters'] - cfg['warmup_iters'])
        assert 0 <= decay_ratio <= 1
        coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))
        return cfg['min_lr'] + coeff * (cfg['learning_rate'] - cfg['min_lr'])
    
    # Logging
    if cfg['wandb_log'] and master_process:
        import wandb
        wandb.init(project=cfg['wandb_project'], name=cfg['wandb_run_name'], config=cfg)
    
    # Training loop
    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:
        # Determine and set the learning rate for this iteration
        lr = get_lr(iter_num) if cfg['decay_lr'] else cfg['learning_rate']
        for param_group in optimizer.param_groups:
            param_group['lr'] = lr
        
        # Evaluate the loss on train/val sets and write checkpoints
        if iter_num % cfg['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 cfg['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 cfg['always_save_checkpoint']:
                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': cfg,
                    }
                    print(f"saving checkpoint to {cfg['out_dir']}")
                    torch.save(checkpoint, os.path.join(cfg['out_dir'], 'ckpt.pt'))
        
        if iter_num == 0 and cfg['eval_only']:
            break
        
        # Forward backward update
        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()
        
        # Clip gradients
        if cfg['grad_clip'] != 0.0:
            scaler.unscale_(optimizer)
            torch.nn.utils.clip_grad_norm_(model.parameters(), cfg['grad_clip'])
        
        # Step the optimizer
        scaler.step(optimizer)
        scaler.update()
        optimizer.zero_grad(set_to_none=True)
        
        # Timing and logging
        t1 = time.time()
        dt = t1 - t0
        t0 = t1
        if iter_num % cfg['log_interval'] == 0 and master_process:
            lossf = loss.item() * gradient_accumulation_steps
            if local_iter_num >= 5:
                mfu = raw_model.estimate_mfu(cfg['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
        
        # Termination conditions
        if iter_num > cfg['max_iters']:
            break
    
    if ddp:
        destroy_process_group()

# ============================================================================
# MODAL SETUP
# ============================================================================

# Create Modal app
app = modal.App("nanogpt-training")

# Build Modal image with all dependencies
image = (
    modal.Image.debian_slim(python_version="3.11")
    .pip_install(
        "numpy",
        "torch",
        "transformers",
        "wandb",
        "requests"
    )
)

# Create Modal volume for persistent storage
volume = modal.Volume.from_name("nanogpt-data", create_if_missing=True)

# Modal entry point function
@app.function(
    gpu=f"{GPU_TYPE}:{N_GPUS}",
    volumes={"/data": volume},
    timeout=60 * 60 * 6,  # 6 hours
    image=image,
    secrets=[modal.Secret.from_name("wandb-secret")] if CONFIG.get("wandb_log", False) else [],
)
def train_modal():
    """Launch distributed training on Modal"""
    
    print(f"Starting Modal training with {N_GPUS} {GPU_TYPE} GPUs")
    print(f"Dataset: {CONFIG['dataset']}")
    
    # Prepare data
    ensure_shakespeare_data("/data")
    
    # Copy this script to a temporary location for torchrun
    script_path = Path(__file__)
    script_content = script_path.read_text()
    temp_script = "/tmp/train_modal.py"
    Path(temp_script).write_text(script_content)
    
    # Launch distributed training with torchrun
    cmd = [
        "torchrun",
        f"--nproc-per-node={N_GPUS}",
        temp_script,
    ]
    
    print(f"Running command: {' '.join(cmd)}")
    
    # Change to temp directory to run
    os.chdir("/tmp")
    
    # Launch distributed training
    subprocess.run(cmd, check=True)
    
    print("Training completed successfully!")
    return "Training completed"

# Main entry point
if __name__ == "__main__":
    # Check if we're running under torchrun
    if "RANK" in os.environ:
        # We're running distributed - execute training
        train()
    else:
        # Not running under torchrun
        print("This script should be run with torchrun or through Modal")
        print("Examples:")
        print("  Local: torchrun --nproc-per-node=4 train_modal_standalone.py")
        print("  Modal: modal run train_modal_standalone.py::train_modal")
        sys.exit(1)

self-contained nanogpt training script for modal - no external dependencies.
all configuration is embedded in this file. no cli args, no external model.py needed.

to run on modal:
$ modal run train_modal_standalone.py::train_modal
in detached mode (local process can terminate but training continues on modal):
$ modal run -d train_modal_standalone.py::train_modal