train_gpt2.py

import os
import math
import time
import inspect
from dataclasses import dataclass
import torch
import torch.nn as nn
from torch.nn import functional as F
# -----------------------------------------------------------------------------

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)
        # output projection
        self.c_proj = nn.Linear(config.n_embd, config.n_embd)
        self.c_proj.NANOGPT_SCALE_INIT = 1
        # regularization
        self.n_head = config.n_head
        self.n_embd = config.n_embd

        self.dropout = config.dropout
        self.resid_dropout = nn.Dropout(config.dropout)

    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
        # nh is "number of heads", hs is "head size", and C (number of channels) = nh * hs
        # e.g. in GPT-2 (124M), n_head=12, hs=64, so nh*hs=C=768 channels in the Transformer
        qkv = self.c_attn(x)
        q, k, v = qkv.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)
        y = F.scaled_dot_product_attention(q, k, v, dropout_p=self.dropout if self.training else 0, is_causal=True) # flash attention
        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)
        self.gelu    = nn.GELU(approximate='tanh')
        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd)
        self.c_proj.NANOGPT_SCALE_INIT = 1
        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 = nn.LayerNorm(config.n_embd)
        self.attn = CausalSelfAttention(config)
        self.ln_2 = nn.LayerNorm(config.n_embd)
        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
    
class ConvBlock(nn.Module):

    def __init__(self, config, block_size):
        super().__init__()
        self.block_size = block_size
        self.block = Block(config, in_embed=config.n_embd)

    def forward_foldable(self, x):
        B, T, C = x.size()
        x = x.unfold(1, self.block_size, self.block_size).transpose(2, 3)
        x = x.view(B * self.block_size, T // self.block_size, C)
        x = self.block(x)
        x = x.view(B, T, C)
        return x

    def forward(self, x):
        B, T, C = x.size()
        if T % self.block_size == 0: return self.forward_foldable(x)

        rem_count = T % self.block_size
        x_rem, x_fold = x[:,0:rem_count,:], x[:,rem_count:,:]

        x_rem = self.block(x_rem)
        x_fold = self.forward_foldable(x_fold) if T - rem_count > 0 else x_fold

        return torch.cat((x_rem, x_fold), dim=1)

@dataclass
class GPTConfig:
    block_size: int = 1024 # max sequence length
    vocab_size: int = 50257 # number of tokens: 50,000 BPE merges + 256 bytes tokens + 1 <|endoftext|> token
    n_layer: int = 12 # number of layers
    n_head: int = 12 # number of heads
    n_embd: int = 768 # embedding dimension
    dropout: int = 0

class GPT(nn.Module):

    def __init__(self, config):
        super().__init__()
        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([ConvBlock(config, block_size=64) for _ in range(config.n_layer)]),
            ln_f = nn.LayerNorm(config.n_embd),
        ))
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)

        # weight sharing scheme
        self.transformer.wte.weight = self.lm_head.weight

        # init params
        self.apply(self._init_weights)

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            std = 0.02
            if hasattr(module, 'NANOGPT_SCALE_INIT'):
                std *= (2 * self.config.n_layer) ** -0.5
            torch.nn.init.normal_(module.weight, mean=0.0, std=std)
            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):
        # idx is of shape (B, T)
        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}"
        # forward the token and posisition embeddings
        pos = torch.arange(0, T, dtype=torch.long, device=idx.device) # shape (T)
        pos_emb = self.transformer.wpe(pos) # position embeddings of shape (T, n_embd)
        tok_emb = self.transformer.wte(idx) # token embeddings of shape (B, T, n_embd)
        x = self.transformer.drop(tok_emb + pos_emb)
        # forward the blocks of the transformer
        for block in self.transformer.h:
            x = block(x)
        # forward the final layernorm and the classifier
        x = self.transformer.ln_f(x)
        logits = self.lm_head(x) # (B, T, vocab_size)
        loss = None
        if targets is not None:
            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
        return logits, loss

    def configure_optimizers(self, weight_decay, learning_rate, device_type, betas=(0.9, 0.95)):
        # start with all of the candidate parameters (that require grad)
        param_dict = {pn: p for pn, p in self.named_parameters()}
        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)
        if master_process:
            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"
        if master_process:
            print(f"using fused AdamW: {use_fused}")
        optimizer = torch.optim.AdamW(optim_groups, lr=learning_rate, betas=betas, eps=1e-8, fused=use_fused)
        return optimizer

# -----------------------------------------------------------------------------
import requests
import numpy as np

# data_url = 'https://gist.githubusercontent.com/blakesanie/dde3a2b7e698f52f389532b4b52bc254/raw/76fe1b5e9efcf0d2afdfd78b0bfaa737ad0a67d3/shakespeare.txt'
data_url = 'https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt'

# download the tiny shakespeare dataset
input_file_path = os.path.join(os.path.abspath(''), 'input.txt')
if not os.path.exists(input_file_path):
    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().replace('$', '')
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 = torch.tensor(np.array(train_ids, dtype=np.uint16).astype(np.int32), dtype=torch.long)
val_ids = torch.tensor(np.array(val_ids, dtype=np.uint16).astype(np.int32), dtype=torch.long)

class ShakespeareDataLoaderLite:
    def __init__(self, B, T, process_rank, num_processes, split):
        self.B, self.T = B, T
        assert split in {'train', 'val'}
        self.data = train_ids if split == 'train' else val_ids

    def reset(self):
        pass

    def next_batch(self):
        B, T = self.B, self.T
        ix = torch.randint(len(self.data) - T, (B,))
        x = torch.stack([self.data[i:i+T] for i in ix])
        y = torch.stack([self.data[i+1:i+1+T] for i in ix])
        return x, y

# -----------------------------------------------------------------------------
# simple launch:
# python train_gpt2.py
# DDP launch for e.g. 8 GPUs:
# torchrun --standalone --nproc_per_node=8 train_gpt2.py

# run the training loop
from torch.distributed import init_process_group, destroy_process_group
from torch.nn.parallel import DistributedDataParallel as DDP
import torch.distributed as dist

# set up DDP (distributed data parallel).
# torchrun command sets the env variables RANK, LOCAL_RANK, and WORLD_SIZE
ddp = int(os.environ.get('RANK', -1)) != -1 # is this a ddp run?
if ddp:
    # use of DDP atm demands CUDA, we set the device appropriately according to rank
    assert torch.cuda.is_available(), "for now i think we need CUDA for DDP"
    init_process_group(backend='nccl')
    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 # this process will do logging, checkpointing etc.
else:
    # vanilla, non-DDP run
    ddp_rank = 0
    ddp_local_rank = 0
    ddp_world_size = 1
    master_process = True
    # attempt to autodetect device
    device = "cpu"
    if torch.cuda.is_available():
        device = "cuda"
    elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
        device = "mps"
    print(f"using device: {device}")

# added after video, pytorch can be serious about it's device vs. device_type distinction
device_type = "cuda" if device.startswith("cuda") else "cpu"

torch.manual_seed(1337)
if torch.cuda.is_available():
    torch.cuda.manual_seed(1337)

total_batch_size = 524288 # 2**19, ~0.5M, in number of tokens
B = 64 # micro batch size
T = 256 # sequence length
assert total_batch_size % (B * T * ddp_world_size) == 0, "make sure total_batch_size is divisible by B * T * ddp_world_size"
grad_accum_steps = total_batch_size // (B * T * ddp_world_size)
grad_accum_steps = 1
if master_process:
    print(f"total desired batch size: {total_batch_size}")
    print(f"=> calculated gradient accumulation steps: {grad_accum_steps}")

train_loader = ShakespeareDataLoaderLite(B=B, T=T, process_rank=ddp_rank, num_processes=ddp_world_size, split="train")
val_loader = ShakespeareDataLoaderLite(B=B, T=T, process_rank=ddp_rank, num_processes=ddp_world_size, split="val")

torch.set_float32_matmul_precision('high')

# create model
model = GPT(GPTConfig(block_size=256, vocab_size=vocab_size, n_layer=6, n_head=6, n_embd=96, dropout=0.20))
# model = GPT.from_pretrained("gpt2") # or init from OpenAI GPT-2
model.to(device)
use_compile = False # torch.compile interferes with HellaSwag eval and Generation. TODO fix
if use_compile:
    model = torch.compile(model)
if ddp:
    model = DDP(model, device_ids=[ddp_local_rank])
raw_model = model.module if ddp else model # always contains the "raw" unwrapped model

max_lr = 1e-3
min_lr = max_lr * 0.1
warmup_steps = 100
max_steps = 2000 # 19,073 steps is ~1 epoch, if data is 10B tokens and batch size 0.5M tokens

beta2 = 0.99

log_steps = 100

def get_lr(it):
    # 1) linear warmup for warmup_iters steps
    if it < warmup_steps:
        return max_lr * (it+1) / warmup_steps
    # 2) if it > lr_decay_iters, return min learning rate
    if it > max_steps:
        return min_lr
    # 3) in between, use cosine decay down to min learning rate
    decay_ratio = (it - warmup_steps) / (max_steps - warmup_steps)
    assert 0 <= decay_ratio <= 1
    coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) # coeff starts at 1 and goes to 0
    return min_lr + coeff * (max_lr - min_lr)

# optimize!
optimizer = raw_model.configure_optimizers(weight_decay=0.1, learning_rate=1e-3, device_type=device_type, betas=(0.9, beta2))

# create the log directory we will write checkpoints to and log to
log_dir = "log"
os.makedirs(log_dir, exist_ok=True)
log_file = os.path.join(log_dir, f"log.txt")
with open(log_file, "w") as f: # open for writing to clear the file
    pass

for step in range(max_steps):
    t0 = time.time()
    last_step = (step == max_steps - 1)

    # once in a while evaluate our validation loss
    if step % log_steps == 0 or last_step:
        model.eval()
        val_loader.reset()
        with torch.no_grad():
            val_loss_accum = 0.0
            val_loss_steps = 20
            for _ in range(val_loss_steps):
                x, y = val_loader.next_batch()
                x, y = x.to(device), y.to(device)
                with torch.autocast(device_type=device_type, dtype=torch.float16):
                    logits, loss = model(x, y)
                loss = loss / val_loss_steps
                val_loss_accum += loss.detach()
        if ddp:
            dist.all_reduce(val_loss_accum, op=dist.ReduceOp.AVG)
        if master_process:
            print(f"validation loss: {val_loss_accum.item():.4f}")
            with open(log_file, "a") as f:
                f.write(f"{step} val {val_loss_accum.item():.4f}\n")
            if step > 0 and (step % 5000 == 0 or last_step):
                # optionally write model checkpoints
                checkpoint_path = os.path.join(log_dir, f"model_{step:05d}.pt")
                checkpoint = {
                    'model': raw_model.state_dict(),
                    'config': raw_model.config,
                    'step': step,
                    'val_loss': val_loss_accum.item()
                }
                # you might also want to add optimizer.state_dict() and
                # rng seeds etc., if you wanted to more exactly resume training
                torch.save(checkpoint, checkpoint_path)

    # once in a while generate from the model (except step 0, which is noise)
    if ((step > 0 and step % log_steps == 0) or last_step):
        model.eval()
        num_return_sequences = 4
        max_length = 64
        tokens = encode("\n")
        tokens = torch.tensor(tokens, dtype=torch.long)
        tokens = tokens.unsqueeze(0).repeat(num_return_sequences, 1)
        xgen = tokens.to(device)
        sample_rng = torch.Generator(device=device)
        sample_rng.manual_seed(42 + ddp_rank)
        while xgen.size(1) < max_length:
            # forward the model to get the logits
            with torch.no_grad():
                with torch.autocast(device_type=device_type, dtype=torch.float16):
                    logits, loss = model(xgen) # (B, T, vocab_size)
                # take the logits at the last position
                logits = logits[:, -1, :] # (B, vocab_size)
                # get the probabilities
                probs = F.softmax(logits, dim=-1)
                # do top-k sampling of 50 (huggingface pipeline default)
                # topk_probs here becomes (5, 50), topk_indices is (5, 50)
                topk_probs, topk_indices = torch.topk(probs, 50, dim=-1)
                # select a token from the top-k probabilities
                # note: multinomial does not demand the input to sum to 1
                ix = torch.multinomial(topk_probs, 1, generator=sample_rng) # (B, 1)
                # gather the corresponding indices
                xcol = torch.gather(topk_indices, -1, ix) # (B, 1)
                # append to the sequence
                xgen = torch.cat((xgen, xcol), dim=1)
        # print the generated text
        for i in range(num_return_sequences):
            tokens = xgen[i, :max_length].tolist()
            decoded = decode(tokens)
            print(decoded)

    # do one step of the optimization
    model.train()
    optimizer.zero_grad()
    loss_accum = 0.0
    for micro_step in range(grad_accum_steps):
        x, y = train_loader.next_batch()
        x, y = x.to(device), y.to(device)
        # added after video, this field is also used by the forward pass.
        if ddp:
            model.require_backward_grad_sync = (micro_step == grad_accum_steps - 1)
        with torch.autocast(device_type=device_type, dtype=torch.float16):
            logits, loss = model(x, y)
        # we have to scale the loss to account for gradient accumulation,
        # because the gradients just add on each successive backward().
        # addition of gradients corresponds to a SUM in the objective, but
        # instead of a SUM we want MEAN. Scale the loss here so it comes out right
        loss = loss / grad_accum_steps
        loss_accum += loss.detach()
        loss.backward()
    if ddp:
        dist.all_reduce(loss_accum, op=dist.ReduceOp.AVG)
    norm = torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
    # determine and set the learning rate for this iteration
    lr = get_lr(step)
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr
    optimizer.step()
    if device_type == "cuda":
        torch.cuda.synchronize() # wait for the GPU to finish work
    t1 = time.time()
    dt = t1 - t0 # time difference in seconds
    tokens_processed = train_loader.B * train_loader.T * grad_accum_steps * ddp_world_size
    tokens_per_sec = tokens_processed / dt
    if master_process:
        print(f"step {step:5d} | loss: {loss_accum.item():.6f} | lr {lr:.4e} | norm: {norm:.4f} | dt: {dt*1000:.2f}ms | tok/sec: {tokens_per_sec:.2f}")
        with open(log_file, "a") as f:
            f.write(f"{step} train {loss_accum.item():.6f}\n")

if ddp:
    destroy_process_group()