chess-mamba-vs-xformer/train_bygame.py

import os
import time
import math
import pickle
from contextlib import nullcontext
import numpy as np
import torch
import torch.nn.functional as F
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.distributed import init_process_group, destroy_process_group
import pyarrow.parquet as pq
import random
from torch.utils.data import Dataset, DataLoader
import glob

# -----------------------------------------------------------------------------
# default config values designed for Mamba model training
# I/O
out_dir = 'out'
eval_interval = 2000
log_interval = 1
eval_iters = 5
eval_only = False
always_save_checkpoint = True
init_from = 'resume'  # 'scratch', 'resume', 'anneal', or Mamba model name
# wandb logging
wandb_log = False
wandb_project = 'mamba'
wandb_run_name = 'mamba_run'  # modify as needed
# data
dataset = 'chess'  # specify your dataset
gradient_accumulation_steps = 5 * 8
batch_size = 12
base_batch_size = batch_size
effective_batch_size = batch_size
max_seq_len = 1024 # For xformer, this is the block size
train_file_update_interval = 7

# model
model_type = 'mamba'
# TODO: add 'xformer' type / model paramers. move model imports to after exec() (when these values finalized)
n_layer = 12
d_model = 768
dt_rank = 'auto'
d_state = 16
expand_factor = 2
bias = False
conv_bias = True
pscan = True
vocab_size = 32
move_num_in_gamestate = True
# xformer-specific params. Note that n_layer, vocab_size, move_num_in_gamestate, and bias are shared by both model types
n_head = 12
n_embd = 768
dropout = 0.0 # for pretraining 0 is good, for finetuning try 0.1+

# optimizer settings
learning_rate = 6e-4
max_iters = 600000  # max_iters is for auto-stopping end of stable phase
weight_decay = 1e-1
beta1 = 0.9
beta2 = 0.95
grad_clip = 0.5
auto_clip = False
auto_clip_max = 0.5
auto_clip_min = 3.333e-3
grad_clip_start_size = 100
grad_clip_max_size = 500
grad_clip_percentile = 10
# learning rate decay settings
decay_lr = True
warmup_iters = 2000
min_lr = 6e-5
# DDP settings
backend = 'nccl'
# system
device = 'cuda' if torch.cuda.is_available() else 'cpu'
dtype = 'bfloat16' if torch.cuda.is_bf16_supported() else 'float32'
compile = False  # set to True if using PyTorch 2.0
# -----------------------------------------------------------------------------

config_keys = [k for k, v in globals().items() if not k.startswith('_') and isinstance(v, (int, float, bool, str))]
exec(open('configurator.py').read())  # overrides from command line or config file
config = {k: globals()[k] for k in config_keys}  # will be useful for logging
# -----------------------------------------------------------------------------


anneal_checkpoint = 'anneal/ckpt.pt'
anneal_dir = os.path.join(out_dir, 'anneal/')
anneal_start_iters = None # Set at init
anneal_decay_iters = None # Set at init

if model_type == 'mamba':
    from mamba_lm import MambaLM, MambaLMConfig
    from mamba_ssm import MambaLMHeadModel
    model_config = MambaLMConfig(
        d_model=d_model,
        #n_layers=n_layer,
        n_layer=n_layer,
        ssm_cfg={
           'dt_rank': dt_rank,
           'd_state': d_state,
           #'expand_factor': expand_factor,
           'bias': bias,
           'conv_bias':conv_bias,
           #'pscan':pscan,
        },
        vocab_size=vocab_size,
        pad_vocab_size_multiple=1
    ).to_mamba_config()
elif model_type == 'xformer':
    from xformer import GPTConfig, GPT
    model_config = GPTConfig(
        n_layer=n_layer,
        n_head=n_head,
        n_embd=n_embd,
        block_size=max_seq_len,
        bias=bias,
        vocab_size=vocab_size,
        dropout=dropout)
else:
    print(f"Unknown model_type {model_type}.")
    exit()

# DDP and other initializations
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

if master_process:
    os.makedirs(out_dir, exist_ok=True)
    os.makedirs(anneal_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'
ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16}[dtype]
ctx = nullcontext() if device_type == 'cpu' else torch.amp.autocast(device_type=device_type, dtype=ptdtype)

# poor man's data loader
data_dir = os.path.join('data', dataset)
current_train_file_index = 0
train_files = glob.glob(os.path.join(data_dir, 'train*.parquet')) + \
              glob.glob(os.path.join(data_dir, 'stable*.parquet')) + \
              glob.glob(os.path.join(data_dir, 'anneal*.parquet'))
train_datasets = []
print("Loading dataset...")
for f in train_files:
    dataset = pq.read_table(f).to_pandas()
    dataset = dataset[dataset['tokenized'].apply(len) >= 8]
    train_datasets.append(dataset)
    print('.',end='',flush=True)
print("\nLoaded.")
#val_data = pq.read_table(os.path.join(data_dir, 'val.parquet')).to_pandas()
#val_data = val_data[val_data['tokenized'].apply(len) >= 8]
truncated_games_count = 0
total_games_count = 0
games_seen = 0
tokens_seen = 0
tokens_seen_padded = 0
def get_batch(split):
    global truncated_games_count, total_games_count, current_train_file_index, tokens_seen, tokens_seen_padded

    # Randomly select batch_size games
    dataset = train_datasets[current_train_file_index] if split == 'train' else None # else val_data  # Use the correct DataFrame based on the split
    sample_df = dataset.sample(batch_size)
    games = sample_df['tokenized'].tolist()

    # Prepare sequences tensor for the batch
    max_length_in_batch = min(max(len(game) for game in games), max_seq_len)
    pad_to = max_length_in_batch #if model_type == 'mamba' else max_seq_len
    sequences = torch.zeros((batch_size, pad_to), dtype=torch.int64)

    for i, game in enumerate(games):
        total_games_count += 1
        game_len = min(len(game), pad_to)
        tokens_seen += game_len
        tokens_seen_padded += pad_to
        sequences[i, :game_len] = torch.tensor(game[:game_len], dtype=torch.int64)
            
    if (total_games_count // batch_size) % train_file_update_interval == 0:
        current_train_file_index = random.randint(0, len(train_files) - 1)
        # print(f"Switched to file: {train_files[current_train_file_index]}")
            
    if device_type == 'cuda':
        sequences = sequences.pin_memory().to(device, non_blocking=True)
    else:
        sequences = sequences.to(device)

    return sequences, max_length_in_batch
    
# init these up here, can override if init_from='resume' (i.e. from a checkpoint)
iter_num = 0
best_val_loss = 1e9

# attempt to derive vocab_size from the dataset
meta_path = os.path.join(data_dir, 'meta.pkl')
meta_vocab_size = None
if not move_num_in_gamestate:
    meta_vocab_size = 28
elif 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 initialization
if init_from == 'scratch':
    print(f"Initializing a new {model_type} model from scratch")
    if meta_vocab_size is None:
        print(f"defaulting to vocab_size of {vocab_size}")
    else:
        model_config.vocab_size = meta_vocab_size
    if model_type == 'mamba':
        #model = MambaLM(model_config)
        model = MambaLMHeadModel(model_config)
    else:
        model = GPT(model_config)
    if auto_clip:
        grad_clip = 0
        config['grad_clip'] = 0
        grad_norm_history = []
elif init_from == 'resume' or init_from == 'anneal':
    print(f"Resuming training from {out_dir}")
    if init_from == 'anneal':
        ckpt_path = os.path.join(out_dir, anneal_checkpoint)
    else:
        ckpt_path = os.path.join(out_dir, 'ckpt.pt')
    checkpoint = torch.load(ckpt_path, map_location=device)
    model_config = checkpoint['model_args']
    if model_type == 'mamba':
        #model = MambaLM(model_config)
        model = MambaLMHeadModel(model_config)
    else:
        model = GPT(model_config)
    state_dict = checkpoint['model']
    # fix the keys of the state dictionary :(
    # honestly no idea how checkpoints sometimes get this prefix, have to debug more
    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)
    if 'effective_batch_size' not in checkpoint['config']:
        print("Checkpoint was saved without `effective_batch_size`, assuming current value (will save with next checkpoint). This is used for correcting `iter_num` when the effetive batch size is changed.")
        checkpoint['config']['effective_batch_size'] = effective_batch_size
    iter_num = int(round(checkpoint['iter_num'] * (checkpoint['config']['effective_batch_size'] / effective_batch_size)))
    if 'games_seen' in checkpoint:
        games_seen = checkpoint['games_seen']
    else:
        games_seen = checkpoint['config']['effective_batch_size'] * checkpoint['iter_num']
        checkpoint['games_seen'] = games_seen
        print(f"Checkpoint was saved without `games_seen`, assuming checkpoint's effective batch size * iters (will save with next checkpoint). {games_seen}")
    tokens_seen = checkpoint.get('tokens_seen', 0)
    tokens_seen_padded = checkpoint.get('tokens_seen_padded', 0)
    best_val_loss = checkpoint['best_val_loss']
    print(f"Best val loss: {best_val_loss}")
    if auto_clip:
        grad_clip = checkpoint['config']['grad_clip']
        config['grad_clip'] = grad_clip
        #grad_norm_history = [t.item() if torch.is_tensor(t) else t for t in checkpoint.get('grad_norm_history', [])]
        grad_norm_history = checkpoint.get('grad_norm_history', [])
    if init_from == 'anneal':
        print(f"\n\nANNEAL STARTING/RESUMING FROM ITERNUM: {iter_num} ({games_seen} games)\n\n")
        anneal_start_iters = iter_num if 'anneal_start_iters' not in checkpoint else checkpoint['anneal_start_iters']
        anneal_decay_iters = iter_num / 8 if 'anneal_decay_iters' not in checkpoint else checkpoint['anneal_decay_iters'] # / 9 is og, but going deeper on lr too (can always take earlier ckpt during anneal if it doesn't keep improving)... have used 6.75
        print(anneal_start_iters)
        print(anneal_decay_iters)
        if 'anneal_start_iters' not in checkpoint:
            grad_clip = 0
            config['grad_clip'] = 0
            grad_norm_history = []
            print(f"Starting anneal. Resumed from {anneal_checkpoint}, will now decay learning rate for {anneal_decay_iters} / until iter_num {anneal_start_iters + anneal_decay_iters}.")
        out_dir = anneal_dir
        weight_decay = weight_decay / 12.5  # / 17.0
        beta2 = np.sqrt(beta2) * beta2
        auto_clip = True
        grad_clip_percentile = 6.75
elif init_from.startswith('state-spaces'):
    print(f"Initializing from Mamba pre-trained weights: {init_from}")
    model = from_pretrained(init_from)
    model_config = model.config
else:
    raise ValueError("Invalid init_from value")

model.to(device)

print(f'Model with {sum([p.numel() for p in model.parameters()])} parameters loaded.')

# Optimizer and GradScaler
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate, weight_decay=weight_decay, betas=(beta1, beta2))
scaler = torch.cuda.amp.GradScaler(enabled=dtype == 'float16')
if init_from == 'resume':
    optimizer.load_state_dict(checkpoint['optimizer'])
checkpoint = None

# Compile the model if using PyTorch 2.0
if compile:
    print("compiling the model... (takes a ~minute)")
    model = torch.compile(model)

# Wrap model in DDP container if necessary
if ddp:
    model = DDP(model, device_ids=[ddp_local_rank])
            
            
def batch_to_loss(sequences, max_length_in_batch):
    if model_type == 'mamba':
        logits = model(sequences[:, :-1]).logits  # Forward pass, exclude last token for input
        # Compute loss (assuming next token prediction task)
        targets = sequences[:, 1:].reshape(-1)  # Shifted by one for next token prediction
        return F.cross_entropy(logits.view(-1, logits.size(-1)), targets)
        #return F.cross_entropy(logits.reshape(-1), targets)
    else:
        inputs = sequences[:, :-1]
        targets = sequences[:, 1:].reshape(-1)
        _, loss = model(inputs, targets)
        return loss
    

@torch.no_grad()
def estimate_loss():
    global tokens_seen, tokens_seen_padded
    out = {}
    model.eval()
    tokens_seen_b4 = tokens_seen
    tokens_seen_padded_b4 = tokens_seen_padded
    for split in ['train']: #['train', 'val']:
        losses = torch.zeros(eval_iters)
        for k in range(eval_iters):
            loss = batch_to_loss(*get_batch(split))
            losses[k] = loss.item()

        split = 'val' # Temporary hack
        out[split] = losses.mean()
    tokens_seen = tokens_seen_b4
    tokens_seen_padded = tokens_seen_padded_b4
    model.train()
    return out


# WSD scheduler
def get_lr(it):
    if init_from == 'anneal':
        # Linear decay from max LR to min LR over (anneal_start_iters / 9) iters
        decay_ratio = min(it - anneal_start_iters, anneal_decay_iters) / anneal_decay_iters
        return learning_rate - decay_ratio * (learning_rate - min_lr)

    if it < warmup_iters:
        # Warmup
        return learning_rate * it / warmup_iters
    
    # Stable max LR
    return learning_rate

# Logging setup
if wandb_log and master_process:
    import wandb
    wandb.init(project=wandb_project, name=wandb_run_name, config=config)

# Training loop
local_iter_num = 0  # Number of iterations in the lifetime of this process
last_crossed_multiple = 0 
save_every_n_games = 150000
raw_model = model.module if ddp else model  # Unwrap DDP container if needed

# initial save
if init_from == 'scratch':
    checkpoint = {
        'model': raw_model.state_dict(),
        'optimizer': optimizer.state_dict(),
        'model_args': model_config,
        'iter_num': 0,
        "games_seen": 0,
        "tokens_seen": 0,
        "tokens_seen_padded": 0,
        'best_val_loss': best_val_loss,
        'config': config,
    }
    checkpoint['grad_norm_history'] = grad_norm_history
    print(f"saving checkpoint to {out_dir}\n")
    torch.save(checkpoint, os.path.join(out_dir, 'ckpt.pt'))

t0 = time.time()
while True:
    # Determine and set the learning rate for this iteration
    lr = get_lr(iter_num) if decay_lr else 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 % eval_interval == 0 and master_process and local_iter_num > 0:
        torch.cuda.empty_cache()
        losses = estimate_loss()
        if init_from == 'anneal':
            print(f"\ngame {games_seen} ({iter_num}, {(iter_num-anneal_start_iters) / anneal_decay_iters:.3%}): 'val' loss {losses['val']:.4f}")
        else:
            print(f"\ngame {games_seen} ({iter_num}, {iter_num / max_iters:.3%}): 'val' loss {losses['val']:.4f}")
        if auto_clip and len(grad_norm_history) >= grad_clip_start_size:
            grad_clip_prev = grad_clip
            grad_clip = np.percentile(grad_norm_history, grad_clip_percentile)
            grad_clip = max(min(grad_clip, auto_clip_max), auto_clip_min)
            # Transition between grad_clips smoothly, weighed to new value
            grad_clip = (grad_clip*9.0 + grad_clip_prev*4.0) / 13.0
            grad_clip = max(min(grad_clip, auto_clip_max), auto_clip_min)  # should never actually clip here
            config['grad_clip'] = grad_clip
            print(f"Auto adjusted grad_clip to {grad_clip}")
        torch.cuda.empty_cache()
        if wandb_log:
            wandb.log({
                "etc/iter": iter_num,
                "etc/games": games_seen,
                "etc/tokens_seen": tokens_seen,
                "etc/tokens_seen_padded": tokens_seen_padded,
                "etc/grad_clip": grad_clip,
                "etc/lr": lr,
                "val/loss": losses['val'],
                
            })
        if losses['val'] < best_val_loss or always_save_checkpoint:
            if iter_num > 0:
                checkpoint = {
                    'model': raw_model.state_dict(),
                    'optimizer': optimizer.state_dict(),
                    'model_args': model_config,
                    'iter_num': iter_num,
                    "games_seen": games_seen,
                    "tokens_seen": tokens_seen,
                    "tokens_seen_padded": tokens_seen_padded,
                    'best_val_loss': min(best_val_loss, losses['val']),
                    'config': config,
                }
                checkpoint['grad_norm_history'] = grad_norm_history
                if init_from == 'anneal':
                    checkpoint['anneal_start_iters'] = anneal_start_iters
                    checkpoint['anneal_decay_iters'] = anneal_decay_iters
                print(f"saving checkpoint to {out_dir}\n")
                torch.save(checkpoint, os.path.join(out_dir, 'ckpt.pt'))
                current_nearest_multiple = (games_seen // save_every_n_games) * save_every_n_games
                if losses['val'] < best_val_loss: # Temporary / only good after it's settled
                    best_val_loss = losses['val']
                    torch.save(checkpoint, os.path.join(out_dir, f'ckpt_{int(games_seen)}b.pt'))
                elif current_nearest_multiple != last_crossed_multiple: # elif so we don't double up
                    last_crossed_multiple = current_nearest_multiple
                    torch.save(checkpoint, os.path.join(out_dir, f'ckpt_{int(games_seen)}.pt'))
                
    if iter_num == 0 and eval_only:
        break

    # Forward and backward pass
    for micro_step in range(gradient_accumulation_steps):
        if ddp:
            model.require_backward_grad_sync = (micro_step == gradient_accumulation_steps - 1)

        sequences, max_length_in_batch = get_batch('train')  # Fetch the training data
        with ctx:
            loss = batch_to_loss(sequences, max_length_in_batch)
            loss = loss / gradient_accumulation_steps

        scaler.scale(loss).backward()
        #print('.', end='')

    # clip the gradient
    if grad_clip != 0.0 or auto_clip:
        scaler.unscale_(optimizer)
        total_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip if grad_clip != 0.0 else 999.9) # The 0 check is for auto_clip enabled but not enough history
        grad_norm_history.append(total_norm.item())
        grad_norm_history = grad_norm_history[-grad_clip_max_size:]
        
    # step the optimizer and scaler if training in fp16
    scaler.step(optimizer)
    scaler.update()
    # flush the gradients as soon as we can, no need for this memory anymore
    optimizer.zero_grad(set_to_none=True)
    #torch.cuda.empty_cache()
    
    # timing and logging
    t1 = time.time()
    dt = t1 - t0
    t0 = t1
    if iter_num % log_interval == 0 and master_process:
        # get loss as float. note: this is a CPU-GPU sync point
        # scale up to undo the division above, approximating the true total loss (exact would have been a sum)
        lossf = loss.item() * gradient_accumulation_steps
        if init_from == 'anneal':
            print(f"game {games_seen} ({iter_num}, {(iter_num-anneal_start_iters) / anneal_decay_iters:.3%}): loss {lossf:.4f}, time {dt*1000:.2f}ms")
        else:
            print(f"game {games_seen} ({iter_num}, {iter_num / max_iters:.3%}): loss {lossf:.4f}, time {dt*1000:.2f}ms")
        if wandb_log:
            wandb.log({
                "etc/iter": iter_num,
                "etc/games": games_seen,
                "etc/tokens_seen": tokens_seen,
                "etc/tokens_seen_padded": tokens_seen_padded,
                "etc/grad_norm": grad_norm_history[-1] if grad_norm_history else 0,
                "etc/lr": lr,
                "train/loss": lossf,
            })
    iter_num += 1
    local_iter_num += 1
    games_seen += effective_batch_size

    # termination conditions
    if iter_num > max_iters and not init_from == 'anneal':  # max iters is for auto-stopping end of stable phase
        checkpoint = {
            'model': raw_model.state_dict(),
            'optimizer': optimizer.state_dict(),
            'model_args': model_config,
            'iter_num': iter_num,
            "games_seen": games_seen,
            "tokens_seen": tokens_seen,
            "tokens_seen_padded": tokens_seen_padded,
            'best_val_loss': best_val_loss,
            'config': config,
        }
        checkpoint['grad_norm_history'] = grad_norm_history
        if init_from == 'anneal':
            checkpoint['anneal_start_iters'] = anneal_start_iters
            checkpoint['anneal_decay_iters'] = anneal_decay_iters
        print(f"Max_iters reached. Saving pre-anneal checkpoint to {anneal_checkpoint}")
        torch.save(checkpoint, os.path.join(out_dir, anneal_checkpoint))
        break
    if init_from == 'anneal' and iter_num >= anneal_start_iters + anneal_decay_iters:
        checkpoint = {
            'model': raw_model.state_dict(),
            'optimizer': optimizer.state_dict(),
            'model_args': model_config,
            'iter_num': iter_num,
            "games_seen": games_seen,
            "tokens_seen": tokens_seen,
            "tokens_seen_padded": tokens_seen_padded,
            'best_val_loss': best_val_loss,
            'config': config,
        }
        checkpoint['grad_norm_history'] = grad_norm_history
        if init_from == 'anneal':
            checkpoint['anneal_start_iters'] = anneal_start_iters
            checkpoint['anneal_decay_iters'] = anneal_decay_iters
        print(f"Anneal complete. Saving checkpoint to {out_dir}")
        torch.save(checkpoint, os.path.join(out_dir, 'anneal_complete.pt'))
        break

    
    
if ddp:
    destroy_process_group()