Training

Overview

The training module provides a complete implementation for training GPT models with support for:

Single GPU and distributed data parallel (DDP) training
Mixed precision training (float16/bfloat16)
Gradient accumulation
Learning rate scheduling with warmup and cosine decay
Checkpointing and resumption
WandB integration for experiment tracking

Training modes

You can run the training script in multiple configurations:

python train.py --batch_size=32 --compile=False

If your cluster does not have Infiniband, prepend NCCL_IB_DISABLE=1 to the commands.

Key functions

get_batch

get_batch(split)

Loads a batch of training or validation data using memory-mapped files.

split

str

required

Data split to load from: 'train' or 'val'

Returns: Tuple of (x, y)

x: Input token sequences of shape (batch_size, block_size)
y: Target token sequences of shape (batch_size, block_size), shifted by one position

Show Implementation details

The function:

Creates a new numpy memmap for each batch to avoid memory leaks
Randomly samples starting positions from the dataset
Extracts sequences of length block_size
Creates input (x) and target (y) sequences where y is x shifted by one token
Pins memory and transfers to GPU asynchronously if using CUDA

Data is expected to be stored as binary files:

data/{dataset}/train.bin
data/{dataset}/val.bin

Each file contains uint16 token indices.

estimate_loss

@torch.no_grad()
estimate_loss()

Computes accurate loss estimates over multiple batches for both training and validation splits. Returns: Dictionary with keys 'train' and 'val', each containing the mean loss

eval_iters

int

default:"200"

Number of iterations to average over (configured globally)

Show Why multiple batches?

Evaluating on multiple batches provides a more stable and accurate loss estimate compared to a single batch, which can be noisy. The function:

Sets model to eval mode
Runs inference on eval_iters batches for each split
Averages the losses
Returns model to train mode

This is called periodically during training (every eval_interval steps) to monitor progress.

get_lr

get_lr(it)

Computes the learning rate for a given iteration using cosine decay with linear warmup.

int

required

Current iteration number

Returns: float - Learning rate for this iteration

Show Learning rate schedule

The schedule has three phases:

Linear warmup (iterations 0 to warmup_iters):

lr = learning_rate * (it + 1) / (warmup_iters + 1)

Cosine decay (iterations warmup_iters to lr_decay_iters):

decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)
coeff = 0.5 * (1.0 + cos(π * decay_ratio))
lr = min_lr + coeff * (learning_rate - min_lr)

Constant minimum (iterations > lr_decay_iters):
```
lr = min_lr
```

This follows the approach from the Chinchilla paper.

Training loop structure

The main training loop performs the following steps:

1. Learning rate scheduling

lr = get_lr(iter_num) if decay_lr else learning_rate
for param_group in optimizer.param_groups:
    param_group['lr'] = lr

2. Periodic evaluation

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}")
    # Save checkpoint if validation loss improved or always_save_checkpoint=True

3. Forward and backward pass

With gradient accumulation to simulate larger batch sizes:

for micro_step in range(gradient_accumulation_steps):
    if ddp:
        # Only sync gradients on the last micro step
        model.require_backward_grad_sync = (micro_step == gradient_accumulation_steps - 1)
    with ctx:  # Automatic mixed precision context
        logits, loss = model(X, Y)
        loss = loss / gradient_accumulation_steps
    # Async prefetch next batch
    X, Y = get_batch('train')
    # Backward pass with gradient scaling for fp16
    scaler.scale(loss).backward()

4. Gradient clipping and optimizer step

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)

5. Logging

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}%")

Configuration parameters

I/O settings

out_dir

str

default:"'out'"

Directory for saving checkpoints

eval_interval

int

default:"2000"

How often to evaluate on val set and save checkpoints

log_interval

int

default:"1"

How often to log training metrics

eval_iters

int

default:"200"

Number of iterations for loss estimation

eval_only

bool

default:"False"

If True, exit after first evaluation (useful for testing)

always_save_checkpoint

bool

default:"True"

If True, save checkpoint after each eval even if val loss didn’t improve

init_from

str

default:"'scratch'"

Initialization mode: 'scratch', 'resume', or a GPT-2 variant ('gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl')

Data settings

dataset

str

default:"'openwebtext'"

Name of dataset (must have corresponding data// directory)

gradient_accumulation_steps

int

default:"40"

Accumulate gradients over this many steps to simulate larger batches

batch_size

int

default:"12"

Micro-batch size (per GPU if using DDP)

block_size

int

default:"1024"

Context length for training sequences

Model architecture

n_layer

int

default:"12"

Number of transformer layers

n_head

int

default:"12"

Number of attention heads

n_embd

int

default:"768"

Embedding dimension

dropout

float

default:"0.0"

Dropout rate (0.0 for pretraining, 0.1+ for finetuning)

bias

bool

default:"False"

Use bias in Linear and LayerNorm layers

Optimizer settings

learning_rate

float

default:"6e-4"

Maximum learning rate

max_iters

int

default:"600000"

Total number of training iterations

weight_decay

float

default:"1e-1"

Weight decay coefficient

beta1

float

default:"0.9"

AdamW beta1 parameter

beta2

float

default:"0.95"

AdamW beta2 parameter

grad_clip

float

default:"1.0"

Gradient clipping threshold (0.0 to disable)

Learning rate decay

decay_lr

bool

default:"True"

Enable learning rate decay

warmup_iters

int

default:"2000"

Number of warmup iterations

lr_decay_iters

int

default:"600000"

Iterations for learning rate decay (should be ~= max_iters)

min_lr

float

default:"6e-5"

Minimum learning rate (should be ~= learning_rate/10)

System settings

device

str

default:"'cuda'"

Device to train on: 'cpu', 'cuda', 'cuda:0', 'cuda:1', 'mps', etc.

dtype

str

default:"'bfloat16' or 'float16'"

Data type for training: 'float32', 'bfloat16', or 'float16'. Automatically selects bfloat16 if supported.

compile

bool

default:"True"

Use PyTorch 2.0 compilation for faster training

DDP settings

backend

str

default:"'nccl'"

DDP backend: 'nccl' (recommended for CUDA) or 'gloo'

Checkpointing

Checkpoints are saved to {out_dir}/ckpt.pt and contain:

{
    'model': model.state_dict(),
    'optimizer': optimizer.state_dict(),
    'model_args': model_args,
    'iter_num': iter_num,
    'best_val_loss': best_val_loss,
    'config': config,
}

To resume training from a checkpoint, set init_from='resume' and ensure the checkpoint exists in out_dir.

WandB integration

wandb_log

bool

default:"False"

Enable Weights & Biases logging

wandb_project

str

default:"'owt'"

WandB project name

wandb_run_name

str

default:"'gpt2'"

WandB run name

When enabled, the following metrics are logged:

Training loss
Validation loss
Learning rate
Model FLOPs Utilization (MFU)

Performance tips

Use gradient accumulation to simulate larger batch sizes without running out of memory. Effective batch size = batch_size * gradient_accumulation_steps * num_gpus.

Enable compilation with compile=True to use PyTorch 2.0’s optimizations for faster training (requires PyTorch >= 2.0).

Use bfloat16 if your GPU supports it (requires Ampere or newer). It provides better numerical stability than float16 without requiring gradient scaling.

Allow TF32 (enabled by default) for ~20% speedup on Ampere GPUs without loss of accuracy.

Core Components

Architecture

Overview

Training modes

Key functions

get_batch

estimate_loss

get_lr

Training loop structure

1. Learning rate scheduling

2. Periodic evaluation

3. Forward and backward pass

4. Gradient clipping and optimizer step

5. Logging

Configuration parameters

I/O settings

Data settings

Model architecture

Optimizer settings

Learning rate decay

System settings

DDP settings

Checkpointing

WandB integration

Performance tips

Build docs developers (and LLMs) love

Core Components

Architecture

​Overview

​Training modes

​Key functions

​get_batch

​estimate_loss

​get_lr

​Training loop structure

​1. Learning rate scheduling

​2. Periodic evaluation

​3. Forward and backward pass

​4. Gradient clipping and optimizer step

​5. Logging

​Configuration parameters

​I/O settings

​Data settings

​Model architecture

​Optimizer settings

​Learning rate decay

​System settings

​DDP settings

​Checkpointing

​WandB integration

​Performance tips

Build docs developers (and LLMs) love

Overview

Training modes

Key functions

get_batch

estimate_loss

get_lr

Training loop structure

1. Learning rate scheduling

2. Periodic evaluation

3. Forward and backward pass

4. Gradient clipping and optimizer step

5. Logging

Configuration parameters

I/O settings

Data settings

Model architecture

Optimizer settings

Learning rate decay

System settings

DDP settings

Checkpointing

WandB integration

Performance tips