Sophia: A Scalable Stochastic Second-order Optimizer for Language Model Pre-training

This is an official implementation of the Sophia-G optimizer in the paper https://arxiv.org/abs/2305.14342 and GPT-2 training scripts. The code is based on nanoGPT and levanter. Please cite the paper and star this repo if you find Sophia useful. Thanks!

@article{liu2023sophia,
 title={Sophia: A Scalable Stochastic Second-order Optimizer for Language Model Pre-training},
 author={Liu, Hong and Li, Zhiyuan and Hall, David and Liang, Percy and Ma, Tengyu},
 journal={arXiv preprint arXiv:2305.14342},
 year={2023}
}

News and Updates

Updated results with latest PyTorch version.

Dependencies

PyTorch 2.1.2
transformers 4.33.0
datasets
tiktoken
wandb

General Usage

Below is an example code snippet for training a general model with NLL loss with SophiaG. Please refer to the next section for guidelines on hyperparameter tuning.

import torch
import torch.nn.functional as F
from sophia import SophiaG

# init model loss function and input data
model = Model()
data_loader = ...

# init the optimizer
optimizer = SophiaG(model.parameters(), lr=2e-4, betas=(0.965, 0.99), rho=0.01, weight_decay=1e-1)

total_bs = len(data_loader)
bs = total_bs * block_size
k = 10
iter_num = -1

# training loop
for epoch in range(epochs):
    for X, Y in data_loader:
        # standard training code
        logits, loss = model(X, Y)
        loss.backward()
        optimizer.step(bs=bs)
        optimizer.zero_grad(set_to_none=True)
        iter_num += 1

        if iter_num % k != k - 1:
            continue
        else:
            # update hessian EMA
            logits, _ = model(X, None)
            samp_dist = torch.distributions.Categorical(logits=logits)
            y_sample = samp_dist.sample()
            loss_sampled = F.cross_entropy(logits.view(-1, logits.size(-1)), y_sample.view(-1), ignore_index=-1)
            loss_sampled.backward()
            optimizer.update_hessian()
            optimizer.zero_grad(set_to_none=True)
            model.zero_grad()

Hyper-parameter Tuning

Definition of learning rate

The update in the code is written as $\theta_{t+1} = \theta_t - lr*\textup{clip}(m_t / (\rho * h_t + \epsilon), 1)$, which is equivalent to the update in the paper up to a re-parameterization. (the $lr$ here corresponds to $\rho \cdot \eta_t$ in the paper). As a result, the learning rate of AdamW and Lion is not directly comparable. Empirically, Adam and Lion with learning rate ratio 5:1 has similar behaviour. The learning rate of SophiaG and Lion is directly comparable. Sophia allows to use much larger learning rate the Lion, and this is why Sophia is much faster.

Tuning the hyperparameter $\rho$

Tune $\rho$ to make the proportion of the clipped coordinates stable and in a proper range. This is tracked as train/win_rate in the GPT-2 training example. train/win_rate should peak in the beginning and remain stable afterwards. train/win_rate should stay in the range of 0.1 - 0.5. Typically a large $\rho$ will lead to a large train/win_rate. An example of typical win_rate behavior in T5 model is provided below.

Tuning the learning rate and weight decay

Choose lr to be slightly smaller than the learning rate that you would use for AdamW or 3 - 5 times the learning rate that you would use for Lion.

If the loss blows up, slightly decrease the learning rate or increase $\rho$.
Always use about 2x larger weight decay than what you would use for AdamW.

Hyperparameters for GPT-2 models

Choose lr to be about the same as the learning rate that you would use for AdamW or 5 - 10 times the learning rate that you would use for Lion.
Tune $\rho$ to make the proportion of the parameters where the update is not clipped stable and in a proper range. This is tracked as train/win_rate in the GPT-2 training example. train/win_rate should peak in the beginning and remain stable afterwards. train/win_rate should stay in the range of 0.1 - 0.5. Typically a large $\rho$ will lead to a large train/win_rate.
Use slightly larger weight decay than AdamW, e.g. 0.2.
Except learning rate, all other hyperparameters are transferable across different model sizes.
See the table below for the hyperparameters for different model sizes.

Model Size	lr for Adam	lr for Lion	lr for Sophia	$\rho$ for Sophia	weight decay for Sophia
125M	6e-4	1e-4	6e-4	0.05	0.2
355M	3e-4	1e-4	7e-4	0.08	0.2
770M	2e-4	8e-5	3e-4	0.05	0.2

Please feel free to let us know what you find out during hyper-parameters tuning. We appreciate your valuable feedback and comments!

Reproduce GPT-2 Results

Prepare the OpenWebText data following nanoGPT:

$ python data/openwebtext/prepare.py

Start pre-training GPT2 Small (125M):

If you have a machine with 10 A5000 (24GB) GPUs,

$ torchrun --standalone --nproc_per_node=10 \
      train_sophiag.py \
      config/train_gpt2_small_sophiag.py \
      --batch_size=8 \
      --gradient_accumulation_steps=6

If you have a machine with 8 A100 (40GB) GPUs,

$ torchrun --standalone --nproc_per_node=8 \
      train_sophiag.py \
      config/train_gpt2_small_sophiag.py \
      --batch_size=12 \
      --gradient_accumulation_steps=5

To reproduce the AdamW baseline following nanoGPT:

$ torchrun --standalone --nproc_per_node=10 \
      train_adam.py \
      config/train_gpt2_small_adam.py \
      --batch_size=8 \
      --gradient_accumulation_steps=6

This will lead to results in the figure below:

Start pre-training GPT2 Medium (355M):

If you have a machine with 8 A100 (40GB) GPUs,

$ torchrun --standalone --nproc_per_node=8 \
      train_sophiag.py \
      config/train_gpt2_medium_sophiag.py \
      --batch_size=6 \
      --gradient_accumulation_steps=10

To reproduce the AdamW baseline:

$ torchrun --standalone --nproc_per_node=8 \
      train_adam.py \
      config/train_gpt2_medium_adam.py \
      --batch_size=6 \
      --gradient_accumulation_steps=10

Please adjust nproc_per_node, batch_size, and gradient_accumulation_steps accordingly if you use other hardware setup. Make sure their product equals 480.