AdamW has long been the go-to optimizer for transformer pretraining. For years, the research community has been searching for faster and more stable optimizers, with a focus on achieving only positive outcomes. In this work, we introduce a simple, single-line modification in PyTorch for any momentum-based optimizer. This modification, termed Cautious Optimizer (e.g., C-AdamW and C-Lion), opens the door to improved training performance.
Our theoretical findings reveal that this modification preserves Adam’s Hamiltonian function and retains its convergence guarantees under Lyapunov analysis. Additionally, a new family of optimizers emerges from this insight. Among these, we select the simplest for empirical experiments, achieving up to 1.47x speed-up on Llama and MAE pretraining.
- [2025-01-23] PPO (Reinforcement Learning)
- [2025-01-14] Post Training experiment on Qwen2.5 1.5B Instruct
- [2024-12-03] 🤗🤗🤗 More validation runs on ViTs timm-optim-caution
- [2024-12-03] 🤗🤗🤗 Caution implemented in huggingface/pytorch-image-models.
- [2024-11-24] Pre-release paper available on arXiv: Cautious Optimizers: Improving Training with One Line of Code.
- [2024-11-24] Official implementation of C-Optim released! Experiment with C-AdamW and C-Lion today.
pip install -r requirements.txttorchrun --standalone --nproc_per_node 1 torchrun_main.py \
--model_config configs/llama_60m.json \
--lr 0.001 \
--batch_size 16 \
--total_batch_size 512 \
--activation_checkpointing \
--num_training_steps 10000 \
--warmup_steps 1000 \
--weight_decay 0 \
--grad_clipping 1.0 \
--dtype bfloat16 \
--eval_every 1000 \
--single_gpu \
--optimizer c-adamw \
--max_length 1024torchrun --standalone --nproc_per_node 4 run_mae.py \
--dataset_name ILSVRC/imagenet-1k \
--output_dir ./vit-mae-c \
--remove_unused_columns False \
--label_names pixel_values \
--mask_ratio 0.75 \
--norm_pix_loss \
--do_train \
--do_eval \
--base_learning_rate 1.5e-4 \
--lr_scheduler_type cosine \
--weight_decay 0.05 \
--num_train_epochs 50 \
--warmup_ratio 0.05 \
--per_device_train_batch_size 256 \
--per_device_eval_batch_size 8 \
--logging_strategy steps \
--logging_steps 10 \
--eval_strategy epoch \
--save_strategy epoch \
--load_best_model_at_end True \
--save_total_limit 3 \
--seed 1337 \
--custom_optim c-adamw \
--trust_remote_code \
--gradient_accumulation_steps 4torchrun \
--rdzv_id=$JOB_ID \
--rdzv-backend=c10d \
--nnodes=1:8 \
--nproc-per-node=1 \
--rdzv_endpoint=$MASTER_ADDR:$MASTER_PORT \
post_training.py --model "Qwen/Qwen2.5-1.5B-Instruct" \
--output_dir cautious_1.5b \
--per_device_train_batch_size 1 \
--gradient_accumulation_steps 2 \
--max_length 8192 \
--cautious
accelerate launch ppo_tldr.py \
--dataset_name trl-internal-testing/tldr-preference-sft-trl-style \
--dataset_test_split validation \
--output_dir models/minimal/ppo_tldr \
--learning_rate 3e-6 \
--per_device_train_batch_size 16 \
--gradient_accumulation_steps 4 \
--total_episodes 1000000 \
--model_name_or_path EleutherAI/pythia-1b-deduped \
--sft_model_path cleanrl/EleutherAI_pythia-1b-deduped__sft__tldr \
--reward_model_path cleanrl/EleutherAI_pythia-1b-deduped__reward__tldr \
--local_rollout_forward_batch_size 16 \
--missing_eos_penalty 1.0 \
--stop_token eos \
--eval_strategy steps \
--eval_steps 100 \
--custom_optim c_adamw \
--num_gpus 8
@article{liang2024cautious,
title={Cautious optimizers: Improving training with one line of code},
author={Liang, Kaizhao and Chen, Lizhang and Liu, Bo and Liu, Qiang},
journal={arXiv preprint arXiv:2411.16085},
year={2024}
}