{"categories":[{"label":"Artificial Intelligence","url":"https://skillfed.io/packages/category/scientific-engineering-artificial-intelligence/8"}],"enrichment":{"capability":"Lion is a PyTorch optimizer that implements an evolved sign momentum algorithm discovered by Google Brain as an alternative to AdamW for training neural networks.","skillfed_tags":["optimizer","pytorch","training"],"use_cases":["Training large language models where a 3x smaller learning rate than AdamW yields better convergence","Text-to-image model training when hyperparameters are carefully tuned to the specific architecture","Vision transformer training with cosine decay learning rate schedules","Scenarios with high batch sizes (64+) where Lion's sign momentum approach may outperform adaptive methods"],"what_it_does":"Lion is a PyTorch optimizer implementing an evolved sign momentum algorithm that the description positions as a potential successor to AdamW. It depends only on torch and integrates directly into PyTorch's optimizer interface. The optimizer requires careful hyperparameter tuning: learning rates should typically be 3\u201310x smaller than AdamW equivalents, weight decay values 3\u201310x larger, and default \u03b21 and \u03b22 values differ from AdamW (0.9 and 0.99 versus 0.9 and 0.999). The description documents mixed empirical results\u2014positive outcomes reported for language modeling and text-to-image training when tuned correctly, but negative results outside the paper's tested domains (reinforcement learning, feedforward networks, hybrid architectures). The author recommends Lion primarily for high batch sizes (64 or above) and notes sensitivity to batch size, data volume, and augmentation.","worth_installing":"Yes, with conditions. Lion is permissively licensed, has low install friction, and is actively maintained. Install it if you are training models in domains the paper evaluated (language modeling, vision transformers, text-to-image) and are willing to invest in hyperparameter tuning. The description's own updates acknowledge it performs worse than Adam without careful learning rate adjustment and shows negative results outside tested architectures. Not recommended for reinforcement learning, standard feedforward networks, or exploratory work where you cannot afford tuning overhead."},"id":"lion-pytorch","links":{"html":"https://skillfed.io/packages/lion-pytorch","md":"https://skillfed.io/packages/lion-pytorch.md","pypi":"https://pypi.org/project/lion-pytorch/"},"maintenance":{"status":"active"},"meta":{"latest_release":"2026-07-09","license_spdx":null,"license_treatment":"permissive","name":"lion-pytorch","python_support":"supports_current","summary":"Lion Optimizer - Pytorch"},"popularity":{"monthly_downloads":159256,"position":10703,"tier":"top_15000"},"security":{"n_vulnerabilities":0},"version":"0.2.5"}