Yandex, the Russian IT giant, has launched a new open-source tool to assist AI firms save money and resources while training new models.

According to Yandex's statement, the new YaFSDP tool can enable quicker and more efficient large language model (LLM) training while conserving up to 20% of GPU resources, saving money and time.

Yandex also made YaFSDP-related information available on GitHub for anybody to read and experiment with.

How to Train Your LLM

Yandex provides a lot of technical information in its release article, claiming that the new open-source application is "the most effective publicly available tool for enhancing GPU communication and reducing memory usage in LLM training" and can give up to 26% quicker rates than the previous version, FSDP. 

"LLM training is a time-consuming and resource-intensive process," Yandex asserts. "Machine learning engineers and firms that create their own LLMs devote substantial time and GPU resources - and hence money - in training these models. The bigger the model, the longer and more expensive it takes to train."

The business predicts that employing YaFSDP to train a model with 70 billion parameters will save the resources of around 150 GPUs, equating to monthly savings of $0.5 to $1.5 million (depending on the virtual GPU provider or platform), assuming everything goes well. 

Yandex claims that using Meta's Llama models, which have been commended for being cutting-edge while embracing open source, results in a "final speedup" of 21% and 26% on Llama 2 70B and Llama 3 70B.

YaFSDP is not Yandex's first excursion into AI tools; the company has already launched CatBoost, YTsaurus, AQLM, and Petals, each of which addresses a different aspect of AI development.

Yandex's decision comes after it was forced to divest itself following Russia's invasion of Ukraine, selling its domestic company to its prior management in the nation.

Many prominent technology businesses are also incorporating AI into their products, most notably with Apple offering Apple Intelligence services as part of the next iOS 18 upgrade.