Yandex Launches YaFSDP, The Most Advanced LLM Training Tool For Generative AI

JAKARTA - Yandex, a global technology company, has just launched YaFSDP, an open-source method to train the Large Language Model (LLM) that allows generative AI to interact with users in real-time.

As part of the company's commitment to contributing to the global artificial intelligence (AI) community, Yandex makes YaFSDP publicly available to LLM developers and AI enthusiasts around the world.

"We hope this can contribute to increased accessibility and efficiency for researchers and generative AI developers around the world," explained Mikhail Khruschev, Yandex Senior Developer and YaFSDP Team in a statement quoted Wednesday, June 12.

Yandex says that YaFSDP is able to optimize learning speed and performance, enabling AI developers around the world to use less computing power and GPU resources while training their models.

Currently, YaFSDP is the most effective method in training LLM compared to other Fully Sharded Parallel Data (FSDP) methods available in public.

With YaFSDP, the speed can be increased by up to 26 percent compared to the usual FSDP, depending on the architecture and the number of parameters. Thus, the training time for LLM can be reduced while saving up to 20 percent of the use of GPU resources.

"YaFSDP has shown impressive results on models that handle 13 to 70 billion parameters, with very strong performance in the range of 30 to 70 billion," Mikhail said.

"Currently, we see that YaFSDP is best suited for widely used open-source models based on LLAMA architecture," he added.