Yandex developed and published in the open access to GitHub the new YAFSDP tool, designed to help companies working with artificial intelligence, optimize resources in teaching large language models (LLM). Key advantages yafsdp :
- accelerates and increases the efficiency of the LLM learning process, saving to 20% of GPU resources, which reduces the costs and time for training.
is the most effective publicly accessible tool for optimizing the use of GPU memory and improving the connection between graphic processors in learning LLM.
Provides up to 26% higher training rate compared to previous versions of the FSDP tool.
“LLM training is a time -consuming and resource -intensive process,” said Yandex. “Machine training engineers and companies developing their own LLM spend considerable time and GPU resources – what is equal to money – to train these models. The more the model, the more time and cost is required to study it.”
The company evaluates that the use of YAFSDP for teaching a model with 70 billion parameters can save resources of approximately 150 GPU, which is about 0.5-1.5 million dollars per month, depending on the supplier of virtual GPU or platforms.
Using META advanced Llama models, known for its innovations and support for open AI, Yandex demonstrated the impressive results of its YAFSDP:
- tool.
on the basis of Llama 2 70b, the final acceleration of learning by 21%
- was achieved
on Llama 3 70b acceleration was 26%
These indicators indicate the high performance of YAFSDP in the optimization of GPU resources and memory when training large language models.
The development of YAFSDP is the next contribution of Yandex to the open ecosystem of AI. Earlier, the company has released such popular tools as:
catboost – an advanced library of gradient boosting on trees of open source solutions
ytsaurus – the main system for storing and processing Yandex data
AQLM – FDDIVATIVE quantization for language models
petals – decentralized conclusion and accurate setting of large language models
Many large technological companies also make the basis of their products, for example, Apple has recently announced its Apple Intelligence services as part of the upcoming IOS 18.
update.
The publication of YAFSDP under an open license demonstrates the commitment of Yandex to the principles of open AI and the desire to make a significant contribution to the development of the industry, providing the community with advanced developments. This will allow other companies and researchers to benefit from the faster and more economical learning of language models.