C++ Developer for YandexGPT (Neuro)

Our team develops and maintains backends based on LLM models (under the YandexGPT/Alice brand). We work directly on inference on GPU accelerators, as well as a wide range of tasks related to product development and support: API interaction with the frontend, implementation of the ML stack at runtime, logs for analytics, and much more. With us, you will be able to work on complex and diverse tasks.

What tasks await you

Inference of heavy generative language models on GPU accelerators The heart of LLM-based products is, of course, the direct computation of LLM models. You will be solving tasks related to the allocation of various components with LLM models, configuring their interaction, release processes, and selecting various parameters for optimization.

Optimization of work distribution methods between compute nodes It is possible to optimize not only the computation itself but also the methods of splitting the incoming flow between nodes to achieve the most advantageous latency-based distribution of work between nodes. We are also experimenting with deferred continuation of computations.

Development of various parts of a multi-component system The answer based on search sources is a complex multi-component product. It's important not only to compute something using an LLM model but also: 1) to bring data to the model's input; 2) to correctly save results for delivery to users; 3) to configure interaction with the frontend (e.g., streaming); 4) to provide product and ML teams with the ability to conduct experiments and improve the product. All of this generates many meaningful and complex tasks. Working on the backend of search LLMs opens up many technical challenges, less common in products that do not work with LLM models. After all, computations lasting not hundreds of milliseconds but many seconds require a revision of established approaches.

More about backend at Yandex — in the channel Yandex for Backend

We expect you to

• Understand basic algorithms and data structures
• Have a good knowledge of C++
• Are familiar with software design patterns and idioms
• Can write reliable and readable multithreaded code
• Have experience developing a large service in a team