LLM pretraining Vlad Savinov Member of Technical Staff @ Reflection AI I work on LLM pretraining. X GitHub LinkedIn Before joining Reflection, I was a Staff DL Engineer and Team Lead for YandexGPT pretraining at Yandex, focusing on distributed LLM training, optimizations, and model architecture. My work included FP8 training recipes, Context Parallel for long-context models, MoE training at scale, and development of a company-wide distributed training framework. Earlier, I led an Applied ML team and worked on large-scale fine-tuning for enterprise search and code assistance. Interests Model Architecture Distributed Training & Efficiency Reinforcement Learning Public Talks & Appearances 2025 LLM Scaling Week - Session 1, Session 2 (November) Speeding up Training with FP8 and Triton - YerevaNN + Yandex Hall, Yerevan (English) (November) The Technology Behind Large Language Models and GPT - Armenian Science Week, Yerevan (English) (October) Moscow State University - Efficient DL Systems (Russian) (September) Benchmarking Seminar FP8 Seminar Distributed Training Lecture 2023 DataFest 2023 - Machine Learning Talk