Advanced Fine-Tuning: Enhancing fine-tuning methodologies – both LoRA-based and full-parameter – for cutting-edge LLMs (e.g., GPT-OSS, Kimi K2.5, DeepSeek V3.1/V3.2, GLM-4.7), focusing on both model quality and training efficiency.
Inference Optimization: Identifying LLM inference bottlenecks to drive production speedups. This involves building model training and evaluation pipelines in JAX for speculative decoding, experimenting with architectures (dense/MoE, auto-regressive/parallel), and deriving scaling laws to guide resource allocation.
Low Precision Training & Inference: Investigating low-precision (FP8, NVFP4/MXFP4) methodologies for supervised fine-tuning and reinforcement learning – spanning both inference and training – optimized for modern hardware
We expect you to have:
A profound understanding of theoretical foundations of machine learning and reinforcement learning.
Deep expertise in modern deep learning for language processing and generation
Experience with training large models on multiple computational nodes
Reasonable understanding of performance aspects of large neural network training (sharding strategies, custom kernels, hardware features etc.)
Strong software engineering skills (we mostly use Python)
Deep experience with modern deep learning frameworks (we use JAX)
Proficiency in contemporary software engineering approaches, including CI/CD, version control and unit testing
Strong communication and leadership abilities
