Pranav Ajit Nair

prof_pic.jpg

Google DeepMind, India

pranavajitnair@google.com

I am a Pre-Doctoral Researcher working at Google DeepMind, India, with Dr. Praneeth Netrapalli, Dr. Arun Suggala and Dr. Prateek Jain. My work involves making LLM inference faster through quantization, speculative decoding, and sparsification. I am also working on speeding up “million-context-attention” through clustering and approximate logit computation.

My interests include but are not limited to i) Making LLM inference faster through next-generation architectures, quantization, sparsification, speculative decoding, KV cache compression, adaptive routing for elastic models, etc., and ii) speeding up LLM pretraining and finetuning through better adapters, novel loss functions, better second-order optimizers and faster checkpointing.

selected publications

  1. ICML
    Tandem Transformers for Inference Efficient LLMs
    S AishwaryaP, Pranav Ajit Nair, Yashas Samaga, and 4 more authors
    ICML, 2024
  2. Arxiv
    CDQuant: Greedy Coordinate Descent for Accurate LLM Quantization
    Pranav Ajit Nair, and Arun Sai Suggala
    Arxiv, 2024