Uday’s Technical Writeups

GPU Kernel DSLs and Abstraction Levels

GPU

programming

DSL

compilers

Executing frontier models efficiently requires high-performance kernel code that runs on the GPU. Traditionally kernels were written for a few compute intensive operations…

Mixture-of-Experts Inference: parallelism, communication, and load balancing

LLM

inference

MoE

Since 2025, Mixture-of-Experts (MoE) has become a popular architecture for recent LLMs. MoE packs more capacity into a model while keeping the compute budget steady. An MoE…

Disaggregated Inference

LLM

inference

disaggregated inference

systems

Disaggregated Inference (DI), also known as disaggregated serving, disaggregated prefilling, or P/D (prefill/decode) disaggregation, is an LLM serving architecture that…

The Foundational Features of LLM Inference Engines

LLM

inference

vLLM

SGLang

systems

LLM inference engines like vLLM and SGLang have become the workhorses of LLM serving, both in the datacenter and increasingly at the edge. They ship new features…

Hello World

general

Hello, this is Uday’s technical writeups. Thank you for stopping by!