Jay Gala (@Jaygala223)

I Wrote a GPU Matmul Kernel From Scratch in Triton. Here's Everything I Learned

I recently started learning Triton, OpenAI's Python-based language for writing GPU kernels. My project: build a matrix multiplication kernel from scratch, step by step, until it's competitive with PyT

Jun 23 · 15 min readgalacodes.hashnode.dev

GPUs: The Hardware That Power AI

You've probably used ChatGPT or Claude. Maybe you've even fine-tuned a small language model on your laptop. But have you ever wondered why training or even inferencing GPT or Claude requires tens of thousands of specialized chips aka GPUs instead of ...

Nov 30, 2025 · 18 min readgalacodes.hashnode.dev

Speculative Decoding: From Theory to Implementation

Let's talk about speculative decoding. One of the most elegant optimization techniques in modern LLM inference. If you've ever wondered how to squeeze 2-3x more throughput from your language models without sacrificing output quality, you're in the ri...

1Nov 5, 2025 · 15 min readgalacodes.hashnode.dev

Bits Don't Lie: Datatypes in Modern LLMs

Let’s talk about the different datatypes that are being used in modern LLMs like GPT, LLaMa and the like. The most common ones that you might have heard: FP32, BFloat16, Float16, INT8, etc. These are all standard data types and available in PyTorch a...

May 30, 2025 · 6 min readgalacodes.hashnode.dev

Introduction to LLM inferencing

Unless you’re living under a rock, you’ve probably heard of Large Language Models (LLMs) and even used a few of the popular applications like ChatGPT, Claude, Perplexity, etc. powered by these LLMs. So without going too deep into what LLMs are, let’s...

Oct 26, 2024 · 7 min readgalacodes.hashnode.dev

Jay Gala

About

Available for

Jay Gala's blogs

I Wrote a GPU Matmul Kernel From Scratch in Triton. Here's Everything I Learned

GPUs: The Hardware That Power AI

Speculative Decoding: From Theory to Implementation

Bits Don't Lie: Datatypes in Modern LLMs

Introduction to LLM inferencing

Jay Gala

About

Available for

Jay Gala's blogs

I Wrote a GPU Matmul Kernel From Scratch in Triton. Here's Everything I Learned

GPUs: The Hardware That Power AI

Speculative Decoding: From Theory to Implementation

Bits Don't Lie: Datatypes in Modern LLMs

Introduction to LLM inferencing

Jay Gala

About

Available for

Jay Gala's blogs

I Wrote a GPU Matmul Kernel From Scratch in Triton. Here's Everything I Learned

GPUs: The Hardware That Power AI

Speculative Decoding: From Theory to Implementation

Bits Don't Lie: Datatypes in Modern LLMs

Introduction to LLM inferencing

Search Hashnode

Jay Gala

About

Available for

Jay Gala's blogs

I Wrote a GPU Matmul Kernel From Scratch in Triton. Here's Everything I Learned

GPUs: The Hardware That Power AI

Speculative Decoding: From Theory to Implementation

Bits Don't Lie: Datatypes in Modern LLMs

Introduction to LLM inferencing