Now writing: inference engines & tool-call parsing
From tokens to kernels.
Notes on AI inference infrastructure.
I'm Satyajit, a systems engineer. I write long-form deep dives with code, math and benchmarks on how inference stacks actually work.
Focus areas
- Inference engines
- Tool-call parsing
- LoRA serving
- Constrained decoding
- Compilers & IR
- Benchmarks
// Featured
- // Compiler internals
Dominator Trees - Implementation and Why are they needed?
A Dominator Tree is a CFG analysis which is used in compilers to aid in other optimizations. This is useful when doing optimizations like code-hoisting or code sinking etc.
- // Compiler internals
Compiler IR - Considerations and Implementation
A Compiler intermediate representation - consideration and implementation details.
- // PyVyM
PyVyM - A python based VM for 'Python'
A virtual machine for python designed in python. It also has an accompanying debugger which will be described in a later post.
// Recent Posts
- // Udacity Self-Driving Car
Udacity's Self Driving Car - Vehicle Detection
This post talks about vehicle detection from video stream
- // Udacity Self-Driving Car
Udacity's Self Driving Car - Advanced Lane Finding Algorithm
This post talks about the Advanced Lane Finding algorithm for extracting lane lines
- // Udacity Self-Driving Car
Udacity's Self Driving Car - Behavior Cloning (P3)
This post talks about the Behavior Cloning project from Udacity where a car is trained to drive in a simulator.
OpenStreetMap rendering and simple Routing.
This post is about a small experiment of loading Openstreetmap files and rendering them with both matplotlib and vispy. Also tried a simple routing with dijkstra and A*.
Incremental text search using tries and ternary trees
While brushing up on algorithms, I decided to reimplement tries and ternary trees and in the process thought it would be a good idea to write a UI accompanying the code so its easier to see. Could further add timing code to this to see the performance too.
Adapting Luajit for Python 3
I have looked at spidermonkey and a few other JIT engines and kept on reading about LuaJit over and over again. I finally decided to look at Luajit closely and realized that it is a pretty complex piece of engineering but done single-handedly by Mike Pall. The number of source files is way smaller compared to spidermonkey or v8 but that's probably because Lua is much simpler than javascript. Since is fairly closer to python I wanted to try my hands out to modify Luajit's code to experiment if I could run basic python on it.