Deependu Jha
Open source @ Lightning AI
I write CUDA kernels, poke at PyTorch compiler internals, and try to make models run faster than they have any right to. This is where I write about how that stuff actually works, and share the projects that come out of it.
Writing
all posts →Projects
all projects →ghost-rfdetr
flagshipCustom CUDA kernels and operator fusion for RF-DETR — faster inference out of the box, no retraining or config required.
DeepTensor
A minimal PyTorch-like deep learning library focused on custom autograd engines, backpropagation, dynamic computation graphs, and efficient tensor operations written from scratch.
DeepLang
A lightweight, experimental compiler & interpreter framework exploring compiler frontends, AST parsing, intermediate representation (IR), optimization passes, and a custom VM.
litdata-fork
Optimized fork of litdata featuring multi-node cluster streaming for deep learning datasets, with specialized caches for CUDA pinned-memory layout.