TVM Relax: ViT Encoder to Transformer Decoder (multimodal architecture) (2025)
Python • TVM • CUDA
Benchmarking a Vision Transformer encoder + transformer decoder in TVM Relax with baseline vs. optimized pipelines.
Learning new stuff
MLIR to PTX: Custom Dialect Fusion and GPU Lowering (2025)
C++ • MLIR • LLVM • CUDA
Tiny playground for a custom MLIR dialect with Add+ReLU fusion and lowering to PTX.
MLIR - Adding Max and Square ReLU operations to the Toy dialect and Lowering to LLVM (2025)
C++ • MLIR • LLVM
Learning by adding new operations to an existing dialect and lowering it down to LLVM.
Wi-Fi Extender LED Fault Detection on Kria KV260 (2025)
Python • PyTorch • ONNX Runtime • Kria KV260
Detect Wi-Fi extender power loss by watching its status LED and sending instant alerts.
Flash Attention in Triton (2025)
Python • PyTorch • Triton
Custom implementation of Flash Attention forward/backward kernels with benchmarking.
CUDA Kernel Microbenchmarks: Naive and Tiled MatMul using RTX 3060 (2025)
C++ • CUDA 12.x • RTX 3060
Benchmarking and comparing Matrix Multiplication operation on shared memory and naive approach.