deep-learning-compiler

SonicBoom is a compiler-first PyTorch alternative for language, vision, audio, and training. It builds on proven kernels and pushes performance further through compiler optimization, with native HTTP serving powered by GNU Artanis.

This is the project repository of our ESEC/FSE 2021 paper: A Comprehensive Study of Deep Learning Compiler Bugs.

GenCoG: A DSL-Based Approach to Generating Computation Graphs for TVM Testing (ISSTA‘23)

A differentiable Machine Learning framework built from scratch using Python and C.

How does torch.compile() achieve massive speedups despite highly optimized torch functions? See how operator fusion works!

⏳ Lightweight deep learning library and compiler suite built for educational purposes.

A Branchless, Zero-Jitter Ingress Router for 32-GPU Distributed Mesh Networks utilizing JAX/XLA and NCCL.

Tiny tensor compiler: one tensor program compiled to a GPU simulator (tiny-gpu) and a systolic-array TPU (tiny-tpu), checked bit-exact against numpy.

Top Deep Learning Compilers (Opensource) Star if you like it!

Occupancy aware C++/CUDA operator fusion decision engine and physical hardware benchmark suite (Tesla T4 calibrated).

Multi-model shared MCTS for Reasoning Compiler-style TVM tensor program optimization.

Use Pytorch Glow deep learning compiler in a Docker container inside your Github workflows

Bare-Metal Deep-Kernel Fused JIT & Memory Bandwidth Optimizer for Frontier AI Accelerators (Tenstorrent, Triton, CUDA, Apple Silicon).

A from-scratch deep learning compiler: it lowers trained ONNX models through two custom MLIR dialects and a full optimization pipeline to a binary instruction stream for a simulated edge NPU, then runs them in a cycle-costed simulator validated against onnxruntime.

TVM BYOC compiler targeting RISC-V RVV on BananaPi BPI-F3 (SpacemiT K1) — from YOLOv8n import to vectorized inference

My upstream C++/OpenMP contributions to PyTorch Internals (Inductor BMM kernels, Autograd, & Compiler correctness).