kernel-fusion

Building an Out-of-Tree PyTorch Backend, End to End

Building a PyTorch backend from scratch: device runtime, caching allocator, dispatcher registration, an IR, fusion and CUDA code generation, and testing.