PyTorch JIT编译器
PyTroch似乎引进了一些有趣的JIT编译功能,用来优化性能,包括做一些fusion。
可以看下的项目:
torchdynymo: https://github.com/pytorch/torchdynamo
似乎用了triton
NVFuser: https://github.com/pytorch/pytorch/blob/release/1.12/torch/csrc/jit/codegen/cuda/README.md
NVFuser is integrated as a backend for TorchScript's Profiling Graph Executor. NVFuser is the default fuser for NVIDIA GPUs.