pytroch-playground

Build a custom op library with triton

Currerntly a todo list:

• Generate op lists by op.yaml and op shim and pointwise op (may also with reduction op)

• Compile kernels (cubin) in python with Triton compile and cache them in cpp

• Add flaggems kernels && flash attention kernels

• Add third party gtest and glog

Done:

• use torch LOG, ENV : TORCH_CPP_LOG_LEVEL=INFO