PyTorch2.0 introduced a compiler--Inductor, and Inductor generage Triton DSL for generating ptx code.
I am curious about why Triton DSL, but not any other DSL that can be compiled to PTX code, was selected as the backend language for Inductor. Is it for extensibility or performance reasons? I cannot find any clue to answer this question. Could anyone help me with this? Thanks.