1. 10
  1.  

  2. 2

    I found it interesting that Triton’s original frontend language was a C dialect unimaginatively named Triton C. This version’s frontend language is a Python EDSL, which shows what people want, I think.

    1. 3

      all I want is static analysis of tensor dimensions :(

    2. 1

      I wonder how this compares to XLA which uses the same LLVM NVPTX backend but also supports compiling to CPUs. On top of that, is it feasible to target GPUs other than Nvidia’s or is everything else really too slow to care about?