Looks like PyTorch nightlies are using a much newer version of Triton relative to what they were a month or two ago. They’ve got their own branch which closely tracks master: https://github.com/openai/triton/tree/torch-inductor-stable.

Packaging that could be difficult given they pull in their own build of MLIR (based on LLVM 17: https://github.com/openai/triton/blob/2c32f4399986045ff25cae201ed3b16d922a9d3b/python/setup.py#L72) and unconditionally grab NVCC from conda (https://github.com/openai/triton/blob/2c32f4399986045ff25cae201ed3b16d922a9d3b/python/setup.py#L107).

We don’t have MLIR packaged yet: https://github.com/NixOS/nixpkgs/pull/163878 (although I think we do build it for ROCm?)

Thoughts?