Vendoring ptxas

[SomeoneSerge]

Hi! Observations first. I noticed that openai/triton vendors a copy of ptxas and distributes it in the published wheels. In NixOS/nixpkgs#222273 and NixOS/nixpkgs#227420 this mean making the triton package as "unfree", which prevents users from building it unless they explicitly opt-in (also, this prevents our CI from building the binary cache for the package). I also see that the pypi page declares the project as MIT, despite the wheels redistributing ptxas.

• Q: Can we make ptxas optional? That'd require a way to inhibit https://github.com/openai/triton/blob/deb2c71fb4f912a5298003fa3fc789885b726607/python/setup.py#L108 as well as ensuring that at runtime any attempts of using the cuda back-end without ptxas result in a sensible error message
• Q: Can we make ptxas location overridable? E.g. triton could check for some special environment variable before proceeding with its default inference logic

[Jokeren]

Q2. You can use the TRITON_PTXAS_PATH env variable.

[ptillet]

Q1. Yeah, I'd accept a PR that does this, as long as it's not the default behavior :)