Generalize code in tutorials: 2, 3, 4

[anmyachev]

No description provided.

[Jokeren]

IMO, it's still not generalized because DEVICE is hardcoded as cuda. Can you describe how users are supposed to use it on intel devices?

[anmyachev]

IMO, it's still not generalized because DEVICE is hardcoded as cuda. Can you describe how users are supposed to use it on intel devices?

I agree, this only allows changing the device in one place from the user's point of view. From our point of view (plugin developers) it is also more convenient, since our code differs less from upstream's, which makes the merge easier.

Perhaps it would be suitable to redefine the device via an environment variable like: DEVICE=xpu python tutorials/02-tutorial. In this case, the user will not need to change the source code at all. What do you think?

[Jokeren]

One potential solution is to use

triton.runtime.driver.active.get_current_target().backend

It works well for amd and cuda, but I haven't tried for xpu. (I don't have any testbed)

[ThomasRaoux]

+1 to what Keren said, I don't think moving the place where we set the device is an improvement. It only makes sense if we can automatically detect triton backend as suggested by Keren