Auto-select index_dtype

[jansel]

Stacked PRs (oldest at bottom):

• ->Auto-select index_dtype #1131

---

Auto-select index_dtype

Fixes #1123

[Chillee]

I think it's worth distinguishing what's happening at what component. For example, if I use kernel.bind will it automatically re-specialize? I guess if I do kernel.bind and pass in new inputs it won't re-specialize?

[jansel]

I think it's worth distinguishing what's happening at what component. For example, if I use kernel.bind will it automatically re-specialize? I guess if I do kernel.bind and pass in new inputs it won't re-specialize?

kernel.bind() specializes on numel>=2**31 if index_dtype=None. So you will get an automatic recompile if the kernel inputs cross that boundary. If you do @helion.kernel(index_dtype=torch.int64, static_shapes=False) it won't specialize.

If you manually call kernel.bind(), you must manually guarantee all inputs you pass to the BoundKernel have the same specialization_key.

This PR also updates the docs. Do you find the docs clear?

[Chillee]

How do I figure out the specialization key for a set of inputs?

[jansel]

kernel.specialization_key([a, b, c]) in code and there is discussion of it in https://helionlang.com/deployment_autotuning.html