Update test_linear_nvfuser

[IvanYashchuk]

Thunder's nvFuser executor got support for prims.linear in #318. This PR updates the test

• to use the sample generator from OpInfos
• to reject linear execution if inputs are not bf16, fp16
• to skip the test if the nvFuser version is not new enough

[nikitaved]

What does that mean?

[IvanYashchuk]

There's a for loop over sample inputs, inside the for loop there's a continue statement to skip inputs that are not 2D. If someone goes ahead and modifies the sample input generator for linear to not generate a 2D sample this test won't test anything silently. With assert on "at least one sample input tested," there would be a failure.

[nikitaved]

Or we could just create a single sample and append to the cases?

[IvanYashchuk]

Yes, it's a viable option.

[jjsjann123]

Wasn't aware there's such restriction.

cc'ing @Priya2698 is this an intended behavior? I thought we are just dispatching to aten now and don't see why it's rejecting full precision.

[Priya2698]

At the moment, we do have that restriction since we use fusedMultiplySum and MmaOp that only accept bf16/fp16.

[Priya2698]

This restriction will although be lifted with the new IR nodes.

[jjsjann123]

Got'ya.

Thanks @IvanYashchuk for catching this. @Priya2698 remember to do a version bump when that restriction is lifted, so we can have a proper behavior guarded with nvfuser_version here.

[Priya2698]

Thanks for creating this example. I'll update the matmul test similarly once we have complete support.