supporting float32 on Scalar values

[jjsjann123]

Tracking the issue for user experience. In the issue comment here.

The to translate a user program like this vvv

# t18 = prims.where(t17, -0.0, 0.0) # t18: "cuda:0 f32[4, 2, 3]"

We are calling nvfuser where on (boolean_tv, scalar_1, scalar_2).
The type promotion/inference logic is that, we determine the output dtype of where based on input[1] and input[2].

We are producing output TV from this operation, but since all scalar types are in double, nvfuser generates a double tensor as output, instead of a float32 as the thunder scripts. (since thunder sees input scalar as in float).

This isn't blocking, as we can always patch the executor in thunder for explicit type inference: Lightning-AI/lightning-thunder#1734

Per our thunder developer's request, we still want an issue to track this so we no longer needed these WAR on user of nvfuser.

[kevinstephano]

@jjsjann123 should we continue to track this issue? There has not been any action 6 months.

[jjsjann123]

I'm closing this as not-planned. Since there's existing WAR on integration already.