✨[Feature] Automatic casting of Long/Double Tensor constants in FX

[gs-olive]

Context

Certain Torch layers, including torch.nn.Embedding handle inputs, weights, and constant Tensors which are required by Torch to be 64-bit types. This causes issues when attempting to build TRT converters for such engines, since the inputs and constants will then be 64-bit, which is not supported by TRT.

Feature Proposal

When users specify the truncate_long_and_double flag, it should have an effect in both FX and Dynamo. A similar feature was implemented for Dynamo compile in a broader context in PR #1983. For this feature, if the user specifies the truncate_long_and_double flag, we should automatically cast the 64-bit constant tensors (weights, constant indices, etc.) to their 32-bit equivalents for use in TRT.

[github-actions]

This issue has not seen activity for 90 days, Remove stale label or comment or this will be closed in 10 days

[gs-olive]

Fixed on main, in the Dynamo path