Fix output error occurring due to check if it is a dict or not

[rnyak]

The Transformers4Rrec end-to-end example gives an error Failed to process the request(s) for model 'mymodel_pt', message: ValueError: output of the forward function should be a dict when we send request to TIS. the reason for that is due to recent changes in the prediction tasks' outputs in TF4Rec repo (see NVIDIA-Merlin/Transformers4Rec#546). This PR provides a quick fix to avoid such error. However, for long term solution Oliver's draft PR (see NVIDIA-Merlin/systems#252) might be preferable.

[karlhigley]

Looks fine but could be cleaned up a bit. Left suggestions that can be merged directly from the comments.

[karlhigley]

The code modified by NVIDIA-Merlin/systems#252 was never intended to be used directly—it was added to Systems as a starting point for refactoring toward the operator API, and some parts of the old code still exist there. If something is using the legacy serving code (e.g. export_pytorch_ensemble) from Systems instead of from NVTabular, it should be reverted to using the NVT version, which is mostly frozen in order avoid regressions where we have insufficient test coverage.

Making minor fixes to the NVT code in order to keep it working with the rest of the Merlin ecosystem is fine if needed, but we've been avoiding making any major changes there. The corresponding code in Systems is open to major changes and therefore unstable; it's also highly likely to be removed in a future Systems version.

[github-actions]

Documentation preview

https://nvidia-merlin.github.io/NVTabular/review/pr-1742

[karlhigley]

See NVIDIA-Merlin/systems#265 for links to the open PRs related to legacy (non-Torchscript Python) serving for PyTorch. Possible that the fixes in NVIDIA-Merlin/systems#252 should be applied to NVTabular instead?

[sararb]

See NVIDIA-Merlin/systems#265 for links to the open PRs related to legacy (non-Torchscript Python) serving for PyTorch. Possible that the fixes in NVIDIA-Merlin/systems#252 should be applied to NVTabular instead?

Thank you for the context @karlhigley, with Ronay we weren't sure if the legacy function will be included in systems or not. So It makes sense to me to include the changes proposed by Oliver in NVIDIA-Merlin/systems#265 in NVTabular instead (as a separate PR from this quick fix).