Test nim inference workflow and make minor updates

[ktangsali]

PhysicsNeMo-CFD Pull Request

Description

Tested the nim inference workflow and made minor edits as a part of the testing. The l2_error_vs_sdf now clip first and then compute the metrics. Also, device is handled better.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.
• The CHANGELOG.md is up to date with these changes.

Dependencies

[ktangsali]

/blossom-ci

[peterdsharpe]

Submitting initial comments on everything except the notebooks (will review locally).

Other than the comments here, is it possible to save the PNG files (specifically resampled_volume_errors.png and variations_due_to_checkpoint.png) in a smaller file size? Right now those are adding ~25 MB to the repo size in total. Maybe JPEG or lower-res PNG?

[ktangsali]

Submitting initial comments on everything except the notebooks (will review locally).

Other than the comments here, is it possible to save the PNG files (specifically resampled_volume_errors.png and variations_due_to_checkpoint.png) in a smaller file size? Right now those are adding ~25 MB to the repo size in total. Maybe JPEG or lower-res PNG?

Thanks! Compressed the images.

[ktangsali]

/blossom-ci

[peterdsharpe]

Overall looks good. Leaving my notebook comments here, since GitHub won't let me add line-by-line annotations within *.ipynb's.

---

General items:

• Please update changelog (warp version bump), Dependencies section.

---

A blocker in our notebooks, relating to us pointing users to deprecated APIs:

• The cell that runs NIM inference uses _create_nbrs_surface and _interpolate. These are explicitly deprecated in the same file, and we direct users to use interpolate_mesh_to_pc or physicsnemo.nn.functional.neighbors.knn instead. Can we fix our tutorial code here to use non-deprecated APIs? The NIM produces a sparse point cloud output_dict["surface_coordinates"][0] with fields. Wrap it as pv.PolyData(coords) with point_data populated, then call interpolate_mesh_to_pc with pc=mesh_cell_centers, mesh=nim_pc, mesh_dtype="point", device="gpu". About 5 lines of glue.

---

Another blocker in our notebooks, related to credentials leaking:

The cells that run !wget "{url}" for drivaer_418.stl and drivaer_420.stl capture the redirect chain, including:

https://cas-bridge.xethub.hf.co/xet-bridge-us/.../drivaer_418.stl?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=cas%2F...&X-Amz-Signature=155cd1bfd12bedc99e9a5c8eafca8602bb1edbbf108495138d9d0af5906e971d&...

These are pre-signed S3-style URLs. They expired ~1 hour after generation (so this PR's URLs are dead now), but committing them is bad hygiene: it normalizes embedding presigned URLs in public artifacts, and future re-runs will commit live ones. Concrete fixes, in order of preference:

1. Replace !wget "{url}" -O "{filename}" with huggingface_hub.hf_hub_download (already a transitive dep via evaluation-hf). It logs nothing comparable.
2. Use requests.get(url, stream=True) with iter_content and a small progress bar that doesn't echo the URL.
3. At minimum, add --quiet to wget and clear the output cells before commit.

[ktangsali]

Overall looks good. Leaving my notebook comments here, since GitHub won't let me add line-by-line annotations within *.ipynb's.

General items:

• Please update changelog (warp version bump), Dependencies section.

A blocker in our notebooks, relating to us pointing users to deprecated APIs:

• The cell that runs NIM inference uses _create_nbrs_surface and _interpolate. These are explicitly deprecated in the same file, and we direct users to use interpolate_mesh_to_pc or physicsnemo.nn.functional.neighbors.knn instead. Can we fix our tutorial code here to use non-deprecated APIs? The NIM produces a sparse point cloud output_dict["surface_coordinates"][0] with fields. Wrap it as pv.PolyData(coords) with point_data populated, then call interpolate_mesh_to_pc with pc=mesh_cell_centers, mesh=nim_pc, mesh_dtype="point", device="gpu". About 5 lines of glue.

Another blocker in our notebooks, related to credentials leaking:

The cells that run !wget "{url}" for drivaer_418.stl and drivaer_420.stl capture the redirect chain, including:

https://cas-bridge.xethub.hf.co/xet-bridge-us/.../drivaer_418.stl?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=cas%2F...&X-Amz-Signature=155cd1bfd12bedc99e9a5c8eafca8602bb1edbbf108495138d9d0af5906e971d&...

These are pre-signed S3-style URLs. They expired ~1 hour after generation (so this PR's URLs are dead now), but committing them is bad hygiene: it normalizes embedding presigned URLs in public artifacts, and future re-runs will commit live ones. Concrete fixes, in order of preference:

1. Replace !wget "{url}" -O "{filename}" with huggingface_hub.hf_hub_download (already a transitive dep via evaluation-hf). It logs nothing comparable.
2. Use requests.get(url, stream=True) with iter_content and a small progress bar that doesn't echo the URL.
3. At minimum, add --quiet to wget and clear the output cells before commit.

Thanks, I have added --quiet flag, updated the interpolation code and added a changelog line.

[ktangsali]

/blossom-ci

[peterdsharpe]

Approving after offline discussion.

One remaining item that we agreed to do before merge - re-running the notebooks to do a) end-to-end integration testing, and b) cleanup outdated output cells that could mislead end-users using this as primary documentation.