2.0.0 rc rebase 2

[ktangsali]

PhysicsNeMo Pull Request

Description

2nd Rebase of RC Branch into Main

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.
• An issue is linked to this pull request.
• If I am implementing a new model or modifying any existing model, I have followed the Models Implementation Coding Standards.

Dependencies

Review Process

All PRs are reviewed by the PhysicsNeMo team before merging.

Depending on which files are changed, GitHub may automatically assign a maintainer for review.

We are also testing AI-based code review tools (e.g., Greptile), which may add automated comments with a confidence score.
This score reflects the AI’s assessment of merge readiness and is not a qualitative judgment of your work, nor is
it an indication that the PR will be accepted / rejected.

AI-generated feedback should be reviewed critically for usefulness.
You are not required to respond to every AI comment, but they are intended to help both authors and reviewers.
Please react to Greptile comments with 👍 or 👎 to provide feedback on their accuracy.

[ktangsali]

/blossom-ci

[greptile-apps]

Greptile Summary

This PR is the second rebase of the 2.0.0 release-candidate branch into main. It bundles several targeted fixes and documentation improvements across the diffusion, model, and checkpoint modules, with the most notable code changes being a breaking API rename and a bug fix in the GALE attention block.

Key changes:

• GALE residual connection bug fix (gale.py): The pre-norm Transformer residual connection now correctly adds the attention output to the unnormalized input fx[i] instead of the normalized normed_inputs[i]. This is a meaningful correctness fix confirmed by updated test-data .pth files.
• DPSDenoiser → DPSScorePredictor rename: The class no longer inherits from Denoiser; it is now a plain callable that implements the Predictor protocol. The __init__.py public API has been updated accordingly, but no backward-compatibility alias is provided — this is a breaking change.
• Inference-mode guards: New RuntimeError guards added to DPSScorePredictor.__call__, ModelConsistencyDPSGuidance.__call__, and DataConsistencyDPSGuidance.__call__ to fail fast when called inside torch.inference_mode() (which prevents autograd from working). The error message wording is slightly ambiguous.
• inference_on_vtk.py tensor construction: Square brackets removed from torch.tensor([air_density]) / torch.tensor([stream_velocity]), changing the result from a 1-D (1,) tensor to a 0-D scalar tensor. This may break downstream batching logic.
• Docstring maintenance: Numerous RST cross-reference fixes (DSMLoss → MSEDSMLoss), LaTeX subscript escaping, and note-block reformatting throughout the diffusion and model subpackages.
• New checkpoint doctest examples and expanded API docs for mesh, datapipes, and diffusion modules.

Important Files Changed

Filename	Overview
physicsnemo/experimental/models/geotransolver/gale.py	Bug fix in GALE_block: residual connection now correctly uses unnormalized fx[i] instead of normed_inputs[i], matching standard pre-norm Transformer semantics. Test data files updated to reflect corrected model outputs.
physicsnemo/diffusion/guidance/dps_guidance.py	Class renamed DPSDenoiser→DPSScorePredictor (no longer inherits Denoiser base class); new inference-mode guard added to three call methods; docstring and math notation updated (γ→ρ, Γ; variance→covariance).
physicsnemo/diffusion/guidance/init.py	DPSDenoiser removed from exports and replaced with DPSScorePredictor. Breaking public API change with no backward compatibility alias.
physicsnemo/utils/checkpoint.py	Docstring improvements: clarified model naming behavior and added runnable doctest examples for save_checkpoint and load_checkpoint. No logic changes.
examples/cfd/external_aerodynamics/transformer_models/src/inference_on_vtk.py	air_density and stream_velocity tensors changed from torch.tensor([scalar]) (shape (1,)) to torch.tensor(scalar) (shape ()). Shape change may affect downstream batching behavior.
test/diffusion/test_preconditioners.py	Added two new pytest fixtures: deterministic_settings (handles CUDA/cuDNN flags safely on CPU-only builds) and tolerances (returns per-device tolerance levels). Clean, well-guarded implementations.

_{Last reviewed commit: eb795b8}

[ktangsali]

/blossom-ci

[ktangsali]

/blossom-ci