Validate checkpoint GPU UUID inputs early

[aryanputta]

Summary

• reject non-UUID gpu_mapping values in cuda.core.checkpoint before they reach driver-facing restore args
• normalize malformed UUID string failures to the documented ValueError path
• add source-level checkpoint helper tests that run without importing the built cuda.core package

Testing

• python3 -m pytest cuda_core/tests/test_checkpoint_helpers.py --noconftest

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[leofang]

Hi @aryanputta as you might have seen in your other PRs, we generally do not accept mock tests like this. Please make sure new/revised tests fit in the existing test style, and GPU tests actually pass locally before submitting.

[leofang]

Actually, reading this more carefully: What is the motivation for accepting CUuuid here? Do you have a use case?

Note that Device.uuid returns a string:
https://nvidia.github.io/cuda-python/cuda-core/latest/generated/cuda.core.Device.html#cuda.core.Device.uuid

[aryanputta]

@leofang You were right on both points here. I over-generalized the helper instead of matching the actual cuda.core API surface.

What I changed in commit 083e411:

• removed the mock-only cuda_core/tests/test_checkpoint_helpers.py
• narrowed _as_cuuuid() so restore mappings now accept UUID strings only
• updated the docs to say gpu_mapping should use the strings returned by Device.uuid

On the motivation question specifically: I do not have a strong public-facing use case for accepting CUuuid objects here. The intended caller-facing path is the string UUID returned by Device.uuid, and the existing real checkpoint tests already follow that contract. For example, cuda_core/tests/test_checkpoint.py builds migration mappings from devices[i].uuid in _build_rotation_mapping() and runs the driver-backed checkpoint/restore scenarios without mocks.

I also checked local validation from this environment. Syntax validation passed, but I could not run the checkpoint pytest module end-to-end here because test import setup fails locally with ModuleNotFoundError: cuda.pathfinder before reaching the GPU-backed scenarios. So I did not keep or add any mock-style substitute tests for this change.