[FEA]: Add a MemoryResource which uses CUDA VMM APIs to allocate memory

[benhg]

Description

This is an implementation for #967.

Summary:

• Implement VMM grow-in-place via:
  • Fast path: contiguous VA extension with fixedAddr + cuMemMap
  • Slow path: remap to new VA, preserve data without memcpy
• Use aligned sizes everywhere and cast pointers for arithmetic
• Return proper enums for CUmemAccessDesc.flags
• Default all VMMConfig fields to common values

References:

NVIDIA VMM blog
CUDA Driver VMM docs

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[leofang]

@benhg the IPC PR was merged, could you resolve conflicts plz? 🙂

[benhg]

@benhg the IPC PR was merged, could you resolve conflicts plz? 🙂

Done, thanks for the reminder.

[leofang]

pre-commit.ci autofix

[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.

[kkraus14]

I agree with @Andy-Jost's comment that it would be nice to parametrize some existing tests with using the VirtualMemoryResource, but I'm okay doing that in a follow up

[leofang]

Thanks, Ben! Sorry I was not as responsive as I'd like for reviews. Keith told me this is ready to go. We have a number of PRs that have conflicts with each other, so what I'd do is to merge all P0 tasks for the release, and then merge this the last, to minimize conflicts. By the time we get to merge this PR, I will resolve all conflicts and rerun tests. No action is needed.

Marking this PR red only to avoid unplanned merge, again no action is needed. Much appreciated, Ben!

[leofang]

I am working on resolving the merge conflicts logically.

[leofang]

/ok to test 6450712

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-968/
https://nvidia.github.io/cuda-python/pr-preview/pr-968/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-968/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-968/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[benhg]

/ok to test 4af54ac

[leofang]

Vetter: a user that has permissions to leave an /ok to test comment on a pull request. This is a user with write access (or greater) for a particular repository or a member that is listed in the additional_vetters configuration list (see note below)

I always forget that not all NV employees can launch the CI... 😞

[leofang]

/ok to test 4af54ac

[leofang]

Ben, Piotr and I discussed with Vishnu from the memory team, and confirmed that VMM is not supported by the TCC mode. I suggested Ben to skip the tests on Windows for now.. perhaps also raise a warning (or exception?) on VirtualMemoryResource constructor. We can revisit this once we turn on WDDM (#462) and MCDM CI (#984).

[leofang]

/ok to test 9db04b1

[leofang]

Thanks Ben for your first contribution to cuda.core and Keith/Andy for review 🎉