fix(cuda.core): pass numa_id to cuMemCreate for host_numa virtual memory by somtri · Pull Request #3149 · NVIDIA/cuda-python

somtri · 2026-10-19T08:31:31Z

Description

VirtualMemoryResource with location_type="host_numa" passed -1 as the location id to cuMemCreate, and the driver rejects that with CUDA_ERROR_INVALID_VALUE. This follows the fix described in the issue comment:

VirtualMemoryResourceOptions has a new numa_id: int | None = None field, added last so the positional order does not change. _check_config requires it for "host_numa", rejects it for every other location type, and rejects negative values with the message PinnedMemoryResourceOptions uses. _fill_prop writes it to prop.location.id.
location_type="host_numa_current" raises ValueError at construction. The enum member stays.
One addition the comment did not ask for: in modify_allocation, _check_config also rejects a per-call config whose numa_id differs from the resource's, as it already does for location_type. _fill_prop takes the location from the resource, so without the check a different value would be ignored silently. I can drop it if you prefer.
test_vmm_host_location_types_report_host_accessible and test_is_ipc_enabled built host_numa without numa_id and built host_numa_current, so both needed updates. Three new tests cover the allocation, the host_numa_current rejection, and the numa_id checks. They carry agent_authored markers.
Release note in 1.3.0-notes.rst.

The allocation test writes and reads the buffer from the host. The resource grants access only to its own location, so cuMemsetD8 on the buffer fails with CUDA_ERROR_INVALID_VALUE unless the device is in peers. That is the host-resource peers gap from #2910, which this PR leaves alone.

Testing

Windows 11, RTX 3080 Laptop GPU (WDDM, driver 617.42), CUDA 13.4, cuda-bindings 13.4.4, Python 3.14. The device reports host_numa_id = 0 and host_numa_virtual_memory_management_supported = True.

The issue's reproducer, extended to host_numa_current and catching each error, before the change:

host allocated 2097152 bytes
host_numa FAILED: CUDAError CUDA_ERROR_INVALID_VALUE: ...
host_numa_current FAILED: CUDAError CUDA_ERROR_INVALID_VALUE: ...

After the change, numa_id=0 allocates 2097152 bytes, and host_numa_current raises ValueError at construction.

$ pytest tests/test_memory.py -k "vmm_host_numa or vmm_numa_id or host_location_types" -v
8 passed
$ pytest tests/test_memory.py tests/memory_ipc/test_vmm_ipc.py -k "vmm or Vmm or VirtualMemory or is_ipc_enabled" -q
50 passed, 29 skipped    (before the change: 45 passed, 30 skipped)

test_is_ipc_enabled skips on Windows ("Sharing virtual memory needs POSIX file descriptor handles"), so the test_vmm_ipc.py change has not run locally. pre-commit run passes on the five changed files.

Claude Code (claude-opus-5-5) wrote this change and ran the tests above on my machine. I reviewed the diff.

Checklist

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

VirtualMemoryResource passed -1 as the location id for host_numa allocations, and cuMemCreate rejects that with CUDA_ERROR_INVALID_VALUE. Add VirtualMemoryResourceOptions.numa_id. It is required for host_numa, rejected for other location types, and must match in a per-call modify_allocation config. Reject host_numa_current at construction, because the driver does not accept it for physical allocations. Closes NVIDIA#2694 Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Som Tripathi <somtri@iastate.edu>

copy-pr-bot · 2026-10-19T08:31:34Z

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.

coderabbitai · 2026-10-19T08:35:34Z

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Source mode: Auto
Repository: NVIDIA/cuda-python
Branch: main
File: .coderabbit.yaml
Commit: d5f37e0
Layers: Repository YAML
Review profile: CHILL
Plan: Enterprise
Run ID: 44cd9ddf-47ae-4747-9aaf-b22f82362b39

📥 Commits

Reviewing files that changed from the base of the PR and between 1e11c4a and e9a2441.

📒 Files selected for processing: 5

cuda_core/cuda/core/_memory/_virtual_memory_resource.pyi
cuda_core/cuda/core/_memory/_virtual_memory_resource.pyx
cuda_core/docs/source/release/1.3.0-notes.rst
cuda_core/tests/memory_ipc/test_vmm_ipc.py
cuda_core/tests/test_memory.py

🚧 Files skipped from review as they are similar to previous changes: 1

cuda_core/docs/source/release/1.3.0-notes.rst

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

📝 Summary

Summary by CodeRabbit

New Features
- Physical memory allocations can target a specific host NUMA node using location_type="host_numa" and a nonnegative numa_id.
- Device properties identify the NUMA node closest to the device. On systems without NUMA, use numa_id=0.
Bug Fixes
- Invalid NUMA configurations are rejected, including missing or misplaced NUMA IDs, negative IDs, per-call growth on a different NUMA node, and host_numa_current for physical allocations.

VirtualMemoryResourceOptions now accepts a NUMA node ID for host NUMA allocations. VirtualMemoryResource validates the configuration and uses that ID for physical allocations. Tests cover allocation, host access, and invalid configurations.

Changes: Host NUMA allocations

Layer / File(s)	Summary
NUMA option contract `cuda_core/cuda/core/_memory/_virtual_memory_resource.pyx`, `cuda_core/cuda/core/_memory/_virtual_memory_resource.pyi`	Adds and documents the optional `numa_id` option. It is required for `host_numa`, rejected for other location types, and must be nonnegative.
NUMA validation and allocation `cuda_core/cuda/core/_memory/_virtual_memory_resource.pyx`	Validates the NUMA configuration, rejects `host_numa_current`, and sets the host NUMA allocation location ID to `numa_id`.
NUMA behavior coverage `cuda_core/tests/test_memory.py`, `cuda_core/tests/memory_ipc/test_vmm_ipc.py`, `cuda_core/docs/source/release/1.3.0-notes.rst`	Tests cover host NUMA allocation, host access, IPC configuration, and invalid settings. Release notes describe the option and its constraints.

Assessment against linked issues

Objective	Addressed	Explanation
Enable `host_numa` allocations by providing the required NUMA node ID [`#2694`]	✅
Reject `host_numa_current` for physical allocations [`#2694`]	✅

Possibly related issues

cuda.core: VirtualMemoryResourceOptions accepts values the driver rejects, and modify_allocation mishandles edge inputs #2910: The PR rejects HOST_NUMA_CURRENT but does not address the issue’s separate host-resource peers behavior.

Possibly related PRs

NVIDIA/cuda-python#3034: Both PRs change the handling of release notes in cuda_core/docs/source/release/1.3.0-notes.rst; #3034 deletes that file and may remove this PR’s note if it is not migrated.

Suggested reviewers: andy-jost

Priority: ➖ Normal

Change: Bug fix · Severity of issue fixed: Medium

Merge Risk: ⚪ Minimal · up to e9a24

The NUMA allocation change appears ready to merge after normal checks; no actionable merge-blocking issue was established.

- [ ] Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

coderabbitai

Actionable comments posted: 1

ℹ️ Review info

⚙️ Run configuration

Configuration used: Repository: NVIDIA/cuda-python/.coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: b60e59cb-b28e-419e-a1f0-4c8405fefed6

📥 Commits

Reviewing files that changed from the base of the PR and between 576ef68 and c33542b.

📒 Files selected for processing: 5

cuda_core/cuda/core/_memory/_virtual_memory_resource.pyi
cuda_core/cuda/core/_memory/_virtual_memory_resource.pyx
cuda_core/docs/source/release/1.3.0-notes.rst
cuda_core/tests/memory_ipc/test_vmm_ipc.py
cuda_core/tests/test_memory.py

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 10 remain after this review.

The numa_id docstring points to Device.properties.host_numa_id, which is -1 on a system without NUMA, and numa_id rejects -1. cuMemCreate wants location id 0 on such a system, so say that in the docstring. Signed-off-by: Som Tripathi <somtri@iastate.edu>

Andy-Jost · 2026-10-23T18:27:40Z

/ok to test 1e11c4a

Andy-Jost

Thank you for the fix, @somtri. The change is small, it matches the scope agreed in #2694, and the new tests cover the host_numa path that was broken. I have one non-blocking nit.

Nits

cuda_core/cuda/core/_memory/_virtual_memory_resource.pyx, line 716 (modify_allocation docstring). The new check at lines 484-488 rejects a per-call config whose numa_id differs from the resource's, but the docstring still describes only location_type and handle_type. Please add numa_id in two places: the config entry under Parameters ("must name the resource's location_type, handle_type, and numa_id") and the ValueError entry under Raises ("if config names a different location type, handle type, or numa_id"). That matches how the existing location_type rule is documented.

Your questions

Keep the per-call numa_id mismatch check in modify_allocation. Yes, keep it. It mirrors the existing location_type rule, and _fill_prop reads the location from the resource's config, so without the check a different per-call numa_id would be ignored silently. We prefer an explicit error over a silent fallback. The docstring update in nit 1 is the only follow-up.
The test_vmm_ipc.py change that has not run locally. That is fine. The module skips on Windows, and CI runs it on Linux, so CI will be its first run. If the host_numa row fails there, we can look at it together.

Next steps

I am adding the missing metadata (assignee, bug and P1 labels, milestone "cuda.core 1.3.0") and starting CI with /ok to test. Once CI is green and the docstring nit is in, I will merge.

…1.0 library names) Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Andy-Jost · 2026-10-23T21:08:42Z

I merged main into this branch to pick up #3188 (cuDSS 1.0 library names), which the pathfinder CI step needs now. No other change. CI runs again on the new head.

Andy-Jost · 2026-10-23T21:08:44Z

/ok to test e9a2441

github-actions · 2026-10-23T21:41:03Z

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-3149/
https://nvidia.github.io/cuda-python/pr-preview/pr-3149/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-3149/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-3149/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

github-actions Bot added the cuda.core Everything related to the cuda.core module label Oct 19, 2026

coderabbitai Bot reviewed Oct 19, 2026

View reviewed changes

Comment thread cuda_core/cuda/core/_memory/_virtual_memory_resource.pyx Outdated

Andy-Jost added bug Something isn't working P1 Medium priority - Should do labels Oct 23, 2026

Andy-Jost assigned somtri and Andy-Jost Oct 23, 2026

Andy-Jost added this to the cuda.core 1.3.0 milestone Oct 23, 2026

Andy-Jost approved these changes Oct 23, 2026

View reviewed changes

Andy-Jost mentioned this pull request Oct 23, 2026

cuda.core: VirtualMemoryResourceOptions accepts values the driver rejects, and modify_allocation mishandles edge inputs #2910

Open

Merge main into fix/2694-vmm-host-numa to pick up NVIDIA#3188 (cuDSS …

e9a2441

…1.0 library names) Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Navigation Menu

Repository navigation

fix(cuda.core): pass numa_id to cuMemCreate for host_numa virtual memory - #3149

fix(cuda.core): pass numa_id to cuMemCreate for host_numa virtual memory#3149
somtri wants to merge 3 commits into
NVIDIA:mainfrom
somtri:fix/2694-vmm-host-numa

somtri commented Oct 19, 2026

Uh oh!

copy-pr-bot Bot commented Oct 19, 2026

Uh oh!

coderabbitai Bot commented Oct 19, 2026 •

edited

Loading

Summary by CodeRabbit

Assessment against linked issues

Uh oh!

coderabbitai Bot left a comment

Uh oh!

Uh oh!

Andy-Jost commented Oct 23, 2026

Uh oh!

Andy-Jost left a comment

Uh oh!

Andy-Jost commented Oct 23, 2026

Uh oh!

Andy-Jost commented Oct 23, 2026

Uh oh!

github-actions Bot commented Oct 23, 2026

Preview will be ready when the GitHub Pages deployment is complete.

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

somtri commented Oct 19, 2026

Description

Testing

Checklist

Uh oh!

copy-pr-bot Bot commented Oct 19, 2026

Uh oh!

coderabbitai Bot commented Oct 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Assessment against linked issues

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Andy-Jost commented Oct 23, 2026

Uh oh!

Andy-Jost left a comment

Choose a reason for hiding this comment

Uh oh!

Andy-Jost commented Oct 23, 2026

Uh oh!

Andy-Jost commented Oct 23, 2026

Uh oh!

github-actions Bot commented Oct 23, 2026

Preview will be ready when the GitHub Pages deployment is complete.

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

coderabbitai Bot commented Oct 19, 2026 •

edited

Loading