test(cuda.core): cover the forkserver start method for VMM IPC buffers

[Andy-Jost]

Summary

The VMM IPC tests from #3028 ran under spawn only. forkserver is safe after CUDA initialization because the forkserver is a fresh interpreter and children fork from it, not from the test process, and it is the Linux default start method since Python 3.14. Its Popen collects file descriptors during pickling and sends them afterwards, like spawn, so the Process-argument path in VirtualMemoryBuffer.__reduce__ has to work there too. This PR adds the tests that confirm it.

Changes

• tests/memory_ipc/test_vmm_ipc.py: new TestForkserver class with two tests under mp.get_context("forkserver"): a VirtualMemoryBuffer passed as a Process argument, and the same buffer sent through a Queue. Each child reads the parent's fill pattern and writes one back. The class skips where forkserver is unavailable (Windows).

Related Work

Closes #3111. Suggested by @juenglin in the #3028 review.

🤖 Generated with Claude Code

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Repository: NVIDIA/cuda-python/.coderabbit.yaml
• Review profile: CHILL
• Plan: Enterprise
• Run ID: 4b14af62-c166-42fb-b4e7-4f96f3859686

📥 Commits

Reviewing files that changed from the base of the PR and between 2224499 and 219be4f.

📒 Files selected for processing: 1

• cuda_core/tests/memory_ipc/test_vmm_ipc.py

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

• Tests
  • Expanded validation of shared-memory communication between processes using the forkserver start method. Coverage now checks that processes can receive shared buffers directly or through queues, read and update shared data, and exit cleanly. These checks run only on platforms that support forkserver, helping ensure consistent behavior across supported environments.

The VMM IPC tests now cover forkserver child processes that receive buffers as process arguments or through a queue. The tests verify child reads, shared writes, and clean child exit. They skip when forkserver is unavailable.

Changes: Forkserver VMM IPC coverage

Layer / File(s)	Summary
Process-argument transfer cuda_core/tests/memory_ipc/test_vmm_ipc.py	A forkserver child receives the buffer as a process argument, reads the parent’s data, writes to the shared buffer, and closes it. The parent checks the write and child exit.
Queue buffer transfer cuda_core/tests/memory_ipc/test_vmm_ipc.py	A forkserver child receives the buffer through a queue, reports its data, writes to the shared buffer, and closes it. The parent checks the write and child exit.

Assessment against linked issues

Objective	Addressed	Explanation
Add forkserver VMM IPC tests for process arguments and queues, skipping where forkserver is unavailable [#3111]	✅

Priority: ⬇️ Low

Change: Other

Merge Risk: ⚪ Minimal · up to 219be

The new forkserver tests are mergeable after normal checks; no blocking issue was identified.

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

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

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-3124/
https://nvidia.github.io/cuda-python/pr-preview/pr-3124/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-3124/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-3124/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[juenglin]

Optional: this test checks the data round trip, but not that the parent's file descriptors are released. The spawn counterpart, test_process_argument_releases_with_the_process_object, also checks that the open file descriptor count returns to its baseline once the Process object is released. The same check under forkserver would catch a stash that outlives the Process.

It would need thread_unsafe (it counts process-wide file descriptors) and _settled_open_fds() for the baseline. Start the forkserver before taking the baseline (for example with one throwaway process), because its first start leaves a permanent file descriptor open in the parent.