cuda.bindings: pass host functions to cuLaunchHostFunc and cudaLaunchHostFunc unchanged

[Andy-Jost]

Summary

driver.cuLaunchHostFunc and runtime.cudaLaunchHostFunc routed the host function through an internal trampoline that freed its (fn, userData) record after the first call. A host function captured into a CUDA graph is called again on every launch of that graph, so the second launch used freed memory and crashed (#790). The wrappers now pass fn and userData to CUDA unchanged, as the _v2 variants already do.

Changes

• cuda/bindings/driver.pyx, cuda/bindings/runtime.pyx: remove the cuHostCallbackWrapper and cudaStreamRtHostCallbackWrapper trampolines and their records; the two wrappers call CUDA with the caller's function pointer and user data. The generated-file seals are updated. The delta equals the output of the cybind generator with the matching fix (cybind MR !587), applied to the current bindings; the generator change reaches the public bindings through the regular regeneration.
• tests/test_cuda.py, tests/test_cudart.py: capture a host function into a graph and launch it five times through the driver API and through the runtime API; the function must run once per launch. Both tests crash on the second launch without the fix.
• 13.5.0 release note.

Related Work

Fixes #790. The generator side is cybind MR !587; review there asked for this PR to carry the output and a regression test that runs real code.

🤖 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: 8b554a4d-c8fe-4dc9-bd05-714cd8169942

📥 Commits

Reviewing files that changed from the base of the PR and between 7a21ea4 and 9d5441d.

📒 Files selected for processing: 4

• cuda_bindings/docs/source/overview.rst
• cuda_bindings/docs/source/release/13.5.0-notes.rst
• cuda_bindings/tests/test_cuda.py
• cuda_bindings/tests/test_cudart.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

• Bug Fixes
  • Fixed host callbacks captured in CUDA graphs so they can run on repeated graph launches.
  • Host callbacks and their user data are passed directly to CUDA through both driver and runtime APIs.
• Documentation
  • Clarified that callbacks and user data must remain available until pending launches finish; for graph captures, retain them until the graph and its executables are destroyed.
• Tests
  • Added coverage verifying callbacks run once per launch across repeated graph replays.

The driver and runtime host-function bindings now pass callbacks and user data directly to CUDA. Tests capture callbacks in CUDA graphs and verify five launches each. Documentation describes callback and user-data lifetime requirements.

Changes: Host Function Graph Replay

Layer / File(s)	Summary
Direct callback forwarding cuda_bindings/cuda/bindings/driver.pyx, cuda_bindings/cuda/bindings/runtime.pyx	Both bindings remove callback-data allocation and trampoline wrappers. They pass the callback and user data directly to CUDA.
Graph-replay coverage and lifetime guidance cuda_bindings/tests/test_cuda.py, cuda_bindings/tests/test_cudart.py, cuda_bindings/docs/source/overview.rst, cuda_bindings/docs/source/release/13.5.0-notes.rst	Tests verify five callback invocations across five graph launches for each API. The overview and release notes describe callback and user-data lifetime requirements.

Assessment against linked issues

Objective	Addressed	Explanation
[#790] Allow cudaLaunchHostFunc callbacks captured in CUDA graphs to execute on repeated graph replays.	✅

Suggested reviewers: mdboom

Priority: ⬆️ High

Change: Bug fix · Severity of issue fixed: Medium

Merge Risk: ⚪ Minimal · up to 9d544

No concrete issue remains that should block merging after normal checks.

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

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

[juenglin]

Does this fix need to be backported to the 12.9.x branch?

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-3061/
https://nvidia.github.io/cuda-python/pr-preview/pr-3061/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-3061/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-3061/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[mdboom]

Does this fix need to be backported to the 12.9.x branch?

Yes. I would apply the to-be-backported label but the bot it triggers doesn't do the right thing. It just needs to be manually backported (with the test + the regenerated bindings with --ctk-target-version 12.9.

[mdboom]

LGTM, assuming tests pass.

[juenglin]

Follow-ups (not blocking):

• Backport to 12.9.x, which has the same trampolines.
• test_cudart_func_callback can't fail: its asserts run inside ctypes callbacks, where exceptions are swallowed. Fix it rather than delete it, since it's the only test of cudaStreamAddCallback.
• Document that fn and the userData memory must outlive any graph exec containing the node.

[mdboom]

Based on this that @juenglin found, I'm retracting my approval, since without that I think the fact that the tests are passing is meaningless.

test_cudart_func_callback can't fail: its asserts run inside ctypes callbacks, where exceptions are swallowed. Fix it rather than delete it, since it's the only test of cudaStreamAddCallback.

[mdboom]

See my comment -- I think we should block on this until the tests don't swallow exceptions.