Repository navigation
test(cuda.core): gate the NVRTC option tests on the cuda-bindings version - #3335
Merged
Merged
Conversation
…sion test_cpp_program_with_various_options checks that cuda.core encodes each ProgramOptions field as a flag that NVRTC accepts. An older NVRTC rejects flags that it predates, so the test failed with NVRTC 12.6 or 12.8 for no_cache, frandom_seed, and ofast_compile, which first appear in 12.9. Run the test only when the NVRTC major.minor is at least the cuda-bindings major.minor. The gate moves with cuda-bindings releases and needs no per-option version table. It makes the per-option PCH gates (12.8) redundant, so remove them. Closes NVIDIA#3333. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Andy-Jost
added
bug
Findings
 |
| Layer / File(s) | Summary |
|---|---|
Gate option tests by NVRTC version cuda_core/tests/test_program.py |
The parameterized test uses the cuda-bindings major/minor version as its NVRTC threshold. Individual PCH parameter skips were removed. |
Assessment against linked issues
| Objective | Addressed | Explanation |
|---|---|---|
Skip NVRTC option cases when the installed NVRTC version is below the required version [#3333] |
✅ | |
Avoid failures from torch CUDA Array Interface v2 samples [#3333] |
❌ | The PR does not change the CUDA Array Interface tests or their handling of v2 samples. |
Possibly related issues
- Issue 2722: The PR changes the version conditions for the program-option test, but does not address the broader CI coverage objective.
- Issue 2077: The PR changes the version conditions for the program-option test, but does not address the broader test coverage objective.
Suggested reviewers: leofang
Change: Other · Severity of issue fixed: Low
Merge Risk: 
69ea3
-
Some supported NVRTC options lose test coverage in mixed-version environments. Narrowing the skips would retain that coverage.
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
| ProgramOptions(instantiate_templates_in_pch=True), | ||
| ], | ||
| ) | ||
| @nvrtc_at_bindings_version |
There was a problem hiding this comment.
important: Gate only unsupported option cases.
With cuda-bindings 13.4 and NVRTC 13.3, this skips every parameter, including ftz, which NVRTC 13.3 supports. (docs.nvidia.com) This removes option coverage from supported mixed-version environments. Apply version gates to the affected parameters instead.
As per path instructions: “skip only for genuine setup or feature-unavailable conditions.”
Source: Path instructions
Contributor
|
Andy-Jost
added
P2
juenglin
approved these changes
Nov 2, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
test_cpp_program_with_various_optionschecks that cuda.core encodes eachProgramOptionsfield as a flag that NVRTC accepts. With an older NVRTC 12.x (for example 12.6, which torch cu126 pins), the test fails forno_cache,frandom_seed, andofast_compile, because these flags first appear in NVRTC 12.9. This change runs the test only when the NVRTC major.minor is at least the major.minor of the installed cuda-bindings, and skips it on an older NVRTC. The gate moves with cuda-bindings releases, so it needs no per-option version table.Changes
cuda_core/tests/test_program.py: add thenvrtc_at_bindings_versionmarker and apply it totest_cpp_program_with_various_options.cuda_core/tests/test_program.py: remove the four per-option PCH gates (NVRTC >= 12.8) in that test; the new gate covers them. Thenumba_debuggate (13.2) and thedevice_float128architecture gate stay.CI lanes whose toolkit matches the cuda-bindings build version (12.9.x and 13.4.x) still run the test; lanes with an older toolkit now skip it.
Related Work
Closes #3333. Item 2 of that issue (torch CAI v2) was fixed by #3033.
🤖 Generated with Claude Code