fix(cuda_core): use nvJitLink-compatible 0/1 values for boolean linker options by lijinf2 · Pull Request #3128 · NVIDIA/cuda-python

lijinf2 · 2026-10-16T00:06:36Z

Summary

Fixes nvbug 6621758.

nvJitLink's input validation was tightened upstream (CUDA Toolkit 13.6) to strictly require 0/1 for boolean-valued options and non-negative integers for -split-compile/-split-compile-extended.

cuda_core's LinkerOptions._prepare_nvjitlink_options() was emitting Python-style true/false strings and raw f"{bool}" ("True"/"False"), which the newer nvJitLink rejects with ERROR_UNRECOGNIZED_OPTION.

This PR changes the nvJitLink backend option formatting to:

-ftz, -prec-div, -prec-sqrt, -fma: emit 1/0 instead of true/false
-split-compile, -split-compile-extended: coerce value with int(...) so a bool is sent as 1/0 rather than the literal string "True"/"False"

The NVRTC backend (--ftz=true/false) and NVVM backend (-ftz=1/0, already correct) are untouched. This only affects the nvJitLink-specific option formatting.

Backward compatibility

Verified directly on nvJitLink 13.2.51, 13.4.92 (both released), and a 13.6.0 nightly build:

Option	13.2 / 13.4 (released)	13.6 (nightly, new validation)
`-ftz=true` / `-ftz=1`	both OK	`true` FAILS, `1` OK
`-split-compile=True` / `=1`	both OK	`True` FAILS, `1` OK

Released toolkits accept both the old and new formats, so always emitting 1/0 is backward compatible and does not regress any currently supported CUDA 13.x toolkit, while fixing compatibility with newer/nightly toolkits.

Testing

Reproduced and verified fix against nightly CUDA 13.6 toolkit on real GPU hardware: test_linker.py + test_program.py (241 tests) and test_program_cache.py (237 tests) pass with zero failures.

…621758)

copy-pr-bot · 2026-10-16T00:06:39Z

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 · 2026-10-16T00:15:01Z

📝 Summary

Summary by CodeRabbit

Bug Fixes
- Boolean nvJitLink options for flush-to-zero, precise division, precise square root, and fused multiply-add are now serialized as 0 or 1 instead of false or true. Split-compilation option values are also formatted numerically. LTO-enabled links using these settings can now complete and return compiled object code.

nvJitLink boolean options now serialize as 0 or 1, and split-compile option values are converted to integers. Tests check the ftz serialization and an LTO-enabled cubin link.

Changes: nvJitLink Option Serialization

Layer / File(s)	Summary
Format nvJitLink options `cuda_core/cuda/core/_linker.pyx`, `cuda_core/tests/test_linker.py`	The linker formats `ftz`, `prec-div`, `prec-sqrt`, and `fma` as `0` or `1`. It converts `split_compile` and `split_compile_extended` values to integers. Tests expect `-ftz=1` and verify that linking with LTO and the four boolean options returns an `ObjectCode`.

Suggested reviewers: andy-jost

Priority: ➖ Normal

Change: Bug fix

Merge Risk: 🔵 Low · up to 746df

The linker formatting changes work as intended, but some serialized values are not pinned by tests. A future regression could escape detection on compatible nvJitLink versions or with boolean split-compile inputs, leaving a bounded test-coverage risk.

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

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

✨ Finishing Touches: 💡 1

🛠️ Fix failing CI checks: 💡

Commit to this branch
Create a new PR

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: 36ecdafb-5184-41f1-a5f6-bad397b13885

📥 Commits

Reviewing files that changed from the base of the PR and between 2224499 and 0e5afde.

📒 Files selected for processing: 2

cuda_core/cuda/core/_linker.pyx
cuda_core/tests/test_linker.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.

lijinf2 · 2026-10-16T00:20:05Z

/ok to test c539cc2

github-actions · 2026-10-16T00:40:00Z

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-3128/
https://nvidia.github.io/cuda-python/pr-preview/pr-3128/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-3128/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-3128/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

leofang

I think this is a breaking change on the nvJitLink side. We should move the bug to their module and ask them to triage/fix.

lijinf2 · 2026-10-16T18:18:08Z

I think this is a breaking change on the nvJitLink side. We should move the bug to their module and ask them to triage/fix.

Agreed, This does look like a behavior change on the nvJitLink side. The old true/false format stopped being accepted starting with a recent build. Let me reach out to the nvJitLink team to understand why and report back before we decide how to proceed.

lijinf2 · 2026-10-19T22:33:12Z

Confirmed with the nvJitLink team: only 0/1 has ever actually worked. Other values (true/false/True) can be set without error, but fail once an LTO-enabled link actually runs. Fix stays as emitting 0/1.

Added an LTO-enabled runtime test, which closes a gap in previous coverage (option-formatting checks alone don't catch this).

Ready for re-review.

lijinf2 · 2026-10-19T22:33:24Z

/ok to test 746df99

coderabbitai

🧹 Nitpick comments: 2

cuda_core/tests/test_linker.py: 1
213-236: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win

suggestion: Assert all four formatted options.

The LTO link checks acceptance only. It does not detect a regression to true/false on versions that accept both spellings. Add focused serialization assertions and keep the runtime link as the end-to-end check.
Suggested test assertion
-    options = LinkerOptions(arch="sm_80", debug=True, ftz=True, max_register_count=32)
+    options = LinkerOptions(
+        arch="sm_80",
+        debug=True,
+        ftz=True,
+        prec_div=True,
+        prec_sqrt=True,
+        fma=True,
+        max_register_count=32,
+    )
...
     assert "-ftz=1" in options_str
+    assert "-prec-div=1" in options_str
+    assert "-prec-sqrt=1" in options_str
+    assert "-fma=1" in options_str
cuda_core/cuda/core/_linker.pyx: 1
430-440: 🗄️ Data Integrity & Integration | 🔵 Trivial | ⚡ Quick win

suggestion: Add direct serialization coverage for boolean split-compile inputs. bool is an int subtype, and the repository already uses ProgramOptions(split_compile=True). The current LinkerOptions tests use only integers and do not inspect serialized flags, so reverting int(...) would leave True and False regressions undetected.
+def test_linker_options_split_compile_serializes_booleans():
+    options = LinkerOptions(split_compile=True, split_compile_extended=False)
+    flags = options.as_bytes("nvjitlink")
+    assert b"-split-compile=1" in flags
+    assert b"-split-compile-extended=0" in flags

ℹ️ Review info

⚙️ Run configuration

Configuration used: Repository: NVIDIA/cuda-python/.coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: a3a7523f-7d13-4280-8400-4bda3c7dce70

📥 Commits

Reviewing files that changed from the base of the PR and between c539cc2 and 746df99.

📒 Files selected for processing: 1

cuda_core/tests/test_linker.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.

fix: use nvJitLink-compatible 0/1 values for boolean options (nvbug 6…

0e5afde

…621758)

github-actions Bot added the cuda.core Everything related to the cuda.core module label Oct 16, 2026

coderabbitai Bot reviewed Oct 16, 2026

View reviewed changes

Comment thread cuda_core/cuda/core/_linker.pyx

lijinf2 self-assigned this Oct 16, 2026

lijinf2 added the bug Something isn't working label Oct 16, 2026

lijinf2 added this to the cuda.core 1.3.0 milestone Oct 16, 2026

refactor: drop explanatory comment on split_compile int coercion

c539cc2

leofang requested changes Oct 16, 2026

View reviewed changes

add lto enabled runtime test for ftz/prec_div/prec_sqrt/fma

746df99

lijinf2 requested a review from leofang October 19, 2026 22:33

coderabbitai Bot reviewed Oct 19, 2026

View reviewed changes

Navigation Menu

Repository navigation

fix(cuda_core): use nvJitLink-compatible 0/1 values for boolean linker options - #3128

fix(cuda_core): use nvJitLink-compatible 0/1 values for boolean linker options#3128
lijinf2 wants to merge 3 commits into
NVIDIA:mainfrom
lijinf2:fix/6621758-nvjitlink-bool-to-int-options

lijinf2 commented Oct 16, 2026 •

edited

Loading

Uh oh!

copy-pr-bot Bot commented Oct 16, 2026

Uh oh!

coderabbitai Bot commented Oct 16, 2026 •

edited

Loading

Summary by CodeRabbit

Uh oh!

coderabbitai Bot left a comment

Uh oh!

Uh oh!

lijinf2 commented Oct 16, 2026

Uh oh!

github-actions Bot commented Oct 16, 2026

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

Uh oh!

leofang left a comment

Uh oh!

lijinf2 commented Oct 16, 2026

Uh oh!

lijinf2 commented Oct 19, 2026

Uh oh!

lijinf2 commented Oct 19, 2026

Uh oh!

coderabbitai Bot left a comment

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

lijinf2 commented Oct 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Backward compatibility

Testing

Uh oh!

copy-pr-bot Bot commented Oct 16, 2026

Uh oh!

coderabbitai Bot commented Oct 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

lijinf2 commented Oct 16, 2026

Uh oh!

github-actions Bot commented Oct 16, 2026

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

Uh oh!

leofang left a comment

Choose a reason for hiding this comment

Uh oh!

lijinf2 commented Oct 16, 2026

Uh oh!

lijinf2 commented Oct 19, 2026

Uh oh!

lijinf2 commented Oct 19, 2026

Uh oh!

coderabbitai Bot left a comment

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

lijinf2 commented Oct 16, 2026 •

edited

Loading

coderabbitai Bot commented Oct 16, 2026 •

edited

Loading