Repository navigation
Add pathfinder support for cupti64_2026.1.2.dll (CTK 13.2.2) - #3279
Conversation
nvidia-cuda-cupti 13.2.86 (CUDA 13.2.2) ships cupti64_2026.1.2.dll on
Windows, which the cupti descriptor does not list, so
load_nvidia_dynamic_lib("cupti") fails on a machine that only has the
13.2 wheels installed.
Found via the PyTorch cu132 Windows nightly, which loads the CUDA wheel
libraries through cuda-pathfinder at import time
(pytorch/pytorch#202206).
Signed-off-by: Klaus Zimmermann <klaus.zimmermann@quansight.com>
|
No actionable comments were generated in the recent review.
|
| Layer / File(s) | Summary |
|---|---|
Windows DLL candidate list cuda_pathfinder/cuda/pathfinder/_dynamic_libs/descriptor_catalog.py, cuda_pathfinder/docs/nv-versions.json, cuda_pathfinder/docs/source/release/1.8.5-notes.rst |
The CUPTI Windows DLL preference list adds cupti64_2026.1.2.dll between the 2026.2.0 and 2026.1.1 entries. The version list adds 1.8.5, and the release notes describe the DLL addition. |
Possibly related issues
- [BUG]: Windows CTK 13.2.2 CUPTI Name Unsupported #3281: Reports that
load_nvidia_dynamic_libcannot loadcupti64_2026.1.2.dllbecause the descriptor lacks its name. This change adds that DLL name.
Possibly related PRs
- NVIDIA/cuda-python#3233: Adds lookup APIs that use the descriptor catalog, which now includes the new DLL candidate.
Suggested labels: enhancement
Suggested reviewers: rwgk
Priority:
Change: Feature
Merge Risk: 
Minimal · up to ed8b3
-
No actionable merge risk remains.
Comment @coderabbitai help to get the list of available commands.
| "cupti64_2026.3.0.dll", | ||
| "cupti64_2026.2.1.dll", | ||
| "cupti64_2026.2.0.dll", | ||
| "cupti64_2026.1.2.dll", |
There was a problem hiding this comment.
Please add cupti discovery unit tests to the prei-existing set. Thank you.
There was a problem hiding this comment.
We have that, but this particular version escaped our testing:
In theory, we'd have to test against all released CTK version, every major.minor release.
Currently we're only testing against the latest major.minor releases.
It's a complex project to ensure full coverage at a reasonable cost (runner time).
There was a problem hiding this comment.
This issue will be fixed with the next CTK release onwards. Windows Cupti will follow the same naming convention as linux
| "cupti64_2026.3.0.dll", | ||
| "cupti64_2026.2.1.dll", | ||
| "cupti64_2026.2.0.dll", | ||
| "cupti64_2026.1.2.dll", |
There was a problem hiding this comment.
We have that, but this particular version escaped our testing:
In theory, we'd have to test against all released CTK version, every major.minor release.
Currently we're only testing against the latest major.minor releases.
It's a complex project to ensure full coverage at a reasonable cost (runner time).
Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
|
I just pushed pathfinder 1.8.5 release notes, so that we can make a pathfinder release immediately after this PR merges. |
|
/ok to test ed8b385 |
|
I set this PR to auto-merge. I'll keep an eye on it and will start the release after the PR has merged. End-to-end it'll probably take 3-4 hours to get the release out (2 x full CI is the bottleneck). |
|
Please open this PR as a draft and mark it ready for review once the description and code are ready to be looked at. Before submitting, please review: - [PR lifecycle](https://github.com/pytorch/pytorch/blob/main/CONTRIBUTING.md#pr-lifecycle) in the contributing guide - [AI-Assisted Development](https://github.com/pytorch/pytorch/blob/main/AI_POLICY.md) policy Go through the following checklist - [x] Passes lint (`spin fixlint`) - [ ] Added/updated tests - [ ] Updated documentation (if applicable) - [ ] Included benchmark results (for PRs impacting perf) --- <!-- Delete this line and everything above it once you've gone through the checklist. --> ## Linked issue or supporting maintainer <!-- Keep the one line that applies and delete the others. Link an issue labeled "actionable", or name the maintainer who agreed to this change beforehand. PRs with neither will be automatically closed, unless the author has write access to the repo. --> Fixes #202565 ## BC-breaking? No ## Summary (human written only) Implements approach 1 from #202565 - Add try/except for every `load_nvidia_dynamic_lib(lib)` call - Pin cuda-pathfinder to 1.8.5 for this fix NVIDIA/cuda-python#3279 ## Agent details (if applicable) <!-- You can add here the details from your agent related to this PR. --> Pull Request resolved: #202589 Approved by: https://github.com/zklaus, https://github.com/atalman
Description
nvidia-cuda-cupti 13.2.86 (CUDA 13.2.2) ships cupti64_2026.1.2.dll on Windows, which the cupti descriptor does not list, so load_nvidia_dynamic_lib("cupti") fails on a machine that only has the 13.2 wheels installed.
Found via the PyTorch cu132 Windows nightly, which loads the CUDA wheel libraries through cuda-pathfinder at import time (pytorch/pytorch#202206).
Checklist
New or existing tests cover these changes.