feat(pathfinder): support "cuda"/"nvml" driver libs and reject unsupported libnames#1602
Merged
rwgk merged 6 commits intoNVIDIA:mainfrom Feb 13, 2026
Merged
Conversation
Contributor
|
Auto-sync is disabled for ready for review pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
cpcloud
force-pushed
the
driver-libs-and-validation
branch
from
182be13 to
058f824
Compare
…orted libnames
Add support for loading NVIDIA driver libraries ("cuda", "nvml") via
load_nvidia_dynamic_lib(). These are part of the display driver (not the
CTK) and use a simplified system-search-only path, skipping site-packages,
conda, CUDA_HOME, and canary probe steps.
Also reject unrecognized libnames with a ValueError instead of silently
falling through to system search, which could produce surprising results
for unsupported libs like "cupti".
Closes NVIDIA#1288, closes NVIDIA#1564.
Co-authored-by: Cursor <cursoragent@cursor.com>
cpcloud
force-pushed
the
driver-libs-and-validation
branch
from
058f824 to
902a7f5
Compare
Contributor
Author
|
/ok to test |
This comment has been minimized.
This comment has been minimized.
Contributor
There was a problem hiding this comment.
Pull request overview
Adds explicit support in cuda.pathfinder for loading NVIDIA driver libraries ("cuda", "nvml") via a system-search-only path, and hard-fails on unknown library names to avoid surprising implicit loads.
Changes:
- Add platform-specific driver library name mappings (
libcuda.so.1/nvcuda.dll,libnvidia-ml.so.1/nvml.dll) to the supported-lib registries. - Introduce a dedicated driver-only loading path (skip wheel/conda/CUDA_HOME/canary) and validate
libnameagainst a platform-aware supported set. - Expand tests to cover driver lib dispatch behavior and unsupported-libname validation.
Reviewed changes
Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
cuda_pathfinder/cuda/pathfinder/_dynamic_libs/supported_nvidia_libs.py |
Adds driver-library mappings and folds them into the platform registries. |
cuda_pathfinder/cuda/pathfinder/_dynamic_libs/load_nvidia_dynamic_lib.py |
Adds supported-name validation and a driver-only load path that bypasses the CTK cascade. |
cuda_pathfinder/tests/test_driver_lib_loading.py |
New targeted unit tests for driver-only behavior and dispatch rules. |
cuda_pathfinder/tests/test_load_nvidia_dynamic_lib.py |
Updates tests to use mocker and adds unsupported-libname validation tests. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
cuda_pathfinder/cuda/pathfinder/_dynamic_libs/load_nvidia_dynamic_lib.py
Show resolved
Hide resolved
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Contributor
Author
|
/ok to test |
rparolin
added
bug
rwgk
reviewed
Feb 13, 2026
Collaborator
rwgk
left a comment
rwgk
left a comment
There was a problem hiding this comment.
Mostly nits, but we need real tests ("all_must_work") or we'll be blind to regressions.
Comment on lines
219
to
220
| # CUDA_HOME). Note the non-standard naming: "cuda" → libcuda.so.1, | ||
| # "nvml" → libnvidia-ml.so.1. |
Collaborator
There was a problem hiding this comment.
Suggested change
| # CUDA_HOME). Note the non-standard naming: "cuda" → libcuda.so.1, | |
| # "nvml" → libnvidia-ml.so.1. | |
| # CUDA_HOME). |
Collaborator
There was a problem hiding this comment.
(seems entirely obvious; avoids duplicating names in comment that may go stale)
Comment on lines
32
to
39
| @pytest.mark.parametrize("libname", ["cuda", "nvml"]) | ||
| def test_driver_only_libnames_contains(libname): | ||
| assert libname in _DRIVER_ONLY_LIBNAMES | ||
|
|
||
|
|
||
| @pytest.mark.parametrize("libname", ["cudart", "nvrtc", "cublas", "nvvm"]) | ||
| def test_driver_only_libnames_excludes_ctk_libs(libname): | ||
| assert libname not in _DRIVER_ONLY_LIBNAMES |
Collaborator
There was a problem hiding this comment.
Best removed: these just seem to exercise in and not in based on made-up lists of specific libnames.
| assert libname not in _DRIVER_ONLY_LIBNAMES | ||
|
|
||
|
|
||
| # --------------------------------------------------------------------------- |
Collaborator
There was a problem hiding this comment.
We need some real tests; mocks are great for hard-to-reach-branch coverage, but other than that, they give a false sense of "oh this works".
We should have tests for real loading, and track them with INFOs, so that we can validate that it actually works by inspecting the test logs (CI, QA).
We should have at least some all_must_work on platforms where we're sure it must work. Without, we'll be blind to regressions.
I realize this is unusual, but that's the nature of a piece of software with the sole purpose of adapting to highly diverse environments that we don't control.
Contributor
Author
There was a problem hiding this comment.
Can we avoid the info logging stuff until we merge #1593? I'd like to get that in so I don't have to redo related changes from that PR.
Contributor
Author
There was a problem hiding this comment.
Ok, nevermind about the INFO logging, I'll just do it here. It's more important to just get this feature in.
| load_nvidia_dynamic_lib("cudart") | ||
|
|
||
|
|
||
| @pytest.mark.parametrize("libname", ["bogus", "not_a_real_lib", "cupti"]) |
Collaborator
There was a problem hiding this comment.
Do we need to include "cupti" here? I'd rather stay clear of mentioning any real names here, if for nothing else but avoiding head scratching when searching the codebase for cupti.
Contributor
Author
There was a problem hiding this comment.
Not really, fixes incoming.
Use only obviously fake names to prevent confusion when searching the codebase for real library names like "cupti". Co-authored-by: Cursor <cursoragent@cursor.com>
The duplicated soname/DLL names in the comment can drift from the dict values below; the dict itself is the source of truth. Co-authored-by: Cursor <cursoragent@cursor.com>
…ibnames These just exercise Python's `in` operator against a hardcoded list and don't provide meaningful coverage. Co-authored-by: Cursor <cursoragent@cursor.com>
Run "cuda" and "nvml" through the actual OS loader in spawned child processes, following the same pattern as test_load_nvidia_dynamic_lib. Results are tracked via INFO summary lines for CI/QA log inspection. Co-authored-by: Cursor <cursoragent@cursor.com>
cpcloud
commented
Feb 13, 2026
|
|
||
|
|
||
| @pytest.mark.parametrize("libname", sorted(_DRIVER_ONLY_LIBNAMES)) | ||
| def test_real_load_driver_lib(info_summary_append, libname): |
Collaborator
|
/ok to test ba23547 |
Collaborator
|
It is working as expected: Using logs_57373437624.zip (1st attempt, two jobs are missing): There are The lists below cover |
This comment has been minimized.
This comment has been minimized.
|
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
3 participants
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
load_nvidia_dynamic_lib()"cuda","nvml"#1288: Add support forload_nvidia_dynamic_lib("cuda")andload_nvidia_dynamic_lib("nvml"). These are NVIDIA driver libraries (not CTK) with non-standard naming (libcuda.so.1/nvcuda.dll,libnvidia-ml.so.1/nvml.dll). They use a simplified system-search-only path, skipping site-packages, conda, CUDA_HOME, and canary probe.ValueErrorinstead of silently falling through to system search. The error message includes the full list of supported names for discoverability.Breaking change
load_nvidia_dynamic_lib()now raisesValueErrorfor unrecognized libnames. Any downstream code passing names like"cupti"that happened to work via system search will break. Adding CUPTI as a supported lib (#1572) before the next release would restore that path through the proper search cascade.Changes
supported_nvidia_libs.py— newSUPPORTED_LINUX_SONAMES_DRIVER/SUPPORTED_WINDOWS_DLLS_DRIVERdictsload_nvidia_dynamic_lib.py—_DRIVER_ONLY_LIBNAMES,_load_driver_lib_no_cache(),_ALL_SUPPORTED_LIBNAMESvalidation, updated docstringtests/test_driver_lib_loading.py— 10 new tests for driver lib flowtests/test_load_nvidia_dynamic_lib.py— 2 new tests for validation, migrated tomockerCloses #1288, closes #1564.
Made with Cursor