[FEA]: Add pathfinder cudla support (.so, .h)#1855
Open
rwgk wants to merge 9 commits intoNVIDIA:mainfrom
Open
Conversation
Add pathfinder support for loading ``libcudla.so.1`` from the ``nvidia-cudla`` package and probing ``libnvcudla.so`` through the existing canary subprocess path. Use that probe in the cudla load test so hosts without the platform runtime are skipped, while real ``libcudla.so.1`` load failures still surface when ``libnvcudla.so`` is available. Made-with: Cursor
Mark cudla and nvcudla as aarch64-only descriptors and derive the supported library tables from the current machine as well as the current OS. This keeps those libraries known to pathfinder while reporting them as unavailable on linux-64, and updates the descriptor-registry tests to match the new current-platform filtering model. Made-with: Cursor
Contributor
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
Skip the cudla and nvcudla load tests on aarch64 hosts when the nvcudla canary probe cannot resolve libnvcudla.so. This keeps non-Tegra linux-aarch64 systems from failing strict test runs while still exercising the real success path on Tegra platforms where the platform runtime is installed. Made-with: Cursor
Remove the machine-architecture gating for cudla and nvcudla so they remain part of the normal Linux descriptor tables. Let the nvcudla canary probe decide whether cudla and nvcudla tests should run, which keeps strict test runs green on hosts without the platform runtime while still exercising real load behavior where libnvcudla.so is available. Made-with: Cursor
Move the libnvcudla.so skip logic into conftest so cudla and nvcudla tests use one shared rule. Keeping the helper in the pytest support layer avoids duplicate test code while still deferring the pathfinder import until the helper runs. Made-with: Cursor
Register cudla as a CTK header so locate_nvidia_header_directory() can find cudla.h in the standard cu13 wheel include directory. In strict header tests, skip cudla on hosts where libnvcudla.so is not available so Tegra setups still exercise the real path without making unsupported hosts fail. Made-with: Cursor
Move cudla into the CTK descriptor block so its packaging classification matches how it is shipped in toolkit installs and the optional nvidia-cudla wheel. This keeps the catalog organization consistent with the current understanding of cudla as a CUDA Toolkit component rather than a third-party add-on. Made-with: Cursor
Collaborator
Author
|
/ok to test |
|
rwgk
changed the title
pathfinder.load_nvidia_dynamic_lib("cudla")
rwgk
changed the title
rwgk
added
P0
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
1 participant
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.
Closes #1857
cuda.pathfindersupport for the cuDLA shared librariescudlaandnvcudla.load_nvidia_dynamic_lib("cudla")to locatelibcudla.so.1from either CUDA Toolkit installs or thenvidia-cudlawheel, and addnvcudlaas the system-search canary/runtime dependency.locate_nvidia_header_directory("cudla")support forcudla.hfrom from either CUDA Toolkit installs or the standardnvidia/cu13/includewheel layout."all_must_work"tests green on hosts without the platform runtime by skipping thecudla/nvcudlachecks whenlibnvcudla.sois not loadable, while still exercising the real success path on Tegra systems.Manual Testing (in addition to CI testing) passed on Tegra Thor (CTK 13.2) and Tegra Orin (unreleased CTK):
WithOUT
nvidia-cudlawheel:With
nvidia-cudlawheel: