Skip to content

[FEA]: Add support for user-opt-in logging in path finder #650

Description

@leofang

Examples:

Activity

  1. added
    enhancementAny code-related improvements
    triageNeeds the team's attention
    on May 20, 2025
  2. changed the title [-]Add support for user-opt-in logging in path finder[/-] [+][FEA]: Add support for user-opt-in logging in path finder[/+] on Mar 4, 2026
  3. added
    featureNew feature or request
    and removed
    enhancementAny code-related improvements
    on Mar 4, 2026
  4. removed this from the cuda.pathfinder backlog milestone on Apr 2, 2026
  5. u7k4rs6 commented on Jun 4, 2026

    @u7k4rs6

    Hi @leofang @rparolin, I'd like to take this one. It's unassigned with no branch or PR, and I have a CUDA dev setup on Ubuntu to test against. Since it was moved out of the pathfinder backlog, could you confirm you'd still accept an external PR before I invest? Assuming yes, here's the approach I'd propose so we can lock the interface up front:

    • Standard-library logging with a dedicated cuda.pathfinder logger, off by default since it sits on the import hot path.
    • Opt-in via an env var, e.g. CUDA_PATHFINDER_LOG_LEVEL mapped to standard logging levels (happy to match whatever naming convention you prefer), plus optionally a small programmatic setter.
    • On each lookup, emit a search trail: candidate paths and their order, env vars consulted (CUDA_HOME, LD_LIBRARY_PATH, conda/pip locations), the resolved path and version actually loaded, fallbacks taken, and on failure the full list of paths tried. This mirrors the CuPy (cupy#9127) and nvmath-python logging you linked.
    • Tests via pytest caplog, most of which run without a GPU.

    Does the interface (logger name, env var, optional API toggle) look right? If so, I'll scope the PR to cuda.pathfinder. Thanks!

  6. rparolin commented on Jul 17, 2026

    @rparolin
    Collaborator

    @rwgk Can you comment on the proposed direction above? Thanks!

  7. u7k4rs6 commented on Aug 5, 2026

    @u7k4rs6

    @rwgk @rparolin friendly ping on the above. Rather than keep the interface question open, I'm happy to just open a draft PR against what I sketched (dedicated cuda.pathfinder logger, CUDA_PATHFINDER_LOG_LEVEL, off by default) so there's something concrete to react to, and adjust naming from review. Planning to start on that this week unless you'd rather settle the interface first.

  8. u7k4rs6 commented on Aug 5, 2026

    @u7k4rs6

    Opened a draft PR implementing the interface above, scoped to load_nvidia_dynamic_lib. Happy to extend to the other four finder families, or change the naming and levels, whatever suits.

  9. added this to the cuda.pathfinder next milestone on Aug 20, 2026
  10. u7k4rs6 commented on Aug 21, 2026

    @u7k4rs6

    @rparolin Thanks for the assign. The draft's ready for a look whenever, and could someone post /ok to test so CI can report on it? Happy to un-draft once it's green, or extend to the other four finder families if you like the shape.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

cuda.bindingsEverything related to the cuda.bindings modulecuda.pathfinderEverything related to the cuda.pathfinder modulefeatureNew feature or request

Projects

  • Status
    In Progress

Relationships

None yet

Development

No branches or pull requests

Issue actions