Repository navigation
Support the driver error logging system #671
Description
xref: https://docs.nvidia.com/cuda/cuda-c-programming-guide/#error-log-management
Current thinking:
- Prioritize for in-memory logging access (the file-based access can already be turned on easily by setting
CUDA_LOG_FILEso only a very sophisticated logging API on the Python side would add value) - Provide a hook to a custom Python logger
- Allow users to opt in via an API in
cuda.core.utils, say,cuda.core.utils.{enable,disable}_logging(). This is a global logging at the driver level, so scoped logging (ex: context manager) does not seem to make much sense.
Activity
lijinf2 commentedon Sep 8, 2026
Updated priority to P0 to track progress. @kkraus14 and @leofang
Relevant to Python logging (info, debug, warning, error, critical, etc).
The first step needs a design doc to describe what it should look like in the context of cuda python.
8 remaining items
refactor(cuda.core): rename register_cuda_log_bridge to register_cuda…
refactor(cuda.core): trim comments in log bridge and its tests
refactor(cuda.core): rename LogCallback to PyLogCallback
refactor(cuda.core): drop redundant length param from PyLogCallback
refactor(cuda.core): drop redundant user_data param from PyLogCallback
test(cuda.core): translate comments to English, tighten assertions
feat(cuda.core): bridge CUDA driver logs into Python logging
test(cuda.core): cover rebind, level filtering, and alert counting
feat(cuda.core): add callback param to register_cuda_log_bridge
refactor(cuda.core): rename register_cuda_log_bridge to register_cuda…
refactor(cuda.core): trim comments in log bridge and its tests
refactor(cuda.core): rename LogCallback to PyLogCallback
refactor(cuda.core): drop redundant length param from PyLogCallback
refactor(cuda.core): drop redundant user_data param from PyLogCallback
test(cuda.core): translate comments to English, tighten assertions