-
Notifications
You must be signed in to change notification settings - Fork 4
Home
hcs edited this page Dec 13, 2024
·
4 revisions
A list of issues you may encounter and ideas for a remedy.
Sometimes it is useful to get the exact version of the code that was compiled and linked into the library. This is possible by setting the environment variable INDEXER_LOG_LEVEL=info and run some code that initializes the library, e.g. running a simple indexing operation. If the code was compiled from a repository clone, you'll see the commit hash
(version) 1.1.1, 03a03032075d9fbd10b8a89431503581bd680246: Fri, 13 Dec 2024 16:10:57 +0100
- cudaErrorInsufficientDriver: Most likely caused by an Nvidia GPU driver that is too old for the CUDA version used by the indexer, see https://docs.nvidia.com/deploy/cuda-compatibility
-
cudaErrorLaunchOutOfResources: Most likely caused by the fact, that there isn't enough GPU memory, for various possible reasons
- When building the library in debug mode, executables can get so large as to exhaust GPU resources. The library should be rebuilt in release mode if that is the case.
- Too many indexer objects are created in parallel. If that is the case, and you have several GPUs, maybe you'll be able to set set the INDEXER_GPU_DEVICE to different values for parallel program runs, or you might be able to set the default GPU device programmatically to different devices before creating indexer objects.
- The parameters for the number of maximum spots or input and output cell numbers might be too high. Consider setting them to lower values.
- cudaErrorUnsupportedPtxVersion: PTX is an intermediate representation produced by the CUDA compiler and transformed to machine code by the Nvidia driver at runtime. The compiler being newer than the driver is the most likely cause of this exception, and the remedy is to update the driver.