vpruthi/doc-exports
1
0
Fork 0
forked from docs/doc-exports
Code Pull Requests Activity
doc-exports/docs/modelarts/umn/modelarts_trouble_0047.html
Lai, Weijian 6aa966a79a ModelArts UMN 24.3.0 version 
Reviewed-by: Pruthi, Vineet <vineet.pruthi@t-systems.com>
Co-authored-by: Lai, Weijian <laiweijian4@huawei.com>
Co-committed-by: Lai, Weijian <laiweijian4@huawei.com>
2024-11-02 09:04:52 +00:00

2.9 KiB
Raw Permalink Blame History

Reinstalled CUDA Version Does Not Match the One in the Target Image

Symptom

An error occurs after the engine version is reinstalled or a new CUDA package is compiled based on the existing image.
1. "RuntimeError: cuda runtime error (11) : invalid argument at /pytorch/aten/src/THC/THCCachingHostAllocator.cpp:278"
2. "libcudart.so.9.0 cannot open shared object file no such file or directory"
3. "Make sure the device specification refers to a valid device. The requested device appears to be a GPU,but CUDA is not enabled"

Possible Causes

The CUDA version of the newly installed package does not match the CUDA version in the image.

Solution

Use the local PyCharm to remotely access notebook for debugging and installation.
  1. Remotely log in to the selected image and run nvcc -V to obtain the CUDA version of the image.
  2. Reinstall Torch. Ensure that the version matches the one obtained in the previous step.

Summary and Suggestions

Before creating a training job, use the ModelArts development environment to debug the training code to maximally eliminate errors in code migration.