-
Notifications
You must be signed in to change notification settings - Fork 679
Replies: 1 comment · 4 replies
-
Can you enable debug level, it will print more information regard engine loading? Anther way to debug is add the following code at beginning of your code:
|
Beta Was this translation helpful? Give feedback.
All reactions
-
Below is the output from the remote environment when I enable debugging: Click to expand!
In contrast, this is the output from my local development machine: Click to expand!
Afterwards, I tried manually transferring the contents of
Please let me know if I can provide any further details. |
Beta Was this translation helpful? Give feedback.
All reactions
-
I think the different is: LD_LIBRARY_PATH: /usr/local/cuda-10.2/lib64 DJL use libcudart.so to detect GPU, which assume JVM can load libcudart.so from either LD_LIBRARY_PATH or java.library.path. |
Beta Was this translation helpful? Give feedback.
All reactions
-
👍 1
-
Can you also show me what in your jar file, I suspect you have multiple mxnet.properties in your fatjar file. |
Beta Was this translation helpful? Give feedback.
All reactions
-
Aha! That was it. I did the following to find
Finally, DJL was able to detect the GPU:
Maybe it is be possible to automatically resolve the |
Beta Was this translation helpful? Give feedback.
-
Since the deployment environment does not support internet connectivity, I built a fat JAR using the following deps:
It uses the GPU on my local development machine, but after deploying, DJL falls back to using the CPU:
Same version of CUDA on both machines. Any ideas what I might be doing wrong?
Beta Was this translation helpful? Give feedback.
All reactions