Running jdl on Windows 11 with NVIDIA 4070ti. Should I install Ubuntu and WSL? Which cuda version? #2738
-
I have Windows 11 with a NVIDIA 4070ti. Neither TrainBertOnCode nor TrainResnetWithCifar10 runs on the gpu. I get: [INFO ] - Ignore mismatching platform from: jar:file:/D:/.gradle/caches/modules-2/files-2.1/ai.djl.mxnet/mxnet-native-cu112mkl/1.9.1/3e0e468585ed841cdcc93d34908a21f55b515950/mxnet-native-cu112mkl-1.9.1-linux-x86_64.jar!/native/lib/mxnet.properties My system has CUDA 12.1 (/c/Program\ Files/NVIDIA\ GPU\ Computing\ Toolkit/CUDA/v12.1/) and a directory with CUDA 11.6 (/d/ProgramFilesD/cuda-11.6.2/toolkit , as seen with git bash shell). Changing paths in build.gradle with systemProperties, jdl still won't run examples on the gpu. I presume that I need to downgrade CUDA to 11.2 on my OS, but that doesn't seem easy to do. Anyway, would it be easier for me to install Ubuntu using WSL(2)? Can I make the examples use something other than MXNet? Clearer instructions and recommended setups would help. |
Beta Was this translation helpful? Give feedback.
Replies: 3 comments
-
MXNet doesn't support CUDA 12, it can only run on CUDA 11.2. You can try to use PyTorch. PyTorch should be able to run with CUDA 12.1. You can set system properties to use PyTorch: |
Beta Was this translation helpful? Give feedback.
-
It works now. Thanks. I installed CUDA 11.7 alongside several other CUDAs in /c/Program Files/NVIDIA GPU Computing Toolkit/CUDA/ (as seen from git bash).: $ cd /c/Program Files/NVIDIA GPU Computing Toolkit/CUDA And in examples/build.gradle I added: run { CUDA_PATH is set to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7 |
Beta Was this translation helpful? Give feedback.
-
It stopped working but I fixed it: I found that I also have to do
Otherwise, it tries to use version 2.01. Both are in my .djl.ai/pytorch/ folder: $ ls -l BTW, this is on Windows 11, using git bash. |
Beta Was this translation helpful? Give feedback.
MXNet doesn't support CUDA 12, it can only run on CUDA 11.2.
You can try to use PyTorch. PyTorch should be able to run with CUDA 12.1.
You can set system properties to use PyTorch:
System.setProperty("ai.djl.default_engine", "PyTorch")