No matching cuda flavor for win found: cu112mkl/sm_86.

[eleven-dimension]

Hi, team. I am trying to run TrainMnist in examples on gpu in windows os but get this warning so DJL is running on CPU. Is this the result of a downloading problem? How can I fix it?

Thank you very much!

[eleven-dimension]

After going through the source code, I find only cuda 11.8 is supported :(

[frankfliu]

@eleven-dimension
PyTorch cu118 is compatible with cu12.1. You should be able to run run PyTorch with cud121.

See: #2485

[frankfliu]

@eleven-dimension
My bad. cu-112 is bit too old. Would you please upgrade to 11.8?

It should still work if you set PYTORCH_FLAVOR, but remember there are issues with 112

[eleven-dimension]

My build passes device test but cpu is still used in TrainMnist example. I set breakpoints and it seems that win/cu118mkl/mxnet_86.dll.gz doesn't exist and it can't be found at the url https://publish.djl.ai/mxnet-1.9.1/win/cu118mkl/mxnet_86.dll.gz as I expected. Where can it be downloaded?

[frankfliu]

Can set default engine to PyTorch to avoid using MXNet?

[eleven-dimension]

Okay, thanks for your time!