During the execution of a Monte-Carlo Markov-Chain, I get the following warning :
]/opt/intel/oneapi/intelpython/python3.9/lib/python3.9/site-packages/numba_dpex/config.py:46: UserWarning: dpctl could not find any non-host SYCL device on the system. A non-host SYCL device is required to use numba_dpex.
warnings.warn(msg, UserWarning)
/opt/intel/oneapi/intelpython/python3.9/lib/python3.9/site-packages/numba/core/dispatcher.py:289: UserWarning: Numba extension module 'numba_dpex.numpy_usm_shared' failed to load due to 'ImportError(No non-host SYCL device found to execute kernels.)'.
entrypoints.init_all()
It seems that error message I'm receiving is related to numba_dpex, which is an extension of Numba for running code on heterogeneous devices via SYCL and DPC++.
The first message is warning me that no non-host SYCL device was found on the system. A non-host device is typically an external compute device like a GPU or another hardware accelerator. This means that my system either doesn't have such devices, or they are not accessible or correctly configured.
The second message is informing me that the extension module 'numba_dpex.numpy_usm_shared' could not be loaded due to 'ImportError(No non-host SYCL device found to execute kernels)'. This module appears to require such a device to function correctly.
This could imply that my MCMC code is attempting to utilize a hardware accelerator that isn't present on my system, or that the necessary libraries to access such an accelerator aren't installed or correctly configured.
To address this issue, I might:
1) Check if my system has a compatible non-host SYCL device (like a GPU), and ensure that the appropriate drivers are installed.
2) If I don't have a non-host device or don't want to use one, I might need to modify my code so it doesn't depend on numba_dpex, or configure numba_dpex to not attempt to use a non-host device.
How could I check 1) and 2) on a Debian 11 ? I suspect that recent upgrade of CUDA drivers has put the mess. I have 2 GPU A6000 cards and I don't use explicitely Numba in my MCMC code.
Any clue is welcome