Frequently Asked Questions

I. How to check whether GPU drivers are installed successfully?

Run import tensorflow and it will show which platform you are running on: Intel(R) Level-Zero(default) or Intel(R) OpenCL.

And the high level API of TensorFlow tf.config.experimental.list_physical_devices() will tell you the device types that are registered to TensorFlow core.

$ python
>>> import tensorflow as tf
2021-07-01 06:40:55.510076: I itex/core/devices/gpu/dpcpp_runtime.cc:116] Selected platform: Intel(R) Level-Zero.
>>> tf.config.experimental.list_physical_devices()
[PhysicalDevice(name='/physical_device:CPU:0', device_type='CPU'), PhysicalDevice(name='/physical_device:XPU:0', device_type='XPU')]

II. How to know the configurations and rate of utilization of local GPU devices?

System Monitoring Utility tool can be used to show the capability (clock frequency, EU count, amount of device memory and so on) of your devices and usage of each sub-module (device memory, GPU engines and so on).

Troubleshooting

This section shows common problems and solutions for compilation and runtime issues you may encounter.

Build from source

Error	Solution	Comments
external/onednn/src/sycl/level_zero_utils.cpp:33:10: fatal error: 'level_zero/ze_api.h' file not found #include ^~~~~~~~~~~~~~~~~~~~~	install level-zero-dev lib	level-zero-dev lib is needed when building from source

Runtime

Error	Solution	Comments
ModuleNotFoundError: No module named 'tensorflow'	install TensorFlow	ITEX depends on TensorFlow
tensorflow.python.framework.errors_impl.NotFoundError: libmkl_sycl.so.2: cannot open shared object file: No such file or directory	source /opt/intel/oneapi/setvars.sh	set env vars of oneAPI Base Toolkit
INTEL MKL ERROR: /opt/intel/oneapi/mkl/latest/lib/intel64/libmkl_avx512.so.2: undefined symbol: mkl_sparse_optimize_bsr_trsm_i8. Intel MKL FATAL ERROR: Cannot load libmkl_avx512.so.2 or libmkl_def.so.2.	export LD_PRELOAD=/opt/intel/oneapi/mkl/latest/lib/intel64/libmkl_core.so:/opt/intel/oneapi/mkl/latest/lib/intel64/libmkl_tbb_thread.so	It's a known issue and will be fixed in the next release.