Troubleshooting

General Usage

• Problem: FP64 data type is unsupported on current platform.

  • Cause: FP64 is not natively supported by the Intel® Arc™ A-Series Graphics platforms. If you run any AI workload on that platform and receive this error message, it means a kernel requires FP64 instructions that are not supported and the execution is stopped.

• Problem: Runtime error invalid device pointer if import horovod.torch as hvd before import intel_extension_for_pytorch.

  • Cause: Intel® Optimization for Horovod* uses utilities provided by Intel® Extension for PyTorch*. The improper import order causes Intel® Extension for PyTorch* to be unloaded before Intel® Optimization for Horovod* at the end of the execution and triggers this error.

  • Solution: Do import intel_extension_for_pytorch before import horovod.torch as hvd.

• Problem: Number of dpcpp devices should be greater than zero.

  • Cause: If you use Intel® Extension for PyTorch* in a conda environment, you might encounter this error. Conda also ships the libstdc++.so dynamic library file that may conflict with the one shipped in the OS.

  • Solution: Export the libstdc++.so file path in the OS to an environment variable LD_PRELOAD.

• Problem: -997 runtime error when running some AI models on Intel® Arc™ Graphics family.

  • Cause: Some of the -997 runtime error are actually out-of-memory errors. As Intel® Arc™ Graphics GPUs have less device memory than Intel® Data Center GPU Flex Series 170 and Intel® Data Center GPU Max Series, running some AI models on them may trigger out-of-memory errors and cause them to report failure such as -997 runtime error most likely. This is expected. Memory usage optimization is working in progress to allow Intel® Arc™ Graphics GPUs to support more AI models.

• Problem: Building from source for Intel® Arc™ A-Series GPUs fails on WSL2 without any error thrown.

  • Cause: Your system probably does not have enough RAM, so Linux kernel’s Out-of-memory killer was invoked. You can verify this by running dmesg on bash (WSL2 terminal).

  • Solution: If the OOM killer had indeed killed the build process, then you can try increasing the swap-size of WSL2, and/or decreasing the number of parallel build jobs with the environment variable MAX_JOBS (by default, it’s equal to the number of logical CPU cores. So, setting MAX_JOBS to 1 is a very conservative approach that would slow things down a lot).

• Problem: Some workloads terminate with an error CL_DEVICE_NOT_FOUND after some time on WSL2.

  • Cause: This issue is due to the TDR feature on Windows.

  • Solution: Try increasing TDRDelay in your Windows Registry to a large value, such as 20 (it is 2 seconds, by default), and reboot.

Library Dependencies

• Problem: Cannot find oneMKL library when building Intel® Extension for PyTorch* without oneMKL.

  /usr/bin/ld: cannot find -lmkl_sycl
  /usr/bin/ld: cannot find -lmkl_intel_ilp64
  /usr/bin/ld: cannot find -lmkl_core
  /usr/bin/ld: cannot find -lmkl_tbb_thread
  dpcpp: error: linker command failed with exit code 1 (use -v to see invocation)
  

  • Cause: When PyTorch* is built with oneMKL library and Intel® Extension for PyTorch* is built without MKL library, this linker issue may occur.

  • Solution: Resolve the issue by setting:

    export USE_ONEMKL=OFF
    export MKL_DPCPP_ROOT=${HOME}/intel/oneapi/mkl/latest
    

  Then clean build Intel® Extension for PyTorch*.

• Problem: Undefined symbol: mkl_lapack_dspevd. Intel MKL FATAL ERROR: cannot load libmkl_vml_avx512.so.2 or `libmkl_vml_def.so.2.

  • Cause: This issue may occur when Intel® Extension for PyTorch* is built with oneMKL library and PyTorch* is not build with any MKL library. The oneMKL kernel may run into CPU backend incorrectly and trigger this issue.

  • Solution: Resolve the issue by installing the oneMKL library from conda:

    conda install mkl
    conda install mkl-include
    

  Then clean build PyTorch*.

• Problem: OSError: libmkl_intel_lp64.so.2: cannot open shared object file: No such file or directory.

  • Cause: Wrong MKL library is used when multiple MKL libraries exist in system.

  • Solution: Preload oneMKL by:

    export LD_PRELOAD=${MKL_DPCPP_ROOT}/lib/intel64/libmkl_intel_lp64.so.2:${MKL_DPCPP_ROOT}/lib/intel64/libmkl_intel_ilp64.so.2:${MKL_DPCPP_ROOT}/lib/intel64/libmkl_gnu_thread.so.2:${MKL_DPCPP_ROOT}/lib/intel64/libmkl_core.so.2:${MKL_DPCPP_ROOT}/lib/intel64/libmkl_sycl.so.2
    

    If you continue seeing similar issues for other shared object files, add the corresponding files under ${MKL_DPCPP_ROOT}/lib/intel64/ by LD_PRELOAD. Note that the suffix of the libraries may change (e.g. from .1 to .2), if more than one oneMKL library is installed on the system.

• Problem: If you encounter issues related to MPI environment variable configuration when running distributed tasks.

  • Cause: MPI environment variable configuration not correct.

  • Solution: conda deactivate and then conda activate to activate the correct MPI environment variable automatically.

    conda deactivate
    conda activate
    

• Problem: If you encounter issues Runtime error related to C++ compiler with torch.compile. Runtime Error: Failed to find C++ compiler. Please specify via CXX environment variable.

  • Cause: Not activate C++ compiler. torch.compile need to find correct cl.exe path.

  • Solution: One could open “Developer Command Prompt for VS 2022” or follow Visual Studio Developer Command Prompt and Developer PowerShell to activate visual studio environment.

• Problem: LoweringException: ImportError: cannot import name ‘intel’ from ‘triton._C.libtriton’

  • Cause: Installing Triton causes pytorch-triton-xpu to stop working.

  • Solution: Resolve the issue with following command:

    pip list | grep triton
    # If triton related packages are listed, remove them
    pip uninstall triton
    pip uninstall pytorch-triton-xpu
    # Reinstall correct version of pytorch-triton-xpu
    pip install pytorch-triton-xpu==3.3.0 --index-url  https://download.pytorch.org/whl/xpu
    

• Problem: RuntimeError: oneCCL: ze_handle_manager.cpp:226 get_ptr: EXCEPTION: unknown memory type, when executing DLRMv2 BF16 training on 4 cards Intel® Data Center GPU Max platform.

  • Cause: Issue exists in the default sycl path of oneCCL 2021.14 which uses two IPC exchanges.

  • Solution: Use export CCL_ATL_TRANSPORT=ofi to work around.

• Problem: Segmentation fault, when executing LLaMa2-70B inference on Intel® Data Center GPU Max platform, base on online quantization.

  • Cause: Issue exists Intel Neural Compressor (INC) v3.3: during the initial import of INC, the accelerator is cached with lru_cache. Subsequently, setting INC_TARGET_DEVICE in INC transformers-like API does not take effect. This results in two devices being present in the model, leading to memory-related errors as seen in the error messages.

  • Solution: Run the workload INC_TARGET_DEVICE="cpu" python to work around, if using online quantization.

Performance Issue

• Problem: Extended durations for data transfers from the host system to the device (H2D) and from the device back to the host system (D2H).

  • Cause: Absence of certain Dynamic Kernel Module Support (DKMS) packages on Ubuntu 22.04 or earlier versions.

  • Solution: For those running Ubuntu 22.04 or below, it’s crucial to follow all the recommended installation procedures, including those labeled as optional. These steps are likely necessary to install the missing DKMS packages and ensure your system is functioning optimally. The Kernel Mode Driver (KMD) package that addresses this issue has been integrated into the Linux kernel for Ubuntu 23.04 and subsequent releases.