The macOS host tools provided are:Nsight Systems - a system profiler and timeline trace tool supporting Pascal and newer GPUsNsight Compute - a CUDA kernel profiler supporting Volta and new GPUsVisual Profiler - a CUDA kernel and system profiler and timeline trace tool supporting older GPUs (see installation instructions, below)cuda-gdb - a GPU and CPU CUDA application debugger (see installation instructions, below)
If you get this message it may be because your GPU is of CUDA compatibility 3.0 (eg. nVidia 750m). Contrary to what appears in the warning, CUDA 3.0 is supported. We can remove these warnings by going to /usr/local/lib/python3.7/site-packages/torch/cuda/__init__.py and commenting out lines 118-119. The location and line numbers can vary but the UserWarning raised indicated the file path and line. The commented lines in my distribution of pytorch (1.1.0) are as follows:
Use vl_compilenn with the cudnnEnable,true option to compile thelibrary; do not forget to use cudaMethod,nvcc as, at it is likely,the CUDA toolkit version is newer than MATLAB's CUDA toolkit. Forexample, on macOS this may look like:
If you are using a shared system, ask your system administrator on how to install or load the NVIDIA driver. Generally, you should be able to find and use the CUDA driver library, called libcuda.so on Linux, libcuda.dylib on macOS and nvcuda64.dll on Windows. You should also be able to execute the nvidia-smi command, which lists all available GPUs you have access to.
Finally, to be able to use all of the Julia GPU stack you need to have permission to profile GPU code. On Linux, that means loading the nvidia kernel module with the NVreg_RestrictProfilingToAdminUsers=0 option configured (e.g., in /etc/modprobe.d). Refer to the following document for more information.
Note: You cannot pass compute_XX as an argument to --cuda-gpu-arch;only sm_XX is currently supported. However, clang always includes PTX inits binaries, so e.g. a binary compiled with --cuda-gpu-arch=sm_30 would beforwards-compatible with e.g. sm_35 GPUs.
Due to the different ways that CUDA support is enabled by project authors, thereis no universal way to detect GPU support in a package. For many GPU-enabledpackages, there is a dependency on the cudatoolkit package. Other packagessuch as Numba do not have a cudatoolkit dependency, because they can be usedwithout the GPU.
GPU-enabled packages are built against a specific version of CUDA. Currentlysupported versions include CUDA 8, 9.0 and 9.2. The NVIDIA drivers are designedto be backward compatible to older CUDA versions, so a system with NVIDIA driverversion 384.81 can support CUDA 9.0 packages and earlier. As a result, if a useris not using the latest NVIDIA driver, they may need to manually pick aparticular CUDA version by selecting the version of the cudatoolkit condapackage in their environment. To select a cudatoolkit version, add aselector such as cudatoolkit=8.0 to the version specification.
Here you will learn how to check NVIDIA CUDA version in 3 ways: nvcc from CUDA toolkit, nvidia-smi from NVIDIA driver, and simply checking a file. Using one of these methods, you will be able to see the CUDA version regardless the software you are using, such as PyTorch, TensorFlow, conda (Miniconda/Anaconda) or inside docker.
If you have installed the cuda-toolkit software either from the official Ubuntu repositories via sudo apt install nvidia-cuda-toolkit, or by downloading and installing it manually from the official NVIDIA website, you will have nvcc in your path (try echo $PATH) and its location will be /usr/bin/nvcc (by running which nvcc).
The second way to check CUDA version is to run nvidia-smi, which comes from downloading the NVIDIA driver, specifically the NVIDIA-utils package. You can install either Nvidia driver from the official repositories of Ubuntu, or from the NVIDIA website.
Importantly, except for CUDA version. There are more details in the nvidia-smi output, driver version (440.100), GPU name, GPU fan percentage, power consumption/capability, memory usage, can also be found here. You can also find the processes which use the GPU at the moment. This is helpful if you want to see if your model or system is using GPU such as PyTorch or TensorFlow.
In the AMUSE root directory a self-help script can be found. If building or testing any of thecodes mentioned above fails and you wonder why, it will hopefully provide you with helpful suggestions.From a command line run the bash script cuda_self_help:
This will install the pytorch build with the latest cudatoolkit version. If you need a higher or lower CUDA XX build (e.g. CUDA 9.0), following the instructions here, to install the desired pytorch build.
By default pip will install the latest pytorch with the latest cudatoolkit. If your hardware doesn't support the latest cudatoolkit, follow the instructions here, to install a pytorch build that fits your hardware.
The only requirement is that you have installed and configured the NVIDIA driver correctly. Usually you can test that by running nvidia-smi. While it's possible that this application is not available on your system, it's very likely that if it doesn't work, than your don't have your NVIDIA drivers configured properly. And remember that a reboot is always required after installing NVIDIA drivers. 2b1af7f3a8