CUDA kernel failed : no kernel image is available for execution on the device, Error when running PyTorch model inside Google Compute VM

Question

I have a docker image of a PyTorch model that returns this error when run inside a google compute engine VM running on debian/Tesla P4 GPU/google deep learning image:

CUDA kernel failed : no kernel image is available for execution on the device

This occurs on the line where my model is called. The PyTorch model includes custom c++ extensions, I'm using this model https://github.com/daveredrum/Pointnet2.ScanNet

My image installs these at runtime

The image runs fine on my local system. Both VM and my system have these versions:

Cuda compilation tools 10.1, V10.1.243

torch 1.4.0

torchvision 0.5.0

The main difference is the GPU as far as I'm aware

Local:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 435.21       Driver Version: 435.21       CUDA Version: 10.1     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 960M    Off  | 00000000:01:00.0 Off |                  N/A |
| N/A   36C    P8    N/A /  N/A |    361MiB /  2004MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

VM:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.87.01    Driver Version: 418.87.01    CUDA Version: 10.1     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla P4            Off  | 00000000:00:04.0 Off |                    0 |
| N/A   42C    P0    23W /  75W |      0MiB /  7611MiB |      3%      Default |

If I ssh into the VM torch.cuda.is_available() returns true

Therefore I suspect it must be something to do with the compilation of the extensions

This is the relevant part of my docker file:

ENV CUDA_HOME "/usr/local/cuda-10.1"
ENV PATH /usr/local/nvidia/bin:/usr/local/cuda-10.1/bin:${PATH}
ENV NVIDIA_VISIBLE_DEVICES all
ENV NVIDIA_DRIVER_CAPABILITIES compute,utility
ENV NVIDIA_REQUIRE_CUDA "cuda>=10.1 brand=tesla,driver>=384,driver<385 brand=tesla,driver>=396,driver<397 brand=tesla,driver>=410,driver<411 brand=tesla,driver>=418,driver<419"
ENV FORCE_CUDA=1

# CUDA 10.1-specific steps
RUN conda install -c open3d-admin open3d
RUN conda install -y -c pytorch \
    cudatoolkit=10.1 \
    "pytorch=1.4.0=py3.6_cuda10.1.243_cudnn7.6.3_0" \
    "torchvision=0.5.0=py36_cu101" \
 && conda clean -ya
RUN pip install -r requirements.txt
RUN pip install flask
RUN pip install plyfile
RUN pip install scipy


# Install OpenCV3 Python bindings
RUN sudo apt-get update && sudo apt-get install -y --no-install-recommends \
    libgtk2.0-0 \
    libcanberra-gtk-module \
    libgl1-mesa-glx \
 && sudo rm -rf /var/lib/apt/lists/*

RUN dir
RUN cd pointnet2 && python setup.py install
RUN cd ..

I have already re-running this line from ssh in the VM:

TORCH_CUDA_ARCH_LIST="6.0 6.1 7.0" python setup.py install

Which I think targets the installation to the Tesla P4 compute capability?

Is there some other setting or troubleshooting step I can try?

I didn't know anything about docker/VMs/pytorch extensions until a couple of days ago, so somewhat shooting in the dark. Also this is my first stackoverflow post, apologies if I'm not following some etiquette, feel free to point out.

Solution

I resolved this in the end by manually deleting all the folders except for "src" in the folder containing setup.py

Then rebuilt the docker image

Then when building the image I ran TORCH_CUDA_ARCH_LIST="6.1" python setup.py install, to install the cuda extensions targeting the correct compute capability for the GPU on the VM

and it worked!

I guess just running setup.py without deleting the folders previously installed doesn't fully overwrite the extension

Comments

[Alexandre]

I would need to know how you are running docker, is it on a single instance, or a cluster? Did you get the Image from the Docker repository? I also need to know if you got your CUDA driver from the NVIDIA dev site? Or was the driver included in the image?