RuntimeError: Found no NVIDIA driver on your system

Question

I am trying to build a docker image that requires GPU CUDA access. I am using an AWS EC2 p2.xlarge instance that supposedly has the correct hardware. My Dockerfile is building but when running my code, PyTorch can't find CUDA.

print(torch.__version__, torch.cuda.is_available())

returns

1.10.0+cu113 False

This is my Dockerfile

FROM nvidia/cuda:11.2.1-cudnn8-devel-ubuntu18.04

ENV DEBIAN_FRONTEND noninteractive
RUN apt-get update && apt-get install -y \
    python3-opencv ca-certificates python3-dev git wget sudo ninja-build
RUN ln -sv /usr/bin/python3 /usr/bin/python

# Update apt packages
RUN apt update
RUN apt upgrade -y
RUN apt-get -y install git

# Install pip
RUN apt install python3-pip -y
RUN python -m pip install --upgrade pip

COPY requirements.txt ./
RUN pip install --no-cache-dir -r requirements.txt
RUN pip install opencv-python
RUN pip install --no-cache-dir awscli
RUN pip install opencv-python
RUN pip install wheel
RUN pip install setuptools
RUN pip install pip install nvidia-ml-py3

############ Detectron2 pre-built binaries Pytorch default install ############
RUN pip3 install torch==1.10.0+cu113 torchvision==0.11.1+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html
RUN pip install cython

RUN pip install --user 'git+https://github.com/facebookresearch/fvcore'

RUN pip install 'git+https://github.com/facebookresearch/detectron2.git'

ARG magma-cuda112
RUN if [ -z "$MAGMA_CUDA_VERSION" ] ; then \
    echo "Building with CPU support ..."; \
    else \
    echo "Building with GPU/CUDA support ..."; \
    conda install -y -c pytorch ${MAGMA_CUDA_VERSION} && conda clean -ya; \
    fi

# set FORCE_CUDA because during `docker build` cuda is not accessible
ENV FORCE_CUDA="1"
# This will by default build detectron2 for all common cuda architectures and take a lot more time,
# because inside `docker build`, there is no way to tell which architecture will be used.
ARG TORCH_CUDA_ARCH_LIST="3.7+PTX;5.0;6.0;6.1;7.0;7.5;8.0;8.6"
ENV TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST}"

COPY . .
RUN aws s3 cp s3://bucket/file/model.pth ./output
CMD [ "python", "impaction_detection.py" ]

I am trying to build this image using a docker-compose.yml file.

Answer 1

For Google Collab only: Just click on RAM/Disk(this will appear once your notebook is connected) button on the notebook and change the runtime type to t4gpu, this will work and your epoch will run way faster.