forked from Kaggle/docker-python
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathgpu.Dockerfile
More file actions
114 lines (101 loc) · 5.67 KB
/
Copy pathgpu.Dockerfile
File metadata and controls
114 lines (101 loc) · 5.67 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
ARG BASE_TAG=staging
FROM nvidia/cuda:10.2-cudnn7-devel-ubuntu18.04 AS nvidia
FROM gcr.io/kaggle-images/python-tensorflow-whl:2.3.1-py37-2 as tensorflow_whl
FROM gcr.io/kaggle-images/python:${BASE_TAG}
ADD clean-layer.sh /tmp/clean-layer.sh
# Cuda support
COPY --from=nvidia /etc/apt/sources.list.d/cuda.list /etc/apt/sources.list.d/
COPY --from=nvidia /etc/apt/sources.list.d/nvidia-ml.list /etc/apt/sources.list.d/
COPY --from=nvidia /etc/apt/trusted.gpg /etc/apt/trusted.gpg.d/cuda.gpg
# See b/142337634#comment28
RUN sed -i 's/deb https:\/\/developer.download.nvidia.com/deb http:\/\/developer.download.nvidia.com/' /etc/apt/sources.list.d/*.list
# Ensure the cuda libraries are compatible with the custom Tensorflow wheels.
# TODO(b/120050292): Use templating to keep in sync or COPY installed binaries from it.
ENV CUDA_MAJOR_VERSION=10
ENV CUDA_MINOR_VERSION=2
ENV CUDA_PATCH_VERSION=89
ENV CUDA_VERSION=$CUDA_MAJOR_VERSION.$CUDA_MINOR_VERSION.$CUDA_PATCH_VERSION
ENV CUDA_PKG_VERSION=$CUDA_MAJOR_VERSION-$CUDA_MINOR_VERSION=$CUDA_VERSION-1
LABEL com.nvidia.volumes.needed="nvidia_driver"
LABEL com.nvidia.cuda.version="${CUDA_VERSION}"
ENV PATH=/usr/local/nvidia/bin:/usr/local/cuda/bin:/opt/bin:${PATH}
# The stub is useful to us both for built-time linking and run-time linking, on CPU-only systems.
# When intended to be used with actual GPUs, make sure to (besides providing access to the host
# CUDA user libraries, either manually or through the use of nvidia-docker) exclude them. One
# convenient way to do so is to obscure its contents by a bind mount:
# docker run .... -v /non-existing-directory:/usr/local/cuda/lib64/stubs:ro ...
ENV LD_LIBRARY_PATH_NO_STUBS="/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH"
ENV LD_LIBRARY_PATH="/usr/local/nvidia/lib64:/usr/local/cuda/lib64:/usr/local/cuda/lib64/stubs:$LD_LIBRARY_PATH"
ENV NVIDIA_VISIBLE_DEVICES=all
ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
ENV NVIDIA_REQUIRE_CUDA="cuda>=$CUDA_MAJOR_VERSION.$CUDA_MINOR_VERSION"
RUN apt-get update && apt-get install -y --no-install-recommends \
cuda-cupti-$CUDA_PKG_VERSION \
cuda-cudart-$CUDA_PKG_VERSION \
cuda-cudart-dev-$CUDA_PKG_VERSION \
cuda-libraries-$CUDA_PKG_VERSION \
cuda-libraries-dev-$CUDA_PKG_VERSION \
cuda-nvml-dev-$CUDA_PKG_VERSION \
cuda-minimal-build-$CUDA_PKG_VERSION \
cuda-command-line-tools-$CUDA_PKG_VERSION \
libcudnn7=7.6.5.32-1+cuda$CUDA_MAJOR_VERSION.$CUDA_MINOR_VERSION \
libcudnn7-dev=7.6.5.32-1+cuda$CUDA_MAJOR_VERSION.$CUDA_MINOR_VERSION \
libnccl2=2.5.6-1+cuda$CUDA_MAJOR_VERSION.$CUDA_MINOR_VERSION \
libnccl-dev=2.5.6-1+cuda$CUDA_MAJOR_VERSION.$CUDA_MINOR_VERSION && \
ln -s /usr/local/cuda-$CUDA_MAJOR_VERSION.$CUDA_MINOR_VERSION /usr/local/cuda && \
ln -s /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/stubs/libcuda.so.1 && \
/tmp/clean-layer.sh
# Install OpenCL & libboost (required by LightGBM GPU version)
RUN apt-get install -y ocl-icd-libopencl1 clinfo libboost-all-dev && \
mkdir -p /etc/OpenCL/vendors && \
echo "libnvidia-opencl.so.1" > /etc/OpenCL/vendors/nvidia.icd && \
/tmp/clean-layer.sh
# When using pip in a conda environment, conda commands should be ran first and then
# the remaining pip commands: https://www.anaconda.com/using-pip-in-a-conda-environment/
# However, because this image is based on the CPU image, this isn't possible but better
# to put them at the top of this file to minize conflicts.
RUN conda remove --force -y pytorch torchvision torchaudio cpuonly && \
conda install "pytorch=1.7" "torchvision=0.8" "torchaudio=0.7" "torchtext=0.8" "cudf=0.16" "cuml=0.16" cudatoolkit=$CUDA_VERSION && \
/tmp/clean-layer.sh
# Install LightGBM with GPU
RUN pip uninstall -y lightgbm && \
cd /usr/local/src && \
git clone --recursive https://github.com/microsoft/LightGBM && \
cd LightGBM && \
git checkout tags/v2.3.1 && \
mkdir build && cd build && \
cmake -DUSE_GPU=1 -DOpenCL_LIBRARY=/usr/local/cuda/lib64/libOpenCL.so -DOpenCL_INCLUDE_DIR=/usr/local/cuda/include/ .. && \
make -j$(nproc) && \
cd /usr/local/src/LightGBM/python-package && \
python setup.py install --precompile && \
mkdir -p /etc/OpenCL/vendors && \
echo "libnvidia-opencl.so.1" > /etc/OpenCL/vendors/nvidia.icd && \
/tmp/clean-layer.sh
# Install JAX
RUN pip install jax==0.2.6 jaxlib==0.1.57+cuda$CUDA_MAJOR_VERSION$CUDA_MINOR_VERSION -f https://storage.googleapis.com/jax-releases/jax_releases.html && \
/tmp/clean-layer.sh
# Reinstall packages with a separate version for GPU support.
COPY --from=tensorflow_whl /tmp/tensorflow_gpu/*.whl /tmp/tensorflow_gpu/
RUN pip uninstall -y tensorflow && \
pip install /tmp/tensorflow_gpu/tensorflow*.whl && \
rm -rf /tmp/tensorflow_gpu && \
pip uninstall -y mxnet && \
# b/126259508 --no-deps prevents numpy from being downgraded.
pip install --no-deps mxnet-cu$CUDA_MAJOR_VERSION$CUDA_MINOR_VERSION && \
/tmp/clean-layer.sh
# Reinstall TensorFlow addons (TFA) with GPU support.
COPY --from=tensorflow_whl /tmp/tfa_gpu/*.whl /tmp/tfa_gpu/
RUN pip install /tmp/tfa_gpu/tensorflow*.whl && \
rm -rf /tmp/tfa_gpu/ && \
/tmp/clean-layer.sh
# Install GPU-only packages
RUN pip install pycuda && \
pip install cupy-cuda$CUDA_MAJOR_VERSION$CUDA_MINOR_VERSION && \
pip install pynvrtc && \
pip install nnabla-ext-cuda$CUDA_MAJOR_VERSION$CUDA_MINOR_VERSION && \
/tmp/clean-layer.sh
# Re-add TensorBoard Jupyter extension patch
# b/139212522 re-enable TensorBoard once solution for slowdown is implemented.
# ADD patches/tensorboard/notebook.py /opt/conda/lib/python3.7/site-packages/tensorboard/notebook.py
# Remove the CUDA stubs.
ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH_NO_STUBS"