Advance nightly docker to 11.6 (#87858) Fixes following: https://github.com/pytorch/pytorch/actions/runs/3242695506/jobs/5316334351 crash in Docker builds introduced by: #82682 The PR seems to introduce some changes not compatible with cuda 11.3 which is used by our Docker builds This is a reland of original pr: https://github.com/pytorch/pytorch/pull/86941 (Created this new PR to start fresh) Which was reverted because conda install, installed wrong version of pytorch. It installed pytorch for cuda 11.3 still rather then 11.6 This should be fixed now with Release 1.13 Pull Request resolved: https://github.com/pytorch/pytorch/pull/87858 Approved by: https://github.com/seemethere, https://github.com/malfet, https://github.com/izaitsevfb
diff --git a/Dockerfile b/Dockerfile index 815a910..e49e0a4 100644 --- a/Dockerfile +++ b/Dockerfile
@@ -59,17 +59,18 @@ FROM conda as conda-installs ARG PYTHON_VERSION=3.8 -ARG CUDA_VERSION=11.3 +ARG CUDA_VERSION=11.6 ARG CUDA_CHANNEL=nvidia ARG INSTALL_CHANNEL=pytorch-nightly -ENV CONDA_OVERRIDE_CUDA=${CUDA_VERSION} # Automatically set by buildx +RUN /opt/conda/bin/conda update -y conda RUN /opt/conda/bin/conda install -c "${INSTALL_CHANNEL}" -y python=${PYTHON_VERSION} ARG TARGETPLATFORM + # On arm64 we can only install wheel packages RUN case ${TARGETPLATFORM} in \ "linux/arm64") pip install --extra-index-url https://download.pytorch.org/whl/cpu/ torch torchvision torchtext ;; \ - *) /opt/conda/bin/conda install -c "${INSTALL_CHANNEL}" -c "${CUDA_CHANNEL}" -y "python=${PYTHON_VERSION}" pytorch torchvision torchtext "cudatoolkit=${CUDA_VERSION}" ;; \ + *) /opt/conda/bin/conda install -c "${INSTALL_CHANNEL}" -c "${CUDA_CHANNEL}" -y "python=${PYTHON_VERSION}" pytorch torchvision torchtext "pytorch-cuda=$(echo $CUDA_VERSION | cut -d'.' -f 1-2)" ;; \ esac && \ /opt/conda/bin/conda clean -ya RUN /opt/conda/bin/pip install torchelastic 