mirror of
https://github.com/netbirdio/gvisor.git
synced 2026-05-22 17:12:49 -07:00
Etienne PerotandgVisor bot
8516598640
The PyTorch benchmarks were removed in https://github.com/google/gvisor/commit/9304ed401fd9604bc36c0436a1132a8b4e3f0851#diff-fd8d6db82d75e1038ed6136c9930c17d6985ff5d22f2ed9e5e8910661de14228 but the Kubernetes PyTorch Kubernetes benchmarks actually depended on them. PiperOrigin-RevId: 706569220
91 lines
3.1 KiB
Docker
91 lines
3.1 KiB
Docker
FROM nvidia/cuda:12.4.0-devel-ubuntu22.04
|
|
|
|
# Used for determining the correct pip index URL below.
|
|
ENV CUDA_VERSION=12.4
|
|
|
|
ENV PYTORCH_DATASETS_DIR=/pytorch-data
|
|
ENV TORCH_HOME=/pytorch-home
|
|
RUN mkdir -p "$TORCH_HOME" && \
|
|
mkdir -p "$PYTORCH_DATASETS_DIR"
|
|
|
|
RUN apt-get update && \
|
|
apt-get install --yes \
|
|
libgl1-mesa-glx libglib2.0-0 \
|
|
pkg-config \
|
|
python3 \
|
|
python3-distutils \
|
|
python3-pip \
|
|
clang \
|
|
wget \
|
|
vim \
|
|
git
|
|
|
|
RUN PIP_INDEX_URL="https://download.pytorch.org/whl/cu$(echo "$CUDA_VERSION" | sed 's~\.~~g')" && \
|
|
python3 -m pip install --ignore-installed \
|
|
boto3 \
|
|
"clang~=$(clang --version | grep -oP 'clang version [.0-9]+' | cut -d' ' -f3)" \
|
|
lightning \
|
|
matplotlib \
|
|
memory_profiler \
|
|
numba && \
|
|
python3 -m pip install --ignore-installed \
|
|
torch \
|
|
torchvision \
|
|
torchaudio \
|
|
numpy \
|
|
--index-url "$PIP_INDEX_URL"
|
|
|
|
COPY download_pytorch_datasets.py /tmp/
|
|
RUN python3 /tmp/download_pytorch_datasets.py && \
|
|
rm /tmp/download_pytorch_datasets.py
|
|
|
|
RUN PYTORCH_EXAMPLES_COMMIT=30b310a977a82dbfc3d8e4a820f3b14d876d3bd2 && \
|
|
mkdir /pytorch-examples && \
|
|
cd /pytorch-examples && \
|
|
git init && \
|
|
git remote add origin https://github.com/pytorch/examples && \
|
|
git fetch --depth 1 origin "$PYTORCH_EXAMPLES_COMMIT" && \
|
|
git checkout FETCH_HEAD && \
|
|
sed -ri "s~(datasets.*)\\(['\"](../)?data['\"],~\\1('$PYTORCH_DATASETS_DIR',~g" **/*.py && \
|
|
sed -ri 's/download=True/download=False/' **/*.py
|
|
|
|
COPY *.py /
|
|
RUN rm /download_pytorch_datasets.py
|
|
|
|
RUN PYTORCH_BENCHMARKS_COMMIT=675fb8f537d302a4fef3ed2a67349209e65046ac && \
|
|
mkdir /pytorch-benchmark && \
|
|
cd /pytorch-benchmark && \
|
|
git init && \
|
|
git remote add origin https://github.com/pytorch/benchmark.git && \
|
|
git fetch --depth 1 origin "$PYTORCH_BENCHMARKS_COMMIT" && \
|
|
git checkout FETCH_HEAD
|
|
|
|
# Note that mobilenet_v2 does not have a requirements.txt file.
|
|
RUN cd /pytorch-benchmark && \
|
|
python3 -m pip install --ignore-installed \
|
|
-r requirements.txt \
|
|
-r torchbenchmark/models/LearningToPaint/requirements.txt \
|
|
-r torchbenchmark/models/fastNLP_Bert/requirements.txt \
|
|
-r torchbenchmark/models/hf_BigBird/requirements.txt \
|
|
-r torchbenchmark/models/speech_transformer/requirements.txt
|
|
|
|
# These benchmarks are chosen based on diversity of the type of model and their
|
|
# profile with respect to using the GPU and moving data. For more context, see
|
|
# this paper: https://arxiv.org/pdf/2304.14226.pdf
|
|
RUN cd /pytorch-benchmark && \
|
|
python3 install.py \
|
|
LearningToPaint \
|
|
fastNLP_Bert \
|
|
hf_BigBird \
|
|
speech_transformer \
|
|
mobilenet_v2
|
|
|
|
# Some of these benchmarks download a dataset at runtime.
|
|
# Run them once on CPU just to get this predownloaded into the image.
|
|
RUN cd /pytorch-benchmark && \
|
|
python3 run.py LearningToPaint --device cpu && \
|
|
python3 run.py fastNLP_Bert --device cpu && \
|
|
python3 run.py hf_BigBird --device cpu && \
|
|
python3 run.py speech_transformer --device cpu && \
|
|
python3 run.py mobilenet_v2 --device cpu
|