Skip to content

Latest commit

 

History

History
81 lines (59 loc) · 2.56 KB

README.md

File metadata and controls

81 lines (59 loc) · 2.56 KB
Raw

jaxlib GPU docker image

This is my attempt at creating a repeatable setup for building jaxlib from source and interactively running tests on a GPU VM.

I don't really know what I'm doing, especially when it comes to docker.

Build image

sudo docker build -t gpu_jaxlib cuda12 # or cuda11

Building GPU jaxlib takes a long time, so do this on a beefy machine (doesn't have to have GPUs attached).

Copy jaxlib wheel out of image

image_name=gpu_jaxlib
jaxlib_path=$(sudo docker run --rm $image_name find /root/jax/dist/ -mindepth 1 | tail -n 1)
id=$(sudo docker create $image_name)
sudo docker cp $id:$jaxlib_path .
sudo docker rm -v $id
ls `basename $jaxlib_path`

Save image to tar file

sudo docker save -o <tar file> gpu_jaxlib:latest

It needs lots of disk space (~25 GB).

One-time setup on GPU VM to run image

These might need tweaking.

# from https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#docker
distribution=$(. /etc/os-release;echo $ID$VERSION_ID) \
      && curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg \
      && curl -s -L https://nvidia.github.io/libnvidia-container/$distribution/libnvidia-container.list | \
            sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | \
            sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list

sudo apt update
sudo apt install nvidia-driver-530 -y
sudo apt install nvidia-docker2 -y
sudo apt install nvidia-container-toolkit-base -y
sudo nvidia-ctk runtime configure
sudo systemctl restart docker
sudo reboot

Load image from tar file

sudo docker load -i <tar file>

Run image

sudo docker run -it --rm --shm-size=16g --gpus all  gpu_jaxlib:latest

The image starts in a jax checkout that's locally installed via pip install -e ., with a jaxlib built from source from ~/xla and installed.

You can run all unit tests with (may need tweaking):

export TF_CPP_MIN_LOG_LEVEL=0
export XLA_PYTHON_CLIENT_ALLOCATOR=platform
export LD_LIBRARY_PATH="/usr/local/cuda:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64:/usr/local/tensorrt/lib"
export NCCL_DEBUG=WARN

python3.10 -m pytest -n 8 --tb=short --maxfail=1 tests examples --deselect=tests/linalg_test.py::LaxLinalgTest::test_tridiagonal_solve1 --deselect=tests/multi_device_test.py::MultiDeviceTest::test_computation_follows_data --deselect=tests/xmap_test.py::XMapTest::testCollectivePermute2D