Skip to content

create_image_alps.md

Latest commit

 

History

History
35 lines (22 loc) · 1.28 KB

create_image_alps.md

File metadata and controls

35 lines (22 loc) · 1.28 KB

This guide assumes you already have successfully logged in to Alps, and added to the a-infra-02 project.

  1. ssh into Clariden and go to the '/capstor/scratch/cscs/$USER' directory

  2. Get a compute node: srun -t 5:00:00 -A a-infra02 --container-writable --pty bash

  3. Clone the sailor repo git clone https://github.com/eth-easl/sailor.git && cd sailor && git checkout sosp25_ae

  4. While in the folder with the Dockerfile, create a new a new image (adjust the name as you want) podman build -t test:v1 .

  5. You can see your image now using this command podman images

  6. use enroot to export the image into a squash file enroot import -o test.sqsh podman://test:v1

  7. Make it readable setfacl -b test.sqsh && chmod 755 test.sqsh

  8. Log out of the node

  9. Replace the 'user' with your username in ae_scripts/clariden_scripts/sailor.toml

  10. To get a container with the image running and get a shell

srun -t 5:00:00 -A a-infra02 --container-writable --environment=/capstor/scratch/cscs/$USER/sailor/ae_scripts/clariden_scripts/sailor.toml --pty bash

  1. Run a simple training job with just 1 GPU to check all works:
cd /root/sailor/third_party/Megatron-DeepSpeed/
export SAILOR_LOGS_DIR=logs
bash run.sh 1 0 127.0.0.1 1234 1 1 1 1 1