A containerized version of LambdaStack is available on the SAIL nodes. This container provides a working installation of PyTorch, TensorFlow, CUDA and cuDNN.
This page is a survival guide to get started with this container and use it with your projects. It assumes that you have an opened SSH session on one of the machines.
Start the container
First check that the container image is available using the docker image ls command:
$ docker image ls REPOSITORY TAG IMAGE ID CREATED SIZE lambda-stack 20.04 abe4a492cee1 6 hours ago 12GB ubuntu latest df5de72bdb3b 3 weeks ago 77.8MB ubuntu 20.04 3bc6e9f30f51 3 weeks ago 72.8MB debian latest 07d9246c53a6 3 weeks ago 124MB nvidia/cuda 11.0.3-base-ubuntu20.04 8017f5c31b74 5 weeks ago 122MB hello-world latest feb5d9fea6a5 11 months ago 13.3kB
You should see the lambda-stack image in the list, this is the one we will use for now.
To start a program in the container, use the docker run --rm command:
$ docker run --rm lambda-stack:20.04 pwd /root
The --rm flag ensures that the container is cleaned-up after usage.
To run an interactive command like bash or python interpreter, add the -it flag:
$ docker run --rm -it lambda-stack:20.04 bash root@fa416f6b82f5:~# pwd /root root@fa416f6b82f5:~# exit $
So far, the container does not have access to the GPUs. To give it access to them, you need to change the runtime to nvidia and explicitly specify a list of GPUs. The following example uses the 4th and 5th GPUs (indices start at 0):
$ docker run --runtime=nvidia -e NVIDIA_VISIBLE_DEVICES=3,4 --rm lambda-stack:20.04 nvidia-smi
Thu Sep 1 22:02:25 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 515.65.01 Driver Version: 515.65.01 CUDA Version: 11.7 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA A100-SXM... On | 00000000:4C:00.0 Off | 0 |
| N/A 60C P0 374W / 400W | 33549MiB / 81920MiB | 47% Default |
| | | Disabled |
+-------------------------------+----------------------+----------------------+
| 1 NVIDIA A100-SXM... On | 00000000:88:00.0 Off | 0 |
| N/A 54C P0 377W / 400W | 33549MiB / 81920MiB | 86% Default |
| | | Disabled |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
+-----------------------------------------------------------------------------+
Note that GPU device numbers are re-mapped to start from 0 in the container.
This method does not prevent multiple containers to access the same GPUs. Therefore, make sure to check with other users which GPUs they are using.
This method does ensure that your container will not use by mistake any other GPU than the one specified.
An alternative method to access GPUs is the --gpus option:
$ docker run --gpus 2 --rm lambda-stack:20.04 nvidia-smi
Unless you are using the 8 GPUs, we strongly recommend not using this syntax, as it does not let you choose precisely which GPUs to select.
Access a folder from the container
A container is isolated from the host environment by default. Bind mounts allow you to mount a folder from the host machine into the container.
Specify the path to the directory on host and the corresponding path inside the container using the -v flag:
-v <path on host>:<path in container>
For example, assuming you have a project folder in $HOME/my_project and want to access it as /my_project in the container, you would use:
$ docker run -v $HOME/my_project:/my_project --rm lambda-stack:20.04 ls /my_project
Any program running in the container can then access and modify files in the /my_project folder.
You can repeat the -v flag to mount multiple folders in the container.
Add packages to the container
The container may not have all the packages you need. To add more packages, you can create a new container based on the LambdaStack one.
To create a container, you need a Dockerfile definition file. It contains the information about the base container and the installations instructions for the additional packages.
In the following Dockerfile example, the Transformers library (PyTorch version) from Hugging Face is added to the LambdaStack container:
FROM lambda-stack:20.04 RUN pip install pip install transformers[torch]
To build the corresponding container, first create an empty folder and save the Dockerfile in it:
$ mkdir hugging_container $ echo "FROM lambda-stack:20.04" > hugging_container/Dockerfile $ echo "RUN pip install transformers[torch]" >> hugging_container/Dockerfile
then use the docker build command to generate the new container:
$ cd hugging_container $ docker build -t pytorch-transformers .
The -t flag is used to tag the container, making it easier to find and use it later.
Use docker image ls to check the availability of the image:
$ docker image ls REPOSITORY TAG IMAGE ID CREATED SIZE pytorch-transformers latest 432c6be0a999 13 seconds ago 12.1GB lambda-stack 20.04 abe4a492cee1 6 days ago 12GB ubuntu latest df5de72bdb3b 4 weeks ago 77.8MB ubuntu 20.04 3bc6e9f30f51 4 weeks ago 72.8MB debian latest 07d9246c53a6 4 weeks ago 124MB nvidia/cuda 11.0.3-base-ubuntu20.04 8017f5c31b74 6 weeks ago 122MB hello-world latest feb5d9fea6a5 11 months ago 13.3kB
You can now use it in place of the LambdaStack container:
$ docker run --rm pytorch-transformers \
python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('I love you'))"
libibverbs: Warning: couldn't open config directory '/etc/libibverbs.d'.
libibverbs: Warning: couldn't open config directory '/etc/libibverbs.d'.
No model was supplied, defaulted to distilbert-base-uncased-finetuned-sst-2-english and revision af0f99b (https://huggingface.co/distilbert-base-uncased-finetuned-sst-2-english).
Using a pipeline without specifying a model name and revision in production is not recommended.
Downloading config.json: 100%|██████████| 629/629 [00:00<00:00, 1.45MB/s]
Downloading pytorch_model.bin: 100%|██████████| 255M/255M [00:10<00:00, 24.5MB/s]
Downloading tokenizer_config.json: 100%|██████████| 48.0/48.0 [00:00<00:00, 72.8kB/s]
Downloading vocab.txt: 100%|██████████| 226k/226k [00:00<00:00, 295kB/s]
[{'label': 'POSITIVE', 'score': 0.9998656511306763}]
Containers can be deleted using the docker image rm command. For example, remove the pytorch-transformers container as follows:
$ docker image rm pytorch-transformers Untagged: pytorch-transformers:latest Deleted: sha256:432c6be0a999484db090c5d9904e5c783454080d8ad8bc39e0499ace479c4559 Deleted: sha256:623ae3b33709c2fc4c40bc2c3959049345fee0087d39b4f53eb95aefd1c16f7d
Next steps
This document is a very short introduction to Docker to use LambdaStack. If you want to know more about Docker in general, we can recommend this workshop material and the associated recording on Youtube.