Setup Cerebras virtual environment#
To launch a job in the Cerebras cluster, you need to create a Python virtual environment with the Cerebras software dependencies. You need to create the virtual environment once (but only once) for every release (R1.8.0) and framework used (PyTorch or TensorFlow)
To setup the Cerebras virtual environment, the default path for Cerebras packages is
To setup the Cerebras virtual environment follow these steps
Set up a Python virtual environment using Python 3.7. Create the environment named
venv_cerebras_ptfor a PyTorch environment or
venv_cerebras_tffor a TensorFlow environment using the following command:
python3.7 -m venv venv_cerebras_pt
python3.7 -m venv venv_cerebras_tf
The name of the python3.7 executable may be different on your system (e.g, python, python3, python3.7, /opt/python3.7/bin/python3.7). Please ensure that you are using the correct name corresponding to the executable for python3.7.
If you install Python from source you need
sqlite-devel installed on Linux or else you have a partial installation of python. That means many common python packages will fail with an error that looks like
ModuleNotFoundError: No module named '_bz2'
Cerebras provides three packages to set up virtual environments:
cerebras_appliancesoftware package, the
cerebras_tensorflowpackage if you are using TensorFlow, and the
cerebras_pytorchpackage if you are using PyTorch. To set up your PyTorch or TensorFlow environment, you need two out of these three packages. Enter the following commands to install the required packages.
source venv_cerebras_pt/bin/activate pip install --upgrade pip pip install <path_to_wheels>/cerebras_pytorch-<Cerebras release version>+<hash>-py3-none-any.whl --find-links=<path_to_wheels>
source venv_cerebras_tf/bin/activate pip install --upgrade pip pip install <path_to_wheels>/cerebras_tensorflow-<Cerebras release version>+<hash>-py3-none-any.whl --find-links=<path_to_wheels>
Note that the appliance wheel will be installed automatically by installing packages
<path_to_wheels> by default is
/opt/cerebras/wheels; if the path of the Cerebras packages is not the default one, use the location provided by your system administrator.
If your user node requires a proxy to access to external packages, you can add the flag
--proxy <proxy address> to the
pip commands to install the Cerebras packages.
find-links command, it finds the correct
cerebras_appliance version if you place all the wheels in the same directory.
You can add custom packages to your Cerebras environment by: 1) pip install the packages in virtual environment, then 2) Copy the package directory from
venv/lib/python3.7/site-packages/<package_name> to a NFS-mountable location in the Cerebras cluster. Make sure to add this location to the
--mount_dirs command line argument and the corresponding parent location to the
--python_paths command line argument when calling run.py.
Now you are all set and ready to train your first model on the Cerebras Wafer-Scale Cluster!