Skip to content

Software Stack

This section describes how to load/create the packages required for specific computing environments for performing your runs. Researchers are expected to work with a particular singularity container for their code runs.

  • What are these Singularity containers?

    These are stand-alone packages holding the software needed to create a very specific computing environment.

  • What container should be used?

    We recommend using the latest version of the main Cerebras container to advance the code compilation. It can be found in the following paths:

    • Inside the reference folder on the Ocean shared file system:

      /ocean/neocortex/cerebras/cbcore_latest.sif

    • inside the Neocortex SDFs:

      /local[1-4]/cerebras/cbcore_latest.sif

    Please note that this is a symlink and not the actual container file. If you want to check the file path or container version, you can use the ll command.

  • What software is compatible?

    • The TensorFlow version certified as compatible with the modelzoo is TensorFlow version 2.2.
    • The PyTorch version certified as compatible with the modelzoo is version 1.11.

HPC Workflows - Cerebras SDK

The required variables should be already present in the system if you started a Slurm job. If not, you can set them like this: 

SDK_INSTALL_LOCATION=/ocean/neocortex/cerebras
SDK_INSTALL_PATH=$SDK_INSTALL_LOCATION/sdk

Then, the commands required for setting the path to the different tools to be used with the SDK are defined like this: 

PATH_CMD='export PATH='$SDK_INSTALL_PATH':$PATH'
eval $PATH_CMD

And if you want to persist the SDK configuration across sessions. Run the following command to update the PATH variable in your .bashrc file: 

echo $PATH_CMD >> ~/.bashrc

Now, for starting the process, you can take a look at the contents under the $SDK_INSTALL_PATH folder:

csl_examples/code-samples: contains multiple simple examples (20). For example: 

cd ${SDK_INSTALL_PATH}/csl_examples/code-samples/01-tasks-and-colors

Compile

For the CS-2s, the flags should be set to --arch=wse2 and --fabric-dims=757,996 like the following:

cslc --arch=wse2 ./code.csl --fabric-dims=757,996 --fabric-offsets=1,1 --colors=x_in:1,b_in:2,y_out:3,Ax_out:4,sentinel:43 -o out

Run

For passing the CS-2 IP addresses, please use the variable CS_IP_ADDR, along with --fabdims 757,996 like the following:

cs_python run.py --name out --cmaddr ${CS_IP_ADDR}:9000 --fabdims 757,996