Cerebras cluster settings#
Credentials path and management address#
The credentials_path and mgmt_address parameters are set by default in /opt/cerebras/config and typically do not require explicit specification.
An example configuration in /opt/cerebras/config might look like this:
{
"clusters": [
{
"name": "system-name",
"server": "1.2.3.4:9000",
"authority": "cluster-server.system-name.example.com",
"certificateAuthority": "/opt/cerebras/certs/tls.crt"
}
],
"contexts": [
{
"cluster": "cluster-name",
"name": "system-name"
}
],
"currentContext": "system-name"
}
Python paths and mount directories#
When running jobs, it’s crucial to include paths required by dataloaders and any external Python packages, as well as the location of the Cerebras Model Zoo.
Consider this directory structure as an example:
/path/to/datasets
my_dataset/
/path/to/modelzoo
modelzoo
/path/to/packages
package_x
package_y
If input workers need to access the “my_dataset” directory and require Python modules from the Model Zoo, “package_x”, and “package_y”, you would specify the command like this:
python run.py \
CSX \
--params params.yaml \
--mode {train,eval,eval_all,train_and_eval} \
--mount_dirs /path/to/datasets /path/to/modelzoo /path/to/packages \
--python_paths /path/to/packages /path/to/modelzoo \
Note
If some paths have a common parent folder, you only need to specify the parent in the --mount_dirs or --python_paths arguments. For instance, if the Model Zoo is at “/cb/home/user/modelzoo” and data is at “/cb/home/user/data”, you only need to specify --mount_dirs /cb/home.
You can also set defaults for mount_dirs and python_paths in a YAML file and export it as an environment variable to streamline command line usage:
export CEREBRAS_WAFER_SCALE_CLUSTER_DEFAULTS=/path/to/defaults/file.yaml
With this environment variable set, you won’t need to specify --mount_dirs and --python_paths when running a Cerebras Model Zoo job, unless you’re adding additional paths.