Supported config parameters#
The Cerebras Model Zoo supports an extensive range of models, each with its own set of parameters. This document provides a comprehensive list of these parameters as defined in the Model Zoo.
With the introduction of Config classes in the Model Zoo, each parameter is defined in the model’s config file
. These classes, implemented as Python dataclasses, organize and validate the parameters necessary for model definition and training.
Common parameters across models#
1. RunConfig Parameters:
2. Sparsity Parameters:
3. Optimizer Parameters:
Model Specific Parameters#
Large Language Model (LLM) parameters#
Vision model parameters#
Multimodal model parameters#
Understanding Config classes structure#
Each of these parameters are present as part of a Config class. A Config class, implemented as a Python dataclass, serves as a container for storing essential settings and parameters needed for defining and training a model.
Each of these class atributes correspond to the respective section in a YAML file which is used to define the parameters for a training run.
A config class looks like this:
class <ConfigClass>:
train_input = Optional[DataConfig] = None
eval_input = Optional[DataConfig] = None
model = <ModelConfigClass> = required
sparsity: Optional[SparsityConfig] = None
optimizer: OptimizerConfig = required
runconfig: RunConfig = required
For more information about Config classes, refer to the Model Zoo config classes documentation.
Additional notes#
Each model has a designated ModelConfigClass
. In cases where a model is a variant of another, it may inherit the ModelConfigClass
from the parent model. To understand more about this inheritance and the hierarchy of config classes, visit the Config class hierarchy documentation.