Model Zoo CLI Overview

Learn how to use the ModelZoo CLI.

Overview

The ModelZoo CLI is a comprehensive command-line interface that serves as a single entry point for all ModelZoo-related tasks. This tool streamlines various machine learning workflows, from data preprocessing to model training and validation.

Commands

Below is a list of commands that can be used with the ModelZoo CLI tool. Expand each section to see examples and more information.

fit

Trains a model using specified configuration

Example: cszoo fit params_model.yaml

validate

Validates a model using specified configuration

Example: cszoo validate params_model.yaml

validate_all

Runs upstream and downstream validation

Example: cszoo validate_all params_model.yaml

checkpoint

Checkpoint converter

convert

Converts a checkpoint between Huggingface and Cerebras formats or between different Cerebras formats.

Example: cszoo checkpoint convert --model gpt2 --src-fmt cs-auto --tgt-fmt hf --config workdir/params_gpt_tiny.yaml model_dir/checkpoint.mdl

convert-config

Converts a checkpoint between Huggingface and Cerebras formats or between different Cerebras formats.

Example: cszoo checkpoint convert-config --model gpt2 --src-fmt cs-auto --tgt-fmt hf workdir/params_gpt_tiny.yaml

list

Lists all available checkpoint converters. Can also list all checkpoint converters for a specified model.

Example: cszoo checkpoint list

diff

Compares two checkpoints to identify differences.

Example: cszoo checkpoint diff checkpoint_a.mdl checkpoint_b.mdl

model

Query information on ModelZoo models

list

Displays all supported models.

Example: cszoo model list

info

Shows detailed model information.

Example: cszoo model info gpt2

describe

Displays model configuration parameters.

Example: cszoo model describe gpt2

init_checkpoint

Creates initial model checkpoint.

Example: cszoo model init_checkpoint <model_name>

data_preprocess

Preprocess data

list

Shows available preprocessing configurations.

Example: cszoo data_preprocess list

pull

Copies a model configuration file to a local directory.

Example: cszoo data_preprocess pull summarization_preprocessing -o workdir

run

Executes preprocessing using specified configuration.

Example: cszoo data_preprocess run --config preprocessing.yaml

data_processor

Query information on ModelZoo data processors

list

Shows available data processors.

Example: cszoo data_processor list

info

Displays data processor information.

Example: cszoo data_processor info GptHDF5DataProcessor

describe

Shows processor configuration parameters.

Example: cszoo data_processor describe GptHDF5DataProcessor

benchmark

Benchmarks a specified dataloader.

Example: cszoo data_processor benchmark params.yaml

config

Save and manage model config files

pull

Copies a model config file to a local directory.

Example: cszoo config pull gpt2_tiny -o workdir

validate

Validates a specified config file.

Example: cszoo config validate params.yaml

convert_legacy

Upgrades V1 config files to V2 YAML.

Example: cszoo config convert_legacy old_config.yaml

stats

Retrieves relevant statistics for a model using the specified configuration file.

Example: cszoo config stats params.yaml

Example Workflow: Pretraining a model using the ModelZoo CLI

This workflow guides you through the steps to pretrain a model using the Cerebras ModelZoo CLI. Follow these steps to set up your environment, preprocess data, and run the pretraining process.

Prerequisite: Before proceeding with the steps below, ensure that you have completed the setup and installation guide found here.

1. Create model directory

   Create a directory to store all the files for this pretraining workflow and copy the necessary configuration files.

   mkdir pretraining_tutorial
   cp modelzoo/src/cerebras/modelzoo/tutorials/pretraining/* pretraining_tutorial
   

2. Preprocess the data

   Preprocess the training and validation datasets using the provided configuration files.

   cszoo data_preprocess run --config pretraining_tutorial/train_data_config.yaml
   cszoo data_preprocess run --config pretraining_tutorial/valid_data_config.yaml
   

3. Run model

   Run the pretraining process using the provided configuration.

   cszoo fit pretraining_tutorial/model_config.yaml
   

4. Convert checkpoint to HuggingFace

   Convert the trained model checkpoint into a HuggingFace-compatible format.

   cszoo checkpoint convert \
     --model llama \
     --src-fmt cs-auto \
     --tgt-fmt hf \
     --config pretraining_tutorial/model_config.yaml \
     --output-dir pretraining_tutorial/to_hf \
     pretraining_tutorial/model/checkpoint_0.mdl
   

Getting Help

For detailed information about any command, use the --help flag:

cszoo --help
cszoo <command> --help

CSZoo Assistant

Need help? Our CSZoo Assistant is an LLM agent you can access from the command line with the assistant subcommand.

Use it to:

• Ask questions: cszoo assistant "what is the checkpoint converter?"

• Perform actions: cszoo assistant "convert my checkpoint from huggingface to cerebras"

CSZoo Assistant will always ask your permission before running a command.

Access to the Cerebras Inference API is required and you’ll need to provide your API key with the following command:

export CEREBRAS_API_KEY=<your api key>

Don’t have an API key? Follow these instructions.

CSZoo Assistant is a beta feature and it may make mistakes. Always double-check its reasoning and be aware of the following limitations:

• CSZoo Assistant can currently only access the help manuals found with cszoo ... -h.

• There are currently no advanced context length management mechanisms in place. The assistant will error out if it overflows the context length.