Trainer Configuration Overview#

The Cerebras Model Zoo comes packaged with a few useful utilities. Namely, It features a way to configure a Trainer class using a YAML configuration file.

On this page you will learn how to write a YAML file to configure an instance of the Trainer class. By the end, you should be comfortable enough with the YAML specification to write your own configuration files from scratch.

Prerequisites#

Please ensure that you have read through the Trainer Overview beforehand. The rest of this page assumes that you already have at least a cursory understanding of what the Cerebras Model Zoo Trainer is and how to use the python API.

Base Specification#

The YAML specification is intentionally designed to map almost exactly one-to-one with the Trainer’s python API.

The Trainer’s constructor can be specified via a YAML configuration file as follows:

trainer:
  init:
    device: "CSX"
    model_dir: "./model_dir"
    model:
      # The remaining arguments to the model class
      vocab_size: 1024
      max_position_embeddings: 1024
      ...
    optimizer:
      # Corresponds to cstorch.optim.SGD
      SGD:
        lr: 0.01
        momentum: 0.9
    loop:
      num_steps: 1000
      eval_steps: 100
      eval_frequency: 100
    checkpoint:
      steps: 100
  fit:
    train_dataloader:
      data_processor: GptHDF5MapDataProcessor
      data_dir: "/path/to/train/data"
      batch_size: 64
      ...
    val_dataloader:
      data_processor: GptHDF5MapDataProcessor
      data_dir: "/path/to/validation/data"
      batch_size: 64
      ...

If you open the Python tab above, you can see the equivalent Python code that the YAML configuration corresponds to. As can be seen, the YAML specification almost directly mirrors the python API. This was an intentional design choice to make it easy to use one if you are familiar with the other.

Breaking It Down#

The YAML specification starts with the top level trainer key.

trainer:
  ...

If this key is not present, then the configuration is not valid.

The trainer accepts the following subkeys:

init#

The init key is used to specify the arguments to the Trainer’s constructor.

The arguments are passed as key-value pairs, where the key is the argument name and the value is the argument value.

Below are all of the accepted keys alongside YAML examples and their equivalent Python counterparts:

device#

The device to train the model on. If provided, it must be one of "CSX", "CPU", or "GPU".

trainer:
  init:
    device: "CSX"
    ...
  ...

backend#

Configures the backend used to train the model. If provided, it is expected to be a dictionary whose keys will be used to construct a cerebras.pytorch.backend instance.

trainer:
  init:
    backend:
      backend_type: "CSX"
      cluster_config:
        num_csx: 4
        mount_dirs:
        - /path/to/dir1
        - /path/to/dir2
        ...
      ...
    ...
  ...

Note

The backend argument is mutually exclusive with device. The functionality it provides is a strict superset of the functionality provided by device. To use a certain backend with all default parameters, you may specify device. To configure anything about the backend, you must specify those parameters via the backend key.

To learn more about the backend argument, you can check out Trainer Backend.

model_dir#

The directory where the model artifacts are saved. Some of the artifacts that may be dumped into the model_dir include (but are not limited to):

  • Client-side logs

  • Checkpoints

  • TensorBoard event files

  • Tensor summaries

  • Validation results

trainer:
  init:
    ...
    model_dir: "./model_dir"
    ...
  ...

model#

Configures the Module to train/validate using the constructed Trainer. All subkeys are passed as arguments to the model class.

trainer:
  init:
    ...
    model:
      vocab_size: 1024
      max_position_embeddings: 1024
      ...
    ...
  ...

To learn more about the model argument, you can check out Trainer Model.

optimizer#

Configures the Optimizer to use to optimize the model’s weights during training.

The value at this key is expected to be a dictionary. This dictionary is expected to contain a single key that specifies the name of the Cerebras optimizer to construct. That is to say, it must be the name of a subclass of Optimizer (see optim for the full list of Optimizer subclasses that come packaged in cerebras.pytorch)

The value of the Optimizer name key is expected to be dictionary of key-value pairs that correspond to the arguments of the optimizer subclass.

trainer:
  init:
    ...
    optimizer:
      # Corresponds to cstorch.optim.SGD
      SGD:
        lr: 0.01
        momentum: 0.9
    ...
  ...

Note

The params argument to the optimizer is automatically passed in and thus is not required.

To learn more about the optimizer argument, you can check out Trainer Optimizer.

schedulers#

Configures the Scheduler instances to use during the training run.

The value at this key is expected to be a dictionary or a list of dictionaries. Each dictionary is expected to have a single key specifying the name of the Scheduler. That is to say, it must be the name of a subclass of Scheduler (see cerebras.pytorch.optim for the full list of Scheduler subclasses that come packaged in cerebras.pytorch)

The corresponding value of the Scheduler name key is expected to be mapping of key-value pairs that are passed as keyword arguments to the Scheduler.

trainer:
  init:
    ...
    schedulers:
    - LinearLR:
        initial_learning_rate: 0.01
        end_learning_rate: 0.001
        total_iters: 100
    ...
  ...

Note

The optimizer argument to the Scheduler is automatically passed in and thus is not required.

To learn more about the schedulers argument, you can check out Trainer Schedulers.

precision#

Configures the Precision instance to use.

Today, the only supported Precision type is MixedPrecision.

So, the value of the precision key is expected to be a dictionary corresponding to the arguments of MixedPrecision

trainer:
  init:
    ...
    precision:
      fp16_type: float16
      precision_opt_level: 1
      loss_scaling_factor: dynamic
      max_gradient_norm: 1.0
      ...
    ...
  ...

To learn more about the precision argument, you can check out Trainer Precision.

sparsity#

Configures the SparsityAlgorithm to use to sparsity the model’s weights and optimizer state.

The value at this key is expected to be a dictionary. At a minimum, this dictionary is expected to contain an algorithm key that specifies the name of the sparsity algorithm to apply as well as a sparsity that specifies the level of sparsity to apply.

trainer:
  init:
    ...
    sparsity:
      algorithm: Static
      sparsity: 0.5
    ...
  ...

To learn more about how sparsity can be configured, see Train a model with weight sparsity.

loop#

Configures a TrainingLoop instance that specifies how many steps to train and validate for.

trainer:
  init:
    ...
    loop:
      num_steps: 1000
      eval_steps: 100
      eval_frequency: 100
    ...
  ...

To learn more about the loop argument, you can check out Training Loop.

checkpoint#

Configures a Checkpoint instance that specifies how frequently the trainer should save checkpoints during training.

trainer:
  init:
    ...
    checkpoint:
      steps: 100
    ...
  ...

To learn more about the checkpoint argument, you can check out Checkpointing.

logging#

Configures a Logging instance that configures the Python logger as well as specify how frequently the trainer should be writing logs.

trainer:
  init:
    ...
    logging:
      log_steps: 10
      log_level: INFO
    ...
  ...

In the above example, the Python logger is configured to allow info logs to be printed and to print logs at every 10 steps.

To learn more about the logging argument, you can check out Trainer Logging.

callbacks#

This key accepts a list of dictionaries, each of which specifies the configuration for some Callback class.

Each dictionary is expected to have a single key specifying the name of the Callback subclass (see callbacks for the full list of Callback subclasses that come packaged in cerebras.modelzoo).

The value at this key are passed in as keyword arguments to the subclass’s constructor.

trainer:
  init:
    ...
    callbacks:
    - CheckLoss: {}
    - ComputeNorm: {}
    - RateProfiler: {}
    - LogOptimizerParamGroup:
        keys:
        - lr
    ...
  ...

You can even include your own custom callbacks here. To learn more, you can read Customizing the Trainer with Callbacks

loggers#

This key accepts a list of dictionaries, each of which specifies the configuration for some Logger class.

Each dictionary is expected to have a single key specifying the name of the Logger subclass (see loggers for the full list of Logger subclasses that come packaged in cerebras.modelzoo).

The value at this key are passed in as keyword arguments to the subclass’s constructor.

trainer:
  init:
    ...
    logger:
    - ProgressLogger: {}
    - TensorboardLogger: {}
    ...
  ...

You can even include your own custom loggers here. To learn more, you can read Loggers

seed#

This key accepts a single integer value to seed the random number generator.

Setting this parameter will seed the PyTorch generator via a call to torch.manual_seed.

trainer:
  init:
    ...
    seed: 2024
  ...

fit#

The fit key is used to specify the arguments to the Trainer’s fit method.

The arguments are passed as key-value pairs, where the key is the argument name and the value is the argument value.

Below are all of the accepted keys alongside YAML examples and their equivalent Python counterparts:

train_dataloader#

This key is used to configure the training dataloader used to train the model.

This value at this key is expected to be a dictionary containing at a minimum the data_processor key which specifies the name of the data processors to use.

All other key-values in the dictionary are passed as argument to cerebras.pytorch.utils.data.DataLoader.

trainer:
  init:
    ...
  fit:
    train_dataloader:
      data_processor: GptHDF5MapDataProcessor
      data_dir: "/path/to/train/data"
      batch_size: 64
    ...
  ...

val_dataloader#

This key is used to configure the validation dataloader(s) used to validate the model.

The dataloader configured here gets run for eval_steps every eval_frequency training steps.

This value at this key is expected to be a dictionary or a list of dictionaries. Each dictionary is expected to contain at a minimum the data_processor key which specifies the name of the data processors to use.

All other key-values in the dictionary are passed as argument to cerebras.pytorch.utils.data.DataLoader.

trainer:
  init:
    ...
  fit:
    ...
    val_dataloader:
    - data_processor: GptHDF5MapDataProcessor
      data_dir: "/path/to/validation/data"
      batch_size: 64
    ...
  ...

ckpt_path#

Specifies the path to the checkpoint to load.

trainer:
  init:
    ...
  fit:
    ...
    ckpt_path: /path/to/checkpoint
  ...

validate#

The validate key is used to specify the arguments to the Trainer’s validate method.

The arguments are passed as key-value pairs, where the key is the argument name and the value is the argument value.

Below are all of the accepted keys alongside YAML examples and their equivalent Python counterparts:

val_dataloader#

This key is used to configure the validation dataloader used to validate the model.

This value at this key is expected to be a dictionary that contains at a minimum the data_processor key which specifies the name of the data processors to use.

All other key-values in the dictionary are passed as argument to cerebras.pytorch.utils.data.DataLoader.

trainer:
  init:
    ...
  validate:
    val_dataloader:
      data_processor: GptHDF5MapDataProcessor
      data_dir: "/path/to/validation/data"
      batch_size: 64
    ...
  ...

The validation dataloader is intended to be used alongside the validation metrics classes. See eval metrics to learn more.

ckpt_path#

Specifies the path to the checkpoint to load.

trainer:
  init:
    ...
  validate:
    ...
    ckpt_path: /path/to/checkpoint
  ...

validate_all#

The validate_all key is used to specify the arguments to the Trainer’s validate_all method.

The arguments are passed as key-value pairs, where the key is the argument name and the value is the argument value.

Below are all of the accepted keys alongside YAML examples and their equivalent Python counterparts:

val_dataloaders#

This key is used to configure the validation dataloader(s) used to validate the model.

This value at this key is expected to be a dictionary or a list of dictionaries. Each dictionary is expected to contain at a minimum the data_processor key which specifies the name of the data processors to use.

All other key-values in the dictionary are passed as argument to cerebras.pytorch.utils.data.DataLoader.

trainer:
  init:
    ...
  validate_all:
    val_dataloaders:
    - data_processor: GptHDF5MapDataProcessor
      data_dir: "/path/to/validation/data1"
      batch_size: 64
      ...
    - data_processor: GptHDF5MapDataProcessor
      data_dir: "/path/to/validation/data2"
      batch_size: 64
      ...
    ...
  ...

ckpt_paths#

Specifies the paths to the checkpoints to load.

trainer:
  init:
    ...
  validate_all:
    ...
    ckpt_paths:
    - /path/to/checkpoint1
    - /glob/path/to/checkpoint*
  ...

Note

Globs are accepted as well.

All validation dataloaders are used to run validation for every checkpoint. So, effectively, validate_all is doing

from cerebras.modelzoo import Trainer

trainer = Trainer(...)
for ckpt_path in ckpt_paths:
    trainer.load_checkpoint(ckpt_path)
    for val_dataloader in val_dataloaders:
        trainer.validate(val_dataloader)

Legacy Specification#

In releases 2.2 and below the YAML specification was different. It used to be of the form:

model:
  ...
optimizer:
  ...
train_input:
  ...
eval_input:
  ...
runconfig:
  ...

The reason we changed it to the way it is today is that the older specification was not general or flexible enough to make full use of the Trainer class.

If you have a legacy YAML configuration lying around, you can still use it. There is a converter available that can be used to convert any legacy YAML configurations into the new trainer YAML configuration:

import yaml

from cerebras.modelzoo.trainer.utils import (
    convert_legacy_params_to_trainer_params
)


with open("/path/to/legacy/params.yaml") as f:
    legacy_params = yaml.load(f)

trainer_params = convert_legacy_params_to_trainer_params(
    legacy_params
)

with open("/path/to/trainer/params.yaml", "w") as f:
    yaml.dump(f)

The training scripts provided in the Cerebras ModelZoo are capable of detecting if you passed in a legacy configuration and will automatically invoke this converter before proceeding to constructing and using the Trainer.

If you are already familiar with the Legacy YAML specification and just want to find out how to specify a specific parameter in the Trainer YAML specification, please refer to the table in Correspondance from Legacy to Trainer.

Conclusion#

By this point, whether you are writing it from scratch or starting from an existing legacy configuration, you should have an understanding of how to configure a Trainer using a YAML configuration file.

What’s next?#

To learn more about how you can use the Trainer in some core workflows, you can check out:

To learn more about how you can extend the capabilities of the Trainer class, you can check out: