PyTorch Runners#

The run function that was described in Port PyTorch to Cerebras exists as a wrapper around the PyTorch runners. The run function’s true purpose is to act as an interface between the user and the PyTorchBaseRunner.

The PyTorchBaseRunner is, as the name suggests, the base runner class. It contains all of the common elements required to train or evaluate any PyTorch model. There are a number of PyTorchBaseRunner subclasses that customize the behavior of the run to fit the needs of the specific run being conducted.

  • PyTorchRunner: The PyTorchBaseRunner subclass that facilitates CPU/GPU runs

  • PyTorchCSCompiler: The PyTorchBaseRunner subclass that facilitates compiling a PyTorch model for the Cerebras system

  • PyTorchCSRunner: The PyTorchBaseRunner subclass that facilitates executing a training/evaluation run of a PyTorch model on a Cerebras system

  • PyTorchCSAppliance: The PyTorchBaseRunner subclass that facilitates compiling and executing a weight streaming training/evaluation run using appliance mode of a PyTorch model on a Cerebras system

The run function’s job is to parse the command line arguments, configure the appropriate runner, and start the actual training or evaluation run. However, the bulk of the training/evaluation loop code lives inside the PyTorchBaseRunner.

Using Custom Models with the PyTorch Runners#

If you have a custom model and optimizer that you wish to use with our PyTorch runners, you can use the following steps to transform your existing model and optimizer with our PyTorchBaseModel.

To create a simple wrapper for your model and optimizer, it is as simple as subclassing PyTorchBaseModel and assigning the model and optimizer accordingly:

import torch
from modelzoo.common.pytorch.PyTorchBaseModel import PyTorchBaseModel

class BaseModel(PyTorchBaseModel):
    def __init__(
        self,
        params: dict,
        model: torch.nn.Module,
        optimizer: torch.optim.Optimizer,

    ):
        self.custom_optimizer = optimizer
        self.loss_fn = ...

        super().__init__(params, model, device=None)

    def _configure_optimizer(self, params):
        """
        Override default optimizer configuration and return custom optimizer
        """
        return self.custom_optimizer

    def _configure_lr_scheduler(self, params):
        """
        Override default lr scheduler configuration and return custom
        lr scheduler
        """
        return None

    # An example implementation of the __call__ function
    def __call__(self, data):
        inputs, targets = data
        outputs = self.model(data)
        loss = self.loss_fn(outputs, targets)
        return loss

With this BaseModel definition, it is easy to then wrap it inside a model_fn function which can then be used with the common run interface:

from modelzoo.common.pytorch.run_utils import run

# Define your model and optimizer here
model: torch.nn.Module = ...
optimizer: torch.optim.Optimizer = ...

def model_fn(params):
    return BaseModel(params, model, optimizer)

run(
    model_fn,
    train_input_dataloader_fn,
    eval_input_dataloader_fn,
)

The common CLI arguments are still expected to be present to be parsed and used in the run. The only difference being, the "optimizer" key from the params.yaml won’t be used to configure the optimizer and learning rate scheduler.

If you are implementing your own custom optimizers and learning rate schedulers, please see the Limitations of PyTorch on Cerebras page to see what constraints your implementations must satisfy

Creating Custom PyTorch Runners (Advanced)#

We highly recommend that the run function be used as it is provided. However, if it is simply insufficient, there is a way to customize the existing PyTorch runners to add custom behaviors such as custom logging and performance profiling.

At the end of the day, all custom runners should subclass either the PyTorchBaseRunner or one of its subclasses. We recommend doing the latter, as they are already custom tailored to facilitate their use case.

There are a number of hooks that are provided that can be overridden in the respective subclass to customize the runs:

on_train_start

A method that is run at the very beginning of training, right before the training loop is entered.

on_train_end

A method that is run at the very end of training, right after the training loop is finished, (but only if the training loop finished successfully and without exceptions).

on_eval_start

A method that is run at the very beginning of evaluation, right before the evaluation loop is entered.

on_eval_end

A method that is run at the very end of evaluation, right after the evaluation loop is finished, (but only if the evaluation loop finished successfully and without exceptions).

on_train_epoch_start

A method that is run at the very beginning of each training epoch, right before the training epoch is started.

on_train_epoch_end

A method that is run at the very end of each training epoch, right after the epoch ends, (but only if the epoch finished successfully and without exceptions).

on_eval_epoch_start

A method that is run at the very beginning of each evaluation epoch, right before the evaluation epoch is started.

on_eval_epoch_end

A method that is run at the very end of each evaluation epoch, right after the epoch ends, (but only if the epoch finished successfully and without exceptions).

on_train_batch_start

A method that is run at the very beginning of each training step, right before the forward/backwards pass and optimizer step is performed.

on_train_batch_end

A method that is run at the very end of each training step, right after the optimizer step is performed.

on_train_batch_start

A method that is run at the very beginning of each evaluation step, right before the forward pass.

on_train_batch_end

A method that is run at the very end of each evaluation step, right after the forward pass.

Note

Many of these hooks already have overrides defined in the PyTorchBaseRunner, PyTorchRunner, PyTorchCSCompiler, and PyTorchCSRunner. We highly recommended to call the super class’s implementation of the hook to prevent any unexpected failures.

Warning

We do not recommended overriding any other method from PyTorchBaseRunner besides these hooks. There is a good chance that not overriding them correctly could lead to failures in either the compile or execution.

Using Custom Runner#

To use a custom PyTorch runner, you must modify the PyTorchBaseRunner’s create static method in https://github.com/Cerebras/modelzoo/blob/master/modelzoo/common/pytorch/pytorch_base_runner.py to make use of your custom runner instead of the predefined ones.

Note

It is important that only the runner initialization is modified. There are other components in the create function that are integral to configuring and ensuring the run on the Cerebras system works.