App Best Practices

Tip

This page covers practical recommendations for TorchX applications: portable storage with fsspec, train loop frameworks, checkpointing, metrics, and testing. See Component Best Practices for component authoring guidance.

Prerequisites: Quickstart and Basic Concepts.

Data Passing and Storage

Use fsspec (Filesystem Spec) for storage access. fsspec provides a unified interface to many storage backends (local, S3, GCS, etc.), so apps run on different infrastructures by changing paths alone.

import fsspec

with fsspec.open("s3://bucket/data.pt", "rb") as f:
    data = torch.load(f, weights_only=True)

Train Loops

Common choices:

• Pure PyTorch

• PyTorch Lightning

• PyTorch Ignite

See Train for more information.

Metrics

Use TensorBoard for logging metrics. TensorBoard supports remote storage (S3, GCS) for viewing metrics during training.

See Metrics for TorchX metric components.

Checkpointing

Periodic checkpoints enable failure recovery and training resumption. PyTorch Lightning provides standardized checkpointing.

Fine Tuning

Provide a command line argument to your app that resumes from a checkpoint file. This enables transfer learning, fine tuning, and failure recovery with a single application.

Interpretability

Use Captum for model interpretability. See Interpret for built-in components.

Model Packaging

Python + Saved Weights

The most common format. Load a model definition from Python, then load weights from a .ckpt or .pt file.

TorchScript Models

Serializable, optimized models executable without Python. See the TorchScript docs.

torch.export

PyTorch’s modern export path for production deployment. Produces ExportedProgram artifacts that work with torch.compile and inference runtimes.

Serving / Inference

Use TorchServe for standard use cases. See Serve for built-in components.

Testing

Since TorchX apps are standard Python, test them like any other Python code:

import unittest
from your.custom.app import main

class CustomAppTest(unittest.TestCase):
    def test_main(self) -> None:
        main(["--src", "src", "--dst", "dst"])
        self.assertTrue(...)

See also

Component Best Practices

Best practices for authoring reusable TorchX components.

Quickstart

Getting started with TorchX from scratch.