xprof

What it is and what it does

XProf is an open-source profiler designed for modern ML workloads, providing a web-based interface to analyze and visualize model performance across multiple devices. It integrates with TensorBoard as a plugin or runs standalone, offering tools like an overview page showing step-time breakdowns, a trace viewer displaying operation-level execution timelines, memory profiling, and HLO graph visualization. The package depends on 13 runtime libraries including protobuf, gcsfs, google-cloud-storage, and werkzeug, indicating support for both local and cloud-based profile storage and serving.

The profiler is actively maintained (latest release 9 days old, first released May 2025) and requires Python 3.10 or later. It has no known security vulnerabilities. Installation is straightforward via pip, though the dependency footprint is moderate. The tool is intended for developers, researchers, and educators working with JAX, PyTorch, TensorFlow, and XLA—frameworks commonly used in ML research and production.

Use it for:

• Analyze step-time performance and identify bottlenecks in multi-device ML training runs.
• Visualize operation-level execution timelines to understand which ops run on host vs. device and inter-device communication.
• Monitor memory usage patterns during model training to detect memory leaks or inefficient allocations.
• Debug distributed training by viewing aggregated performance metrics across all devices and hosts.
• Integrate profiling into TensorBoard workflows for unified experiment tracking and analysis.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

XProf is a profiler for ML workloads that visualizes performance across multiple devices, showing execution timelines, memory usage, and operation-level details through a web interface.

Yes, if you train ML models on multiple devices or use JAX, PyTorch, TensorFlow, or XLA. The package is actively maintained, has no security issues, and provides essential multi-device profiling tools. Medium install friction is acceptable for the capability gained. Not necessary for single-device or non-ML workloads.

Install

xprof on PyPI

pip install xprof

uv add xprof

poetry add xprof

Installing xprof

Before you install

Medium install friction due to 13 runtime dependencies including protobuf, gcsfs, and google-cloud-storage. Package is actively maintained with recent releases (9 days old) and has no known vulnerabilities. Python 3.10+ required; users on Python 3.12+ may need to downgrade setuptools below version 70.

License in practice

Licensed under Apache 2.0 (permissive), allowing free use, modification, and distribution with minimal restrictions—suitable for both open-source and commercial projects.

Quickstart

pip install xprof
xprof --logdir=profiler/demo --port=6006
# Then open browser to localhost:6006 to view profile data

Requires Internet access to load Google Chart library; some visualizations may be unavailable offline or behind corporate firewalls. Python 3.10+ required.

Verify before relying

• Whether the 13 runtime dependencies (including gcsfs and google-cloud-storage) are all required for basic profiling or only for specific cloud storage features.
• Performance overhead of the profiler itself on large-scale distributed training runs.
• Compatibility with frameworks beyond JAX, PyTorch, TensorFlow, and XLA mentioned in keywords.

Package facts

Evidence: xprof-2.23.1-py3-none-macosx_11_0_arm64.whl; xprof-2.23.1-py3-none-manylinux_2_27_x86_64.whl; xprof-2.23.1-py3-none-manylinux_2_35_aarch64.whl; xprof-2.23.1-py3-none-win_amd64.whl

Keywords: jax, pytorch, xla, tensorflow, tensorboard, xprof, profile, plugin

Tags

More Libraries packages