---
id: torch-tb-profiler
version: "0.4.3"
license: BSD-3
license_treatment: permissive
maintenance: active
---

# torch-tb-profiler — PyTorch Profiler TensorBoard Plugin

License: permissive · Maintenance: active · Downloads: 376.1K/mo

## What it is and what it does

torch-tb-profiler is a TensorBoard plugin that integrates Libkineto, PyTorch's profiling library, to visualize and analyze ML workload performance. It focuses on low-overhead GPU timeline tracing and provides actionable recommendations for common performance bottlenecks in deep learning models. The plugin runs as an in-process profiler within PyTorch and surfaces results directly in TensorBoard's interface.

The package depends on pandas for data handling and tensorboard for visualization. It's designed for developers and researchers who need to diagnose performance issues in PyTorch models, particularly GPU-related bottlenecks. Installation is straightforward with no compiled dependencies beyond what PyTorch itself requires.

Use it for:

- Visualize GPU kernel execution timelines and identify which operations consume the most compute time.
- Profile data loading, model forward/backward passes, and synchronization overhead in training loops.
- Analyze memory usage patterns and detect GPU memory bottlenecks during model execution.
- Compare profiling results across different hardware or model configurations using TensorBoard's interface.
- Generate performance recommendations for common ML training issues like GPU underutilization.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Integrates PyTorch profiling data with TensorBoard, providing GPU timeline tracing and performance diagnostics for ML workloads through a TensorBoard plugin interface.

Yes, if you actively profile PyTorch models and use TensorBoard. The low install friction, permissive license, and active maintenance make it a straightforward addition to a profiling workflow. However, the last release was in October 2023; verify compatibility with your current PyTorch version before relying on it for new projects.

## Install

pip install torch-tb-profiler
uv add torch-tb-profiler
poetry add torch-tb-profiler

## Installing torch-tb-profiler

Before you install: Low friction installation with only two runtime dependencies (pandas and tensorboard). The project maintains active status with recent commits and follows PyTorch's release schedule on a 3-month cadence.

License in practice: BSD-3 permissive license allows commercial and private use with minimal restrictions; you must retain the license notice in distributions.

Quickstart:

pip install torch-tb-profiler

import torch_tb_profiler
from torch.profiler import profile, record_function

with profile(activities=[...], on_trace_ready=torch_tb_profiler.trace_handler):
    # profile your model

Requires PyTorch and TensorBoard installed; GPU profiling requires CUDA-capable hardware and appropriate drivers.

Verify before relying:

- Whether the package works with current PyTorch versions (last release was 2023-10-06, over 2 years old).
- Specific GPU architectures and CUDA versions supported by Libkineto.
- Whether low-overhead GPU tracing is maintained across different hardware configurations.

## Package facts

- License: BSD-3 (permissive)
- Python support: supports_current
- Install friction: low
- Maintenance: active
- Downloads: 376.1K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

pytorch profiler tensorboard, gpu profiling visualization, pytorch performance diagnostics, ml bottleneck analysis, kineto profiler plugin, pytorch gpu timeline tracing, tensorboard profiling extension, profiling, pytorch-ecosystem, performance-analysis

[View on SkillFed](https://skillfed.io/packages/torch-tb-profiler) · [View on PyPI](https://pypi.org/project/torch-tb-profiler/)