TraceLens documentation

2026-07-22

1 min read time

TraceLens is an open-source Python library developed by AMD that automates performance analysis from GPU trace files. Instead of manually inspecting raw profiling data in tools such as Perfetto or Chrome Trace Viewer, TraceLens parses traces from PyTorch, JAX, and rocprofv3 and produces structured performance reports — including hierarchical GPU-timeline breakdowns, per-operator roofline analysis (TFLOP/s, TB/s), and multi-GPU communication diagnostics.

The TraceLens source code is hosted at github.com/AMD-AGI/TraceLens.

Install

• Install TraceLens

How to

• Generate reports

  • PyTorch performance report

  • PyTorch inference performance report

  • JAX performance report

  • rocprof performance report

  • Collective-communication report

  • Compare performance reports

• Compare two traces

• Replay a single operation

• Fuse multi-rank traces

• Analyze traces with the SDK

• Model op performance without a trace

• Analyze collective communication

• Estimate kernel times with Origami

• Agentic performance analysis with the TraceLens Agent

Concepts

• The Trace2Tree data model

• Understanding PyTorch traces

• Tensor shape metadata

• GEMM analysis

• Inference performance analysis

• Triton kernel performance model

Tutorials

• PyTorch profiling walkthrough

• Distributed profiling walkthrough

• Anomaly detection walkthrough

Reference

• API reference

• Performance report columns

• Compatibility matrix

For information on contributing to TraceLens, see the Contributing guide.

TraceLens is released under the MIT License. For details, see the License page.