TraceLens Analysis Orchestrator

TraceLens Analysis Orchestrator gives an agent a repeatable way to investigate complex PyTorch profiler data. It organizes the work around shared trace preparation, focused performance reviews, output checks, and a final report that ranks the most useful findings. Both individual runs and comparisons between two traces are supported.

Highlights

• Examines system activity alongside compute-heavy GPU operations
• Supports training, inference, and comparative performance investigations
• Separates specialized analysis into parallel agent tasks where appropriate
• Consolidates findings, metadata, tables, and visual evidence into one report
• Checks intermediate results before presenting prioritized recommendations

The workflow is useful for machine-learning performance engineers, framework developers, and infrastructure teams that need more than a raw profiler timeline. It helps turn trace evidence into an organized explanation of utilization gaps, expensive operations, communication overhead, memory transfers, and kernel-level opportunities. The skill is distributed by AMD under the MIT license and is available through the official Claude Code plugin marketplace as part of the AMD skills package. It does not declare an MCP dependency or a restricted frontmatter tool list.