Analyze TPU pretraining compute efficiency from XSpace/XPlane profile data.
npx skills add https://github.com/primatrix/skills --skill compute-breakdown
Please help me install this Agent Skill. Skill: compute-breakdown Source: https://github.com/primatrix/skills/tree/main/plugins/tpu-perf/skills/compute-breakdown Command: npx skills add https://github.com/primatrix/skills --skill compute-breakdown
This Skill analyzes TPU pretraining compute efficiency, providing insights into performance bottlenecks and resource utilization.
Use the compute-breakdown skill to analyze the compute efficiency of the TPU pretraining profile located at '/path/to/profile.xplane.pb'.
High-intent search queries and answers about installing and using this skill.
To analyze TPU pretraining compute efficiency, process XSpace profile data to break down compute time, identify bottlenecks, and evaluate resource utilization. The skill parses XPlane and XLine data to pinpoint performance issues.
Roofline analysis evaluates TPU compute and memory utilization against theoretical peak performance. By comparing actual operations to peak limits, it identifies bottlenecks and reveals opportunities for optimizing pretraining workloads.
To identify non-compute operations slowing TPU training, analyze the profile for data movement and communication overhead. The skill isolates these non-compute operations to provide a comprehensive view of resource usage beyond pure computation.
Yes, xplane parsing and analysis tools are required dependencies. The compute breakdown skill relies on these tools to process underlying XEvent and XStat data for accurate layer-wise performance scoping.
Yes, you can pinpoint performance issues within specific layers using layer scoping. The skill analyzes compute time by layer or module, helping you locate exactly where performance bottlenecks occur during TPU pretraining.
ben-mad-jlpValidate Kodex topic claims against code and update inaccuracies.
haJ1tDiagnose and resolve application performance bottlenecks across frontend metrics, database queries, and caching.
leonling-llCollect timing, hardware counters, ATT traces, and memory-hierarchy data from AMD GPU kernels.
alibabaAnalyzes GPU kernel performance using Roofline models and profiling data.
