skills.rest
primatrix
primatrixOFFICIAL
Explore all skills by this author →
GitHub

xprof-profiling-analysis

Analyze TPU/GPU performance with operator breakdowns, memory profiles, and A/B comparisons.

5|Updated Feb 4, 2026
Search Tags:#machine learning#profiling#performance analysis#TPU/GPU#XProf
One-click install
npx skills add https://github.com/primatrix/skills --skill xprof-profiling-analysis
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: xprof-profiling-analysis
Source: https://github.com/primatrix/skills/tree/main/plugins/xprof-profiling-analysis/skills/xprof-profiling-analysis
Command: npx skills add https://github.com/primatrix/skills --skill xprof-profiling-analysis

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) and assets (resource) components.

What problem does it solve?

This Skill allows for detailed analysis of TPU/GPU training performance using XProf MCP tools, providing insights into operator breakdowns, memory usage, and performance comparisons.

Core Features & Use Cases

  • Performance Analysis: Utilizes XProf MCP tools to perform live queries and deep domain analysis for trace parsing and interpretation.
  • Framework Op Breakdown: Provides detailed JAX-level operation statistics with the ability to filter by operation category.
  • Memory Analysis: Analyzes HBM usage, fragmentation, and buffer allocations to identify memory issues.
  • Quick Recipes: Offers concise code snippets for common analysis tasks.
  • Offline CLI Tool: Utilizes xprof.py for additional deep analysis when MCP data is insufficient.
  • Use Case: A user may run this skill after training a machine learning model to diagnose bottlenecks, optimize memory usage, and enhance overall performance.

Quick Start

Run the xprof_profiling_analysis skill with the latest profiling run to get an overview of MXU utilization and identify key performance metrics.

Frequently Asked Questions about xprof-profiling-analysis

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I analyze TPU performance bottlenecks using XProf trace data?▼

To analyze TPU performance bottlenecks, this skill parses XProf trace data to provide operator breakdowns and calculate MXU utilization, identifying exactly where training bottlenecks occur. It offers live queries and deep domain analysis to interpret the trace.

Can I break down JAX operations by category during GPU profiling?▼

Yes, you can break down JAX operations by category during GPU profiling. The skill provides detailed framework operation statistics, allowing you to filter operation categories to isolate specific JAX-level performance metrics.

What is the best way to diagnose HBM memory fragmentation and buffer allocations on a TPU?▼

The best way to diagnose HBM memory fragmentation is by analyzing HBM usage and buffer allocations. This skill performs memory analysis to identify fragmentation issues and buffer limits during your model training.

How do I calculate MFU and compare performance across different machine learning training runs?▼

You calculate MFU and compare performance across runs by utilizing the skill's A/B comparison feature. It supports MFU calculation and direct performance comparisons to evaluate different training configurations.

What should I do if XProf MCP data is insufficient for my performance analysis?▼

If XProf MCP data is insufficient for performance analysis, you can utilize the offline CLI tool. The skill incorporates an `xprof.py` CLI tool to perform additional deep analysis when live queries fall short.

Related Skills

Explore More Data & Analytics Skills→
ByteDance-SeedByteDance-Seed
2.1k

veomni-profile

Profile deep learning training runs by parsing Chrome traces and memory snapshots.

Official
Intermediate
goupup-aigoupup-ai
2

system-profile

Profile scripts, processes, GPUs, memory, and interconnects for compute bottlenecks.

Community
Intermediate
doanchienthangdevdoanchienthangdev
4

ai-accelerators

Optimizes machine learning workloads on GPUs, TPUs, and edge accelerators with CUDA, TensorRT, and TFLite.

Community
Advanced
aayushsoamaayushsoam
1

pytorch-lightning

Train scalable PyTorch models with automatic distributed training across hardware configurations.

Community
Advanced