skills.rest
ProgMastermind
ProgMastermindCOMMUNITY
Explore all skills by this author →
Website𝕏@PrakashCollymo1GitHub

capture-trace

Capture and analyze PyTorch profiler and kineto traces from ATOM servers.

Updated Jul 15, 2026
Search Tags:#performance analysis#performance tracing#ATOM server#PyTorch profiler#kineto trace
One-click install
npx skills add https://github.com/ProgMastermind/ATOM --skill capture-trace
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: capture-trace
Source: https://github.com/ProgMastermind/ATOM/tree/main/.claude/skills/capture-trace
Command: npx skills add https://github.com/ProgMastermind/ATOM --skill capture-trace

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires torch, kineto, and includes scripts (resource) and references (resource) components.

What problem does it solve?

This Skill helps users capture and analyze performance traces from running ATOM servers, aiding in performance investigation and optimization.

Core Features & Use Cases

  • Trace Capture: Automatically capture PyTorch profiler/kineto traces for performance analysis.
  • Short Benchmark Window: Supports capturing traces from short benchmark windows to ensure readability and prevent OOM errors.
  • Pre-flight Checks: Ensures that the server is stopped cleanly, a short workload is chosen, and the trace directory is empty.
  • Required Tools: Provides a list of required tools for successful trace capture.
  • Workflow Guide: Offers a step-by-step guide to launch the server, drive the benchmark, wait for exporter completion, and inspect the trace layout.
  • Inspecting Traces: Includes commands to verify the layout of the captured traces and inspect kernel events.

Quick Start

To capture a trace, first launch the server with the profiler enabled, then run the benchmark, and finally wait for the exporter to finish.

Frequently Asked Questions about capture-trace

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

FAQPage Schema
How do I capture PyTorch profiler traces for performance investigation?▼

To capture PyTorch profiler traces, launch the ATOM server with the profiler enabled, run a short benchmark workload, and wait for the exporter to finish generating the kineto trace files.

What is the best way to prevent out-of-memory errors during kineto trace capture?▼

The best way to prevent OOM errors during kineto trace capture is to use a short benchmark window, which ensures the resulting trace files remain readable and manageable for performance analysis.

Do I need to run pre-flight checks before capturing a performance trace?▼

Yes, pre-flight checks are required to ensure the server is stopped cleanly, a short workload is chosen, and the trace directory is empty before capturing performance traces.

Can I use this performance tracing workflow on ROCm platforms with ATOM models?▼

Yes, performance tracing supports various ATOM models and ROCm platforms, requiring the PyTorch profiler and kineto Python libraries to capture and analyze traces successfully.

How do I inspect kernel events after capturing a kineto trace?▼

You inspect kernel events by running specific commands to verify the layout of the captured kineto traces, allowing you to analyze the performance data for optimization.

What Python dependencies are required for ATOM server performance tracing?▼

ATOM server performance tracing requires the torch and kineto Python libraries to capture PyTorch profiler traces and analyze the resulting performance data for investigation.

Related Skills

Explore More Software Engineering Skills→
Fridge003Fridge003
9

sglang-torch-profiler-analysis

Analyze SGLang torch-profiler traces to map GPU kernels to Python frames and CPU ops.

Community
Advanced
michaeldubumichaeldubu

perf-analyzer

Parse and analyze CUDA, CPU, and memory benchmark outputs across CSV, JSON, and text formats.

Community
Advanced
ProgMastermindProgMastermind

run-atom-workload

Automate ATOM workload execution for accuracy evaluation, performance benchmarking, and fault reproduction.

Community
Advanced
nishide-devnishide-dev

ml-profile

Profile PyTorch Lightning training to diagnose data loading, computation, and GPU memory bottlenecks.

Community
Intermediate