vllm-ascend-profiler-diff

Compare two Ascend profiler outputs to quantify performance deltas and regressions.

3|1|Updated Feb 2, 2026
One-click install
npx skills add https://github.com/sunchendd/good_skills --skill vllm-ascend-profiler-diff
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: vllm-ascend-profiler-diff
Source: https://github.com/sunchendd/good_skills/tree/main/vllm-ascend-profiler-diff
Command: npx skills add https://github.com/sunchendd/good_skills --skill vllm-ascend-profiler-diff

SYSTEM DOCUMENTATION & REQUIREMENTS

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

What problem does it solve?

This skill provides a structured workflow to compare two Ascend Profiler runs, quantify performance deltas, and optionally trace findings back to source and PRs. The design supports evaluating prefill/decoding, kernel/op deltas, Host/sync behavior, and trace distribution to pinpoint regressions across runs.

Core Features & Use Cases

  • Compare two ASCEND_PROFILER_OUTPUT directories to surface performance deltas and regressions.
  • Step through sanity checks, overall metrics alignment, stage-level attribution, kernel/op dimension, API signals, and optional source/PR attribution.
  • Use cases include debugging performance regressions between baseline and newer runs, attributing changes to specific modules, and generating evidence-backed reports.

Quick Start

Provide Profile A and Profile B ASCEND_PROFILER_OUTPUT directories and run the workflow to generate a comparative report.

Frequently Asked Questions about vllm-ascend-profiler-diff

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

FAQPage Schema
How do I compare two Ascend profiler outputs to find performance regressions?

To compare Ascend profiler outputs and find regressions, provide two ASCEND_PROFILER_OUTPUT directories. The analysis quantifies performance deltas across prefill, decode, kernel, and operator dimensions to generate a structured report.

What is Ascend profiler diff analysis and when do I need it?

Ascend profiler diff analysis evaluates performance deltas between two profiling runs. You need it when debugging regressions between a baseline and a newer run, attributing changes to specific modules, or generating evidence-backed reports.

How do I attribute Ascend profiler regressions to specific source code and PRs?

To attribute Ascend profiler regressions to source and PRs, provide a path to the source repository and version tags alongside the two ASCEND_PROFILER_OUTPUT directories to anchor findings and trace changes to specific commits.

Can I use benchmark JSON files to align overall metrics during an Ascend profiler comparison?

Yes, you can provide optional Benchmark JSONs alongside the Profile A and Profile B ASCEND_PROFILER_OUTPUT directories. These files support overall metrics alignment and sanity checks during the comparative analysis workflow.

What dimensions does the Ascend profiler comparison analyze for performance deltas?

The Ascend profiler comparison analyzes performance deltas across prefill and decode phases, kernel and operator deltas, host synchronization behavior, and trace distribution to pinpoint regressions.