Analyse de traces TorchProfiler

Analyser les fichiers de trace PyTorch profiler et vLLM avec des preuves SQL Perfetto. Utiliser pour les hotspots GPU, attentes de synchronisation, déséquilibre de rang, goulots préfill/décodage.

Spar Skills Guide Bot
Data & IAAvancé
2022/07/2026
Claude Code
#pytorch-profiler#vllm#perfetto#trace-analysis#gpu-profiling

Recommandé pour


name: torchprofiler-trace-analysis description: Analyze PyTorch profiler and vLLM trace files with Perfetto SQL evidence. Use when working with torch profiler traces, merged_trace files, vLLM profiling logs, Perfetto trace analysis, GPU kernel hotspots, synchronization waits, rank imbalance, prefill/decode bottlenecks, or optimization reports.

TorchProfiler Trace Analysis

Purpose

Use this skill to analyze PyTorch profiler traces from vLLM/VAP runs. Prefer evidence from Perfetto SQL and structured trace metadata over raw JSON snippets.

Attribution

This project is inspired by the evidence-driven workflow design of:

  • Gracker/Perfetto-Skills: standard Agent Skill structure, workflow routing, SQL-backed evidence, and report contracts.
  • Gracker/SmartPerfetto: AI-assisted Perfetto analysis, evidence workflows, reports, and trace-processor-backed SQL analysis.

No code or SQL is copied from those projects. The SQL presets and workflows in this project are original and specialized for PyTorch profiler / vLLM traces.

Workflow

  1. Identify the latest or requested trace.
  2. Prefer merged traces named like *-merged_trace.json or *-merged_trace.json.gz.
  3. Run trace overview queries first.
  4. Collect evidence for:
    • synchronization waits
    • GPU kernel hotspots
    • CPU operator hotspots
    • rank imbalance
    • memory copies
    • prefill/decode spans
    • idle gaps
  5. Summarize findings with evidence.
  6. Separate confirmed evidence from hypotheses.
  7. Recommend next inspections in Perfetto/TensorBoard.

Report Format

Use this structure:

# TorchProfiler Trace Report

## Executive Summary
Short summary of the most likely bottleneck.

## Trace Metadata
- Trace file:
- Merged trace:
- Event count:
- Time span:
- Ranks:

## Evidence
| Area | Evidence | Interpretation |
|---|---|---|

## Bottleneck Hypotheses
1. Hypothesis with supporting evidence.

## Perfetto Inspection Guide
- Tracks/events to inspect next.

## Optimization Suggestions
- Concrete model/config/benchmark changes to try.

## Evidence Gaps
- Missing data or uncertainty.

Safety

Do not infer causal conclusions from a query that only parsed successfully. Label uncertain conclusions as hypotheses. Do not request raw trace JSON unless a focused preview is necessary.

Utilities

Skills similaires