Overview

This tool generates interactive HTML dashboards for analyzing PyTorch profiler trace files (.pt.trace.json.gz) from vLLM workloads. It provides visualization capabilities for GPU kernel profiling and performance analysis.

Features

• Interactive HTML Dashboard: Generate a standalone HTML report to zoom, pan, and inspect individual kernel execution events.
• Side-by-Side Trace Comparison: Compare two trace files to easily spot regressions or improvements in kernel-level performance.
• Lightweight and Portable: Outputs a self-contained HTML file viewable in any modern browser. You can open it in your phone, laptop and tablet.

Prerequisites

• OS: Linux
• Python: 3.10 - 3.12
• vLLM profiling refer to https://docs.vllm.ai/en/v0.5.5/dev/profiling/profiling_index.html

Environment Setup

Option 1: Using venv (Recommended)

1. Create virtual environment:

   python3 -m venv tracemap
   source tracemap/bin/activate

2. Install dependencies:

   pip install --upgrade pip
   pip install bokeh pandas numpy

Usage

Basic Usage

Run the profiling dashboard generator with default trace files:

python3 main.py

Custom Trace Files

Specify your own trace files for comparison:

python3 main.py \
    --trace1 /path/to/first_trace.pt.trace.json.gz \
    --trace2 /path/to/second_trace.pt.trace.json.gz \
    --name1 "trace1 name" \
    --name2 "trace2 name" \
    --output custom_dashboard.html

Command Line Arguments

• --trace1: Path to first trace file (default: ./trace_file/examples/trace1.pt.trace.json.gz)
• --trace2: Path to second trace file (default: ./trace_file/examples/trace2.pt.trace.json.gz)
• --name1: Name for first trace (default: Trace_A)
• --name2: Name for second trace (default: Trace_B)
• --output: Output HTML file name (default: gpu_trace_profiling.html)

Output

The tool generates an interactive HTML dashboard that includes:

• GPU kernel execution timelines
• Performance comparisons between different traces
• Interactive visualizations for detailed analysis
• Summary statistics and profiling metrics

Open the generated HTML file in your web browser to explore the profiling results.