skills.rest
PaddlePaddle
PaddlePaddleOFFICIAL
Explore all skills by this author →
WebsiteGitHub

nsys-capture

Inject nsys profiling into FastDeploy GPU inference startup and collect .nsys-rep files.

3.7k|758|Updated Jun 27, 2022
Search Tags:#nsys#gpu-profiling#fastdeploy#nvprof#profiling-automation#start_nsys
One-click install
npx skills add https://github.com/PaddlePaddle/FastDeploy --skill nsys-capture
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: nsys-capture
Source: https://github.com/PaddlePaddle/FastDeploy/tree/main/.claude/skills/nsys-capture
Command: npx skills add https://github.com/PaddlePaddle/FastDeploy --skill nsys-capture

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires openai.

What problem does it solve?

This Skill provides a deterministic GPU profiling workflow for FastDeploy inference services using Nsight Systems (nsys). It automates injection of profiling hooks into the codebase, script generation for launching with nsys, and end-to-end workflows to collect .nsys-rep files for performance analysis.

Core Features & Use Cases

  • Automatic detection and injection of nvprof_start / nvprof_stop calls into fastdeploy/worker/gpu_model_runner.py
  • Automated generation of a start_nsys.sh script that wraps the actual startup command with NSYS_CMD
  • End-to-end profiling pipeline: information collection, profiling, script generation, user confirmation, execution, and file collection
  • Works in scenarios where users want to profile GPU inference services and analyze operator-level and CUDA graph traces

Quick Start

Run the end-to-end workflow to profile your FastDeploy GPU inference service with nsight systems and generate a .nsys-rep file.

Frequently Asked Questions about nsys-capture

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

FAQPage Schema
How do I profile a FastDeploy GPU inference service with nsys?▼

Profiling FastDeploy GPU inference with nsys involves automatically injecting nvprof hooks into the startup code and generating a shell script to capture operator-level traces into a .nsys-rep file.

What is the process to capture operator-level CUDA traces for local GPU model deployments?▼

Capturing operator-level CUDA traces involves an automated pipeline that injects configurable nvprof_start and nvprof_stop hooks into the model runner, then executes an nsys command to generate trace files.

Does GPU profiling automation work with multiple local model deployments simultaneously?▼

GPU profiling automation supports local inference servers across multiple models and deployments, guiding the workflow from initial information collection through to final output file generation.

Do I need to manually edit gpu_model_runner.py to start nsys profiling?▼

Manual editing of gpu_model_runner.py is unnecessary for nsys profiling because the workflow automatically detects the file and injects the required nvprof_start and nvprof_stop calls at configurable steps.

What is the best way to generate a start_nsys.sh script for profiling automation?▼

Generating a start_nsys.sh script for profiling automation is best handled by this workflow, which wraps your FastDeploy startup command with the appropriate NSYS_CMD to capture profiling data.

Related Skills

Explore More Software Engineering Skills→
PaddlePaddlePaddlePaddle
3.7k

fastdeploy-llm-integration

Automate adding inference deployment support for open-source LLMs to FastDeploy.

Official
Advanced
mlc-aimlc-ai
327

capture-nsys-profile

Capture Nsight Systems traces for PithTrain runs across parallelism configurations.

Official
Intermediate
NickCrewNickCrew
29

python-performance-optimization

Profile Python code with cProfile and line_profiler to optimize performance.

Community
Advanced
LancerLabLancerLab

croq-dsl-triton

Tunes and optimizes Triton GPU kernels with build, run, and profiling templates.

Official
Advanced