regression

Compare and profile SGLang serving configurations with automated accuracy and performance tests.

1|Updated Jun 2, 2026

Search Tags:#profiling #regression testing #model evaluation #SGLang #serving configuration

Ready to run

npx skills add https://github.com/yushengsu-thu/tune-lora-perf --skill regression-yushengsu-thu

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: regression
Source: https://github.com/yushengsu-thu/tune-lora-perf/tree/main/regression
Command: npx skills add https://github.com/yushengsu-thu/tune-lora-perf --skill regression-yushengsu-thu

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires python, docker, torch, torch-profiler, and includes scripts (resource) and references (resource) components.

What problem does it solve?

This Skill automates the comparison and profiling of SGLang serving configurations, streamlining the process of testing and validating changes.

Core Features & Use Cases

Base vs Variant Comparison: Compare a base (control) serving configuration against a variant, including LoRA toggles, backend swaps, and environment variable changes.
Automated Testing: Execute a suite of tests (accuracy, performance, prompt-check, profiling) for each configuration.
Profiling: Generate detailed profiling data for CPU and GPU usage.
Use Case: Use this Skill to test a new LoRA configuration for an SGLang model, comparing its accuracy, performance, and decode health against a baseline configuration.

Quick Start

Run the regression skill with the desired model and configuration:

run_regression.sh <model>

Frequently Asked Questions about regression

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

FAQPage Schema

How do I automate SGLang serving configuration comparisons?▼

SGLang serving configuration comparisons are automated by running a regression script that tests a base configuration against a variant, evaluating accuracy, performance, and decode health. This streamlines validating changes like LoRA toggles or backend swaps.

What is the best way to profile SGLang model performance?▼

To profile SGLang model performance, you can execute an automated testing suite that generates detailed profiling data for both CPU and GPU usage. This requires specific Python libraries and Docker to capture resource utilization metrics.

Do I need Docker to run SGLang regression testing?▼

Yes, Docker is required to run SGLang regression testing. The automated profiling and comparison workflows depend on Docker alongside specific Python libraries like torch and torch-profiler to execute the test suites and generate resource metrics.

Can I compare a new LoRA configuration against an SGLang baseline?▼

Yes, you can compare a new LoRA configuration against an SGLang baseline. The regression process executes accuracy, performance, prompt-check, and profiling tests to validate the variant serving configuration against the control.

How does automated profiling work for SGLang serving configurations?▼

Automated profiling for SGLang serving configurations works by executing a suite of tests that generates detailed CPU and GPU usage data. This process applies to model evaluation and serving configuration optimization workflows to validate changes.

What should I do if my SGLang serving configuration regression test fails?▼

If your SGLang serving configuration regression test fails, you should review the generated accuracy, performance, and profiling metrics. Comparing these detailed CPU and GPU outputs against your baseline configuration helps identify the specific degradation.