sglang-deepseek-v31-optimization

Optimize DeepSeek V3.1 tool calling, thinking mode, chat templates, and streaming parser behavior in SGLang.

721|65|Updated Apr 1, 2026
One-click install
npx skills add https://github.com/BBuf/AI-Infra-Auto-Driven-SKILLS --skill sglang-deepseek-v31-optimization
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: sglang-deepseek-v31-optimization
Source: https://github.com/BBuf/AI-Infra-Auto-Driven-SKILLS/tree/main/skills/model-optimization/sglang/sglang-deepseek-v31-optimization
Command: npx skills add https://github.com/BBuf/AI-Infra-Auto-Driven-SKILLS --skill sglang-deepseek-v31-optimization

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

The DeepSeek V3.1 optimization manual guides engineers in auditing, recovering, extending, and debugging tool calling, thinking mode, chat templates, streaming parser behavior, loading fixes, and MTP validation for SGLang deployments.

Core Features & Use Cases

  • PR-backed optimization guidance for DeepSeek V3.1 and V3.1-Terminus across SGLang.
  • Cross-surface guidance covering parser, runtime, and template interactions, including MoE backend configs and backend-specific tests.
  • Use Case: When Codex or SGLang needs to recover or audit DeepSeek V3.1 on a new PR, this playbook provides steps and checks.

Quick Start

Run a focused optimization pass on DeepSeek V3.1 tool calling, thinking mode, chat templates, and streaming parser behavior.

Frequently Asked Questions about sglang-deepseek-v31-optimization

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

FAQPage Schema
How do I optimize DeepSeek V3.1 tool calling and thinking mode in SGLang?

To optimize DeepSeek V3.1 tool calling and thinking mode in SGLang, apply PR-backed playbook steps covering chat templates, streaming parser behavior, and MTP validation to ensure robust runtime and template interactions across current-main and Terminus variants.

What is the best way to debug DeepSeek V3.1 streaming parser regressions in SGLang?

Debug DeepSeek V3.1 streaming parser regressions in SGLang by cross-referencing PR histories and applying backend-specific tests. This playbook guides debugging across parser, runtime, and template surfaces within CI and nightly tests to recover robust behavior.

Does the SGLang optimization playbook support DeepSeek V3.1-Terminus variants?

Yes, the SGLang optimization playbook explicitly supports both current-main and DeepSeek-V3.1-Terminus variants. It provides cross-surface guidance covering MoE backend configs, model loading fixes, and template interactions for these specific deployments.

How do I validate fused MoE configs and MTP loading paths for DeepSeek V3.1?

Validate fused MoE configs and MTP loading paths for DeepSeek V3.1 by referencing PR histories and backend-specific tests. The playbook provides validation steps to ensure reproducible improvements and robust model loading behavior in SGLang.

Why does my DeepSeek V3.1 chat template break after a new SGLang PR?

A DeepSeek V3.1 chat template may break after a new SGLang PR due to back-compat issues across parser and template surfaces. Use this playbook to audit PR-driven changes, recover regressions, and apply targeted fixes for template interactions.