SGLang DeepSeek V3.1 Optimization

Overview

This skill covers the DeepSeek V3.1 support and optimization ladder that is active in SGLang main. V3.1 shares the DeepSeek V3/R1 model backbone, but its tool-call format, chat template, thinking flag, streaming parser, and validation lanes are separate enough to require an independent skill.

Current-main snapshot:

• SGLang origin/main: 929e00eea on 2026-04-21
• sgl-cookbook origin/main: 8ec4d03 on 2026-04-21
• core runtime entry: python/sglang/srt/models/deepseek_v2.py
• V3.1 tool parser: python/sglang/srt/function_call/deepseekv31_detector.py
• V3.1 tool template: examples/chat_template/tool_chat_template_deepseekv31.jinja

The historical evidence lives in:

• references/pr-history.md: chronological PR evidence and code-level notes
• references/playbook.md: investigation order, symptom mapping, validation commands