skills.rest
BBuf
BBufCOMMUNITY
Explore all skills by this author →
WebsiteGitHub

sglang-hunyuan3-preview-optimization

Generates audited optimization dossier for Hunyuan3 Preview BF16 MoE across GPUs.

721|65|Updated Apr 1, 2026
Search Tags:#documentation#optimization#sglang#gpu-sizing#pr-dossier#hunyuan3
One-click install
npx skills add https://github.com/BBuf/AI-Infra-Auto-Driven-SKILLS --skill sglang-hunyuan3-preview-optimization
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: sglang-hunyuan3-preview-optimization
Source: https://github.com/BBuf/AI-Infra-Auto-Driven-SKILLS/tree/main/skills/model-optimization/sglang/sglang-hunyuan3-preview-optimization
Command: npx skills add https://github.com/BBuf/AI-Infra-Auto-Driven-SKILLS --skill sglang-hunyuan3-preview-optimization

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This document consolidates PR-backed optimization guidance for Tencent Hunyuan 3 Preview in SGLang, enabling engineers to audit, extend, and deploy efficient BF16 MoE configurations with precise hardware sizing, parser flags, and attention-backend requirements for production readiness.

Core Features & Use Cases

  • PR-dossier guided optimization: links to per-PR evidence and diff-based justification, ensuring changes are traceable.
  • Hardware sizing and configuration: BF16 weight and TP planning across H200/H100 and B300/GB300 GPUs with model-specific constraints.
  • Safe integration and deployment: includes parser/tool flags, MTP options, EAGLE toggles, and trust-remote-code considerations.
  • Use Case: Create and review a deployment recipe for Hy3 Preview with tested GPUs and Blackwell attention backend settings.

Quick Start

Review the latest Hunyuan3 Preview optimization PR diffs and tailor the guidance to your hardware setup.

Frequently Asked Questions about sglang-hunyuan3-preview-optimization

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

FAQPage Schema
How do I optimize Hunyuan3 Preview for BF16 MoE deployment in SGLang?▼

Optimizing Hunyuan3 Preview for BF16 MoE in SGLang involves applying PR-backed guidance for hardware sizing, parser flags, and attention backend requirements to generate a production-ready deployment dossier.

What GPU hardware sizing is required for Tencent Hunyuan3 Preview?▼

Hardware sizing for Hunyuan3 Preview requires BF16 weight and tensor parallelism planning tailored for H200, H100, B200, B300, and GB300 GPUs to meet model-specific production constraints.

Does Hunyuan3 Preview deployment require the Blackwell attention backend?▼

Yes, deploying Hunyuan3 Preview on Blackwell architecture requires specific attention backend settings to ensure safe integration and efficient execution on B200, B300, and GB300 GPUs.

How do I configure MTP and EAGLE toggles for Hunyuan3 Preview?▼

Configuring MTP options and EAGLE toggles for Hunyuan3 Preview involves setting specific parser and tool flags during deployment to ensure safe integration and correct production behavior.

Do I need trust-remote-code enabled to run Hunyuan3 Preview in SGLang?▼

Trust-remote-code considerations are necessary when running Hunyuan3 Preview in SGLang, requiring specific parser flag configurations to maintain safe integration during deployment.

Why validate Hunyuan3 Preview optimization changes against PR history?▼

Validating optimization changes against PR history ensures diff-backed justification and traceability, maintaining auditability and alignment with documented production rules for Hunyuan3 Preview.

Related Skills

Explore More Software Engineering Skills→
BBufBBuf
721

sglang-mimo-v2-flash-optimization

Generate reproducible optimization plans for MiMo-V2-Flash in SGLang from PR diffs.

Community
Advanced
ovachieverovachiever
52

tensorrt-llm

Optimize LLM inference on NVIDIA GPUs with TensorRT-LLM.

Community
Advanced
jstzwjjstzwj
4

xformers

Provide optimized attention and sparse tensor operations for Transformer models.

Community
Advanced
zechenzhangAGIzechenzhangAGI
11.5k

moe-training

Train Mixture-of-Experts models with DeepSpeed MoE and HuggingFace Transformers.

Community
Advanced