sglang-glm46-glm47-optimization

Document SGLang optimizations for GLM-4.6/4.7 including shared-expert fusion and MoE GEMM overlap.

721|65|Updated Apr 1, 2026
One-click install
npx skills add https://github.com/BBuf/AI-Infra-Auto-Driven-SKILLS --skill sglang-glm46-glm47-optimization
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: sglang-glm46-glm47-optimization
Source: https://github.com/BBuf/AI-Infra-Auto-Driven-SKILLS/tree/main/skills/model-optimization/sglang/sglang-glm46-glm47-optimization
Command: npx skills add https://github.com/BBuf/AI-Infra-Auto-Driven-SKILLS --skill sglang-glm46-glm47-optimization

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

PR-backed and current-main optimization manual for GLM-4.6, GLM-4.6V-adjacent text paths, GLM-4.7, and GLM-4.7-Flash in SGLang. Use when Codex needs to recover, extend, or audit GLM shared-expert fusion, dual-stream MoE GEMM overlap, GLM-4.7 tool parser, NVFP4/MTP, or GLM4-MoE-Lite/Flash loading, AMD/NPU validation.

Core Features & Use Cases

  • PR-based optimization dossiers and diff auditing for GLM-4.6/4.7 lanes.
  • Guidance for GLM-4.7-Flash/Lite loading, MTP/draft quant config, and hardware backend validation.
  • Use cases include debugging, extension, and production validation of shared-expert fusion and MoE pathways across GLM-4.6/4.7.

Quick Start

Consult the PR history, run through the audit steps, and validate GLM-4.6/4.7 optimizations against runtime tests.

Frequently Asked Questions about sglang-glm46-glm47-optimization

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

FAQPage Schema
How do I optimize GLM-4.7 shared-expert fusion in SGLang?

To optimize GLM-4.7 shared-expert fusion in SGLang, consult the PR-backed optimization manual to recover and extend fusion pathways. This involves auditing PR-level changes and validating against runtime tests for production-ready workflows.

What is dual-stream MoE GEMM overlap for GLM-4.6?

Dual-stream MoE GEMM overlap for GLM-4.6 is an optimization technique that overlaps Matrix Multiplication operations in Mixture of Experts pathways. It is documented through PR-based dossiers and diff auditing to ensure efficient production execution.

How do I handle the GLM-4.7 tool parser in SGLang?

Handling the GLM-4.7 tool parser in SGLang requires consulting the PR-backed optimization documentation. This manual captures parser handling logic, ensuring accurate, audit-ready documentation validated against runtime tests and hardware backends.

Does SGLang support NVFP4 and MTP validation for GLM-4.7-Flash?

SGLang supports NVFP4 and MTP validation for GLM-4.7-Flash. The optimization playbook provides guidance for loading, MTP and draft quant configuration, alongside validation implications for AMD and NPU hardware backends.

Why does GLM-4.7-Flash loading fail during quantization configuration?

GLM-4.7-Flash loading issues during quantization often stem from incorrect MTP or draft quant configuration. Use the optimization playbook to audit loading procedures, verify PR histories, and validate hardware backend compatibility.

Can I validate GLM-4.6 MoE optimizations on AMD or NPU hardware backends?

Yes, you can validate GLM-4.6 MoE optimizations on AMD or NPU hardware backends. The playbook includes validation implications and audit-ready documentation to ensure shared-expert fusion and MoE pathways function correctly across hardware.