vllm-ascend Workflow Skills

English | 中文

A collection of OpenCode Agent Skills providing development workflow support for the vllm-ascend (vLLM Ascend NPU plugin) project.

Skills Overview

Skill	Status	Description
vllm-ascend-feature-learning	Available	Generate Chinese technical learning docs for vllm-ascend features
vllm-ascend-feature-develop	Available	Assist vllm-ascend feature development and code implementation

vllm-ascend-feature-learning

Generate comprehensive Chinese code walkthrough learning documents for features/modules in the vllm-ascend repository.

Trigger

Request to learn a vllm-ascend feature in the OpenCode TUI, e.g.:

我想了解 vllm-ascend 的 ACL Graph
帮我生成 vllm-ascend patch 系统的学习文档
学习 vllm-ascend 的 MLA attention 后端

Output Format

Generated learning documents include:

• Mermaid architecture/flow/sequence/class diagrams — Visualize system architecture and execution flow
• GPU vs NPU comparison tables — Compare upstream vLLM and vllm-ascend implementation differences
• Source code walkthrough — Reference real vllm-ascend code with file paths
• Configuration reference — List all relevant environment variables and config parameters
• Plugin integration analysis — Analyze patch/inheritance/entry point integration mechanisms

Documents are output to the docs/feature_learning/ directory.

Covered Features

Domain	Features
Plugin Architecture	Plugin Registration, Patch System, NPUPlatform, AscendConfig
Attention Backends	MLA, SFA, DSA, FA3, Context Parallelism
Distributed	FlashComm, MC2, HCCL, PD Disaggregation, KV Transfer
Graph Optimization	ACL Graph, npugraph_ex, Graph Fusion Pass
Quantization	W8A8, W4A16, FP8, KV C8, W4A4 series
MoE	Fused MoE, EPLB, Token Dispatcher
Speculative Decoding	Eagle, Medusa, N-gram, Rejection Sampler
Memory Management	CaMem Allocator, KV Offload, NZ Format
Scheduling	Dynamic Batch, Profiling Chunk, Balance Scheduling
Hardware-Specific	310P Support, Xlite, CPU Binding

Full feature list: feature-catalog.md.

vllm-ascend-feature-develop

Assist vllm-ascend feature development with code implementation, design document generation, and technical solutions.

Trigger

帮我在 vllm-ascend 中实现 W4A16 量化支持
为 vllm-ascend 添加 xxx 特性
vllm-ascend xxx 特性开发

References

• patch-system.md — Patch system usage guide
• feature-patterns.md — Feature development patterns and examples
• focus-areas.md — Key focus areas

Design documents are output to the docs/feature_develop/ directory.

Generated Documents

Learning Documents (docs/feature_learning/)

Document	Topic
llm_inference_pipeline.md	LLM Inference Pipeline
vllm_inference_pipeline.md	vLLM Inference Pipeline
pipeline_step.md	Inference Pipeline Steps
pooling.md	Pooling Feature
quantization.md	Quantization Overview
quantization_vllm_ascend.md	vllm-ascend Quantization
speculative_decoding.md	Speculative Decoding

Design Documents (docs/feature_develop/)

Document	Topic
design-w4a16-quantization-adaptation-guide.md	W4A16 Quantization Adaptation Design

Directory Structure

workflow/
├── code/                                       # Code workspace (reserved)
├── docs/                                       # Generated document output
│   ├── feature_develop/                        # Feature design docs
│   ├── feature_insight/                        # Feature deep insights
│   └── feature_learning/                       # Feature learning docs
├── sessions/                                   # Session records
├── skills/                                     # Agent Skills definitions
│   ├── basic-skills/                           # Basic skills
│   │   ├── vllm-ascend-feature-develop/        # Feature development
│   │   │   ├── SKILL.md
│   │   │   └── references/
│   │   │       ├── feature-patterns.md
│   │   │       ├── focus-areas.md
│   │   │       └── patch-system.md
│   │   └── vllm-ascend-feature-learning/       # Feature learning docs
│   │       ├── SKILL.md
│   │       └── references/
│   │           ├── feature-catalog.md
│   │           └── style-guide.md
│   ├── kv-pooling-skills/                      # KV pooling skills (reserved)
│   ├── quantization-skills/                    # Quantization skills (reserved)
│   └── speculative-decoding-skills/            # Spec decode skills (reserved)
└── README.md

Prerequisites

• Install OpenCode
• Configure LLM Provider API Key
• Clone vllm-ascend repository locally (for source code walkthrough)

Installation

Configure the skills directory in OpenCode's skills path. Add to opencode.json:

{
  "skills": {
    "paths": ["./skills"]
  }
}

License

This project follows the same license as vllm-ascend.