vllm-ascend-model-adapter

Adapt Hugging Face/local ML models for vLLM Ascend hardware with validated, signed commits and documentation.

3|17|Updated Mar 21, 2026
One-click install
npx skills add https://github.com/intellistream/vllm-ascend-hust --skill vllm-ascend-model-adapter-intellistream
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: vllm-ascend-model-adapter
Source: https://github.com/intellistream/vllm-ascend-hust/tree/main/.agents/skills/vllm-ascend-model-adapter
Command: npx skills add https://github.com/intellistream/vllm-ascend-hust --skill vllm-ascend-model-adapter-intellistream

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This Skill provides a structured workflow to adapt and validate ML models for vLLM on Ascend hardware, ensuring compatibility, deterministic startup, and a clean single-commit delivery.

Core Features & Use Cases

  • Model adaptation: prepare Hugging Face or local models for vLLM Ascend integration with fixed implementation roots and clear validation steps.
  • Deterministic validation: stage-based checks (dummy and real-weight gates) to verify startup, inference, and feature support (ACLGraph, EP, MTP, multimodal where applicable).
  • Handoff-ready delivery: produce a single signed commit with minimal, auditable changes and documentation in Chinese per project requirements.

Quick Start

Run the two-stage vLLM Ascend adaptation workflow from /workspace and verify startup, smoke tests, and a single signed commit deliverable.

Frequently Asked Questions about vllm-ascend-model-adapter

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

FAQPage Schema
How do I adapt a Hugging Face model for vLLM Ascend inference?

To adapt a Hugging Face model for vLLM Ascend, this Skill applies a structured workflow that validates deterministic startup and inference, delivering a single signed commit for integration.

What is the best way to validate model inference on Ascend hardware?

Validating model inference on Ascend hardware requires stage-based checks using dummy and real-weight gates to verify startup, smoke tests, and feature support across ACLGraph and EP paths.

Does vLLM Ascend support multimodal model adaptation?

vLLM Ascend adaptation supports multimodal paths where applicable, running stage-based validations to ensure deterministic startup, inference, and feature compatibility for the target model.

How to ensure deterministic startup when deploying models on vLLM Ascend?

Ensuring deterministic startup for vLLM Ascend deployment involves running a two-stage adaptation workflow from the workspace that executes fixed implementation roots and validates startup gates.

What are the limitations of adapting models for vLLM Ascend?

Adapting models for vLLM Ascend requires fixed implementation roots and specific feature support validation for ACLGraph, EP, and MTP, meaning unsupported features will fail the stage-based validation gates.

Do I need local model weights to validate vLLM Ascend integration?

You do not need real weights initially; vLLM Ascend integration validation uses a two-stage process starting with dummy weight gates before proceeding to real-weight inference checks.