skills.rest
JiashuWu
JiashuWuCOMMUNITY
Explore all skills by this author →
GitHub

vllm-ascend-model-adapter

Guide vLLM model adaptation and validation for Ascend NPUs.

Updated Aug 23, 2026
Search Tags:#vllm#ascend#fp8 dequantization#model adapter#checkpoint remapping#multimodal validation#e2e documentation
One-click install
npx skills add https://github.com/JiashuWu/vllm-ascend-sur-20260529 --skill vllm-ascend-model-adapter-jiashuwu
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: vllm-ascend-model-adapter
Source: https://github.com/JiashuWu/vllm-ascend-sur-20260529/tree/main/vllm-ascend-base-v0.18.0rc1-kvquant-feature-3b042fa/.agents/skills/vllm-ascend-model-adapter
Command: npx skills add https://github.com/JiashuWu/vllm-ascend-sur-20260529 --skill vllm-ascend-model-adapter-jiashuwu

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill reduces the uncertainty and iteration cost of bringing existing or new model architectures to vLLM on Ascend NPUs by guiding minimal, auditable code changes plus deterministic dummy-first and real-weight validation.

Core Features & Use Cases

  • Model adaptation for vLLM on Ascend: maps/implements model adapters, processors, registry registration, and explicit checkpoint weight remap rules without upgrading transformers.
  • Deterministic validation workflow: performs a staged verification flow—first a dummy fast gate for architecture/operator/API path, then a mandatory real-weight gate for load-time and first-request stability.
  • Feature-first troubleshooting: attempts ACLGraph, EP (MoE only), flashcomm1 (MoE only), MTP, and multimodal smoke checks with evidence capture and a fallback ladder when failures occur.
  • Delivery-ready documentation and artifacts: generates Chinese analysis, Chinese runbook, an e2e test config YAML, a model tutorial doc, and enforces a single signed commit containing all changes.

Quick Start

Ask the assistant to adapt your target checkpoint on Ascend by following the dummy-first then real-weight mandatory validation steps and producing the final single signed commit plus Chinese report/runbook and the required test config and tutorial.

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 vLLM model to run on Ascend NPUs?▼

To adapt vLLM models for Ascend NPUs, you need to implement model adapters, processors, and checkpoint weight remapping, followed by a mandatory two-stage dummy and real-weight verification to ensure stability.

What is the dummy-first validation workflow for vLLM on Ascend?▼

The dummy-first validation workflow for vLLM on Ascend performs a fast dummy gate to check architecture and API paths, followed by a mandatory real-weight gate to verify load-time and first-request stability before committing.

How do I handle fp8 dequantization when porting models to vLLM Ascend?▼

Handling fp8 dequantization for vLLM Ascend requires explicit checkpoint weight remap rules and adapter changes, ensuring proper fp8-origin handling without upgrading transformers during the model adaptation process.

Does vLLM on Ascend support multimodal model validation?▼

Yes, vLLM on Ascend supports multimodal validation by executing multimodal smoke checks during the feature-first troubleshooting stage to capture evidence and verify functionality before finalizing the adaptation.

What are the limitations of adapting vLLM architectures for Ascend NPUs?▼

Limitations include strict implementation roots confined to /vllm-workspace/vllm and /vllm-workspace/vllm-ascend directories, and the requirement to bundle all changes into a single signed commit.

Can I use ACLGraph and MoE features with vLLM on Ascend?▼

Yes, vLLM on Ascend attempts feature-first troubleshooting using ACLGraph, EP for MoE only, and flashcomm1 for MoE only, utilizing a fallback ladder to capture evidence when failures occur.

Related Skills

Explore More Software Engineering Skills→
nv-actionnv-action
4

vllm-ascend-model-adapter

Adapt and validate models for vLLM Ascend NPU with a signed commit.

Official
Advanced
thoreinsteinthoreinstein
5

add-test-coverage

Analyze the latest HEAD commit and generate a test coverage plan.

Community
Advanced
vllm-projectvllm-project
2.6k

vllm-ascend-model-adapter

Adapts ML models for vLLM-Ascend NPUs with deterministic validation and signing.

Official
Advanced
vLLM-HUSTvLLM-HUST
3

vllm-ascend-model-adapter

Adapt and validate models for vLLM on Ascend hardware with a signed commit.

Official
Advanced