Right-Hand Support — Feasibility Verification Plan for Ascend NPU Adaptation

Core Conclusion: Ascend NPU adaptation does not require starting from scratch. The CANN open-source community has released a complete Agent Skill system (ascend-agent-skills), supporting the entire process of model migration, optimization, and verification. The Qwen series models already have numerous deployment verification cases on Ascend NPU. Right-Hand Support has existing experience with edge-side model adaptation for Qualcomm Snapdragon and Moore Threads, making migration to Ascend NPU an engineering replication of a proven path.

Document Version: v1.0 | Update Date: 2026-06-22 | Belongs to: Right-Hand Support Project · Agent Hackathon 2026

1. Verification Approach: Not "Whether It's Possible," but "How to Verify"

Strategy: Utilize the existing toolchain and public cases from the Ascend community to quickly close the loop from "claim" to "proof."

已有验证（高通/摩尔线程）→ 昇腾社区工具链（开箱即用）→ 快速复现验证 → 形成可展示证据

二、可用的验证工具与资源

2.1 昇腾Model-Agent Skill（大赛指定工具）

大赛明确提供了昇腾NPU模型适配/调优Skill的使用方式：

# Register marketplace（克隆 git 仓库到本地）
/plugin marketplace add https://gitcode.com/gmq123/ascend-model-agent-plugin

# 安装 plugin
/plugin install ascend-model-agent-plugin@ascend-model-agent-plugin

该Skill的核心能力：

• Check Compatibility: Automatically detect model compatibility with Ascend NPU.
• Perform Optimization: Automated inference optimization (parallelization, KVCache, operator fusion, graph mode).
• Quick Deployment: One-click deployment of inference service.
• Stable Launch: Full verification and acceptance process.

2.2 CANN Open-Source Community Agent Skill System

The CANN open-source community has released an Agent Skill system for Ascend NPU inference optimization scenarios:

| Capability | Description | Validation Data |

|------|------|----------|

| Model Analysis & Baseline Establishment | Automatically analyzes model architecture and establishes performance baselines | Simple models optimized within 2 hours |

| Parallelization Adaptation | Automatically adapts to NPU parallel computing | Complex models complete closed-loop optimization in 6-8 hours |

| KVCache & FA Optimization | Memory optimization | Covers models like Qwen3.5-0.8B |

| Fused Operator Replacement | Operator-level optimization | Decode speedup 5.1x |

| Graph Mode Adaptation | Graph breakpoint fixes | End-to-end throughput improved 2.8x |

Measured Results (Qwen3.5-0.8B):

| Metric | Without Skill | With Skill | Improvement |

|------|------------|-----------|------|

| Decode Speedup | 1.1x | 5.1x | +360% |

| End-to-End Throughput | 11.16 tok/s | 31.59 tok/s | +183% |

| Fused Operator Coverage | 3 categories | Full coverage of 7 categories | +133% |

2.3 ascend-agent-skills Official Repository

The Ascend official Agent Skills repository already includes complete migration and validation Skills:

• Main Orchestration Skill: End-to-end workflow orchestration
• Model Migration Skill: Model format conversion and adaptation
• Data Migration Skill: Dataset adaptation
• Configuration Migration Skill: Configuration file conversion
• Validation & Acceptance Skill: Accuracy and performance validation

The repository currently contains 52 Skills, with 13 belonging to the migration and adaptation category.

III. Existing Validation Cases for Qwen Series Models on Ascend NPU

3.1 Directly Related Cases

3.2 Key Conclusions

1. The Qwen series models have numerous successful deployment cases on Ascend NPUs, covering various scales from 8B to 235B.
2. vLLM-Ascend and SGLang natively support Ascend NPUs, enabling Qwen models to run directly.
3. CANN 8.0 + PyTorch-Ascend already covers mainstream Transformer operators.
4. The SGLang inference engine natively supports Ascend as a backend, allowing models like DeepSeek, Qwen, and GLM to run directly.

IV. Specific Verification Steps (Executable Directly)

4.1 Environment Preparation (1-2 hours)

4.2 Model Adaptation and Verification (2-8 hours)

| Step | Operation | Estimated Duration |

|------|------|----------|

| 1 | Perform model analysis using Ascend Model-Agent Skill | 30 minutes |

| 2 | Execute model migration (using ascend-agent-skills) | 1-2 hours |

| 3 | Perform inference optimization (using CANN Skill system) | 2-4 hours |

| 4 | Execute verification and acceptance (using verification-acceptance Skill) | 1-2 hours |

Expected Outcome:

• Simple model: Complete full optimization loop within 2 hours
• Complex model: Complete basic optimization loop within 6-8 hours

4.3 Verification and Acceptance Criteria

| Verification Item | Criterion | Verification Method |

|--------|------|----------|

| Model Loading | Model weights can be correctly loaded and parsed by the Ascend platform | Run inference test |

| Inference Accuracy | Error vs. GPU baseline

| Inference Performance | End-to-end throughput ≥ 80% of GPU baseline | Performance evaluation |

| Stability | 7x24 hours stable operation | Stress test |

5. How to use existing adaptation experience to enhance persuasiveness

5.1 Qualcomm Snapdragon adaptation experience → Analogy to Ascend NPU

5.2 Moore Threads adaptation experience → Analogy to Ascend NPU

5.3 Core Talking Points

"Zuobang Youbi has completed adaptation verification for two edge-side chips, Qualcomm Snapdragon and Moore Threads — proving our technology stack is chip-agnostic. Adapting to the Ascend NPU is an engineering replication following the same path, and the Ascend community already provides a complete Agent Skill toolchain (ascend-agent-skills + CANN Skill system). The Qwen series models have numerous successful deployment cases. We estimate 2-4 weeks to complete the adaptation and verification for the Ascend NPU."

VI. Directly Citable Evidence List

VII. Summary

The Ascend NPU adaptation by ZuoBangYouBi is not "theoretically feasible" but rather "tools ready, path clear, and rapidly verifiable."

Three levels of assurance:

1. Tooling layer: Ascend Model-Agent Skill + ascend-agent-skills + CANN Skill system, ready out-of-the-box
2. Case layer: Numerous successful deployment cases of Qwen series models on Ascend NPU
3. Experience layer: ZuoBangYouBi has completed edge-side model adaptation on Qualcomm Snapdragon and Moore Threads, path validated

Estimated validation time: 2-4 weeks to complete Ascend NPU adaptation and verification

Validation criteria: Model loads successfully + inference accuracy meets standard + inference performance meets standard + stability meets standard