💡 Lumina-Agent
Huawei Cloud HarmonyOS Agent Competition Champion Solution. A data-centric, memory-aware end-to-end voice command system.
📖 Introduction
Lumina-Agent is a solution designed for the "Huawei Cloud HarmonyOS System Control Agent Competition," addressing strict resource constraints (5GB NPU memory) and complex multi-turn logic. We bypassed complex RAG or Hierarchical architectures, proposing a Minimalist, Data-Centric, and NPU-Optimized End-to-End Architecture.
By leveraging Semantic Prompt Compression, State Differential Reasoning, and a Mixture-of-Experts (MoE) Data Pipeline, we achieved 100% accuracy on the local smoke test set and ranked No.1 in the finals using only a 1.7B parameter model.
👥 Core Team
- Lv Yuze: Team Leader; Architecture Design & Algorithm Optimization
- Shi Yixi: Data Engineering & Anti-Hallucination
- Liu Xiaorui: System Deployment & Memory Optimization
🚀 Key Innovations
🧠 Flat-Direct Architecture
Allowed us to avoid double inference latency and error cascading. We used a Direct Agent for one-pass inference, maximizing global attention visibility and minimizing latency.
⚡ Memory-Aware Compression
To fit 100+ tools into the 5GB NPU limit: Extracted "Verb + Object" cores via Semantic Distillation. Compressed context from 20k to 5k tokens, speeding up inference by 40%.
🏗️ Training on Consumer GPUs
Trained 6k context on RTX 3090 (24GB): Gradient Checkpointing reduced memory by 60%. Full-Linear LoRA fine-tuning enhanced logic w/o forgetting. Full BF16 pipeline aligned with NPU.
📊 MoE Data Pipeline
Constructed a 20,000+ sample dataset with "Resampling Layer", "Logic Layer" (GPT-4 logic), and "Concurrency Layer", effectively solving multi-instruction scenarios and hallucinations.
📈 Competition Results
- Final Rank: 1st Place (Champion)
- Local Accuracy: 100% (Smoke Test)
- Complex Logic Acc: >90%
