端侧大模型应用性能调优清单生成器

You are a performance optimization expert specializing in on-device / edge LLM deployment. Generate a comprehensive performance tuning checklist for my setup. ## My Setup - Hardware: [e.g., MacBook M4 Max 128GB / RTX 4090 / Jetson Orin / iPhone 16 Pro] - Model: [e.g., Llama 3.3 70B / Qwen3 32B / Phi-4 / Gemma 3] - Framework: [e.g., llama.cpp / MLX / vLLM / TensorRT-LLM / LiteRT] - Use case: [e.g., code completion / RAG chatbot / real-time translation] - Target latency: [e.g., < 200ms first token, > 30 tok/s generation] ## Generate ### 1. Quantization Audit - Current quantization level and recommended alternatives - Quality vs speed tradeoff analysis for my specific use case - Recommended quant methods (GGUF Q4_K_M, AWQ, GPTQ, etc.) ### 2. Memory Optimization - KV cache configuration (size, type, quantization) - Context length vs memory budget calculator - Batch size recommendations - Memory-mapped vs fully loaded tradeoffs ### 3. Compute Optimization - Thread/core allocation strategy - GPU layer offloading recommendations - Flash attention / paged attention configuration - Speculative decoding feasibility ### 4. System-Level Tuning - OS-level settings (huge pages, CPU governor, thermal management) - I/O optimization for model loading - Concurrent request handling strategy ### 5. Benchmark Commands - Provide exact CLI commands to benchmark before/after each optimization - Include expected improvement ranges Output as a prioritized checklist with [HIGH/MED/LOW] impact ratings and estimated effort.