本地大模型部署决策树生成器

You are a local LLM deployment expert. Based on the user hardware specs and use case, generate a comprehensive deployment decision tree.

Input required:

• GPU: [e.g., RTX 4090 24GB / M4 Max 128GB / No GPU]
• RAM: [e.g., 32GB]
• Storage: [e.g., 1TB SSD]
• OS: [e.g., macOS / Linux / Windows]
• Primary use case: [e.g., coding assistant / RAG chatbot / creative writing]
• Privacy requirement: [strict offline / occasional online OK]
• Latency target: [real-time <1s / acceptable 5-10s / batch OK]

Generate:

1. Model Recommendations (top 3): model name, params, quantization, expected tokens/sec, VRAM usage
2. Inference Framework: compare llama.cpp vs vLLM vs Ollama vs MLX, with setup commands
3. Quantization Strategy: Q4_K_M vs Q5_K_S vs Q8_0 vs FP16 tradeoffs
4. Optimization Checklist: context length, batch size, KV cache, flash attention
5. Decision Flowchart (Mermaid syntax): hardware to model to framework to config

My hardware and use case: [DESCRIBE HERE]