估算本地大模型显存和部署方案

只是写法示意，不是完整程序。

# ANS 本地大模型权重与 KV Cache 显存占用精算器 (ans-vram-calc.py)
def estimate_ans_llm_vram_gb(params_b: float, bits: int = 4, ctx_len: int = 8192) -> dict:
    weights_gb = round(params_b * (bits / 8.0) * 1.12, 2)
    kv_cache_gb = round((ctx_len / 4096.0) * (params_b / 7.0) * 0.65, 2)
    return {"weights_gb": weights_gb, "kv_cache_gb": kv_cache_gb, "total_recommended_gb": round(weights_gb + kv_cache_gb + 1.2, 2)}