{"id":"4Vjt9LxAXN","url":"https://pastebin.ca/4Vjt9LxAXN","raw_url":"https://raw.anybin.ca/4Vjt9LxAXN","visibility":"public","access":"public","created_at":1794201897035,"expires_at":1796793897035,"fetch_limit":null,"fetches_used":0,"reads_remaining":null,"size_bytes":2189,"syntax_hint":null,"title":null,"filename":null,"change_note":null,"cipher":null,"cipher_meta":null,"parent_id":null,"root_id":"4Vjt9LxAXN","version":1,"owner_id":null,"recipient_id":null,"body":"# lab02_calculator.py —— H3与Qwen3.8 128B 显存及方案性价比计算器\n\ndef calc_model_size(param_billion, bits):\n    \"\"\"计算模型权重体积 (GB)\"\"\"\n    return (param_billion * 1e9 * bits) / 8 / (1024**3)\n\ndef calc_kv_cache(context_len, batch_size=1):\n    \"\"\"粗略估算KV Cache占用 (GB)，按经验值折算\"\"\"\n    # 假设 32K 上下文单请求约 1GB，按比例和并发数放大\n    base_32k_gb = 1.0\n    return (context_len / 32768) * base_32k_gb * batch_size\n\ndef calc_value(price_wan, speed_tok_s):\n    \"\"\"计算性价比 (tok/s/万元)\"\"\"\n    return speed_tok_s / price_wan\n\nif __name__ == '__main__':\n    print(\"=== 1. 模型体积与显存需求推导 ===\")\n    params = 128 # 128B\n    size_int4 = calc_model_size(params, 4)\n    size_int8 = calc_model_size(params, 8)\n    size_fp16 = calc_model_size(params, 16)\n    \n    print(f\"128B 模型体积：\")\n    print(f\"  INT4: {size_int4:.1f} GB\")\n    print(f\"  INT8: {size_int8:.1f} GB\")\n    print(f\"  FP16: {size_fp16:.1f} GB\")\n    \n    kv_cache = calc_kv_cache(context_len=32768, batch_size=4)\n    h3_size = 12.0\n    print(f\"\\nKV Cache (32K上下文, 4并发): 约 {kv_cache:.1f} GB\")\n    print(f\"MiniMax H3 (pruned INT8): {h3_size:.1f} GB\")\n    \n    print(\"\\n=== 2. 三方案性价比对比 ===\")\n    # 方案数据: (名称, 显存+内存, 出词速度, H3耗时, 总价, 性价比)\n    plans = [\n        {\"name\": \"主推: 5090 32G + 128G DDR5\", \"speed\": 20, \"price\": 8.54},\n        {\"name\": \"对照A: 128G 统一内存\", \"speed\": 10, \"price\": 2.50},\n        {\"name\": \"对照B: 双卡 64G 工作站\", \"speed\": 48, \"price\": 14.00}\n    ]\n    \n    for p in plans:\n        cpb = calc_value(p[\"price\"], p[\"speed\"])\n        print(f\"{p['name']}:\")\n        print(f\"  价格: {p['price']} 万元 | 出词: {p['speed']} tok/s\")\n        print(f\"  性价比: {cpb:.2f} tok/s/万元\\n\")\n\n    print(\"=== 3. 预算分配核对 ===\")\n    total_budget = 8.54\n    gpu_price = 5.24\n    mem_price = 2.20\n    print(f\"总预算: {total_budget} 万\")\n    print(f\"GPU占比: {(gpu_price/total_budget)*100:.1f}% (目标55-60%)\")\n    print(f\"内存占比: {(mem_price/total_budget)*100:.1f}% (目标10%，因近期内存暴涨偏高)\")"}