Meta · Dense · GQA
Llama 3.3 70B Instruct：各量化与上下文下的显存
70.6B 参数、80 层、8 个 KV 头。换算成 GB 是这个结果。
架构参数
| Parameters | 70.6B | Vendor model card |
|---|---|---|
| Layers | 80 | config.json |
| Hidden size | 8192 | config.json |
| Attention heads | 64 | config.json |
| KV heads | 8 | config.json (GQA) |
| Head dim | 128 | config.json |
| Max context | 131072 tokens | config.json |
Source: Meta Llama 3 architecture (published); repo gated
各量化精度下的显存
Total includes 8K of KV cache and 1.5 GB runtime overhead.
| 量化 | Weights | KV @ 8K | 总需求 | 可运行于 |
|---|---|---|---|---|
| FP16 / BF16 2.0 B/param | 141.2 GB | 2.7 GB | 145.4 GB | 不行 |
| INT8 / Q8_0 1.06 B/param | 74.8 GB | 2.7 GB | 79.0 GB | Mac Studio M4 Max 128GB |
| Q4_K_M (4-bit) 0.6 B/param | 42.4 GB | 2.7 GB | 46.5 GB | Mac Studio M4 Max 128GB |
| Q5_K_M (5-bit) 0.72 B/param | 50.8 GB | 2.7 GB | 55.0 GB | Mac Studio M4 Max 128GB |
各上下文长度下的 KV 缓存
Q4_K_M weights plus 1.5 GB overhead. KV cache at FP16.
| 上下文 | KV 缓存 | 总需求 | 可运行于 |
|---|---|---|---|
| 4k | 1.3 GB | 45.2 GB | Mac Studio M4 Max 128GB |
| 8k | 2.7 GB | 46.5 GB | Mac Studio M4 Max 128GB |
| 16k | 5.4 GB | 49.2 GB | Mac Studio M4 Max 128GB |
| 32k | 10.7 GB | 54.6 GB | Mac Studio M4 Max 128GB |
| 64k | 21.5 GB | 65.3 GB | Mac Studio M4 Max 128GB |
| 128k | 42.9 GB | 86.8 GB | Mac Studio M4 Max 128GB |
结论
At Q4_K_M the weights alone are about 42.4 GB. With 8K of context and 1.5 GB of runtime overhead it lands near 46.5 GB. It fits: Mac Studio M4 Max 128GB.
所有吞吐数字均为基于内存带宽与公开实测的估算值，显存数字按模型架构与量化假设推算。购买硬件前请以你自己的运行环境为准。