mistralai/Mixtral-8x7B-v0.1 跑在 H100
mistralai/Mixtral-8x7B-v0.1 在 H100 上需要多少 GPU。
架构
| Field |
Value |
model_type |
mixtral |
attention |
GQA (heads=32, kv_heads=8, hd=128) |
moe |
8 routed + 0 shared, top-2 |
权重
| Field |
Value |
Label |
| safetensors 字节 |
86.99 GB |
[已验证] |
| 参数量 |
46.7B |
[估算] |
| 量化方案 |
BF16 [已验证] |
|
量化反演
| Scheme |
Predicted |
Δ |
Error |
| FP16 |
86.99 GB |
133.09 KB 偏多 |
0.0% |
| BF16 ✓ |
86.99 GB |
133.09 KB 偏多 |
0.0% |
| FP8 |
43.50 GB |
43.50 GB 偏多 |
100.0% |
| INT8 |
43.50 GB |
43.50 GB 偏多 |
100.0% |
| FP4_FP8_MIXED |
23.92 GB |
63.07 GB 偏多 |
263.6% |
Best: BF16 — safetensors header: all 48 weight tensors are BF16 (predicts 93,405,577,216 bytes, 0.0% error)
KV 缓存（每请求）
| Context tokens |
KV bytes |
| 4,096 |
512.00 MB |
| 32,768 |
4.00 GB |
推荐集群
| Tier |
GPUs |
Weight/GPU |
Headroom/GPU |
Concurrent @ 128K |
| min |
2 |
43.50 GB |
23.56 GB |
2 |
| dev ★ |
4 |
21.75 GB |
45.31 GB |
11 |
| prod |
8 |
10.87 GB |
56.18 GB |
28 |
性能
- Prefill latency 118 ms @ 2000 input tokens
[估算]
- Cluster decode throughput 258 tok/s
[估算]
- Max concurrent users 8
- Bottleneck
memory_bandwidth
生成命令
vllm serve mistralai/Mixtral-8x7B-v0.1 \
--tensor-parallel-size 4 \
--max-model-len 32768 \
--trust-remote-code \
--gpu-memory-utilization 0.9
生成方式:
llm-cal mistralai/Mixtral-8x7B-v0.1 --gpu H100 --engine vllm --lang zh