mistralai/Mixtral-8x7B-v0.1 on A100-80G
How many A100-80G GPUs to run mistralai/Mixtral-8x7B-v0.1.
Architecture
| Field |
Value |
model_type |
mixtral |
attention |
GQA (heads=32, kv_heads=8, hd=128) |
moe |
8 routed + 0 shared, top-2 |
Weights
| Field |
Value |
Label |
| safetensors bytes |
86.99 GB |
[verified] |
| params |
46.7B |
[estimated] |
| quantization |
BF16 [verified] |
|
Quantization reconciliation
| Scheme |
Predicted |
Δ |
Error |
| FP16 |
86.99 GB |
133.09 KB over |
0.0% |
| BF16 ✓ |
86.99 GB |
133.09 KB over |
0.0% |
| FP8 |
43.50 GB |
43.50 GB over |
100.0% |
| INT8 |
43.50 GB |
43.50 GB over |
100.0% |
| FP4_FP8_MIXED |
23.92 GB |
63.07 GB over |
263.6% |
Best: BF16 — safetensors header: all 48 weight tensors are BF16 (predicts 93,405,577,216 bytes, 0.0% error)
KV cache per request
| Context tokens |
KV bytes |
| 4,096 |
512.00 MB |
| 32,768 |
4.00 GB |
Recommended fleet
| 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 |
- Prefill latency 374 ms @ 2000 input tokens
[estimated]
- Cluster decode throughput 149 tok/s
[estimated]
- Max concurrent users 4
- Bottleneck
memory_bandwidth
Generated command
vllm serve mistralai/Mixtral-8x7B-v0.1 \
--tensor-parallel-size 4 \
--max-model-len 32768 \
--trust-remote-code \
--gpu-memory-utilization 0.9
Generated by:
llm-cal mistralai/Mixtral-8x7B-v0.1 --gpu A100-80G --engine vllm --lang en