Qwen/Qwen2.5-7B 跑在 RTX4090
Qwen/Qwen2.5-7B 在 RTX4090 上需要多少 GPU。
架构
| Field |
Value |
model_type |
qwen2 |
attention |
GQA (heads=28, kv_heads=4, hd=128) |
sliding_window |
131072 |
权重
| Field |
Value |
Label |
| safetensors 字节 |
14.19 GB |
[已验证] |
| 参数量 |
7.6B |
[估算] |
| 量化方案 |
BF16 [已验证] |
|
量化反演
| Scheme |
Predicted |
Δ |
Error |
| FP16 |
14.18 GB |
296.95 KB 偏多 |
0.0% |
| BF16 ✓ |
14.18 GB |
296.95 KB 偏多 |
0.0% |
| FP8 |
7.09 GB |
7.09 GB 偏多 |
100.0% |
| INT8 |
7.09 GB |
7.09 GB 偏多 |
100.0% |
| FP4_FP8_MIXED |
3.90 GB |
10.28 GB 偏多 |
263.6% |
Best: BF16 — safetensors header: all 73 weight tensors are BF16 (predicts 15,230,967,808 bytes, 0.0% error)
KV 缓存(每请求)
| Context tokens |
KV bytes |
| 4,096 |
224.00 MB |
| 32,768 |
1.75 GB |
| 131,072 |
7.00 GB |
推荐集群
| Tier |
GPUs |
Weight/GPU |
Headroom/GPU |
Concurrent @ 128K |
| min |
2 |
7.09 GB |
13.02 GB |
3 |
| dev ★ |
4 |
3.55 GB |
16.57 GB |
9 |
| prod |
7 |
2.03 GB |
18.09 GB |
10 |
性能
- Prefill latency 115 ms @ 2000 input tokens
[估算]
- Cluster decode throughput 381 tok/s
[估算]
- Max concurrent users 9
- Bottleneck
memory_capacity
生成命令
vllm serve Qwen/Qwen2.5-7B \
--tensor-parallel-size 4 \
--max-model-len 131072 \
--gpu-memory-utilization 0.9
生成方式:
llm-cal Qwen/Qwen2.5-7B --gpu RTX4090 --engine vllm --lang zh