Qwen/Qwen2.5-7B on L40S
How many L40S GPUs to run Qwen/Qwen2.5-7B.
Architecture
| Field |
Value |
model_type |
qwen2 |
attention |
GQA (heads=28, kv_heads=4, hd=128) |
sliding_window |
131072 |
Weights
| Field |
Value |
Label |
| safetensors bytes |
14.19 GB |
[verified] |
| params |
7.6B |
[estimated] |
| quantization |
BF16 [verified] |
|
Quantization reconciliation
| Scheme |
Predicted |
Δ |
Error |
| FP16 |
14.18 GB |
296.95 KB over |
0.0% |
| BF16 ✓ |
14.18 GB |
296.95 KB over |
0.0% |
| FP8 |
7.09 GB |
7.09 GB over |
100.0% |
| INT8 |
7.09 GB |
7.09 GB over |
100.0% |
| FP4_FP8_MIXED |
3.90 GB |
10.28 GB over |
263.6% |
Best: BF16 — safetensors header: all 73 weight tensors are BF16 (predicts 15,230,967,808 bytes, 0.0% error)
KV cache per request
| Context tokens |
KV bytes |
| 4,096 |
224.00 MB |
| 32,768 |
1.75 GB |
| 131,072 |
7.00 GB |
Recommended fleet
| Tier |
GPUs |
Weight/GPU |
Headroom/GPU |
Concurrent @ 128K |
| min |
1 |
14.19 GB |
26.05 GB |
3 |
| dev ★ |
2 |
7.09 GB |
33.14 GB |
9 |
| prod |
4 |
3.55 GB |
36.69 GB |
20 |
- Prefill latency 105 ms @ 2000 input tokens
[estimated]
- Cluster decode throughput 82 tok/s
[estimated]
- Max concurrent users 2
- Bottleneck
memory_bandwidth
Generated command
vllm serve Qwen/Qwen2.5-7B \
--tensor-parallel-size 2 \
--max-model-len 131072 \
--gpu-memory-utilization 0.9
Generated by:
llm-cal Qwen/Qwen2.5-7B --gpu L40S --engine vllm --lang en