The sweep finds the knee: output peaks at 32 users on the L4
Output tokens per second against users at the same time for vLLM 0.31.0 on an NVIDIA L4 with Qwen3-8B, three seeds per point. Output rises from 13.5 at one user to 275.1 at 32 users, then falls to 242.3 at 64 users, where requests queue for KV cache space.
0
100
200
300
1
2
4
8
16
32
64
users at the same time
output tokens per second
knee, 275
242
median of three runs, band from lowest to highest