Performance Speed Drops and image description

[Maytown]

Hi, i am using MinerU with the vllm setup and the hybrid-http-client. The Output is wonderful, but the performance speed drops quite significantly when sending multiple requests or just very big files.
For example a 20 page file takes 14 seconds alone, when i send the same file twice we already drop to a 30 second extraction. And big files with 200-300 pages, take around 3 min 30 sec. I am running this on an A100 with 0,9 vllm gpu usage using the fast_api.py.
Is there anything I can do to improve the async performance?

Furthermore is there something planned for image/chart/figure description while extraction?

Thanks in advance!