Performance Issue – Aphrodite Slower Than vLLM & Concurrency Failures

[mishradibyajyoti]

Hi Team,

We have been evaluating Aphrodite in our environment and are currently experiencing performance issues compared to our existing vLLM setup. Specifically:
Aphrodite appears to be significantly slower in terms of response time.
We are facing issues when making concurrent requests, where the system fails to handle parallel calls properly.

Could you please help us understand:

Whether there are recommended configuration settings for improving throughput and concurrency?
If there are known limitations for high-concurrency workloads?
Any best practices for deployment to achieve performance comparable to vLLM?

[AlpinDale]

Aphrodite hasn't been updated in a while, so it'll lag behind vLLM a bit. I plan to wrap-up the update + upstream sync this week, so please look forward to that.

[onzag]

The DRY and XTC sampler make up the main reason I am looking upon this; sadly not implemented in vllm.

They seemed to have rejected your PR automatically.