vLLM 0.31.0: a separate active-request limit helps control load
vLLM 0.31.0 adds a separate limit for concurrently running requests and changes how client-supplied multimodal parameters are handled. This gives operators clearer load control, but latency and application compatibility need validation.
What changed
The release also changes handling of client-supplied multimodal parameters: they are rejected unless the operator explicitly allows them with --trust-request-mm-kwargs. Check how your integration passes parameters and whether it depends on the previous behavior.