A request finds a reusable prefix, so the scheduler treats it as ready. In a hybrid language model, the recurrent state it needs may still be under construction elsewhere in the same round.
RadixFlow's October 13 paper identifies resulting head-of-line blocking. Its traversal skips the locally blocked consumer and fills available slots with independent work, reducing latency in the evaluated hybrid-model workloads. Read the research paper.
Reuse and readiness are different facts
The cache lookup answers whether matching computation exists in the execution graph. It does not necessarily answer whether all required state can be consumed now.
That distinction is less visible in full-attention systems, helping explain why a scheduling assumption can survive until the architecture changes. The paper includes a full-attention control where the intervention adds little.
Fewer worker rounds do not translate proportionally into lower latency if the remaining rounds become heavier. Measure end-to-end completion under realistic arrival patterns.
RadixFlow is an example of an architectural transition exposing old software assumptions. The model does not need better weights to benefit. The serving stack needs a more accurate account of which work is blocked and which merely happened to stand behind it in a queue.
Questions, news tips or corrections? Contact [email protected].
