Currently, when requests are cancelled while executing their final
step, "completion" is handled based on normal stop processing (e.g.
length or stop token), so the abort has no effect. This is typically
not a problem, but when a kv connector is involved it thinks the
request completed successfully rather than being aborted.
This is problematic for disaggregated prefill which will free kv
cache blocks if the request was aborted but not if it completed
successfully—since the cancelled request will never be sent to
the decode side, kv cache blocks remain pinned until the fall-back
timeout expires. The problem is exacerbated when many requests
are cancelled and/or there are large prefills whose forward pass
takes a long time (since the window is bigger).
This PR fixes the problem by processing pending aborts
immediately prior to processing model output each step; we process
only aborts, not new requests, since it's preferable for latency to
process model outputs before new incoming requests.
Fixes#26400.
Signed-off-by: Nick Hill <nhill@redhat.com>
