[MemCache] Skip SWA request-cap pool sizing when chunked prefill is off (#43203)

This commit is contained in:
Liangsheng Yin
2026-10-08 16:19:35 -07:00
committed by GitHub
parent 71584b62c0
commit ae4da8cd78
2 changed files with 12 additions and 3 deletions
@@ -948,7 +948,12 @@ class SWAChunkCapPoolConfigurator(HybridSWAPoolConfigurator):
return False
if not get_memory().disable_radix_cache:
return False
if get_schedule().chunked_prefill_size is None:
# The cap holds at most one prefill chunk per batch in flight; an
# unchunked prefill writes the whole prompt into the SWA pool.
chunked_prefill_size = get_schedule().chunked_prefill_size
if get_disagg().disaggregation_mode != "decode" and (
chunked_prefill_size is None or chunked_prefill_size <= 0
):
return False
if kvc.sliding_window_size is None:
return False