[Bugfix][V1] Free out-of-window blocks on the processed-token basis under async scheduling (#47728)

Signed-off-by: Saddss <28726669061@qq.com>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Signed-off-by: Yifan Qiao <yifanqiao@inferact.ai>
Co-authored-by: Yifan Qiao <yifanqiao@inferact.ai>
Co-authored-by: Saddss <28726669061@qq.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Saddss
2026-07-08 13:34:19 +01:00
committed by GitHub
co-authored by Yifan Qiao Saddss Nick Hill
parent 934eeaecfb
commit cd0de48d08
14 changed files with 549 additions and 143 deletions
@@ -81,7 +81,7 @@ def assert_encoder_kv_cache_spec(engine: LLM) -> None:
assert spec.sliding_window == cdiv(750, 4) + 1 == 189
assert (
spec.max_admission_blocks_per_request(
max_num_batched_tokens=1,
max_in_flight_tokens=1,
max_model_len=vllm_config.model_config.max_model_len,
)
== 13