gahow/agentic-kvc
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Gate evict_sent_blocks behind VLLM_EVICT_SENT_BLOCKS

Browse Source 
Fork commit e13391e unconditionally evicts sent blocks from the prefix
cache on every KV transfer. That is correct only for session MIGRATION
(source won't see the session again); for plain PD-disagg producer->
consumer transfers it destroys cross-turn producer reuse and contaminates
PD reuse experiments. Default OFF; enable for migration runs via
VLLM_EVICT_SENT_BLOCKS=1.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Gahow Wang
2026-05-29 18:18:59 +08:00
parent a0db3cbe77
commit 657cd36f3d
1 changed files with 6 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
third_party/vllm/vllm/v1/core/sched/scheduler.py vendored
View File

@@ -2218,7 +2218,12 @@ class Scheduler(SchedulerInterface):
for group in self.kv_cache_manager.get_block_ids(req_id):
sent_block_ids.update(group)
self._free_blocks(self.requests[req_id])
if sent_block_ids:
# e13391e: evict sent blocks from the prefix cache. Intended ONLY for
# session MIGRATION (the source worker won't see the session again).
# For plain PD-disagg producer->consumer transfers this destroys
# cross-turn producer reuse, so it is gated OFF by default and enabled
# only for migration experiments via VLLM_EVICT_SENT_BLOCKS=1.
if sent_block_ids and __import__("os").environ.get("VLLM_EVICT_SENT_BLOCKS") == "1":
self.kv_cache_manager.evict_blocks(sent_block_ids)
def _update_requests_with_invalid_blocks(