[Bugfix] Fix orphaned aborted prefill bootstrap requests in PP disaggregation (#27028)
This commit is contained in:
@@ -333,10 +333,15 @@ class PrefillBootstrapQueue:
|
|||||||
)
|
)
|
||||||
|
|
||||||
for i, (req, poll) in enumerate(zip(self.queue, polls)):
|
for i, (req, poll) in enumerate(zip(self.queue, polls)):
|
||||||
if rids_to_check is not None:
|
if (
|
||||||
# if req not in reqs_info_to_check, skip
|
rids_to_check is not None
|
||||||
if req.rid not in rids_to_check:
|
and req.rid not in rids_to_check
|
||||||
continue
|
and poll != KVPoll.Failed
|
||||||
|
):
|
||||||
|
# In PP mode, successful bootstrap still requires cross-rank
|
||||||
|
# consensus. Local failures are terminal and must be drained
|
||||||
|
# even if an earlier PP rank has already removed the request.
|
||||||
|
continue
|
||||||
|
|
||||||
if poll == KVPoll.Failed:
|
if poll == KVPoll.Failed:
|
||||||
self.scheduler.handle_bootstrap_failure(req)
|
self.scheduler.handle_bootstrap_failure(req)
|
||||||
|
|||||||
Reference in New Issue
Block a user