fix(disagg): poll receivers during decode preallocation (#37483)
This commit is contained in:
@@ -869,17 +869,8 @@ class DecodePreallocQueue(DecodeHiCachePreallocMixin):
|
||||
if not self.queue:
|
||||
return
|
||||
|
||||
# Still poll if any receiver was aborted, otherwise it stays stuck.
|
||||
if (
|
||||
self.pp_size <= 1
|
||||
and all(decode_req.waiting_for_input for decode_req in self.queue)
|
||||
and not any(
|
||||
decode_req.kv_receiver.conclude_state == KVPoll.Failed
|
||||
for decode_req in self.queue
|
||||
)
|
||||
):
|
||||
return
|
||||
|
||||
# Receiver polling observes asynchronous failures while KV allocation
|
||||
# is blocked.
|
||||
if self.pp_size > 1:
|
||||
polls = poll_and_all_reduce_pp(
|
||||
(decode_req.req.rid for decode_req in self.queue),
|
||||
|
||||
Reference in New Issue
Block a user