Fix swa chunk req deferred (#24318)

This commit is contained in:
Ke Bao
2026-05-04 14:52:15 +08:00
committed by GitHub
parent c2d60c3667
commit aea527afdc
2 changed files with 181 additions and 1 deletions
+20 -1
View File
@@ -1080,6 +1080,15 @@ class Scheduler(
elif self.chunked_prefill_size is not None and self.chunked_prefill_size <= 0:
self.chunked_prefill_size = None
self.chunked_req = None
# Tracks whether the current self.chunked_req was actually scheduled
# into last iteration's batch (i.e., in can_run_list -> got a fresh
# req_pool_idx from prepare_for_extend). Used to gate the
# stash_chunked_request call at the top of get_next_batch_to_run:
# if add_chunked_req early-returned under hybrid-SWA pressure,
# the req_pool_idx was already freed and fill_ids was reset by
# init_next_round_input, so running stash would double-free and
# corrupt prefix_indices.
self._chunked_req_scheduled_last_iter = False
self.is_mixed_chunk = (
self.chunked_prefill_size is not None
and self.server_args.enable_mixed_chunk
@@ -2455,7 +2464,9 @@ class Scheduler(
# Move the chunked request out of the batch so that we can merge
# only finished requests to running_batch.
chunked_req_to_exclude.add(self.chunked_req)
self.stash_chunked_request(self.chunked_req)
if self._chunked_req_scheduled_last_iter:
self.stash_chunked_request(self.chunked_req)
# HiSparse has its own prefill-to-decode transition; skip last_batch merge.
if self.enable_hisparse:
@@ -2650,6 +2661,11 @@ class Scheduler(
if self.chunked_req is not None:
self.chunked_req.init_next_round_input()
self.chunked_req = adder.add_chunked_req(self.chunked_req)
self._chunked_req_scheduled_last_iter = (
self.chunked_req in adder.can_run_list
)
else:
self._chunked_req_scheduled_last_iter = False
if self.enable_lora:
running_loras = {req.lora_id for req in self.running_batch.reqs}
@@ -2741,6 +2757,9 @@ class Scheduler(
# Update chunked prefill
assert self.chunked_req is None
self.chunked_req = adder.new_chunked_req
# new_chunked_req is added to can_run_list by add_one_req,
# so it will be scheduled this iter -> stash is needed next iter.
self._chunked_req_scheduled_last_iter = True
if self.chunked_req is not None:
self.chunked_req.is_chunked += 1