[PD] Support fake decode for PD disaggregation without prefill node (#14628)
Co-authored-by: sunhailiang <sunhailiang@baidu.com>
This commit is contained in:
@@ -489,6 +489,16 @@ class DataParallelController:
|
||||
self.workers
|
||||
)
|
||||
else:
|
||||
# Set default bootstrap_room if in FAKE auto mode and room is None
|
||||
if (
|
||||
req.bootstrap_room is None
|
||||
and self.server_args.disaggregation_decode_enable_fake_auto
|
||||
):
|
||||
req.bootstrap_room = self.round_robin_counter
|
||||
self.round_robin_counter = (self.round_robin_counter + 1) % len(
|
||||
self.workers
|
||||
)
|
||||
|
||||
assert (
|
||||
req.bootstrap_room is not None
|
||||
), "req.bootstrap_room should not be None. Do not send requests directly to prefill or decode instances, but send to the router instead."
|
||||
|
||||
Reference in New Issue
Block a user