chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Support PP for zmq_to_scheduler (#15312)

Browse Source
This commit is contained in:
Tianyu Guo
2025-12-23 17:07:55 +08:00
committed by GitHub
parent f9dd90ac35
commit fa2966983a
5 changed files with 28 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
python/sglang/srt/managers/scheduler.py
View File

@@ -955,6 +955,7 @@ class Scheduler(
hf_config=self.model_config.hf_config,
tp_rank=self.tp_rank,
pp_rank=self.pp_rank,
tp_group=self.tp_group,
)
def init_overlap(self):
@@ -1239,6 +1240,14 @@ class Scheduler(
src=self.tp_group.ranks[0],
)
# Process MM requests under EPD-disaggregation mode
if (
self.pp_rank == 0
and self.server_args.language_only
and self.server_args.encoder_transfer_backend == "zmq_to_scheduler"
):
recv_reqs = self.mm_receiver.process_waiting_requests(recv_reqs)
if self.enable_trace:
for req in recv_reqs:
if isinstance(
@@ -1279,12 +1288,6 @@ class Scheduler(
return work_reqs, control_reqs
def process_input_requests(self, recv_reqs: List):
# Process MM requests under EPD-disaggregation mode
if (
self.server_args.language_only
and self.server_args.encoder_transfer_backend == "zmq_to_scheduler"
):
recv_reqs = self.mm_receiver.process_waiting_requests(recv_reqs)
for recv_req in recv_reqs:
# If it is a health check generation request and there are running requests, ignore it.