EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Auto Sync] Update scheduler.py (20251009) (#11350)

Browse Source 
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Junxiong Wang <junxiong@together.ai>
This commit is contained in:
Yineng Zhang
2025-10-08 17:39:04 -07:00
committed by GitHub
parent a3c2ea4451
commit e22b13c569
1 changed files with 45 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

83
python/sglang/srt/managers/scheduler.py
View File

@@ -273,6 +273,48 @@ class Scheduler(
):
"""A scheduler that manages a tensor parallel GPU worker."""
def launch_draft_worker(
self, gpu_id, tp_rank, moe_ep_rank, server_args, port_args, dp_rank
):
if self.spec_algorithm.is_eagle():
from sglang.srt.speculative.eagle_worker import EAGLEWorker
self.draft_worker = EAGLEWorker(
gpu_id=gpu_id,
tp_rank=tp_rank,
moe_ep_rank=moe_ep_rank,
server_args=server_args,
nccl_port=port_args.nccl_port,
target_worker=self.tp_worker,
dp_rank=dp_rank,
)
elif self.spec_algorithm.is_standalone():
from sglang.srt.speculative.standalone_worker import StandaloneWorker
self.draft_worker = StandaloneWorker(
gpu_id=gpu_id,
tp_rank=tp_rank,
moe_ep_rank=moe_ep_rank,
server_args=server_args,
nccl_port=port_args.nccl_port,
target_worker=self.tp_worker,
dp_rank=dp_rank,
)
elif self.spec_algorithm.is_ngram():
from sglang.srt.speculative.ngram_worker import NGRAMWorker
self.draft_worker = NGRAMWorker(
gpu_id=gpu_id,
tp_rank=tp_rank,
moe_ep_rank=moe_ep_rank,
server_args=server_args,
nccl_port=port_args.nccl_port,
target_worker=self.tp_worker,
dp_rank=dp_rank,
)
else:
self.draft_worker = None
def __init__(
self,
server_args: ServerArgs,
@@ -412,44 +454,9 @@ class Scheduler(
)
# Launch a draft worker for speculative decoding
if self.spec_algorithm.is_eagle():
from sglang.srt.speculative.eagle_worker import EAGLEWorker
self.draft_worker = EAGLEWorker(
gpu_id=gpu_id,
tp_rank=tp_rank,
moe_ep_rank=moe_ep_rank,
server_args=server_args,
nccl_port=port_args.nccl_port,
target_worker=self.tp_worker,
dp_rank=dp_rank,
)
elif self.spec_algorithm.is_standalone():
from sglang.srt.speculative.standalone_worker import StandaloneWorker
self.draft_worker = StandaloneWorker(
gpu_id=gpu_id,
tp_rank=tp_rank,
moe_ep_rank=moe_ep_rank,
server_args=server_args,
nccl_port=port_args.nccl_port,
target_worker=self.tp_worker,
dp_rank=dp_rank,
)
elif self.spec_algorithm.is_ngram():
from sglang.srt.speculative.ngram_worker import NGRAMWorker
self.draft_worker = NGRAMWorker(
gpu_id=gpu_id,
tp_rank=tp_rank,
moe_ep_rank=moe_ep_rank,
server_args=server_args,
nccl_port=port_args.nccl_port,
target_worker=self.tp_worker,
dp_rank=dp_rank,
)
else:
self.draft_worker = None
self.launch_draft_worker(
gpu_id, tp_rank, moe_ep_rank, server_args, port_args, dp_rank
)
# Dispatch the model worker
if self.spec_algorithm.is_none():