minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[NPU] Support nccl backend for --remote-instance-weight-loader (#39413)

Browse Source
This commit is contained in:
silencejade
2026-09-17 17:28:15 +08:00
committed by GitHub
parent e970453b43
commit 575759d90a
2 changed files with 6 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/connector/remote_instance.py
+2 -1
View File
@@ -9,6 +9,7 @@ import torch.distributed as dist
from sglang.srt.connector import BaseConnector from sglang.srt.connector import BaseConnector
from sglang.srt.utils import init_custom_process_group from sglang.srt.utils import init_custom_process_group
from sglang.srt.utils.common import is_npu
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
@@ -43,7 +44,7 @@ class RemoteInstanceConnector(BaseConnector):
master_address = parsed_url.hostname master_address = parsed_url.hostname
master_port = parsed_url.port master_port = parsed_url.port
group_name = f"send_weights_{instance_ip}_{master_port}_{tp_rank}" group_name = f"send_weights_{instance_ip}_{master_port}_{tp_rank}"
backend = "nccl" backend = "nccl" if not is_npu() else "hccl"
logger.info( logger.info(
f"init custom process group: master_address={master_address}, master_port={master_port}, " f"init custom process group: master_address={master_address}, master_port={master_port}, "
python/sglang/srt/model_executor/model_runner_components/weight_exporter.py
+4 -1
View File
@@ -9,6 +9,7 @@ import torch.distributed as dist
from sglang.srt.platforms import current_platform from sglang.srt.platforms import current_platform
from sglang.srt.utils import init_custom_process_group from sglang.srt.utils import init_custom_process_group
from sglang.srt.utils.common import is_npu
from sglang.srt.utils.network import NetworkAddress from sglang.srt.utils.network import NetworkAddress
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
@@ -43,6 +44,8 @@ class WeightExporter:
) )
group_port = ports_list[self.tp_rank] group_port = ports_list[self.tp_rank]
group_name = f"{group_name}_{group_port}_{self.tp_rank}" group_name = f"{group_name}_{group_port}_{self.tp_rank}"
backend = "hccl" if backend == "nccl" and is_npu() else backend
device_type = "cuda" if not is_npu() else "npu"
logger.info( logger.info(
f"init custom process group: tp_rank={self.tp_rank}, gpu_id={self.gpu_id}, master_address={master_address}, master_port={group_port}, " f"init custom process group: tp_rank={self.tp_rank}, gpu_id={self.gpu_id}, master_address={master_address}, master_port={group_port}, "
@@ -60,7 +63,7 @@ class WeightExporter:
world_size=world_size, world_size=world_size,
rank=group_rank, rank=group_rank,
group_name=group_name, group_name=group_name,
device_id=torch.device("cuda", self.gpu_id), device_id=torch.device(device_type, self.gpu_id),
) )
dist.barrier(group=self._weights_send_group[group_name]) dist.barrier(group=self._weights_send_group[group_name])
success = True success = True