From 575759d90af942eeff89f1c8c33a1fcdb2da4181 Mon Sep 17 00:00:00 2001
From: silencejade <222910030+silencejade@users.noreply.github.com>
Date: Thu, 17 Sep 2026 17:28:15 +0800
Subject: [PATCH] [NPU] Support nccl backend for
 --remote-instance-weight-loader (#39413)

---
 python/sglang/srt/connector/remote_instance.py               | 3 ++-
 .../model_runner_components/weight_exporter.py               | 5 ++++-
 2 files changed, 6 insertions(+), 2 deletions(-)

diff --git a/python/sglang/srt/connector/remote_instance.py b/python/sglang/srt/connector/remote_instance.py
index 875063bd4..f28270368 100644
--- a/python/sglang/srt/connector/remote_instance.py
+++ b/python/sglang/srt/connector/remote_instance.py
@@ -9,6 +9,7 @@ import torch.distributed as dist
 
 from sglang.srt.connector import BaseConnector
 from sglang.srt.utils import init_custom_process_group
+from sglang.srt.utils.common import is_npu
 
 logger = logging.getLogger(__name__)
 
@@ -43,7 +44,7 @@ class RemoteInstanceConnector(BaseConnector):
         master_address = parsed_url.hostname
         master_port = parsed_url.port
         group_name = f"send_weights_{instance_ip}_{master_port}_{tp_rank}"
-        backend = "nccl"
+        backend = "nccl" if not is_npu() else "hccl"
 
         logger.info(
             f"init custom process group: master_address={master_address}, master_port={master_port}, "
diff --git a/python/sglang/srt/model_executor/model_runner_components/weight_exporter.py b/python/sglang/srt/model_executor/model_runner_components/weight_exporter.py
index d9f181e56..0d3983731 100644
--- a/python/sglang/srt/model_executor/model_runner_components/weight_exporter.py
+++ b/python/sglang/srt/model_executor/model_runner_components/weight_exporter.py
@@ -9,6 +9,7 @@ import torch.distributed as dist
 
 from sglang.srt.platforms import current_platform
 from sglang.srt.utils import init_custom_process_group
+from sglang.srt.utils.common import is_npu
 from sglang.srt.utils.network import NetworkAddress
 
 logger = logging.getLogger(__name__)
@@ -43,6 +44,8 @@ class WeightExporter:
         )
         group_port = ports_list[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(
             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,
                 rank=group_rank,
                 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])
             success = True