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

[NPU][feature adapt]remote load weight feature adp npu (#17968)

Browse Source 
Co-authored-by: littleYellowBicycle <liguo29@huawei.com>
Co-authored-by: sglang-npu-bot <sglangnpu@163.com>
This commit is contained in:
littleyellowbicycle
2026-02-27 10:10:36 +08:00
committed by GitHub
parent e567215e44
commit 5b5c509480
3 changed files with 442 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/connector/remote_instance.py
View File

@@ -17,7 +17,7 @@ class RemoteInstanceConnector(BaseConnector):
def __init__(self, url: str, device: torch.device = "cpu"):
assert (
device.type == "cuda"
device.type == "cuda" or device.type == "npu"
), "RemoteInstanceConnector only supports cuda device."
super().__init__(url)
self.url = url
@@ -32,7 +32,7 @@ class RemoteInstanceConnector(BaseConnector):
world_size: int = 2,
):
assert (
self.device.type == "cuda"
self.device.type == "cuda" or self.device.type == "npu"
), "RemoteInstanceConnector only supports cuda device."
assert (
gpu_id != -1 and tp_rank != -1

4
python/sglang/srt/server_args.py
View File

@@ -2563,8 +2563,8 @@ class ServerArgs:
)
self.load_format = "auto"
elif (
not self.validate_transfer_engine()
and self.remote_instance_weight_loader_backend == "transfer_engine"
self.remote_instance_weight_loader_backend == "transfer_engine"
and not self.validate_transfer_engine()
):
logger.warning(
"Fallback load_format to 'auto' due to 'transfer_engine' backend is not supported."