yolov11使用torch_npu进行相关推理，在npu选择了设备之后推理会失败

一、问题现象（附报错日志上下文）：
我使用yolov11，利用torch_npu进行相关推理，如果利用model.to("npu:0")选择了npu设备的话，推理过程会报错，如果不选择设备的话，不会报错，但是推理时间过长，一张图片需要3S才能推理陈工

二、软件版本:
-- CANN 版本 (e.g., CANN 3.0.x，5.x.x):  8.0.RC2.alpha003
--Tensorflow/Pytorch/MindSpore 版本:
--Python 版本 (e.g., Python 3.7.5):3.11
-- torch和torch_npu的版本 :2.1.0；2.1.0.post10
三、测试步骤：
使用的官方的yolov11s的模型在npu上的跑，推理时间过长，一张图片在3s左右，相关代码如下
```
# Copyright (c) 2025 Huawei Technologies Co., Ltd
# [Software Name] is licensed under Mulan PSL v2.
# You can use this software according to the terms and conditions of the Mulan PSL v2.
# You may obtain a copy of Mulan PSL v2 at:
#          http://license.coscl.org.cn/MulanPSL2
# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
# See the Mulan PSL v2 for more details.
import argparse
import torch
import torch_npu
from ultralytics import YOLO
from ultralytics.utils.plotting import Annotator, colors, save_one_box
import time
import cv2
def main(args):
    # torch_npu.npu.set_compile_mode(jit_compile=False)
    torch.npu.set_compile_mode(jit_compile=False)
    torch_npu.npu.set_device(0)
    model = YOLO(args.pth)
    # 取消掉的话，推理时间过长
    # model.to("npu:0")
    model.model.eval()
    print("加载模型成功")
    cap = cv2.VideoCapture("./test2.mp4")
        # 检查视频是否成功打开
    print("加载视频成功")
    if not cap.isOpened():
        print("无法打开视频文件")
    is_obb = False
    print("开始推理")
    while cap.isOpened():
        # 逐帧读取视频
        ret, frame = cap.read()
        # 检查帧是否读取成功
        if not ret:
            print("视频播放结束或读取失败")
            return
        images = cv2.resize(frame, (640, 640))
        # results = model.predict(source=frame, conf=0.25, device=torch.device("npu:0"), half=False)
        results = model.predict(source=frame, conf=0.25, imgsz=640, half=False)
        # for result in results:
        #     pred_boxes = result.boxes  # Boxes object for bounding box outputs
        #     print(pred_boxes)
            # for i, d in enumerate(reversed(pred_boxes)):
            #     box = d.xyxyxyxy.reshape(-1, 4, 2).squeeze() if is_obb else d.xyxy.squeeze()
            #     box = box.numpy()
            #     x1, y1, x2, y2 = box[0], box[1], box[2], box[3]
                # cv2.rectangle(frame, (int(x1), int(y1)), (int(x2), int(y2)), (0, 255, 0), 2)  # 绘制边界框，颜色为绿色
        # cv2.imshow("frame", frame)
        # cv2.waitKey(1)
    cap.release()
    cap.destroyAllWindows()
if __name__ == '__main__':
    parser = argparse.ArgumentParser(description="yolov11 dataset convert")
    parser.add_argument("--pth", type=str, help="weight", default="yolo11s.pt")
    parser.add_argument("--dataset", type=str, help="dataset")
    parser.add_argument("--batchsize", type=int, help="batchsize")
    args = parser.parse_args()
    main(args)
```
使用这个代码能够跑通，但是推理时间比较长，感觉像是没有在npu上运行一样
![输入图片说明](https://foruda.gitee.com/images/1756802310804448146/977b5abe_11427886.png "5d0a48fa-c18e-4a6d-8c8e-71fb2e7bc08e.png")
如果把代码的model.to("npu:0")给打开的话，会直接报错：
Traceback (most recent call last):
  File "/home/app/Yolov11_for_PyTorch/infer.py", line 62, in <module>
    main(args)
  File "/home/app/Yolov11_for_PyTorch/infer.py", line 43, in main
    results = model.predict(source=frame, conf=0.25, imgsz=640, half=False)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/app/Yolov11_for_PyTorch/ultralytics/engine/model.py", line 553, in predict
    self.predictor.setup_model(model=self.model, verbose=is_cli)
  File "/home/app/Yolov11_for_PyTorch/ultralytics/engine/predictor.py", line 310, in setup_model
    self.model = AutoBackend(
                 ^^^^^^^^^^^^
  File "/root/miniconda3/envs/yolo11/lib/python3.11/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/app/Yolov11_for_PyTorch/ultralytics/nn/autobackend.py", line 152, in __init__
    model = model.fuse(verbose=verbose)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/app/Yolov11_for_PyTorch/ultralytics/nn/tasks.py", line 206, in fuse
    m.conv = fuse_conv_and_bn(m.conv, m.bn)  # update conv
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/app/Yolov11_for_PyTorch/ultralytics/utils/torch_utils.py", line 261, in fuse_conv_and_bn
    w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
                                               ~~~~~~~^~~~~~~~~~~~~~~~
RuntimeError: call aclnnAdds failed, detail:EZ9999: Inner Error!
EZ9999: 2025-09-02-16:41:22.122.179  Parse dynamic kernel config fail.
        TraceBack (most recent call last):
        AclOpKernelInit failed opType
        Op Add does not has any binary.
        Kernel Run failed. opType: 3, Add
        launch failed for Add, errno:561000.

[ERROR] 2025-09-02-16:41:22 (PID:3108812, Device:0, RankID:-1) ERR01100 OPS call acl api failed
![输入图片说明](https://foruda.gitee.com/images/1756802470399228959/7876c531_11427886.png "3727b504-5532-47ed-aae0-188a9e838b2e.png")

Ascend/pytorch
暂停

内容风险标识

评论 (11)

Ascend/pytorch暂停 .gitee-modal { width: 500px !important; }

内容风险标识