paddle 在昇腾Atlas 300I Duo 通过 Docker 安装后，运行报错 无法使用 NPU

[lezhizhe]

安装Paddle Docker镜像后，执行命令报错：

python -c "import paddle; paddle.utils.run_check()"

I0325 10:26:09.328598   449 init.cc:236] ENV [CUSTOM_DEVICE_ROOT]=/usr/local/lib/python3.10/dist-packages/paddle_custom_device
I0325 10:26:09.328639   449 init.cc:145] Try loading custom device libs from: [/usr/local/lib/python3.10/dist-packages/paddle_custom_device]
I0325 10:26:09.949532   449 custom_device.cc:1099] Succeed in loading custom runtime in lib: /usr/local/lib/python3.10/dist-packages/paddle_custom_device/libpaddle-custom-npu.so
I0325 10:26:09.954984   449 custom_kernel.cc:63] Succeed in loading 357 custom kernel(s) from loaded lib(s), will be used like native ones.
I0325 10:26:09.955219   449 init.cc:157] Finished in LoadCustomDevice with libs_path: [/usr/local/lib/python3.10/dist-packages/paddle_custom_device]
I0325 10:26:09.955257   449 init.cc:242] CustomDevice: npu, visible devices count: 8
Running verify PaddlePaddle program ... 
I0325 10:26:10.864704   449 pir_interpreter.cc:1480] New Executor is Running ...
I0325 10:26:10.867235   449 pir_interpreter.cc:1506] pir interpreter is running by multi-thread mode ...
W0325 10:26:12.564280   639 pir_interpreter.cc:1980] Instruction OP id: 7, Ir OP id: 66, pd_op.sum_grad raises an EnforceNotMet exception common::enforce::EnforceNotMet
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/usr/local/lib/python3.10/dist-packages/paddle/utils/install_check.py", line 273, in run_check
    _run_static_single(use_cuda, use_xpu, use_custom, custom_device_name)
  File "/usr/local/lib/python3.10/dist-packages/paddle/utils/install_check.py", line 153, in _run_static_single
    exe.run(
  File "/usr/local/lib/python3.10/dist-packages/paddle/base/executor.py", line 1896, in run
    res = self._run_pir_impl(
  File "/usr/local/lib/python3.10/dist-packages/paddle/base/executor.py", line 2242, in _run_pir_impl
    ret = new_exe.run(list(feed.keys()), return_numpy)
  File "/usr/local/lib/python3.10/dist-packages/paddle/base/executor.py", line 845, in run
    tensors = self._new_exe.run(
OSError: In user code:


    ExternalError:  ACL error, the error code is : 500001.  (at /paddle/backends/npu/kernels/funcs/npu_op_runner.cc:445)
      [operator < pd_kernel.phi_kernel > error]

docker-compose.yaml 文件如下：

version: '3' 

services:
  paddle-npu-dev:
    container_name: paddle-npu  # 指定容器名称
    #image: ccr-2vdh3abv-pub.cnc.bj.baidubce.com/device/paddle-npu:cann80RC2-ubuntu20-npu-base-aarch64-gcc84  # 基础镜像（需替换实际架构）
    build: 
        ./
    stdin_open: true  # 保持 STDIN 打开（对应 -i 参数）
    tty: true         # 分配伪终端（对应 -t 参数）
    privileged: true  # 特权模式（访问所有设备）
    network_mode: host  # 使用宿主机网络
    working_dir: /work  # 容器默认工作目录
    shm_size: 128G     # 共享内存大小
    environment:
      - ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7  # 指定可见 NPU 设备
    volumes:
      # 挂载当前目录到容器 /work（注意替换为实际路径）
      - /www/paddlespeech/work:/work
      # 挂载 NPU 驱动相关文件
      - /usr/local/Ascend/driver:/usr/local/Ascend/driver
      - /usr/local/bin/npu-smi:/usr/local/bin/npu-smi
      - /usr/local/dcmi:/usr/local/dcmi
    command: /bin/bash  # 容器启动命令

Dockerfile 文件如下：

# 使用指定的基础镜像
FROM ccr-2vdh3abv-pub.cnc.bj.baidubce.com/device/paddle-npu:cann80RC2-ubuntu20-npu-base-aarch64-gcc84

# 设置工作目录
WORKDIR /work

# 配置pip使用阿里云镜像源
RUN pip config set global.index-url https://mirrors.aliyun.com/pypi/simple

RUN python -m pip install paddlepaddle==3.0.0b2 -i https://www.paddlepaddle.org.cn/packages/stable/cpu/ && \
    python -m pip install paddle-custom-npu==3.0.0b2 -i https://www.paddlepaddle.org.cn/packages/stable/npu/

# 设置工作目录
WORKDIR /work

服务器信息：

服务器配置：
CPU：2Kunpeng 920-5250
GPU：4300I Duo
内存：32*32GB

操作系统：BCLinux-for-Euler-21.10-dvd-aarch64-230324

其他补充信息 Additional Supplementary Information

No response

[z-one]

是不是意味着Atlas 300I Duo 不支持 paddle