add --device

[shink]

• dcgan
• gat
• gcn
• language_translation

BACKEND_DEVICE=npu ./run_python_examples.sh "run_all"

Finished run_all, status 0
Some python examples failed:
saved models not found
mnist hogwild failed
graph convolutional network failed

Errors

1. gcn

root@3dfeb58e2e6e:~/pytorch-examples/gcn# python main.py --device npu
[W920 02:46:29.176903976 OperatorEntry.cpp:155] Warning: Warning only once for all operators,  other operators may also be overridden.
  Overriding a previously registered kernel for the same operator and the same dispatch key
  operator: aten::empty.memory_format(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, MemoryFormat? memory_format=None) -> Tensor
    registered at /pytorch/build/aten/src/ATen/RegisterSchema.cpp:6
  dispatch key: CPU
  previous kernel: registered at build/CMakeFiles/torch_npu.dir/compiler_depend.ts:497
       new kernel: registered at build/CMakeFiles/torch_npu.dir/compiler_depend.ts:100 (function operator())
Using npu device
Downloading dataset...
Loading dataset...
Traceback (most recent call last):
  File "/root/pytorch-examples/gcn/main.py", line 260, in <module>
    train_iter(epoch + 1, gcn, optimizer, criterion, (features, adj_mat), labels, idx_train, idx_val, args.val_every)
  File "/root/pytorch-examples/gcn/main.py", line 175, in train_iter
    output = model(*input)
  File "/usr/local/python3.9/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/usr/local/python3.9/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
    return forward_call(*args, **kwargs)
  File "/root/pytorch-examples/gcn/main.py", line 104, in forward
    x = self.gc1(input_tensor, adj_mat)
  File "/usr/local/python3.9/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/usr/local/python3.9/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
    return forward_call(*args, **kwargs)
  File "/root/pytorch-examples/gcn/main.py", line 58, in forward
    support = torch.mm(input_tensor, self.kernel) # Matrix multiplication between input and weight matrix
RuntimeError: CAUTION: The operator 'aten::addmm' is not currently supported on the NPU backend.
[ERROR] 2024-09-20-02:46:48 (PID:146560, Device:0, RankID:-1) ERR01007 OPS feature not supported

2. language_translation

上游 CI 已经不跑这个 example 了

no arm64:

3. mnist_hogwild

root@3dfeb58e2e6e:~/pytorch-examples/mnist_hogwild# python main.py --device npu
[W920 06:22:43.455693712 OperatorEntry.cpp:155] Warning: Warning only once for all operators,  other operators may also be overridden.
  Overriding a previously registered kernel for the same operator and the same dispatch key
  operator: aten::empty.memory_format(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, MemoryFormat? memory_format=None) -> Tensor
    registered at /pytorch/build/aten/src/ATen/RegisterSchema.cpp:6
  dispatch key: CPU
  previous kernel: registered at build/CMakeFiles/torch_npu.dir/compiler_depend.ts:497
       new kernel: registered at build/CMakeFiles/torch_npu.dir/compiler_depend.ts:100 (function operator())
Traceback (most recent call last):
  File "/root/pytorch-examples/mnist_hogwild/main.py", line 91, in <module>
    model.share_memory() # gradients are allocated lazily, so they are not shared here
  File "/usr/local/python3.9/lib/python3.9/site-packages/torch_npu/utils/npu_intercept.py", line 78, in wrapper
    raise RuntimeError(f"{str(func)} is not supported in npu." + pta_error(ErrCode.NOT_SUPPORT))
RuntimeError: <function Module.share_memory at 0xfffdff9013a0> is not supported in npu.
[ERROR] 2024-09-20-06:22:53 (PID:150151, Device:0, RankID:-1) ERR00007 PTA feature not supported