RuntimeError: 0 INTERNAL ASSERT FAILED We don‘t have an op for aten::to but it isn‘t a special case.

在使用nni 工具对PyTorch 模型进行剪枝的时候，出现了如下的 Runtime Error，查询并解决后，记录在此，给出现相同错误的人一个参考。

RuntimeError: 0 INTERNAL ASSERT FAILED at /opt/conda/conda-bld/pytorch_1587428398394/work/torch/csrc/jit/ir/alias_analysis.cpp:318, please report a bug to PyTorch. We don’t have an op for aten::to but it isn’t a special case. Argument types: Tensor, None, int, Device, bool, bool, bool, int,

代码以及环境如下：

'''
python 3.7
pytorch 1.5.0
nni 2.3
'''
pruner = FPGMPruner(net_stu.module, config_list,dependency_aware=True, dummy_input=dummy_input)

报错信息栈如下：

[11] Traceback (most recent call last):
[12]   File "/disheng/BMP-mo/train_pruner.py", line 332, in <module>
[13]     train(args)
[14]   File "/disheng/BMP-mo/train_pruner.py", line 139, in train
[15]     pruner = FPGMPruner(net_stu, config_list,dependency_aware=True, dummy_input=dummy_input)
[16]   File "/opt/conda/lib/python3.7/site-packages/nni/algorithms/compression/pytorch/pruning/one_shot_pruner.py", line 166, in __init__
[17]     dummy_input=dummy_input)
[18]   File "/opt/conda/lib/python3.7/site-packages/nni/algorithms/compression/pytorch/pruning/one_shot_pruner.py", line 40, in __init__
[19]     super().__init__(model, config_list, None, pruning_algorithm, dependency_aware, dummy_input, **algo_kwargs)
[20]   File "/opt/conda/lib/python3.7/site-packages/nni/algorithms/compression/pytorch/pruning/dependency_aware_pruner.py", line 45, in __init__
[21]     self.graph = TorchModuleGraph(model, dummy_input)
[22]   File "/opt/conda/lib/python3.7/site-packages/nni/common/graph_utils.py", line 248, in __init__
[23]     super().__init__(model, dummy_input, traced_model)
[24]   File "/opt/conda/lib/python3.7/site-packages/nni/common/graph_utils.py", line 66, in __init__
[25]     self._trace(model, dummy_input)
[26]   File "/opt/conda/lib/python3.7/site-packages/nni/common/graph_utils.py", line 74, in _trace
[27]     self.trace = torch.jit.trace(model, dummy_input)
[28]   File "/opt/conda/lib/python3.7/site-packages/torch/jit/__init__.py", line 875, in trace
[29]     check_tolerance, _force_outplace, _module_class)
[30]   File "/opt/conda/lib/python3.7/site-packages/torch/jit/__init__.py", line 1027, in trace_module
[31]     module._c._create_method_from_trace(method_name, func, example_inputs, var_lookup_fn, _force_outplace)
[32] RuntimeError: 0 INTERNAL ASSERT FAILED at /opt/conda/conda-bld/pytorch_1587428398394/work/torch/csrc/jit/ir/alias_analysis.cpp:318, please report a bug to PyTorch. We don't have an op for aten::to but it isn't a special case.  Argument types: Tensor, None, int, Device, bool, bool, bool, int, 

从上面的报错信息，可以分析出，是FPGMPruner 在调用 self.trace = torch.jit.trace(model, dummy_input)的时候出现错误，原因是 net_stu 是 DataParallel 类型，这种类型还不能作为 torch.jit.trace的参数，这是一个官方bug，还没有在这个版本中修复。

解决办法：

在这种情况下，传入的参数从net_stu 替换为 net_stu.module，本质是是换了传入参数的类型。

pruner = FPGMPruner(net_stu.module, config_list,dependency_aware=True, dummy_input=dummy_input)

参考：
hasSpecialCase INTERNAL ASSERT FAILED: We don’t have an op for aten::to but it isn’t a special case.