fusion in fx graph mode did not take care of direct attribute access

[jerryzh168]

🐛 Bug

att

To Reproduce

from torch.ao.quantization.quantize_fx import prepare_fx, convert_fx
import torch

class M(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = torch.nn.Linear(5, 5)
    
    def forward(self, x):
        b = self.linear.bias
        return torch.nn.functional.relu(self.linear(x)) + b

m = M().eval()
m = prepare_fx(m, {"": torch.ao.quantization.default_qconfig})
print(m)
m = convert_fx(m)

Output

'LinearReLU' object has no attribute 'bias'
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-2-e6547926d9ec> in <module>
     12 
     13 m = M().eval()
---> 14 m = prepare_fx(m, {"": torch.ao.quantization.default_qconfig})
     15 print(m)
     16 m = convert_fx(m)
/mnt/xarfuse/uid-127034/b9cbaed7-seed-nspid4026531836_cgpid354053-ns-4026531840/torch/ao/quantization/quantize_fx.py in prepare_fx(model, qconfig_dict, prepare_custom_config_dict, equalization_qconfig_dict, backend_config_dict)
    498     torch._C._log_api_usage_once("quantization_api.quantize_fx.prepare_fx")
    499     assert not model.training, "prepare_fx only works for models in " + "eval mode"
--> 500     return _prepare_fx(
    501         model,
    502         qconfig_dict,
/mnt/xarfuse/uid-127034/b9cbaed7-seed-nspid4026531836_cgpid354053-ns-4026531840/torch/ao/quantization/quantize_fx.py in _prepare_fx(model, qconfig_dict, prepare_custom_config_dict, equalization_qconfig_dict, backend_config_dict, is_standalone_module)
    232     for attr_name in preserved_attributes:
    233         setattr(graph_module, attr_name, getattr(model, attr_name))
--> 234     graph_module = _fuse_fx(graph_module, prepare_custom_config_dict)
    235     prepared = prepare(
    236         graph_module,
/mnt/xarfuse/uid-127034/b9cbaed7-seed-nspid4026531836_cgpid354053-ns-4026531840/torch/ao/quantization/quantize_fx.py in _fuse_fx(graph_module, fuse_custom_config_dict)
     56     _check_is_graph_module(graph_module)
     57     fuser = Fuser()
---> 58     return fuser.fuse(graph_module, fuse_custom_config_dict)
     59 
     60 
/mnt/xarfuse/uid-127034/b9cbaed7-seed-nspid4026531836_cgpid354053-ns-4026531840/torch/ao/quantization/fx/fuse.py in fuse(self, model, fuse_custom_config_dict)
     60 
     61         preserved_attributes = set(fuse_custom_config_dict.get("preserved_attributes", []))
---> 62         model = FusedGraphModule(input_root, self.fused_graph, preserved_attributes)
     63         return model
     64 
/mnt/xarfuse/uid-127034/b9cbaed7-seed-nspid4026531836_cgpid354053-ns-4026531840/torch/ao/quantization/fx/graph_module.py in __init__(self, root, graph, preserved_attr_names)
      9         self.preserved_attr_names = preserved_attr_names
     10         preserved_attrs = {attr: getattr(root, attr) for attr in self.preserved_attr_names if hasattr(root, attr)}
---> 11         super().__init__(root, graph)
     12         for attr in preserved_attrs:
     13             setattr(self, attr, preserved_attrs[attr])
/mnt/xarfuse/uid-127034/b9cbaed7-seed-nspid4026531836_cgpid354053-ns-4026531840/torch/fx/graph_module.py in __init__(self, root, graph, class_name)
    288                 if node.op in ['get_attr', 'call_module']:
    289                     assert isinstance(node.target, str)
--> 290                     _copy_attr(root, self, node.target)
    291         elif isinstance(root, dict):
    292             targets_to_copy = []
/mnt/xarfuse/uid-127034/b9cbaed7-seed-nspid4026531836_cgpid354053-ns-4026531840/torch/fx/graph_module.py in _copy_attr(from_module, to_module, target)
    196         from_module, to_module = f, t
    197 
--> 198     orig = getattr(from_module, field)
    199     # If it is a tensor and not a parameter attribute of a module, it should be a named buffer.
    200     # So, we register it as a named buffer in the target module.
/mnt/xarfuse/uid-127034/b9cbaed7-seed-nspid4026531836_cgpid354053-ns-4026531840/torch/nn/modules/module.py in __getattr__(self, name)
   1183             if name in modules:
   1184                 return modules[name]
-> 1185         raise AttributeError("'{}' object has no attribute '{}'".format(
   1186             type(self).__name__, name))
   1187 
AttributeError: 'LinearReLU' object has no attribute 'bias'

Expected behavior

Should run successfully.

Environment

pytorch master

cc @jianyuh @raghuramank100 @jamesr66a @vkuzo @jgong5 @Xia-Weiwen @leslie-fang-intel @jerryzh168

[parakh08]

I am also facing similar issue. Did you get a solution?

[AndroidDevelopersTools]

I am also facing similar issue. Did you get a solution?

Do you find any solutions？