merge parallel-adapter succeed

Author: ShengdingHu

docs/source/notes/faq.md

@@ -7,3 +7,8 @@
 2. **Available Models with default configurations are ..., Please manually add the delta models by speicifying 'modified_modules' based on the visualization of your model structure**
 
     Although most pre-trained models (PTMs) use the transformers archtecture, they are implemented differently. For example, the attention module in GPT2 and BERT is not only named differently, but also implemented in different ways. Common structure mapping mapps the different name conventions of different PTMs into a unified name convention. But there are many PTMs that we do not currently cover. But don't worry! For these models, you can figure out which modules should you modify by simply [visualizing the PTMs](visualization), and then specify the `modified modules` manually (See [name-based addressing](namebasedaddr)). 
+
+
+3. **Requires a dummy_inputs to be passed through the model to understand the dimensionality of each tensor in the computation graph. The {module.__class__.__name__} Class has no dummy_inputs, and automatically created dummy_inputs failed.**
+
+    The `dummy_inputs` can be any data that make `backbone_model.forward(**dummy_inputs)` succeed. Only the form and shape of the `dummy_inputs` matter. To set dummy_inputs for your model, please use: `setattr(backbone_model, 'dummy_inputs', some_dummy_inputs)` before initializing `{self.__class__.__name__}`.

opendelta/basemodel.py

@@ -372,7 +372,7 @@ def _pseudo_data_to_instantiate_module(self, module: Optional[nn.Module]=None):
         except:
             _auto_dummy_fail = True
         if _auto_dummy_fail:  
-            raise AttributeError(f"\nThe {self.__class__.__name__} requires a pseudo-data to be passed through the model to understand the dimensionality of each tensor in the computation graph. \nThe automatically created dummy inputs failed.\nThe `dummy_inputs` can be any data that make `backbone_model.forward(**dummy_inputs)` succeed. Only the form and shape of the `dummy_inputs` matter.\n\tTo set dummy_inputs for your model, please use: `setattr(backbone_model, 'dummy_inputs', some_dummy_inputs)` before initializing `{self.__class__.__name__}` ")
+            raise AttributeError(f"\n\tThe {self.__class__.__name__} requires a dummy_inputs to be passed through the model to understand the dimensionality of each tensor in the computation graph. \n\t The {module.__class__.__name__} Class has no dummy_inputs, and automatically created dummy_inputs failed.\n\t Refer to `https://opendelta.readthedocs.io/en/latest/notes/faq.html` for detail.")
 
 
 
@@ -804,13 +804,7 @@ def detach(self, module: Optional[nn.Module]=None, reset_state_dict=True):
 
                     if _delta_info['method'] == "replace":
                         setattr(submodule, _delta_info["child_name"], _delta_info['org_module'])
-                    elif _delta_info['method'] == "insert_sequential":
-                        if hasattr(submodule.forward, "__wrapped__"):
-                            submodule.forward = submodule.forward.__wrapped__
-                            delattr(submodule, _delta_info["delta_name"])
-                        else:
-                            raise AttributeError("submodule {}'s forward has no attribute __wrapped__. It's not a wrapped function.".format(name))
-                    elif _delta_info['method'] == "insert_parallel":
+                    elif _delta_info['method'] in ["sequential", "before", "after", "parallel"]:
                         if hasattr(submodule.forward, "__wrapped__"):
                             submodule.forward = submodule.forward.__wrapped__
                             delattr(submodule, _delta_info["delta_name"])

opendelta/delta_models/parallel_adapter.py

@@ -5,12 +5,9 @@
 from opendelta.utils.name_based_addressing import *
 from opendelta.utils.cuda import get_device
 from opendelta.basemodel import DeltaBase
-import loralib as lora
 import torch.nn as nn
 import torch
-import math
 from opendelta.delta_models.layers.activations import Activations
-import inspect
 from opendelta import BaseDeltaConfig
 import opendelta.utils.logging as logging
 logger = logging.get_logger(__name__)
@@ -147,20 +144,24 @@ class attributes:
 
     """
     config_class = ParallelAdapterConfig
-    delta_type = "adapter"
-    default_modified_modules = ["attn", "attn", "ff.w1", "ff.w2"]
+    delta_type = "parallel_adapter"
+    default_modified_modules = ["attn@", "attn@", "ff@.w1@", "ff@.w2@"]
+    # default_modified_modules = ["attn", "attn", "ff.w1", "ff.w2"]
+    _need_pseudo_data = True
     def __init__(self,
                  backbone_model: nn.Module, 
                  bottleneck_dim: Optional[int]=24, 
                  non_linearity: Optional[str]='gelu_new',
                  modified_modules: Optional[bool] = None,
+                 exclude_modules: Optional[List[str]] = None,
                  unfrozen_modules: Optional[bool] = None,
                  common_structure: Optional[bool] = None,
                  interactive_modify: Optional[Union[bool, int]] = False,
                  ):
         DeltaBase.__init__(self, 
                            backbone_model, 
                            modified_modules=modified_modules,
+                           exclude_modules=exclude_modules,
                            unfrozen_modules=unfrozen_modules,
                            common_structure=common_structure,
                            interactive_modify=interactive_modify,