File size: 10,839 Bytes
8124a18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
import torch
from .utils import parent_module, brackets_to_periods
import transformers
import os
os.environ['CUDA_LAUNCH_BLOCKING'] = "1"

def euc(query, key):
    # Euclidean distance
    if len(key.shape) < 2:
        key = key.view(1, -1)
    return torch.cdist(key, query, p=2)

def perturb_values(chosen_value, num_pert, device):
    # Create a bunch of noised versions of the value, then create batch, then train value
    chosen_value = chosen_value
    noise = torch.normal(0, 1, chosen_value.shape, device=device)
    noise[0] = noise[0]*0
    noise.requires_grad = True
    chosen_value = chosen_value + noise
    return chosen_value

class GRACE(torch.nn.Module):
    def __init__(self, config, model, device):
        super(GRACE, self).__init__()
        self.config = config
        self.log_dict = {}
        self.model = model
        # self.tokenizer = model.tokenizer
        layer = config.inner_params[0]
        self.device = device

        # --- ensure proper formatting (GRACE edits ~layers~ not weights matrices) ---        
        suffixes = [".weight", ".bias"]
        self.layer = layer.rsplit(".", 1)[0] if any(layer.endswith(x) for x in suffixes) else layer
        
        for n, p in self.model.named_parameters():
            p.requires_grad = False
        
        if isinstance(self.model, transformers.models.gpt2.modeling_gpt2.GPT2LMHeadModel):
            transpose = False
        else:
            transpose = True

        # --- Add GRACE to chosen layers ---
        edit_module = parent_module(self.model, brackets_to_periods(self.layer))
        layer_name = self.layer.rsplit(".", 1)[-1]
        original_layer = getattr(edit_module, layer_name)

        if type(original_layer) is not GRACEAdapter:
            setattr(edit_module, layer_name, GRACEAdapter(config, original_layer, transpose=transpose).to(self.device))
        
    def __call__(self, **kwargs):
        # if self.config.task == "hallucination":
        #     print(kwargs)
        #     key_id = (kwargs["labels"] == -100).sum() - 1
        #     setattr(eval(f"self.model.{self.layer}"), "key_id", key_id) # Tell GRACE which token to use for its query (default is the last token)
        return self.model(**kwargs)
    
    def generate(self, *args, **kwargs):
        setattr(eval(f"self.model.{self.layer}"), "key_id", -1)
        return self.model.generate(*args, **kwargs)
        
    def edit(self, config, tokens):
        key_id = (tokens["labels"] == -100).sum() - 1
        setattr(eval(f"self.model.{self.layer}"), "key_id", key_id)
        
        # --- pass edit label, training mode, and key_id into GRACE ---
        setattr(eval(f"self.model.{self.layer}"), "training", True)
        setattr(eval(f"self.model.{self.layer}"), "edit_label", tokens["labels"])
                
        self.losses = []
        # --- train GRACE value ---
        for i in range(config.n_iter):
            # --- insert iteration into each layer (only initiate keys on iteration 1) ---
            setattr(eval(f"self.model.{self.layer}"), "iter", i)
            
            # --- pass tokens through model (including through the GRACE layer) ---
            outputs = self.model(**tokens)
            if i == 0:
                # --- we only need to create an optimizer for the first iteration (but forward pass instantiates the key, so optimzer is passed after first inference) ---
                optimizer = torch.optim.Adam(self.model.parameters(), config.edit_lr)
            loss = outputs.loss
            loss.backward()
            optimizer.step()
            optimizer.zero_grad()
            self.losses.append(loss.detach().cpu().numpy())
        
        self.loss = loss # Log final loss

        # --- pull out info we want to log from the GRACE layer ---
        setattr(eval(f"self.model.{self.layer}"), "training", False)
        chosen_key = getattr(eval(f"self.model.{self.layer}"), "chosen_key")
        nkeys = len(getattr(eval(f"self.model.{self.layer}"), "keys"))
            
        self.log_dict["chosen_key"] =  chosen_key
        self.log_dict["nkeys"] = nkeys

class GRACEAdapter(torch.nn.Module):
    def __init__(self, config, layer, transpose):
        super(GRACEAdapter, self).__init__()

        self.layer = layer
        self.weight = self.layer.weight
        self.init_epsilon = config.eps
        self.dist_fn = config.dist_fn
        self.replacement = config.replacement
        self.device = layer.weight.device
        self.config = config
        self.num_pert = config.num_pert
        self.key_id = -1
        self.ensure_replace_token_loc = False
    
        if transpose:
            self.key_shape = layer.weight.shape[1]
            self.value_shape = layer.weight.shape[0]
        else:
            self.key_shape = layer.weight.shape[0]
            self.value_shape = layer.weight.shape[1]
        self.training = False

    def add_key(self, new_key, new_value):
        keys = torch.vstack([self.keys, new_key.detach()]) # Add new key to list of keys

        values = torch.nn.Parameter(torch.vstack([self.values, new_value]), requires_grad=True) # Add new value to list of values

        new_epsilon = torch.tensor(self.init_epsilon, device=self.device).view(1)
        epsilons = torch.vstack([self.epsilons, new_epsilon]) # Add new epsilon to list of epsilons

        key_labels = self.key_labels + [self.edit_label] # Add new key_label to list of key_labels

        return keys, values, epsilons, key_labels

    def init_key_value(self, query, value):
        key = query.detach()
        epsilon = torch.tensor(self.init_epsilon, device=self.device, requires_grad=False).view(1)
        key_label = [self.edit_label]
        return key, value, epsilon, key_label

    def label_match(self, edit_label, key_label):
        return edit_label.float().mean() == key_label.float().mean()

    def split_epsilons_in_half(self, nearest_key, smallest_distance):
        self.epsilons[nearest_key] = (smallest_distance / 2) - 1e-5 # Cut nearest epsilon in half
        self.epsilons[-1] = smallest_distance / 2 # Cut new epsilon in half
    
    def forward(self, *args):
        # Run layer forward and save what it would have returned for this instance
        layer_out = self.layer(*args)

        ### If training, we need to modify the codebook
        if (not self.training) & ('keys' not in self.__dict__):
            # If it's not training time and we haven't added any keys yet (this is before doing any editing)
            # print(self.__dict__)
            return layer_out
        else:
            if not self.training and not self.ensure_replace_token_loc and self.key_id == -1:
                token_to_edit = args[0].shape[1]-1
                self.key_id = args[0].shape[1]-1
                self.ensure_replace_token_loc = True
            else:
                token_to_edit = min(self.key_id, args[0].shape[1]-1) # args[0].shape[1] - 1 is sequence length
            query = args[0][:, token_to_edit, :] # Just use activation for last token
            if self.config.val_init == "cold":
                new_value = torch.nn.Parameter(torch.rand(1, self.value_shape, requires_grad=True, device=self.device))
            elif self.config.val_init == "warm":
                new_value = torch.nn.Parameter(layer_out[:, token_to_edit, :].detach(), requires_grad=True)

            if 'keys' not in self.__dict__:
                # If no keys exist, initialize keys, values, epsilons, and key labels
                self.keys, self.values, self.epsilons, self.key_labels = self.init_key_value(query, new_value)
            elif self.iter == 0:
                # Keys exist, so we have decide whether or not to update them (the fact that we've made it to this point means there was an error!)

                # --- search through keys for a match for query ---
                dists = torch.cdist(self.keys, query, p=2).view(-1, len(query))
                smallest_distance, nearest_key = dists.min(0)

                if smallest_distance > (self.init_epsilon + self.epsilons[nearest_key]):
                    # If there's no close key, make a new key                    
                    self.keys, self.values, self.epsilons, self.key_labels = self.add_key(query, new_value)
                else:
                    # If there is a close key, we need to handle conflicts
                    if not self.label_match(self.edit_label, self.key_labels[nearest_key]):
                        self.keys, self.values, self.epsilons, self.key_labels = self.add_key(query, new_value)
                        self.split_epsilons_in_half(nearest_key, smallest_distance)
                    else:
                        # If the current label is the SAME as the nearest label, just make the nearest epsilon bigger
                        if smallest_distance > self.epsilons[nearest_key]:
                            if self.config.eps_expand== "coverage":
                                self.epsilons[nearest_key] = smallest_distance # Replace nearest epsilon with dist between old key and new key
                            elif self.config.eps_expand == "moving_average":
                                a = 0.5
                                self.keys[nearest_key] = a*self.keys[nearest_key] + (1-a)*query # Move old key to be halfway between
                                self.epsilons[nearest_key] = smallest_distance
                                # self.epsilons[nearest_key] = smallest_distance + self.init_epsilon
            else:
                # If not iter 0, we don't need to change keys, we just need to learn the value
                pass
        # print(token_to_edit)
        # compute distance from query to all keys and find the closest keys
        dists = torch.cdist(self.keys, query, p=2).view(-1, len(query))
        smallest_dist, self.chosen_key = dists.min(0)
        smallest_dist = smallest_dist.view(-1, 1)
        chosen_value = self.values[self.chosen_key]
        eps = self.epsilons[self.chosen_key].view(-1, 1)

        if (self.config.val_train == "adv") and (self.training):
            chosen_value = perturb_values(chosen_value, self.num_pert, self.device)

        if self.replacement == "replace_all":
            layer_out = torch.where((smallest_dist <= eps).view(-1, 1, 1), chosen_value.unsqueeze(1).repeat_interleave(layer_out.shape[1], 1), layer_out)
        elif self.replacement == "replace_last":
            layer_out[:, token_to_edit] = torch.where((smallest_dist <= eps), chosen_value, layer_out[:, token_to_edit])
        elif self.replacement == "replace_prompt":
            layer_out[:, :token_to_edit] = torch.where((smallest_dist <= eps), chosen_value, layer_out[:, :token_to_edit])
        else:
            print("token replacement choice not found")
        return layer_out