ltg
/

davda54 commited on
Commit
c6d1bab
1 Parent(s): 89a7e2f

Upload folder using huggingface_hub

Browse files
config.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "NorT5ForConditionalGeneration"
4
+ ],
5
+ "auto_map": {
6
+ "AutoConfig": "configuration_nort5.NorT5Config",
7
+ "AutoModel": "modeling_nort5.NorT5Model",
8
+ "AutoModelForSeq2SeqLM": "modeling_nort5.NorT5ForConditionalGeneration",
9
+ "AutoModelForConditionalGeneration": "modeling_nort5.NorT5ForConditionalGeneration"
10
+ },
11
+ "attention_probs_dropout_prob": 0.0,
12
+ "bos_token_id": 5,
13
+ "cls_token_id": 1,
14
+ "eos_token_id": 6,
15
+ "hidden_dropout_prob": 0.0,
16
+ "hidden_size": 512,
17
+ "initializer_range": 0.02,
18
+ "intermediate_size": 1365,
19
+ "layer_norm_eps": 1e-07,
20
+ "max_position_embeddings": 512,
21
+ "num_attention_heads": 8,
22
+ "num_hidden_layers": 24,
23
+ "output_all_encoded_layers": true,
24
+ "pad_token_id": 3,
25
+ "position_bucket_size": 32,
26
+ "sep_token_id": 2,
27
+ "torch_dtype": "float32",
28
+ "transformers_version": "4.24.0",
29
+ "vocab_size": 65536,
30
+ "max_length": 512,
31
+ "max_new_tokens": 256,
32
+ "is_encoder_decoder": true
33
+ }
34
+
configuration_nort5.py ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers.configuration_utils import PretrainedConfig
2
+
3
+
4
+ class NorT5Config(PretrainedConfig):
5
+ """Configuration class to store the configuration of a `NorT5`.
6
+ """
7
+ def __init__(
8
+ self,
9
+ vocab_size=50000,
10
+ attention_probs_dropout_prob=0.1,
11
+ hidden_dropout_prob=0.1,
12
+ hidden_size=768,
13
+ intermediate_size=2048,
14
+ max_position_embeddings=512,
15
+ position_bucket_size=32,
16
+ num_attention_heads=12,
17
+ num_hidden_layers=12,
18
+ layer_norm_eps=1.0e-7,
19
+ output_all_encoded_layers=True,
20
+ pad_token_id=3,
21
+ cls_token_id=1,
22
+ sep_token_id=2,
23
+ bos_token_id=5,
24
+ eos_token_id=6,
25
+ **kwargs,
26
+ ):
27
+ super().__init__(**kwargs)
28
+
29
+ self.vocab_size = vocab_size
30
+ self.hidden_size = hidden_size
31
+ self.num_hidden_layers = num_hidden_layers
32
+ self.num_attention_heads = num_attention_heads
33
+ self.intermediate_size = intermediate_size
34
+ self.hidden_dropout_prob = hidden_dropout_prob
35
+ self.attention_probs_dropout_prob = attention_probs_dropout_prob
36
+ self.max_position_embeddings = max_position_embeddings
37
+ self.output_all_encoded_layers = output_all_encoded_layers
38
+ self.position_bucket_size = position_bucket_size
39
+ self.layer_norm_eps = layer_norm_eps
40
+ self.pad_token_id = pad_token_id
41
+ self.cls_token_id = cls_token_id
42
+ self.sep_token_id = sep_token_id
43
+ self.bos_token_id = bos_token_id
44
+ self.eos_token_id = eos_token_id
generation_config.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "decoder_start_token_id": 5,
4
+ "eos_token_id": 6,
5
+ "pad_token_id": 3,
6
+ "transformers_version": "4.27.0.dev0"
7
+ }
8
+
modeling_nort5.py ADDED
@@ -0,0 +1,711 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ from typing import List, Optional, Tuple, Union
3
+
4
+ import torch
5
+ import torch.nn as nn
6
+ import torch.nn.functional as F
7
+ from transformers.pytorch_utils import softmax_backward_data
8
+ from torch.utils import checkpoint
9
+
10
+ from .configuration_nort5 import NorT5Config
11
+ from transformers.modeling_utils import PreTrainedModel
12
+ from transformers.activations import gelu_new
13
+ from transformers.modeling_outputs import (
14
+ Seq2SeqModelOutput, Seq2SeqLMOutput, BaseModelOutput, BaseModelOutputWithPastAndCrossAttentions
15
+ )
16
+
17
+
18
+ class Encoder(nn.Module):
19
+ def __init__(self, config, activation_checkpointing=False):
20
+ super().__init__()
21
+ self.main_input_name = "input_ids"
22
+
23
+ self.relative_embedding = RelativeEmbedding(config)
24
+ self.layers = nn.ModuleList([EncoderLayer(config) for _ in range(config.num_hidden_layers)])
25
+
26
+ for i, layer in enumerate(self.layers):
27
+ layer.mlp.mlp[1].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
28
+ layer.mlp.mlp[-2].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
29
+
30
+ self.activation_checkpointing = activation_checkpointing
31
+
32
+ def forward(self, hidden_states, attention_mask):
33
+ relative_embedding = self.relative_embedding()
34
+ hidden_states, attention_probs = [hidden_states], []
35
+
36
+ for layer in self.layers:
37
+ if self.activation_checkpointing:
38
+ hidden_state, attention_p = checkpoint.checkpoint(layer, hidden_states[-1], attention_mask, relative_embedding)
39
+ else:
40
+ hidden_state, attention_p = layer(hidden_states[-1], attention_mask, relative_embedding)
41
+
42
+ hidden_states.append(hidden_state)
43
+ attention_probs.append(attention_p)
44
+
45
+ return hidden_states, attention_probs
46
+
47
+
48
+ class Decoder(nn.Module):
49
+ def __init__(self, config, activation_checkpointing=False):
50
+ super().__init__()
51
+ self.self_relative_embedding = RelativeEmbedding(config)
52
+ self.cross_relative_embedding = RelativeEmbedding(config)
53
+ self.layers = nn.ModuleList([DecoderLayer(config) for _ in range(config.num_hidden_layers)])
54
+
55
+ for i, layer in enumerate(self.layers):
56
+ layer.mlp.mlp[1].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
57
+ layer.mlp.mlp[-2].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
58
+
59
+ self.activation_checkpointing = activation_checkpointing
60
+
61
+ def forward(self, x, encoder_output, encoder_padding_mask, past_key_values=None):
62
+ self_relative_embedding = self.self_relative_embedding()
63
+ cross_relative_embedding = self.cross_relative_embedding()
64
+
65
+ if past_key_values is None:
66
+ autoreg_mask = torch.triu(
67
+ torch.full((x.size(0), x.size(0)), True, device=x.device),
68
+ diagonal=1
69
+ )
70
+ else:
71
+ autoreg_mask = None
72
+
73
+ # initialize past_key_values with `None` if past does not exist
74
+ if past_key_values is None:
75
+ past_key_values = [None] * len(self.layers)
76
+
77
+ hidden_states, self_attention_probs, cross_attention_probs, key_value_states = [x], [], [], []
78
+ for layer, past_key_value in zip(self.layers, past_key_values):
79
+ if self.activation_checkpointing:
80
+ hidden_state, self_attention_p, cross_attention_p, key_value_state = checkpoint.checkpoint(layer, hidden_states[-1], autoreg_mask, encoder_output, encoder_padding_mask, self_relative_embedding, cross_relative_embedding, past_key_value=None)
81
+ else:
82
+ hidden_state, self_attention_p, cross_attention_p, key_value_state = layer(hidden_states[-1], autoreg_mask, encoder_output, encoder_padding_mask, self_relative_embedding, cross_relative_embedding, past_key_value=past_key_value)
83
+
84
+ hidden_states.append(hidden_state)
85
+ self_attention_probs.append(self_attention_p)
86
+ cross_attention_probs.append(cross_attention_p)
87
+ key_value_states.append(key_value_state)
88
+
89
+ return hidden_states, self_attention_probs, cross_attention_probs, key_value_states
90
+
91
+
92
+ class MaskClassifier(nn.Module):
93
+ def __init__(self, config):
94
+ super().__init__()
95
+ self.nonlinearity = nn.Sequential(
96
+ nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=False),
97
+ nn.Dropout(config.hidden_dropout_prob),
98
+ nn.Linear(config.hidden_size, config.vocab_size)
99
+ )
100
+ self.initialize(config.hidden_size)
101
+
102
+ def initialize(self, hidden_size):
103
+ std = math.sqrt(2.0 / (5.0 * hidden_size))
104
+ nn.init.trunc_normal_(self.nonlinearity[-1].weight, mean=0.0, std=std, a=-2*std, b=2*std)
105
+ self.nonlinearity[-1].bias.data.zero_()
106
+
107
+ def forward(self, x):
108
+ x = self.nonlinearity(x)
109
+ return x
110
+
111
+
112
+ class EncoderLayer(nn.Module):
113
+ def __init__(self, config):
114
+ super().__init__()
115
+ self.attention = Attention(config, is_cross_attention=False)
116
+ self.mlp = FeedForward(config)
117
+
118
+ def forward(self, x, padding_mask, relative_embedding):
119
+ attention_output, attention_probs, _ = self.attention(x, x, padding_mask, relative_embedding)
120
+ x = x + attention_output
121
+ x = x + self.mlp(x)
122
+ return x, attention_probs
123
+
124
+
125
+ class DecoderLayer(nn.Module):
126
+ def __init__(self, config):
127
+ super().__init__()
128
+ self.self_attention = Attention(config, is_cross_attention=False)
129
+ self.cross_attention = Attention(config, is_cross_attention=True)
130
+ self.mlp = FeedForward(config)
131
+
132
+ def forward(self, x, autoreg_mask, encoder_output, encoder_padding_mask, self_relative_embedding, cross_relative_embedding, past_key_value=None):
133
+ query_offset = 0
134
+ if past_key_value is not None:
135
+ self_attn_past_key_value = past_key_value[:2]
136
+ cross_attn_past_key_value = past_key_value[2:]
137
+ query_offset = self_attn_past_key_value[0].size(2)
138
+ else:
139
+ self_attn_past_key_value, cross_attn_past_key_value = None, None
140
+
141
+ x_, self_attention_probs, self_key_value_state = self.self_attention(x, x, autoreg_mask, self_relative_embedding, past_key_value=self_attn_past_key_value, query_offset=query_offset)
142
+ x = x + x_
143
+ x_, cross_attention_probs, cross_key_value_state = self.cross_attention(x, encoder_output, encoder_padding_mask, cross_relative_embedding, past_key_value=cross_attn_past_key_value, query_offset=query_offset)
144
+ x = x + x_
145
+ x = x + self.mlp(x)
146
+
147
+ return x, self_attention_probs, cross_attention_probs, self_key_value_state + cross_key_value_state
148
+
149
+
150
+ class GeGLU(nn.Module):
151
+ def forward(self, x):
152
+ x, gate = x.chunk(2, dim=-1)
153
+ x = x * gelu_new(gate)
154
+ return x
155
+
156
+
157
+ class FeedForward(nn.Module):
158
+ def __init__(self, config):
159
+ super().__init__()
160
+ self.mlp = nn.Sequential(
161
+ nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, elementwise_affine=False),
162
+ nn.Linear(config.hidden_size, 2*config.intermediate_size, bias=False),
163
+ GeGLU(),
164
+ nn.LayerNorm(config.intermediate_size, eps=config.layer_norm_eps, elementwise_affine=False),
165
+ nn.Linear(config.intermediate_size, config.hidden_size, bias=False),
166
+ nn.Dropout(config.hidden_dropout_prob)
167
+ )
168
+ self.initialize(config.hidden_size)
169
+
170
+ def initialize(self, hidden_size):
171
+ std = math.sqrt(2.0 / (5.0 * hidden_size))
172
+ nn.init.trunc_normal_(self.mlp[1].weight, mean=0.0, std=std, a=-2*std, b=2*std)
173
+ nn.init.trunc_normal_(self.mlp[-2].weight, mean=0.0, std=std, a=-2*std, b=2*std)
174
+
175
+ def forward(self, x):
176
+ return self.mlp(x)
177
+
178
+
179
+ class MaskedSoftmax(torch.autograd.Function):
180
+ @staticmethod
181
+ def forward(self, x, mask, dim):
182
+ self.dim = dim
183
+ if mask is not None:
184
+ x.masked_fill_(mask, float('-inf'))
185
+ x = torch.softmax(x, self.dim)
186
+ if mask is not None:
187
+ x.masked_fill_(mask, 0.0)
188
+ self.save_for_backward(x)
189
+ return x
190
+
191
+ @staticmethod
192
+ def backward(self, grad_output):
193
+ output, = self.saved_tensors
194
+ input_grad = softmax_backward_data(self, grad_output, output, self.dim, output)
195
+ return input_grad, None, None
196
+
197
+
198
+ class Attention(nn.Module):
199
+ def __init__(self, config, is_cross_attention=False):
200
+ super().__init__()
201
+
202
+ self.config = config
203
+ self.is_cross_attention = is_cross_attention
204
+
205
+ if config.hidden_size % config.num_attention_heads != 0:
206
+ raise ValueError(f"The hidden size {config.hidden_size} is not a multiple of the number of attention heads {config.num_attention_heads}")
207
+
208
+ self.hidden_size = config.hidden_size
209
+ self.num_heads = config.num_attention_heads
210
+ self.head_size = config.hidden_size // config.num_attention_heads
211
+
212
+ self.in_proj_q = nn.Linear(config.hidden_size, config.hidden_size, bias=True)
213
+ self.in_proj_k = nn.Linear(config.hidden_size, config.hidden_size, bias=True)
214
+ self.in_proj_v = nn.Linear(config.hidden_size, config.hidden_size, bias=True)
215
+ self.out_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=True)
216
+
217
+ self.pre_layer_norm = nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=False)
218
+ self.post_layer_norm = nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=True)
219
+
220
+ position_indices = torch.arange(512, dtype=torch.long).unsqueeze(1) \
221
+ - torch.arange(512, dtype=torch.long).unsqueeze(0)
222
+ position_indices = self.make_log_bucket_position(position_indices, config.position_bucket_size, 512)
223
+ position_indices = config.position_bucket_size - 1 + position_indices
224
+ self.register_buffer("position_indices", position_indices, persistent=True)
225
+
226
+ self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
227
+ self.scale = 1.0 / math.sqrt(3 * self.head_size)
228
+ self.initialize()
229
+
230
+ def make_log_bucket_position(self, relative_pos, bucket_size, max_position):
231
+ sign = torch.sign(relative_pos)
232
+ mid = bucket_size // 2
233
+ abs_pos = torch.where((relative_pos < mid) & (relative_pos > -mid), mid - 1, torch.abs(relative_pos).clamp(max=max_position - 1))
234
+ log_pos = torch.ceil(torch.log(abs_pos / mid) / math.log((max_position-1) / mid) * (mid - 1)).int() + mid
235
+ bucket_pos = torch.where(abs_pos <= mid, relative_pos, log_pos * sign).long()
236
+ return bucket_pos
237
+
238
+ def initialize(self):
239
+ std = math.sqrt(2.0 / (5.0 * self.hidden_size))
240
+ nn.init.trunc_normal_(self.in_proj_q.weight, mean=0.0, std=std, a=-2*std, b=2*std)
241
+ nn.init.trunc_normal_(self.in_proj_k.weight, mean=0.0, std=std, a=-2*std, b=2*std)
242
+ nn.init.trunc_normal_(self.in_proj_v.weight, mean=0.0, std=std, a=-2*std, b=2*std)
243
+ nn.init.trunc_normal_(self.out_proj.weight, mean=0.0, std=std, a=-2*std, b=2*std)
244
+ self.in_proj_q.bias.data.zero_()
245
+ self.in_proj_k.bias.data.zero_()
246
+ self.in_proj_v.bias.data.zero_()
247
+ self.out_proj.bias.data.zero_()
248
+
249
+ def forward(self, q, kv, attention_mask, relative_embedding, past_key_value=None, query_offset=0):
250
+ key_len, batch_size, _ = kv.size()
251
+ query_len, _, _ = q.size()
252
+
253
+ if not self.is_cross_attention or past_key_value is None or past_key_value[0].size(1) != kv.size(0):
254
+ kv = self.pre_layer_norm(kv)
255
+ key = self.in_proj_k(kv) # shape: [T, B, D]
256
+ value = self.in_proj_v(kv) # shape: [T, B, D]
257
+ key = key.reshape(key_len, batch_size * self.num_heads, self.head_size).transpose(0, 1) # shape: [BxH, T, D]
258
+ value = value.view(key_len, batch_size * self.num_heads, self.head_size).transpose(0, 1) # shape: [BxH, T, D]
259
+
260
+ if past_key_value is not None:
261
+ if not self.is_cross_attention:
262
+ key = torch.cat([past_key_value[0].flatten(0, 1), key], dim=1)
263
+ value = torch.cat([past_key_value[1].flatten(0, 1), value], dim=1)
264
+ key_len = key.size(1)
265
+ elif past_key_value[0].size(1) == kv.size(0):
266
+ key = past_key_value[0].flatten(0, 1)
267
+ value = past_key_value[1].flatten(0, 1)
268
+
269
+ if self.position_indices.size(0) < max(query_len, key_len):
270
+ position_indices = torch.arange(max(query_len, key_len), dtype=torch.long).unsqueeze(1) \
271
+ - torch.arange(max(query_len, key_len), dtype=torch.long).unsqueeze(0)
272
+ position_indices = self.make_log_bucket_position(position_indices, self.config.position_bucket_size, 512)
273
+ position_indices = self.config.position_bucket_size - 1 + position_indices
274
+ self.register_buffer("position_indices", position_indices.to(q.device), persistent=True)
275
+
276
+ q = self.pre_layer_norm(q)
277
+ query = self.in_proj_q(q) # shape: [T, B, D]
278
+ query = query.reshape(query_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)
279
+
280
+ attention_scores = torch.bmm(query, key.transpose(1, 2) * self.scale)
281
+
282
+ query_pos = self.in_proj_q(self.dropout(relative_embedding)) # shape: [2T-1, D]
283
+ query_pos = query_pos.view(-1, self.num_heads, self.head_size) # shape: [2T-1, H, D]
284
+ key_pos = self.in_proj_k(self.dropout(relative_embedding)) # shape: [2T-1, D]
285
+ key_pos = key_pos.view(-1, self.num_heads, self.head_size) # shape: [2T-1, H, D]
286
+
287
+ query_ = query.view(batch_size, self.num_heads, query_len, self.head_size)
288
+ key_ = key.view(batch_size, self.num_heads, key_len, self.head_size)
289
+
290
+ attention_c_p = torch.einsum("bhqd,khd->bhqk", query_, key_pos.squeeze(1) * self.scale)
291
+ attention_p_c = torch.einsum("bhkd,qhd->bhqk", key_ * self.scale, query_pos.squeeze(1))
292
+ position_indices = self.position_indices[query_offset:query_offset+query_len, :key_len].expand(batch_size, self.num_heads, -1, -1)
293
+ attention_c_p = attention_c_p.gather(3, position_indices)
294
+ attention_p_c = attention_p_c.gather(2, position_indices)
295
+
296
+ attention_scores = attention_scores.view(batch_size, self.num_heads, query_len, key_len)
297
+ attention_scores.add_(attention_c_p)
298
+ attention_scores.add_(attention_p_c)
299
+
300
+ attention_probs = MaskedSoftmax.apply(attention_scores, attention_mask, -1)
301
+
302
+ attention_probs = self.dropout(attention_probs)
303
+ context = torch.bmm(attention_probs.flatten(0, 1), value) # shape: [B*H, Q, D]
304
+ context = context.transpose(0, 1).reshape(context.size(1), -1, self.hidden_size) # shape: [Q, B, H*D]
305
+ context = self.out_proj(context)
306
+ context = self.post_layer_norm(context)
307
+ context = self.dropout(context)
308
+
309
+ key = key.detach().unflatten(0, (-1, self.num_heads))
310
+ value = value.detach().unflatten(0, (-1, self.num_heads))
311
+
312
+ return context, attention_probs.detach(), (key, value)
313
+
314
+
315
+ class WordEmbedding(nn.Module):
316
+ def __init__(self, config):
317
+ super().__init__()
318
+ self.hidden_size = config.hidden_size
319
+
320
+ self.word_embedding = nn.Embedding(config.vocab_size, config.hidden_size)
321
+ self.word_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, elementwise_affine=False)
322
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
323
+
324
+ self.initialize()
325
+
326
+ def initialize(self):
327
+ std = math.sqrt(2.0 / (5.0 * self.hidden_size))
328
+ nn.init.trunc_normal_(self.word_embedding.weight, mean=0.0, std=std, a=-2*std, b=2*std)
329
+
330
+ def forward(self, input_ids):
331
+ return self.dropout(self.word_layer_norm(self.word_embedding(input_ids)))
332
+
333
+
334
+ class RelativeEmbedding(nn.Module):
335
+ def __init__(self, config):
336
+ super().__init__()
337
+ self.relative_embedding = nn.Parameter(torch.empty(2 * config.position_bucket_size - 1, config.hidden_size))
338
+ self.relative_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
339
+
340
+ self.initialize(config.hidden_size)
341
+
342
+ def initialize(self, hidden_size):
343
+ std = math.sqrt(2.0 / (5.0 * hidden_size))
344
+ nn.init.trunc_normal_(self.relative_embedding, mean=0.0, std=std, a=-2*std, b=2*std)
345
+
346
+ def forward(self):
347
+ return self.relative_layer_norm(self.relative_embedding)
348
+
349
+
350
+ #
351
+ # HuggingFace wrappers
352
+ #
353
+
354
+ class NorT5PreTrainedModel(PreTrainedModel):
355
+ config_class = NorT5Config
356
+ base_model_prefix = "norT5"
357
+ supports_gradient_checkpointing = True
358
+
359
+ def _set_gradient_checkpointing(self, module, value=False):
360
+ if isinstance(module, Encoder):
361
+ module.activation_checkpointing = value
362
+
363
+ def _init_weights(self, module):
364
+ pass # everything is already initialized
365
+
366
+
367
+ class NorT5Model(NorT5PreTrainedModel):
368
+ def __init__(self, config, add_lm_layer=False, add_decoder=True):
369
+ super().__init__(config)
370
+ self.config = config
371
+
372
+ self.cls_token_id = config.cls_token_id
373
+ self.sep_token_id = config.sep_token_id
374
+ self.bos_token_id = config.bos_token_id
375
+ self.eos_token_id = config.eos_token_id
376
+ self.pad_token_id = config.pad_token_id
377
+
378
+ self.embedding = WordEmbedding(config)
379
+ self.encoder = Encoder(config, activation_checkpointing=False)
380
+ self.decoder = Decoder(config, activation_checkpointing=False) if add_decoder else None
381
+ self.classifier = MaskClassifier(config) if add_lm_layer else None
382
+
383
+ def get_input_embeddings(self):
384
+ return self.embedding.word_embedding
385
+
386
+ def set_input_embeddings(self, value):
387
+ self.embedding.word_embedding = value
388
+
389
+ def get_encoder(self):
390
+ class EncoderWrapper:
391
+ def __call__(cls, *args, **kwargs):
392
+ return cls.forward(*args, **kwargs)
393
+
394
+ def forward(
395
+ cls,
396
+ input_ids: Optional[torch.Tensor] = None,
397
+ attention_mask: Optional[torch.Tensor] = None,
398
+ output_hidden_states: Optional[bool] = None,
399
+ output_attentions: Optional[bool] = None,
400
+ return_dict: Optional[bool] = None,
401
+ ):
402
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
403
+
404
+ return self.get_encoder_output(
405
+ input_ids, attention_mask, output_hidden_states, output_attentions, return_dict=return_dict
406
+ )
407
+ return EncoderWrapper()
408
+
409
+ def get_decoder(self):
410
+ return self.get_decoder_output
411
+
412
+ def set_decoder_special_tokens(self, target_id):
413
+ target_id.masked_fill_(target_id == self.cls_token_id, self.bos_token_id)
414
+ target_id.masked_fill_(target_id == self.sep_token_id, self.eos_token_id)
415
+ return target_id
416
+
417
+ def _shift_right(self, input_ids):
418
+ shifted_input_ids = input_ids.new_zeros(input_ids.shape)
419
+ shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
420
+ shifted_input_ids[..., 0] = self.bos_token_id
421
+ shifted_input_ids.masked_fill_(shifted_input_ids == -100, self.pad_token_id)
422
+
423
+ return shifted_input_ids
424
+
425
+ def get_encoder_output(
426
+ self,
427
+ input_ids: torch.Tensor = None,
428
+ attention_mask: Optional[torch.Tensor] = None,
429
+ output_hidden_states: Optional[bool] = None,
430
+ output_attentions: Optional[bool] = None,
431
+ return_dict = False
432
+ ):
433
+ if input_ids is not None:
434
+ input_shape = input_ids.size()
435
+ else:
436
+ raise ValueError("You have to specify input_ids")
437
+
438
+ batch_size, seq_length = input_shape
439
+ device = input_ids.device
440
+
441
+ if attention_mask is None:
442
+ attention_mask = torch.zeros(batch_size, seq_length, dtype=torch.bool, device=device)
443
+ else:
444
+ attention_mask = ~attention_mask.bool()
445
+ attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
446
+
447
+ static_embeddings = self.embedding(input_ids.t())
448
+ contextualized_embeddings, attention_probs = self.encoder(static_embeddings, attention_mask)
449
+ contextualized_embeddings = [e.transpose(0, 1) for e in contextualized_embeddings]
450
+ last_layer = contextualized_embeddings[-1]
451
+ contextualized_embeddings = [contextualized_embeddings[0]] + [
452
+ contextualized_embeddings[i] - contextualized_embeddings[i - 1]
453
+ for i in range(1, len(contextualized_embeddings))
454
+ ]
455
+
456
+ if not return_dict:
457
+ return (
458
+ last_layer,
459
+ *([contextualized_embeddings] if output_hidden_states else []),
460
+ *([attention_probs] if output_attentions else [])
461
+ )
462
+
463
+ return BaseModelOutput(
464
+ last_hidden_state=last_layer,
465
+ hidden_states=contextualized_embeddings if output_hidden_states else None,
466
+ attentions=attention_probs if output_attentions else None
467
+ )
468
+
469
+ def get_decoder_output(
470
+ self,
471
+ target_ids: torch.Tensor = None,
472
+ encoder_output: torch.Tensor = None,
473
+ attention_mask: Optional[torch.Tensor] = None,
474
+ past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
475
+ use_cache: Optional[bool] = None,
476
+ output_hidden_states: Optional[bool] = None,
477
+ output_attentions: Optional[bool] = None,
478
+ return_dict = False
479
+ ):
480
+ batch_size, seq_length, _ = encoder_output.shape
481
+ device = target_ids.device
482
+
483
+ if attention_mask is None:
484
+ attention_mask = torch.zeros(batch_size, seq_length, dtype=torch.bool, device=device)
485
+ else:
486
+ attention_mask = ~attention_mask.bool()
487
+ attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
488
+
489
+ hidden_states, self_attention_p, cross_attention_p, key_value_states = self.decoder(
490
+ self.embedding(target_ids.t()),
491
+ encoder_output.transpose(0, 1),
492
+ attention_mask,
493
+ past_key_values
494
+ )
495
+
496
+ hidden_states = [e.transpose(0, 1) for e in hidden_states]
497
+ last_layer = hidden_states[-1]
498
+ hidden_states = [hidden_states[0]] + [
499
+ hidden_states[i] - hidden_states[i - 1]
500
+ for i in range(1, len(hidden_states))
501
+ ]
502
+
503
+ if not return_dict:
504
+ return (
505
+ last_layer,
506
+ *([key_value_states] if use_cache else []),
507
+ *([hidden_states] if output_hidden_states else []),
508
+ *([self_attention_p] if output_attentions else []),
509
+ *([cross_attention_p] if output_attentions else []),
510
+ )
511
+
512
+ return BaseModelOutputWithPastAndCrossAttentions(
513
+ last_hidden_state=last_layer,
514
+ past_key_values=key_value_states if use_cache else None,
515
+ hidden_states=hidden_states if output_hidden_states else None,
516
+ attentions=self_attention_p if output_attentions else None,
517
+ cross_attentions=cross_attention_p if output_attentions else None
518
+ )
519
+
520
+
521
+ def forward(
522
+ self,
523
+ input_ids: Optional[torch.LongTensor] = None,
524
+ attention_mask: Optional[torch.FloatTensor] = None,
525
+ decoder_input_ids: Optional[torch.LongTensor] = None,
526
+ decoder_attention_mask: Optional[torch.BoolTensor] = None,
527
+ encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
528
+ past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
529
+ use_cache: Optional[bool] = None,
530
+ output_attentions: Optional[bool] = None,
531
+ output_hidden_states: Optional[bool] = None,
532
+ return_dict: Optional[bool] = None
533
+ ):
534
+
535
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
536
+
537
+ decoder_input_ids = self.set_decoder_special_tokens(decoder_input_ids)
538
+
539
+ if encoder_outputs is None:
540
+ encoder_outputs = self.get_encoder_output(
541
+ input_ids, attention_mask, output_hidden_states, output_attentions, return_dict
542
+ )
543
+ elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
544
+ encoder_outputs = BaseModelOutput(
545
+ last_hidden_state=encoder_outputs[0],
546
+ hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
547
+ attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
548
+ )
549
+
550
+ decoder_outputs = self.get_decoder_output(
551
+ decoder_input_ids, encoder_outputs[0], attention_mask, past_key_values, use_cache, output_hidden_states, output_attentions, return_dict
552
+ )
553
+
554
+ if not return_dict:
555
+ return decoder_outputs + encoder_outputs
556
+
557
+ return Seq2SeqModelOutput(
558
+ last_hidden_state=decoder_outputs.last_hidden_state,
559
+ past_key_values=decoder_outputs.past_key_values,
560
+ decoder_hidden_states=decoder_outputs.hidden_states,
561
+ decoder_attentions=decoder_outputs.attentions,
562
+ cross_attentions=decoder_outputs.cross_attentions,
563
+ encoder_last_hidden_state=encoder_outputs.last_hidden_state,
564
+ encoder_hidden_states=encoder_outputs.hidden_states,
565
+ encoder_attentions=encoder_outputs.attentions,
566
+ )
567
+
568
+
569
+ class NorT5ForConditionalGeneration(NorT5Model):
570
+
571
+ def __init__(self, config):
572
+ super().__init__(config, add_lm_layer=True)
573
+
574
+ def forward(
575
+ self,
576
+ input_ids: Optional[torch.LongTensor] = None,
577
+ attention_mask: Optional[torch.FloatTensor] = None,
578
+ decoder_input_ids: Optional[torch.LongTensor] = None,
579
+ decoder_attention_mask: Optional[torch.BoolTensor] = None,
580
+ head_mask: Optional[torch.FloatTensor] = None,
581
+ decoder_head_mask: Optional[torch.FloatTensor] = None,
582
+ cross_attn_head_mask: Optional[torch.Tensor] = None,
583
+ encoder_outputs: Optional[Tuple[Tuple[torch.Tensor]]] = None,
584
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
585
+ inputs_embeds: Optional[torch.FloatTensor] = None,
586
+ decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
587
+ labels: Optional[torch.LongTensor] = None,
588
+ use_cache: Optional[bool] = None,
589
+ output_attentions: Optional[bool] = None,
590
+ output_hidden_states: Optional[bool] = None,
591
+ return_dict: Optional[bool] = None,
592
+ ):
593
+ use_cache = use_cache if use_cache is not None else getattr(self.config, "use_cache", False)
594
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
595
+
596
+ if encoder_outputs is None:
597
+ encoder_outputs = self.get_encoder_output(
598
+ input_ids, attention_mask, output_hidden_states, output_attentions, return_dict
599
+ )
600
+ elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
601
+ encoder_outputs = BaseModelOutput(
602
+ last_hidden_state=encoder_outputs[0],
603
+ hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
604
+ attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
605
+ )
606
+
607
+ if labels is not None:
608
+ labels = self.set_decoder_special_tokens(labels)
609
+
610
+ if labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None:
611
+ decoder_input_ids = self._shift_right(labels)
612
+ elif decoder_input_ids is not None:
613
+ decoder_input_ids = self.set_decoder_special_tokens(decoder_input_ids)
614
+
615
+ decoder_outputs = self.get_decoder_output(
616
+ decoder_input_ids, encoder_outputs[0], attention_mask, past_key_values, use_cache, output_hidden_states, output_attentions, return_dict
617
+ )
618
+ lm_logits = self.classifier(decoder_outputs[0])
619
+
620
+ loss = None
621
+ if labels is not None:
622
+ labels.masked_fill_(labels == self.pad_token_id, -100)
623
+ loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
624
+ loss = loss_fct(lm_logits.flatten(0, 1), labels.flatten())
625
+
626
+ if not return_dict:
627
+ output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs
628
+ return ((loss,) + output) if loss is not None else output
629
+
630
+ return Seq2SeqLMOutput(
631
+ loss=loss,
632
+ logits=lm_logits,
633
+ past_key_values=decoder_outputs.past_key_values,
634
+ decoder_hidden_states=decoder_outputs.hidden_states,
635
+ decoder_attentions=decoder_outputs.attentions,
636
+ cross_attentions=decoder_outputs.cross_attentions,
637
+ encoder_last_hidden_state=encoder_outputs.last_hidden_state,
638
+ encoder_hidden_states=encoder_outputs.hidden_states,
639
+ encoder_attentions=encoder_outputs.attentions,
640
+ )
641
+
642
+ def prepare_inputs_for_generation(
643
+ self,
644
+ input_ids,
645
+ past_key_values=None,
646
+ attention_mask=None,
647
+ head_mask=None,
648
+ decoder_head_mask=None,
649
+ cross_attn_head_mask=None,
650
+ use_cache=None,
651
+ encoder_outputs=None,
652
+ **kwargs,
653
+ ):
654
+ if past_key_values is not None:
655
+ input_ids = input_ids[:, -1:]
656
+
657
+ return {
658
+ "decoder_input_ids": input_ids,
659
+ "past_key_values": past_key_values,
660
+ "encoder_outputs": encoder_outputs,
661
+ "attention_mask": attention_mask,
662
+ "head_mask": head_mask,
663
+ "decoder_head_mask": decoder_head_mask,
664
+ "cross_attn_head_mask": cross_attn_head_mask,
665
+ "use_cache": use_cache,
666
+ }
667
+
668
+ def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
669
+ return self._shift_right(labels)
670
+
671
+ def _reorder_cache(self, past_key_values, beam_idx):
672
+ # if decoder past is not included in output
673
+ # speedy decoding is disabled and no need to reorder
674
+ if past_key_values is None:
675
+ print("You might want to consider setting `use_cache=True` to speed up decoding")
676
+ return past_key_values
677
+
678
+ reordered_decoder_past = ()
679
+ for layer_past_states in past_key_values:
680
+ # get the correct batch idx from layer past batch dim
681
+ # batch dim of `past` is at 2nd position
682
+ reordered_layer_past_states = ()
683
+ for layer_past_state in layer_past_states:
684
+ # need to set correct `past` for each of the four key / value states
685
+ layer_past_state = layer_past_state.index_select(0, beam_idx.to(layer_past_state.device))
686
+ reordered_layer_past_states = reordered_layer_past_states + (layer_past_state,)
687
+
688
+ assert reordered_layer_past_states[0].shape == layer_past_states[0].shape
689
+ assert len(reordered_layer_past_states) == len(layer_past_states)
690
+
691
+ reordered_decoder_past = reordered_decoder_past + (reordered_layer_past_states,)
692
+ return reordered_decoder_past
693
+
694
+
695
+ class NorT5Encoder(NorT5Model):
696
+ def __init__(self, config):
697
+ super().__init__(config, add_lm_layer=False, add_decoder=True)
698
+
699
+ def forward(
700
+ self,
701
+ input_ids: Optional[torch.Tensor] = None,
702
+ attention_mask: Optional[torch.Tensor] = None,
703
+ output_hidden_states: Optional[bool] = None,
704
+ output_attentions: Optional[bool] = None,
705
+ return_dict: Optional[bool] = None,
706
+ ):
707
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
708
+
709
+ return self.get_encoder_output(
710
+ input_ids, attention_mask, output_hidden_states, output_attentions, return_dict=return_dict
711
+ )
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0fb08bd3a0fea16f20137417c1ea2b991be43aa2ac0f1c4a3d84cdf5eb708a0d
3
+ size 1125877210
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"bos_token": "[BOS]", "eos_token": "[EOS]", "unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]"}
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ {
2
+ "tokenizer_class": "PreTrainedTokenizerFast"
3
+ }