SaiedAlshahrani commited on
Commit
319e604
1 Parent(s): 114d90f

Copy modeling_jais.py from source

Browse files
Files changed (1) hide show
  1. modeling_jais.py +1522 -0
modeling_jais.py ADDED
@@ -0,0 +1,1522 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2023 The OpenAI Team Authors and HuggingFace Inc. team.
3
+ # Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
4
+ # Copyright 2023 G42 Systems.
5
+ #
6
+ # Licensed under the Apache License, Version 2.0 (the "License");
7
+ # you may not use this file except in compliance with the License.
8
+ # You may obtain a copy of the License at
9
+ #
10
+ # http://www.apache.org/licenses/LICENSE-2.0
11
+ #
12
+ # Unless required by applicable law or agreed to in writing, software
13
+ # distributed under the License is distributed on an "AS IS" BASIS,
14
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15
+ # See the License for the specific language governing permissions and
16
+ # limitations under the License.
17
+ """ PyTorch JAIS model."""
18
+
19
+ import math
20
+ import os
21
+ import warnings
22
+ from typing import Optional, Tuple, Union
23
+
24
+ import torch
25
+ from torch import Tensor, nn
26
+ from torch.cuda.amp import autocast
27
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
28
+
29
+ from transformers.activations import ACT2FN
30
+ from transformers.modeling_outputs import (
31
+ BaseModelOutputWithPastAndCrossAttentions,
32
+ CausalLMOutputWithCrossAttentions,
33
+ QuestionAnsweringModelOutput,
34
+ SequenceClassifierOutputWithPast,
35
+ TokenClassifierOutput,
36
+ )
37
+ from transformers.modeling_utils import PreTrainedModel
38
+ from transformers.pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
39
+ from transformers.utils import (
40
+ add_code_sample_docstrings,
41
+ add_start_docstrings,
42
+ add_start_docstrings_to_model_forward,
43
+ logging,
44
+ )
45
+ from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
46
+ from .configuration_jais import JAISConfig
47
+
48
+
49
+ logger = logging.get_logger(__name__)
50
+
51
+ _CHECKPOINT_FOR_DOC = "IIAI/checkpoint"
52
+ _CONFIG_FOR_DOC = "JAISConfig"
53
+
54
+
55
+ class SwiGLUActivation(nn.Module):
56
+ def forward(self, x1: Tensor, x2: Tensor) -> Tensor:
57
+ return x1 * nn.functional.silu(x2)
58
+
59
+
60
+ class AlibiPositionEmbeddingLayer(nn.Module):
61
+ def __init__(self, num_heads):
62
+ super(AlibiPositionEmbeddingLayer, self).__init__()
63
+
64
+ self.num_heads = num_heads
65
+ slopes = torch.tensor(
66
+ AlibiPositionEmbeddingLayer._get_alibi_slopes(num_heads)
67
+ ).unsqueeze(-1)
68
+ self.slopes = nn.parameter.Parameter(slopes, requires_grad=False)
69
+
70
+ def forward(self, seq_length, key_length, cached_qk_len):
71
+ context_position = torch.arange(
72
+ cached_qk_len, cached_qk_len + seq_length, device=self.slopes.device
73
+ )[:, None]
74
+ memory_position = torch.arange(
75
+ key_length + cached_qk_len, device=self.slopes.device
76
+ )[None, :]
77
+ relative_position = memory_position - context_position
78
+ relative_position = torch.abs(relative_position).unsqueeze(0).expand(self.num_heads, -1, -1)
79
+ alibi = (self.slopes * -1.0).unsqueeze(1) * relative_position
80
+ return alibi
81
+
82
+ @staticmethod
83
+ def _get_alibi_slopes(n):
84
+ def get_slopes_power_of_2(n):
85
+ start = 2 ** (-(2 ** -(math.log2(n) - 3)))
86
+ ratio = start
87
+ return [start * ratio ** i for i in range(n)]
88
+
89
+ if math.log2(n).is_integer():
90
+ return get_slopes_power_of_2(
91
+ n
92
+ ) # In the paper, we only train models that have 2^a heads for some a. This function has
93
+ else: # some good properties that only occur when the input is a power of 2. To maintain that even
94
+ closest_power_of_2 = 2 ** math.floor(
95
+ math.log2(n)
96
+ ) # when the number of heads is not a power of 2, we use this workaround.
97
+ return (
98
+ get_slopes_power_of_2(closest_power_of_2)
99
+ + AlibiPositionEmbeddingLayer._get_alibi_slopes(
100
+ 2 * closest_power_of_2
101
+ )[0::2][: n - closest_power_of_2]
102
+ )
103
+
104
+
105
+ def load_tf_weights_in_jais(model, config, jais_checkpoint_path):
106
+ """Load tf checkpoints in a pytorch model"""
107
+ try:
108
+ import re
109
+
110
+ import tensorflow as tf
111
+ except ImportError:
112
+ logger.error(
113
+ "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
114
+ "https://www.tensorflow.org/install/ for installation instructions."
115
+ )
116
+ raise
117
+ tf_path = os.path.abspath(jais_checkpoint_path)
118
+ logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
119
+ # Load weights from TF model
120
+ init_vars = tf.train.list_variables(tf_path)
121
+ names = []
122
+ arrays = []
123
+ for name, shape in init_vars:
124
+ logger.info(f"Loading TF weight {name} with shape {shape}")
125
+ array = tf.train.load_variable(tf_path, name)
126
+ names.append(name)
127
+ arrays.append(array.squeeze())
128
+
129
+ for name, array in zip(names, arrays):
130
+ name = name[6:] # skip "model/"
131
+ name = name.split("/")
132
+ pointer = model
133
+ for m_name in name:
134
+ if re.fullmatch(r"[A-Za-z]+\d+", m_name):
135
+ scope_names = re.split(r"(\d+)", m_name)
136
+ else:
137
+ scope_names = [m_name]
138
+ if scope_names[0] == "w" or scope_names[0] == "g":
139
+ pointer = getattr(pointer, "weight")
140
+ elif scope_names[0] == "b":
141
+ pointer = getattr(pointer, "bias")
142
+ elif scope_names[0] == "wpe" or scope_names[0] == "wte":
143
+ pointer = getattr(pointer, scope_names[0])
144
+ pointer = getattr(pointer, "weight")
145
+ else:
146
+ pointer = getattr(pointer, scope_names[0])
147
+ if len(scope_names) >= 2:
148
+ num = int(scope_names[1])
149
+ pointer = pointer[num]
150
+ try:
151
+ assert (
152
+ pointer.shape == array.shape
153
+ ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
154
+ except AssertionError as e:
155
+ e.args += (pointer.shape, array.shape)
156
+ raise
157
+ logger.info(f"Initialize PyTorch weight {name}")
158
+ pointer.data = torch.from_numpy(array)
159
+ return model
160
+
161
+
162
+ class JAISAttention(nn.Module):
163
+ def __init__(self, config, is_cross_attention=False, layer_idx=None):
164
+ super().__init__()
165
+
166
+ max_positions = config.max_position_embeddings
167
+ self.register_buffer(
168
+ "bias",
169
+ torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view(
170
+ 1, 1, max_positions, max_positions
171
+ ),
172
+ persistent=False,
173
+ )
174
+ self.register_buffer("masked_bias", torch.tensor(-1e4), persistent=False)
175
+
176
+ self.embed_dim = config.hidden_size
177
+ self.num_heads = config.num_attention_heads
178
+ self.head_dim = self.embed_dim // self.num_heads
179
+ self.split_size = self.embed_dim
180
+ if self.head_dim * self.num_heads != self.embed_dim:
181
+ raise ValueError(
182
+ f"`embed_dim` must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
183
+ f" {self.num_heads})."
184
+ )
185
+
186
+ self.scale_attn_weights = config.scale_attn_weights
187
+ self.is_cross_attention = is_cross_attention
188
+
189
+ # Layer-wise attention scaling, reordering, and upcasting
190
+ self.scale_attn_by_inverse_layer_idx = config.scale_attn_by_inverse_layer_idx
191
+ self.layer_idx = layer_idx
192
+ self.reorder_and_upcast_attn = config.reorder_and_upcast_attn
193
+
194
+ if self.is_cross_attention:
195
+ self.c_attn = Conv1D(2 * self.embed_dim, self.embed_dim)
196
+ self.q_attn = Conv1D(self.embed_dim, self.embed_dim)
197
+ else:
198
+ self.c_attn = Conv1D(3 * self.embed_dim, self.embed_dim)
199
+ self.c_proj = Conv1D(self.embed_dim, self.embed_dim)
200
+
201
+ self.attn_dropout = nn.Dropout(config.attn_pdrop)
202
+ self.resid_dropout = nn.Dropout(config.resid_pdrop)
203
+
204
+ self.pruned_heads = set()
205
+
206
+ self.attn_scale_power = 1.0 if config.scale_qk_dot_by_d else 0.5
207
+
208
+ def prune_heads(self, heads):
209
+ if len(heads) == 0:
210
+ return
211
+ heads, index = find_pruneable_heads_and_indices(heads, self.num_heads, self.head_dim, self.pruned_heads)
212
+ index_attn = torch.cat([index, index + self.split_size, index + (2 * self.split_size)])
213
+
214
+ # Prune conv1d layers
215
+ self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
216
+ self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
217
+
218
+ # Update hyper params
219
+ self.split_size = (self.split_size // self.num_heads) * (self.num_heads - len(heads))
220
+ self.num_heads = self.num_heads - len(heads)
221
+ self.pruned_heads = self.pruned_heads.union(heads)
222
+
223
+ def _attn(self, query, key, value, attention_mask=None, head_mask=None, position_bias=None):
224
+ attn_weights = torch.matmul(query, key.transpose(-1, -2))
225
+
226
+ if self.scale_attn_weights:
227
+ attn_weights = attn_weights / torch.full(
228
+ [], value.size(-1) ** self.attn_scale_power, dtype=attn_weights.dtype, device=attn_weights.device
229
+ )
230
+
231
+ # Layer-wise attention scaling
232
+ if self.scale_attn_by_inverse_layer_idx:
233
+ attn_weights = attn_weights / float(self.layer_idx + 1)
234
+
235
+ if not self.is_cross_attention:
236
+ # if only "normal" attention layer implements causal mask
237
+ query_length, key_length = query.size(-2), key.size(-2)
238
+ causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]
239
+ mask_value = torch.finfo(attn_weights.dtype).min
240
+ # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
241
+ # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
242
+ mask_value = torch.full([], mask_value, dtype=attn_weights.dtype).to(attn_weights.device)
243
+ attn_weights = torch.where(causal_mask, attn_weights.to(attn_weights.dtype), mask_value)
244
+
245
+ if attention_mask is not None:
246
+ # Apply the attention mask
247
+ attn_weights = attn_weights + attention_mask
248
+
249
+ if position_bias is not None:
250
+ attn_weights += position_bias.type_as(attn_weights).unsqueeze(0)
251
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1)
252
+
253
+ # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op otherwise
254
+ attn_weights = attn_weights.type(value.dtype)
255
+ attn_weights = self.attn_dropout(attn_weights)
256
+
257
+ # Mask heads if we want to
258
+ if head_mask is not None:
259
+ attn_weights = attn_weights * head_mask
260
+
261
+ attn_output = torch.matmul(attn_weights, value)
262
+
263
+ return attn_output, attn_weights
264
+
265
+ def _upcast_and_reordered_attn(self, query, key, value, attention_mask=None, head_mask=None, position_bias=None):
266
+ # Use `torch.baddbmm` (a bit more efficient w/ alpha param for scaling -- from Megatron-LM)
267
+ bsz, num_heads, q_seq_len, dk = query.size()
268
+ _, _, k_seq_len, _ = key.size()
269
+
270
+ # Preallocate attn_weights for `baddbmm`
271
+ attn_weights = torch.empty(bsz * num_heads, q_seq_len, k_seq_len, dtype=torch.float32, device=query.device)
272
+
273
+ # Compute Scale Factor
274
+ scale_factor = 1.0
275
+ if self.scale_attn_weights:
276
+ scale_factor /= float(value.size(-1)) ** self.attn_scale_power
277
+
278
+ if self.scale_attn_by_inverse_layer_idx:
279
+ scale_factor /= float(self.layer_idx + 1)
280
+
281
+ # Upcast (turn off autocast) and reorder (Scale K by 1 / root(dk))
282
+ with autocast(enabled=False):
283
+ q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(-1, dk, k_seq_len)
284
+ attn_weights = torch.baddbmm(attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor)
285
+ attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)
286
+
287
+ if not self.is_cross_attention:
288
+ # if only "normal" attention layer implements causal mask
289
+ query_length, key_length = query.size(-2), key.size(-2)
290
+ causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]
291
+ mask_value = torch.finfo(attn_weights.dtype).min
292
+ # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
293
+ # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
294
+ mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(attn_weights.device)
295
+ attn_weights = torch.where(causal_mask, attn_weights, mask_value)
296
+
297
+ if attention_mask is not None:
298
+ # Apply the attention mask
299
+ attn_weights = attn_weights + attention_mask
300
+
301
+ if position_bias is not None:
302
+ attn_weights += position_bias.type_as(attn_weights).unsqueeze(0)
303
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1)
304
+
305
+ # Downcast (if necessary) back to V's dtype (if in mixed-precision) -- No-Op if otherwise
306
+ if attn_weights.dtype != torch.float32:
307
+ raise RuntimeError("Error with upcasting, attn_weights does not have dtype torch.float32")
308
+ attn_weights = attn_weights.type(value.dtype)
309
+ attn_weights = self.attn_dropout(attn_weights)
310
+
311
+ # Mask heads if we want to
312
+ if head_mask is not None:
313
+ attn_weights = attn_weights * head_mask
314
+
315
+ attn_output = torch.matmul(attn_weights, value)
316
+
317
+ return attn_output, attn_weights
318
+
319
+ def _split_heads(self, tensor, num_heads, attn_head_size):
320
+ """
321
+ Splits hidden_size dim into attn_head_size and num_heads
322
+ """
323
+ new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
324
+ tensor = tensor.view(new_shape)
325
+ return tensor.permute(0, 2, 1, 3) # (batch, head, seq_length, head_features)
326
+
327
+ def _merge_heads(self, tensor, num_heads, attn_head_size):
328
+ """
329
+ Merges attn_head_size dim and num_attn_heads dim into hidden_size
330
+ """
331
+ tensor = tensor.permute(0, 2, 1, 3).contiguous()
332
+ new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
333
+ return tensor.view(new_shape)
334
+
335
+ def forward(
336
+ self,
337
+ hidden_states: Optional[Tuple[torch.FloatTensor]],
338
+ layer_past: Optional[Tuple[torch.Tensor]] = None,
339
+ attention_mask: Optional[torch.FloatTensor] = None,
340
+ head_mask: Optional[torch.FloatTensor] = None,
341
+ encoder_hidden_states: Optional[torch.Tensor] = None,
342
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
343
+ use_cache: Optional[bool] = False,
344
+ output_attentions: Optional[bool] = False,
345
+ position_bias: Optional[torch.FloatTensor] = None,
346
+ ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]], ...]:
347
+ if encoder_hidden_states is not None:
348
+ if not hasattr(self, "q_attn"):
349
+ raise ValueError(
350
+ "If class is used as cross attention, the weights `q_attn` have to be defined. "
351
+ "Please make sure to instantiate class with `JAISAttention(..., is_cross_attention=True)`."
352
+ )
353
+
354
+ query = self.q_attn(hidden_states)
355
+ key, value = self.c_attn(encoder_hidden_states).split(self.split_size, dim=2)
356
+ attention_mask = encoder_attention_mask
357
+ else:
358
+ query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
359
+
360
+ query = self._split_heads(query, self.num_heads, self.head_dim)
361
+ key = self._split_heads(key, self.num_heads, self.head_dim)
362
+ value = self._split_heads(value, self.num_heads, self.head_dim)
363
+
364
+ if layer_past is not None:
365
+ past_key, past_value = layer_past
366
+ key = torch.cat((past_key, key), dim=-2)
367
+ value = torch.cat((past_value, value), dim=-2)
368
+
369
+ if use_cache is True:
370
+ present = (key, value)
371
+ else:
372
+ present = None
373
+
374
+ if self.reorder_and_upcast_attn:
375
+ attn_output, attn_weights = self._upcast_and_reordered_attn(query, key, value, attention_mask, head_mask, position_bias)
376
+ else:
377
+ attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask, position_bias)
378
+
379
+ attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)
380
+ attn_output = self.c_proj(attn_output)
381
+ attn_output = self.resid_dropout(attn_output)
382
+
383
+ outputs = (attn_output, present)
384
+ if output_attentions:
385
+ outputs += (attn_weights,)
386
+
387
+ return outputs # a, present, (attentions)
388
+
389
+
390
+ class JAISMLP(nn.Module):
391
+ def __init__(self, intermediate_size, config):
392
+ super().__init__()
393
+ embed_dim = config.hidden_size
394
+ self.swiglu = config.activation_function == "swiglu"
395
+ self.c_fc = Conv1D(intermediate_size, embed_dim)
396
+ self.c_fc2 = Conv1D(intermediate_size, embed_dim) if self.swiglu else None
397
+ self.c_proj = Conv1D(embed_dim, intermediate_size)
398
+ self.act = SwiGLUActivation() if self.swiglu else ACT2FN[config.activation_function]
399
+ self.dropout = nn.Dropout(config.resid_pdrop)
400
+
401
+ def forward(self, hidden_states: Optional[Tuple[torch.FloatTensor]]) -> torch.FloatTensor:
402
+ if self.swiglu:
403
+ hidden_states2 = self.c_fc2(hidden_states)
404
+ hidden_states = self.c_fc(hidden_states)
405
+ hidden_states = self.act(hidden_states, hidden_states2) if self.swiglu else self.act(hidden_states)
406
+ hidden_states = self.c_proj(hidden_states)
407
+ hidden_states = self.dropout(hidden_states)
408
+ return hidden_states
409
+
410
+
411
+ class JAISBlock(nn.Module):
412
+ def __init__(self, config, layer_idx=None):
413
+ super().__init__()
414
+ hidden_size = config.hidden_size
415
+ inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
416
+
417
+ self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
418
+ self.attn = JAISAttention(config, layer_idx=layer_idx)
419
+ self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
420
+
421
+ if config.add_cross_attention:
422
+ self.crossattention = JAISAttention(config, is_cross_attention=True, layer_idx=layer_idx)
423
+ self.ln_cross_attn = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
424
+
425
+ self.mlp = JAISMLP(inner_dim, config)
426
+
427
+ def forward(
428
+ self,
429
+ hidden_states: Optional[Tuple[torch.FloatTensor]],
430
+ layer_past: Optional[Tuple[torch.Tensor]] = None,
431
+ attention_mask: Optional[torch.FloatTensor] = None,
432
+ head_mask: Optional[torch.FloatTensor] = None,
433
+ encoder_hidden_states: Optional[torch.Tensor] = None,
434
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
435
+ use_cache: Optional[bool] = False,
436
+ output_attentions: Optional[bool] = False,
437
+ position_bias: Optional[torch.FloatTensor] = None,
438
+ ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
439
+ residual = hidden_states
440
+ hidden_states = self.ln_1(hidden_states)
441
+ attn_outputs = self.attn(
442
+ hidden_states,
443
+ layer_past=layer_past,
444
+ attention_mask=attention_mask,
445
+ head_mask=head_mask,
446
+ use_cache=use_cache,
447
+ output_attentions=output_attentions,
448
+ position_bias=position_bias,
449
+ )
450
+ attn_output = attn_outputs[0] # output_attn: a, present, (attentions)
451
+ outputs = attn_outputs[1:]
452
+ # residual connection
453
+ hidden_states = attn_output + residual
454
+
455
+ if encoder_hidden_states is not None:
456
+ # add one self-attention block for cross-attention
457
+ if not hasattr(self, "crossattention"):
458
+ raise ValueError(
459
+ f"If `encoder_hidden_states` are passed, {self} has to be instantiated with "
460
+ "cross-attention layers by setting `config.add_cross_attention=True`"
461
+ )
462
+ residual = hidden_states
463
+ hidden_states = self.ln_cross_attn(hidden_states)
464
+ cross_attn_outputs = self.crossattention(
465
+ hidden_states,
466
+ attention_mask=attention_mask,
467
+ head_mask=head_mask,
468
+ encoder_hidden_states=encoder_hidden_states,
469
+ encoder_attention_mask=encoder_attention_mask,
470
+ output_attentions=output_attentions,
471
+ position_bias=position_bias,
472
+ )
473
+ attn_output = cross_attn_outputs[0]
474
+ # residual connection
475
+ hidden_states = residual + attn_output
476
+ outputs = outputs + cross_attn_outputs[2:] # add cross attentions if we output attention weights
477
+
478
+ residual = hidden_states
479
+ hidden_states = self.ln_2(hidden_states)
480
+ feed_forward_hidden_states = self.mlp(hidden_states)
481
+ # residual connection
482
+ hidden_states = residual + feed_forward_hidden_states
483
+
484
+ if use_cache:
485
+ outputs = (hidden_states,) + outputs
486
+ else:
487
+ outputs = (hidden_states,) + outputs[1:]
488
+
489
+ return outputs # hidden_states, present, (attentions, cross_attentions)
490
+
491
+
492
+ class JAISPreTrainedModel(PreTrainedModel):
493
+ """
494
+ An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
495
+ models.
496
+ """
497
+
498
+ config_class = JAISConfig
499
+ load_tf_weights = load_tf_weights_in_jais
500
+ base_model_prefix = "transformer"
501
+ is_parallelizable = True
502
+ supports_gradient_checkpointing = True
503
+ _no_split_modules = ["JAISBlock"]
504
+ _skip_keys_device_placement = "past_key_values"
505
+
506
+ def __init__(self, *inputs, **kwargs):
507
+ super().__init__(*inputs, **kwargs)
508
+
509
+ def _init_weights(self, module):
510
+ """Initialize the weights."""
511
+ mup_init_scale = math.sqrt(self.config.width_scale)
512
+ if isinstance(module, (nn.Linear, Conv1D)):
513
+ # Slightly different from the TF version which uses truncated_normal for initialization
514
+ # cf https://github.com/pytorch/pytorch/pull/5617
515
+ module.weight.data.normal_(mean=0.0, std=(self.config.initializer_range * mup_init_scale))
516
+ if module.bias is not None:
517
+ module.bias.data.zero_()
518
+ elif isinstance(module, nn.Embedding):
519
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
520
+ if module.padding_idx is not None:
521
+ module.weight.data[module.padding_idx].zero_()
522
+ elif isinstance(module, nn.LayerNorm):
523
+ module.bias.data.zero_()
524
+ module.weight.data.fill_(1.0)
525
+
526
+ # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
527
+ # > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
528
+ # > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
529
+ # > -- GPT-2 :: https://openai.com/blog/better-language-models/
530
+ #
531
+ # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
532
+ for name, p in module.named_parameters():
533
+ if name == "c_proj.weight":
534
+ # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
535
+ stddev = self.config.initializer_range * mup_init_scale / math.sqrt(2 * self.config.n_layer)
536
+ p.data.normal_(mean=0.0, std=stddev)
537
+
538
+ def _set_gradient_checkpointing(self, module, value=False):
539
+ if isinstance(module, JAISModel):
540
+ module.gradient_checkpointing = value
541
+
542
+
543
+ JAIS_START_DOCSTRING = r"""
544
+
545
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
546
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
547
+ etc.)
548
+
549
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
550
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
551
+ and behavior.
552
+
553
+ Parameters:
554
+ config ([`JAISConfig`]): Model configuration class with all the parameters of the model.
555
+ Initializing with a config file does not load the weights associated with the model, only the
556
+ configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
557
+ """
558
+
559
+ JAIS_INPUTS_DOCSTRING = r"""
560
+ Args:
561
+ input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
562
+ `input_ids_length` = `sequence_length` if `past_key_values` is `None` else
563
+ `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
564
+ sequence tokens in the vocabulary.
565
+
566
+ If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
567
+ `input_ids`.
568
+
569
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
570
+ [`PreTrainedTokenizer.__call__`] for details.
571
+
572
+ [What are input IDs?](../glossary#input-ids)
573
+ past_key_values (`Tuple[Tuple[torch.Tensor]]` of length `config.n_layers`):
574
+ Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
575
+ `past_key_values` output below). Can be used to speed up sequential decoding. The `input_ids` which have
576
+ their past given to this model should not be passed as `input_ids` as they have already been computed.
577
+ attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
578
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
579
+
580
+ - 1 for tokens that are **not masked**,
581
+ - 0 for tokens that are **masked**.
582
+
583
+ If `past_key_values` is used, `attention_mask` needs to contain the masking strategy that was used for
584
+ `past_key_values`. In other words, the `attention_mask` always has to have the length:
585
+ `len(past_key_values) + len(input_ids)`
586
+
587
+ [What are attention masks?](../glossary#attention-mask)
588
+ token_type_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
589
+ Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
590
+ 1]`:
591
+
592
+ - 0 corresponds to a *sentence A* token,
593
+ - 1 corresponds to a *sentence B* token.
594
+
595
+ [What are token type IDs?](../glossary#token-type-ids)
596
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
597
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
598
+ config.max_position_embeddings - 1]`.
599
+
600
+ [What are position IDs?](../glossary#position-ids)
601
+ head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
602
+ Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
603
+
604
+ - 1 indicates the head is **not masked**,
605
+ - 0 indicates the head is **masked**.
606
+
607
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
608
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
609
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
610
+ model's internal embedding lookup matrix.
611
+
612
+ If `past_key_values` is used, optionally only the last `inputs_embeds` have to be input (see
613
+ `past_key_values`).
614
+ use_cache (`bool`, *optional*):
615
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
616
+ `past_key_values`).
617
+ output_attentions (`bool`, *optional*):
618
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
619
+ tensors for more detail.
620
+ output_hidden_states (`bool`, *optional*):
621
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
622
+ more detail.
623
+ return_dict (`bool`, *optional*):
624
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
625
+ """
626
+ PARALLELIZE_DOCSTRING = r"""
627
+ This is an experimental feature and is a subject to change at a moment's notice.
628
+
629
+ Uses a device map to distribute attention modules of the model across several devices. If no device map is given,
630
+ it will evenly distribute blocks across all devices.
631
+
632
+ Args:
633
+ device_map (`Dict[int, list]`, optional, defaults to None):
634
+ A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
635
+ automatically mapped to the first device (for esoteric reasons). That means that the first device should
636
+ have fewer attention modules mapped to it than other devices. For reference, the gpt2 models have the
637
+ following number of attention modules:
638
+
639
+ - gpt2: 12
640
+ - gpt2-medium: 24
641
+ - gpt2-large: 36
642
+ - gpt2-xl: 48
643
+
644
+ Example:
645
+
646
+ ```python
647
+ # Here is an example of a device map on a machine with 4 GPUs using gpt2-xl, which has a total of 48 attention modules:
648
+ model = GPT2LMHeadModel.from_pretrained("gpt2-xl")
649
+ device_map = {
650
+ 0: [0, 1, 2, 3, 4, 5, 6, 7, 8],
651
+ 1: [9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21],
652
+ 2: [22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34],
653
+ 3: [35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47],
654
+ }
655
+ model.parallelize(device_map)
656
+ ```
657
+ """
658
+ DEPARALLELIZE_DOCSTRING = r"""
659
+ Moves the model to cpu from a model parallel state.
660
+
661
+ Example:
662
+
663
+ ```python
664
+ # On a 4 GPU machine with gpt2-large:
665
+ model = GPT2LMHeadModel.from_pretrained("gpt2-large")
666
+ device_map = {
667
+ 0: [0, 1, 2, 3, 4, 5, 6, 7],
668
+ 1: [8, 9, 10, 11, 12, 13, 14, 15],
669
+ 2: [16, 17, 18, 19, 20, 21, 22, 23],
670
+ 3: [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35],
671
+ }
672
+ model.parallelize(device_map) # Splits the model across several devices
673
+ model.deparallelize() # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()
674
+ ```
675
+ """
676
+
677
+
678
+ @add_start_docstrings(
679
+ "The bare JAIS Model transformer outputting raw hidden-states without any specific head on top.",
680
+ JAIS_START_DOCSTRING,
681
+ )
682
+ class JAISModel(JAISPreTrainedModel):
683
+ _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"]
684
+ _keys_to_ignore_on_load_missing = [r"attn.masked_bias", r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias"]
685
+
686
+ def __init__(self, config):
687
+ super().__init__(config)
688
+
689
+ self.embed_dim = config.hidden_size
690
+
691
+ self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
692
+ self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim) if config.position_embedding_type != "alibi" else None
693
+ self.embeddings_scale = config.embeddings_scale
694
+
695
+ self.drop = nn.Dropout(config.embd_pdrop)
696
+ self.h = nn.ModuleList([JAISBlock(config, layer_idx=i) for i in range(config.num_hidden_layers)])
697
+ self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
698
+
699
+ self.relative_pe = AlibiPositionEmbeddingLayer(config.num_attention_heads) if config.position_embedding_type == "alibi" else None
700
+
701
+ # Model parallel
702
+ self.model_parallel = False
703
+ self.device_map = None
704
+ self.gradient_checkpointing = False
705
+
706
+ # Initialize weights and apply final processing
707
+ self.post_init()
708
+
709
+ @add_start_docstrings(PARALLELIZE_DOCSTRING)
710
+ def parallelize(self, device_map=None):
711
+ # Check validity of device_map
712
+ warnings.warn(
713
+ "`JAISModel.parallelize` is deprecated and will be removed in v5 of Transformers, you should load your"
714
+ " model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
715
+ " `device_map` but it needs to be a dictionary module_name to device, so for instance {'h.0': 0, 'h.1': 1,"
716
+ " ...}",
717
+ FutureWarning,
718
+ )
719
+ self.device_map = (
720
+ get_device_map(len(self.h), range(torch.cuda.device_count())) if device_map is None else device_map
721
+ )
722
+ assert_device_map(self.device_map, len(self.h))
723
+ self.model_parallel = True
724
+ self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))
725
+ self.last_device = "cuda:" + str(max(self.device_map.keys()))
726
+ self.wte = self.wte.to(self.first_device)
727
+ if self.wpe is not None:
728
+ self.wpe = self.wpe.to(self.first_device)
729
+ # Load onto devices
730
+ for k, v in self.device_map.items():
731
+ for block in v:
732
+ cuda_device = "cuda:" + str(k)
733
+ self.h[block] = self.h[block].to(cuda_device)
734
+ # ln_f to last
735
+ self.ln_f = self.ln_f.to(self.last_device)
736
+
737
+ @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
738
+ def deparallelize(self):
739
+ warnings.warn(
740
+ "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
741
+ FutureWarning,
742
+ )
743
+ self.model_parallel = False
744
+ self.device_map = None
745
+ self.first_device = "cpu"
746
+ self.last_device = "cpu"
747
+ self.wte = self.wte.to("cpu")
748
+ if self.wpe is not None:
749
+ self.wpe = self.wpe.to("cpu")
750
+ for index in range(len(self.h)):
751
+ self.h[index] = self.h[index].to("cpu")
752
+ self.ln_f = self.ln_f.to("cpu")
753
+ torch.cuda.empty_cache()
754
+
755
+ def get_input_embeddings(self):
756
+ return self.wte
757
+
758
+ def set_input_embeddings(self, new_embeddings):
759
+ self.wte = new_embeddings
760
+
761
+ def _prune_heads(self, heads_to_prune):
762
+ """
763
+ Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
764
+ """
765
+ for layer, heads in heads_to_prune.items():
766
+ self.h[layer].attn.prune_heads(heads)
767
+
768
+ @add_start_docstrings_to_model_forward(JAIS_INPUTS_DOCSTRING)
769
+ @add_code_sample_docstrings(
770
+ checkpoint=_CHECKPOINT_FOR_DOC,
771
+ output_type=BaseModelOutputWithPastAndCrossAttentions,
772
+ config_class=_CONFIG_FOR_DOC,
773
+ )
774
+ def forward(
775
+ self,
776
+ input_ids: Optional[torch.LongTensor] = None,
777
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
778
+ attention_mask: Optional[torch.FloatTensor] = None,
779
+ token_type_ids: Optional[torch.LongTensor] = None,
780
+ position_ids: Optional[torch.LongTensor] = None,
781
+ head_mask: Optional[torch.FloatTensor] = None,
782
+ inputs_embeds: Optional[torch.FloatTensor] = None,
783
+ encoder_hidden_states: Optional[torch.Tensor] = None,
784
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
785
+ use_cache: Optional[bool] = None,
786
+ output_attentions: Optional[bool] = None,
787
+ output_hidden_states: Optional[bool] = None,
788
+ return_dict: Optional[bool] = None,
789
+ ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
790
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
791
+ output_hidden_states = (
792
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
793
+ )
794
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
795
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
796
+
797
+ if input_ids is not None and inputs_embeds is not None:
798
+ raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
799
+ elif input_ids is not None:
800
+ input_shape = input_ids.size()
801
+ input_ids = input_ids.view(-1, input_shape[-1])
802
+ batch_size = input_ids.shape[0]
803
+ elif inputs_embeds is not None:
804
+ input_shape = inputs_embeds.size()[:-1]
805
+ batch_size = inputs_embeds.shape[0]
806
+ else:
807
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
808
+
809
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
810
+
811
+ if token_type_ids is not None:
812
+ token_type_ids = token_type_ids.view(-1, input_shape[-1])
813
+ if position_ids is not None:
814
+ position_ids = position_ids.view(-1, input_shape[-1])
815
+
816
+ if past_key_values is None:
817
+ past_length = 0
818
+ past_key_values = tuple([None] * len(self.h))
819
+ else:
820
+ past_length = past_key_values[0][0].size(-2)
821
+ if position_ids is None:
822
+ position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
823
+ position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
824
+
825
+ # JAISAttention mask.
826
+ if attention_mask is not None:
827
+ if batch_size <= 0:
828
+ raise ValueError("batch_size has to be defined and > 0")
829
+ attention_mask = attention_mask.view(batch_size, -1)
830
+ # We create a 3D attention mask from a 2D tensor mask.
831
+ # Sizes are [batch_size, 1, 1, to_seq_length]
832
+ # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
833
+ # this attention mask is more simple than the triangular masking of causal attention
834
+ # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
835
+ attention_mask = attention_mask[:, None, None, :]
836
+
837
+ # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
838
+ # masked positions, this operation will create a tensor which is 0.0 for
839
+ # positions we want to attend and the dtype's smallest value for masked positions.
840
+ # Since we are adding it to the raw scores before the softmax, this is
841
+ # effectively the same as removing these entirely.
842
+ attention_mask = attention_mask.to(dtype=self.dtype) # fp16 compatibility
843
+ attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
844
+
845
+ # If a 2D or 3D attention mask is provided for the cross-attention
846
+ # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
847
+ if self.config.add_cross_attention and encoder_hidden_states is not None:
848
+ encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
849
+ encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
850
+ if encoder_attention_mask is None:
851
+ encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
852
+ encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)
853
+ else:
854
+ encoder_attention_mask = None
855
+
856
+ # Prepare head mask if needed
857
+ # 1.0 in head_mask indicate we keep the head
858
+ # attention_probs has shape bsz x n_heads x N x N
859
+ # head_mask has shape n_layer x batch x n_heads x N x N
860
+ head_mask = self.get_head_mask(head_mask, self.config.n_layer)
861
+
862
+ if inputs_embeds is None:
863
+ inputs_embeds = self.wte(input_ids)
864
+ if self.wpe is not None:
865
+ position_embeds = self.wpe(position_ids)
866
+ hidden_states = inputs_embeds + position_embeds
867
+ else:
868
+ hidden_states = inputs_embeds
869
+ hidden_states *= torch.tensor(
870
+ float(self.embeddings_scale), dtype=hidden_states.dtype, device=hidden_states.device
871
+ )
872
+
873
+ if token_type_ids is not None:
874
+ token_type_embeds = self.wte(token_type_ids)
875
+ hidden_states = hidden_states + token_type_embeds
876
+
877
+ hidden_states = self.drop(hidden_states)
878
+
879
+ if self.relative_pe is not None:
880
+ length = input_ids.shape[1]
881
+ cached_kv_length = 0
882
+ cached_kv = past_key_values[0]
883
+ if cached_kv is not None:
884
+ cached_kv_length = cached_kv[0].shape[-2]
885
+ position_bias = self.relative_pe(length, length, cached_kv_length)
886
+ else:
887
+ position_bias = None
888
+
889
+ output_shape = input_shape + (hidden_states.size(-1),)
890
+
891
+ if self.gradient_checkpointing and self.training:
892
+ if use_cache:
893
+ logger.warning_once(
894
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
895
+ )
896
+ use_cache = False
897
+
898
+ presents = () if use_cache else None
899
+ all_self_attentions = () if output_attentions else None
900
+ all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
901
+ all_hidden_states = () if output_hidden_states else None
902
+ for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
903
+ # Model parallel
904
+ if self.model_parallel:
905
+ torch.cuda.set_device(hidden_states.device)
906
+ # Ensure layer_past is on same device as hidden_states (might not be correct)
907
+ if layer_past is not None:
908
+ layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
909
+ # Ensure that attention_mask is always on the same device as hidden_states
910
+ if attention_mask is not None:
911
+ attention_mask = attention_mask.to(hidden_states.device)
912
+ if isinstance(head_mask, torch.Tensor):
913
+ head_mask = head_mask.to(hidden_states.device)
914
+ if output_hidden_states:
915
+ all_hidden_states = all_hidden_states + (hidden_states,)
916
+
917
+ if self.gradient_checkpointing and self.training:
918
+
919
+ def create_custom_forward(module):
920
+ def custom_forward(*inputs):
921
+ # None for past_key_value
922
+ return module(*inputs, use_cache, output_attentions)
923
+
924
+ return custom_forward
925
+
926
+ outputs = torch.utils.checkpoint.checkpoint(
927
+ create_custom_forward(block),
928
+ hidden_states,
929
+ None,
930
+ attention_mask,
931
+ head_mask[i],
932
+ encoder_hidden_states,
933
+ encoder_attention_mask,
934
+ )
935
+ else:
936
+ outputs = block(
937
+ hidden_states,
938
+ layer_past=layer_past,
939
+ attention_mask=attention_mask,
940
+ head_mask=head_mask[i],
941
+ encoder_hidden_states=encoder_hidden_states,
942
+ encoder_attention_mask=encoder_attention_mask,
943
+ use_cache=use_cache,
944
+ output_attentions=output_attentions,
945
+ position_bias=position_bias,
946
+ )
947
+
948
+ hidden_states = outputs[0]
949
+ if use_cache is True:
950
+ presents = presents + (outputs[1],)
951
+
952
+ if output_attentions:
953
+ all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
954
+ if self.config.add_cross_attention:
955
+ all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],)
956
+
957
+ # Model Parallel: If it's the last layer for that device, put things on the next device
958
+ if self.model_parallel:
959
+ for k, v in self.device_map.items():
960
+ if i == v[-1] and "cuda:" + str(k) != self.last_device:
961
+ hidden_states = hidden_states.to("cuda:" + str(k + 1))
962
+
963
+ hidden_states = self.ln_f(hidden_states)
964
+
965
+ hidden_states = hidden_states.view(output_shape)
966
+ # Add last hidden state
967
+ if output_hidden_states:
968
+ all_hidden_states = all_hidden_states + (hidden_states,)
969
+
970
+ if not return_dict:
971
+ return tuple(
972
+ v
973
+ for v in [hidden_states, presents, all_hidden_states, all_self_attentions, all_cross_attentions]
974
+ if v is not None
975
+ )
976
+
977
+ return BaseModelOutputWithPastAndCrossAttentions(
978
+ last_hidden_state=hidden_states,
979
+ past_key_values=presents,
980
+ hidden_states=all_hidden_states,
981
+ attentions=all_self_attentions,
982
+ cross_attentions=all_cross_attentions,
983
+ )
984
+
985
+
986
+ @add_start_docstrings(
987
+ """
988
+ The JAIS Model transformer with a language modeling head on top (linear layer with weights tied to the input
989
+ embeddings).
990
+ """,
991
+ JAIS_START_DOCSTRING,
992
+ )
993
+ class JAISLMHeadModel(JAISPreTrainedModel):
994
+ _keys_to_ignore_on_load_missing = [r"lm_head.weight"]
995
+ _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias"]
996
+
997
+ def __init__(self, config):
998
+ super().__init__(config)
999
+ self.transformer = JAISModel(config)
1000
+ self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
1001
+ self.output_logits_scale = config.width_scale
1002
+
1003
+ # Model parallel
1004
+ self.model_parallel = False
1005
+ self.device_map = None
1006
+
1007
+ # Initialize weights and apply final processing
1008
+ self.post_init()
1009
+
1010
+ @add_start_docstrings(PARALLELIZE_DOCSTRING)
1011
+ def parallelize(self, device_map=None):
1012
+ warnings.warn(
1013
+ "`JAISLMHeadModel.parallelize` is deprecated and will be removed in v5 of Transformers, you should load"
1014
+ " your model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
1015
+ " `device_map` but it needs to be a dictionary module_name to device, so for instance {'transformer.h.0':"
1016
+ " 0, 'transformer.h.1': 1, ...}",
1017
+ FutureWarning,
1018
+ )
1019
+ self.device_map = (
1020
+ get_device_map(len(self.transformer.h), range(torch.cuda.device_count()))
1021
+ if device_map is None
1022
+ else device_map
1023
+ )
1024
+ assert_device_map(self.device_map, len(self.transformer.h))
1025
+ self.transformer.parallelize(self.device_map)
1026
+ self.lm_head = self.lm_head.to(self.transformer.first_device)
1027
+ self.model_parallel = True
1028
+
1029
+ @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
1030
+ def deparallelize(self):
1031
+ warnings.warn(
1032
+ "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
1033
+ FutureWarning,
1034
+ )
1035
+ self.transformer.deparallelize()
1036
+ self.transformer = self.transformer.to("cpu")
1037
+ self.lm_head = self.lm_head.to("cpu")
1038
+ self.model_parallel = False
1039
+ torch.cuda.empty_cache()
1040
+
1041
+ def get_output_embeddings(self):
1042
+ return self.lm_head
1043
+
1044
+ def set_output_embeddings(self, new_embeddings):
1045
+ self.lm_head = new_embeddings
1046
+
1047
+ def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
1048
+ token_type_ids = kwargs.get("token_type_ids", None)
1049
+ # only last token for inputs_ids if past is defined in kwargs
1050
+ if past_key_values:
1051
+ input_ids = input_ids[:, -1].unsqueeze(-1)
1052
+ if token_type_ids is not None:
1053
+ token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
1054
+
1055
+ attention_mask = kwargs.get("attention_mask", None)
1056
+ position_ids = kwargs.get("position_ids", None)
1057
+
1058
+ if attention_mask is not None and position_ids is None:
1059
+ # create position_ids on the fly for batch generation
1060
+ position_ids = attention_mask.long().cumsum(-1) - 1
1061
+ position_ids.masked_fill_(attention_mask == 0, 1)
1062
+ if past_key_values:
1063
+ position_ids = position_ids[:, -1].unsqueeze(-1)
1064
+ else:
1065
+ position_ids = None
1066
+
1067
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
1068
+ if inputs_embeds is not None and past_key_values is None:
1069
+ model_inputs = {"inputs_embeds": inputs_embeds}
1070
+ else:
1071
+ model_inputs = {"input_ids": input_ids}
1072
+
1073
+ model_inputs.update(
1074
+ {
1075
+ "past_key_values": past_key_values,
1076
+ "use_cache": kwargs.get("use_cache"),
1077
+ "position_ids": position_ids,
1078
+ "attention_mask": attention_mask,
1079
+ "token_type_ids": token_type_ids,
1080
+ }
1081
+ )
1082
+ return model_inputs
1083
+
1084
+ @add_start_docstrings_to_model_forward(JAIS_INPUTS_DOCSTRING)
1085
+ @add_code_sample_docstrings(
1086
+ checkpoint=_CHECKPOINT_FOR_DOC,
1087
+ output_type=CausalLMOutputWithCrossAttentions,
1088
+ config_class=_CONFIG_FOR_DOC,
1089
+ )
1090
+ def forward(
1091
+ self,
1092
+ input_ids: Optional[torch.LongTensor] = None,
1093
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
1094
+ attention_mask: Optional[torch.FloatTensor] = None,
1095
+ token_type_ids: Optional[torch.LongTensor] = None,
1096
+ position_ids: Optional[torch.LongTensor] = None,
1097
+ head_mask: Optional[torch.FloatTensor] = None,
1098
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1099
+ encoder_hidden_states: Optional[torch.Tensor] = None,
1100
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
1101
+ labels: Optional[torch.LongTensor] = None,
1102
+ use_cache: Optional[bool] = None,
1103
+ output_attentions: Optional[bool] = None,
1104
+ output_hidden_states: Optional[bool] = None,
1105
+ return_dict: Optional[bool] = None,
1106
+ ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
1107
+ r"""
1108
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1109
+ Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
1110
+ `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
1111
+ are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
1112
+ """
1113
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1114
+
1115
+ transformer_outputs = self.transformer(
1116
+ input_ids,
1117
+ past_key_values=past_key_values,
1118
+ attention_mask=attention_mask,
1119
+ token_type_ids=token_type_ids,
1120
+ position_ids=position_ids,
1121
+ head_mask=head_mask,
1122
+ inputs_embeds=inputs_embeds,
1123
+ encoder_hidden_states=encoder_hidden_states,
1124
+ encoder_attention_mask=encoder_attention_mask,
1125
+ use_cache=use_cache,
1126
+ output_attentions=output_attentions,
1127
+ output_hidden_states=output_hidden_states,
1128
+ return_dict=return_dict,
1129
+ )
1130
+ hidden_states = transformer_outputs[0]
1131
+
1132
+ # Set device for model parallelism
1133
+ if self.model_parallel:
1134
+ torch.cuda.set_device(self.transformer.first_device)
1135
+ hidden_states = hidden_states.to(self.lm_head.weight.device)
1136
+
1137
+ lm_logits = self.lm_head(hidden_states)
1138
+ lm_logits *= torch.tensor(
1139
+ float(self.output_logits_scale), dtype=lm_logits.dtype, device=lm_logits.device
1140
+ )
1141
+
1142
+ loss = None
1143
+ if labels is not None:
1144
+ # move labels to correct device to enable model parallelism
1145
+ labels = labels.to(lm_logits.device)
1146
+ # Shift so that tokens < n predict n
1147
+ shift_logits = lm_logits[..., :-1, :].contiguous()
1148
+ shift_labels = labels[..., 1:].contiguous()
1149
+ # Flatten the tokens
1150
+ loss_fct = CrossEntropyLoss()
1151
+ loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
1152
+
1153
+ if not return_dict:
1154
+ output = (lm_logits,) + transformer_outputs[1:]
1155
+ return ((loss,) + output) if loss is not None else output
1156
+
1157
+ return CausalLMOutputWithCrossAttentions(
1158
+ loss=loss,
1159
+ logits=lm_logits,
1160
+ past_key_values=transformer_outputs.past_key_values,
1161
+ hidden_states=transformer_outputs.hidden_states,
1162
+ attentions=transformer_outputs.attentions,
1163
+ cross_attentions=transformer_outputs.cross_attentions,
1164
+ )
1165
+
1166
+ @staticmethod
1167
+ def _reorder_cache(
1168
+ past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
1169
+ ) -> Tuple[Tuple[torch.Tensor]]:
1170
+ """
1171
+ This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
1172
+ [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
1173
+ beam_idx at every generation step.
1174
+ """
1175
+ return tuple(
1176
+ tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
1177
+ for layer_past in past_key_values
1178
+ )
1179
+
1180
+
1181
+ @add_start_docstrings(
1182
+ """
1183
+ The JAIS Model transformer with a sequence classification head on top (linear layer).
1184
+
1185
+ [`JAISForSequenceClassification`] uses the last token in order to do the classification, as other causal models
1186
+ (e.g. GPT-1) do.
1187
+
1188
+ Since it does classification on the last token, it requires to know the position of the last token. If a
1189
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
1190
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
1191
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
1192
+ each row of the batch).
1193
+ """,
1194
+ JAIS_START_DOCSTRING,
1195
+ )
1196
+ class JAISForSequenceClassification(JAISPreTrainedModel):
1197
+ _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"]
1198
+ _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"lm_head.weight"]
1199
+
1200
+ def __init__(self, config):
1201
+ super().__init__(config)
1202
+ self.num_labels = config.num_labels
1203
+ self.transformer = JAISModel(config)
1204
+ self.score = nn.Linear(config.n_embd, self.num_labels, bias=False)
1205
+ self.output_logits_scale = config.width_scale
1206
+
1207
+ # Model parallel
1208
+ self.model_parallel = False
1209
+ self.device_map = None
1210
+
1211
+ # Initialize weights and apply final processing
1212
+ self.post_init()
1213
+
1214
+ @add_start_docstrings_to_model_forward(JAIS_INPUTS_DOCSTRING)
1215
+ @add_code_sample_docstrings(
1216
+ checkpoint="microsoft/DialogRPT-updown",
1217
+ output_type=SequenceClassifierOutputWithPast,
1218
+ config_class=_CONFIG_FOR_DOC,
1219
+ )
1220
+ def forward(
1221
+ self,
1222
+ input_ids: Optional[torch.LongTensor] = None,
1223
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
1224
+ attention_mask: Optional[torch.FloatTensor] = None,
1225
+ token_type_ids: Optional[torch.LongTensor] = None,
1226
+ position_ids: Optional[torch.LongTensor] = None,
1227
+ head_mask: Optional[torch.FloatTensor] = None,
1228
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1229
+ labels: Optional[torch.LongTensor] = None,
1230
+ use_cache: Optional[bool] = None,
1231
+ output_attentions: Optional[bool] = None,
1232
+ output_hidden_states: Optional[bool] = None,
1233
+ return_dict: Optional[bool] = None,
1234
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
1235
+ r"""
1236
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1237
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1238
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1239
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1240
+ """
1241
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1242
+
1243
+ transformer_outputs = self.transformer(
1244
+ input_ids,
1245
+ past_key_values=past_key_values,
1246
+ attention_mask=attention_mask,
1247
+ token_type_ids=token_type_ids,
1248
+ position_ids=position_ids,
1249
+ head_mask=head_mask,
1250
+ inputs_embeds=inputs_embeds,
1251
+ use_cache=use_cache,
1252
+ output_attentions=output_attentions,
1253
+ output_hidden_states=output_hidden_states,
1254
+ return_dict=return_dict,
1255
+ )
1256
+ hidden_states = transformer_outputs[0]
1257
+ logits = self.score(hidden_states)
1258
+ logits *= torch.tensor(
1259
+ float(self.output_logits_scale), dtype=logits.dtype, device=logits.device
1260
+ )
1261
+
1262
+ if input_ids is not None:
1263
+ batch_size, sequence_length = input_ids.shape[:2]
1264
+ else:
1265
+ batch_size, sequence_length = inputs_embeds.shape[:2]
1266
+
1267
+ assert (
1268
+ self.config.pad_token_id is not None or batch_size == 1
1269
+ ), "Cannot handle batch sizes > 1 if no padding token is defined."
1270
+ if self.config.pad_token_id is None:
1271
+ sequence_lengths = -1
1272
+ else:
1273
+ if input_ids is not None:
1274
+ sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
1275
+ else:
1276
+ sequence_lengths = -1
1277
+ logger.warning(
1278
+ f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
1279
+ "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
1280
+ )
1281
+
1282
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
1283
+
1284
+ loss = None
1285
+ if labels is not None:
1286
+ if self.config.problem_type is None:
1287
+ if self.num_labels == 1:
1288
+ self.config.problem_type = "regression"
1289
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1290
+ self.config.problem_type = "single_label_classification"
1291
+ else:
1292
+ self.config.problem_type = "multi_label_classification"
1293
+
1294
+ if self.config.problem_type == "regression":
1295
+ loss_fct = MSELoss()
1296
+ if self.num_labels == 1:
1297
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1298
+ else:
1299
+ loss = loss_fct(pooled_logits, labels)
1300
+ elif self.config.problem_type == "single_label_classification":
1301
+ loss_fct = CrossEntropyLoss()
1302
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
1303
+ elif self.config.problem_type == "multi_label_classification":
1304
+ loss_fct = BCEWithLogitsLoss()
1305
+ loss = loss_fct(pooled_logits, labels)
1306
+ if not return_dict:
1307
+ output = (pooled_logits,) + transformer_outputs[1:]
1308
+ return ((loss,) + output) if loss is not None else output
1309
+
1310
+ return SequenceClassifierOutputWithPast(
1311
+ loss=loss,
1312
+ logits=pooled_logits,
1313
+ past_key_values=transformer_outputs.past_key_values,
1314
+ hidden_states=transformer_outputs.hidden_states,
1315
+ attentions=transformer_outputs.attentions,
1316
+ )
1317
+
1318
+
1319
+ @add_start_docstrings(
1320
+ """
1321
+ JAIS Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
1322
+ Named-Entity-Recognition (NER) tasks.
1323
+ """,
1324
+ JAIS_START_DOCSTRING,
1325
+ )
1326
+ class JAISForTokenClassification(JAISPreTrainedModel):
1327
+ def __init__(self, config):
1328
+ super().__init__(config)
1329
+ self.num_labels = config.num_labels
1330
+
1331
+ self.transformer = JAISModel(config)
1332
+ if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
1333
+ classifier_dropout = config.classifier_dropout
1334
+ elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
1335
+ classifier_dropout = config.hidden_dropout
1336
+ else:
1337
+ classifier_dropout = 0.1
1338
+ self.dropout = nn.Dropout(classifier_dropout)
1339
+ self.classifier = nn.Linear(config.hidden_size, config.num_labels)
1340
+ self.output_logits_scale = config.width_scale
1341
+
1342
+ # Model parallel
1343
+ self.model_parallel = False
1344
+ self.device_map = None
1345
+
1346
+ # Initialize weights and apply final processing
1347
+ self.post_init()
1348
+
1349
+ @add_start_docstrings_to_model_forward(JAIS_INPUTS_DOCSTRING)
1350
+ # fmt: off
1351
+ @add_code_sample_docstrings(
1352
+ checkpoint="brad1141/gpt2-finetuned-comp2",
1353
+ output_type=TokenClassifierOutput,
1354
+ config_class=_CONFIG_FOR_DOC,
1355
+ expected_loss=0.25,
1356
+ expected_output=["Lead", "Lead", "Lead", "Position", "Lead", "Lead", "Lead", "Lead", "Lead", "Lead", "Lead", "Lead"],
1357
+ )
1358
+ # fmt: on
1359
+ def forward(
1360
+ self,
1361
+ input_ids: Optional[torch.LongTensor] = None,
1362
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
1363
+ attention_mask: Optional[torch.FloatTensor] = None,
1364
+ token_type_ids: Optional[torch.LongTensor] = None,
1365
+ position_ids: Optional[torch.LongTensor] = None,
1366
+ head_mask: Optional[torch.FloatTensor] = None,
1367
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1368
+ labels: Optional[torch.LongTensor] = None,
1369
+ use_cache: Optional[bool] = None,
1370
+ output_attentions: Optional[bool] = None,
1371
+ output_hidden_states: Optional[bool] = None,
1372
+ return_dict: Optional[bool] = None,
1373
+ ) -> Union[Tuple, TokenClassifierOutput]:
1374
+ r"""
1375
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1376
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1377
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1378
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1379
+ """
1380
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1381
+
1382
+ transformer_outputs = self.transformer(
1383
+ input_ids,
1384
+ past_key_values=past_key_values,
1385
+ attention_mask=attention_mask,
1386
+ token_type_ids=token_type_ids,
1387
+ position_ids=position_ids,
1388
+ head_mask=head_mask,
1389
+ inputs_embeds=inputs_embeds,
1390
+ use_cache=use_cache,
1391
+ output_attentions=output_attentions,
1392
+ output_hidden_states=output_hidden_states,
1393
+ return_dict=return_dict,
1394
+ )
1395
+
1396
+ hidden_states = transformer_outputs[0]
1397
+ hidden_states = self.dropout(hidden_states)
1398
+ logits = self.classifier(hidden_states)
1399
+ logits *= torch.tensor(
1400
+ float(self.output_logits_scale), dtype=logits.dtype, device=logits.device
1401
+ )
1402
+
1403
+ loss = None
1404
+ if labels is not None:
1405
+ labels = labels.to(logits.device)
1406
+ loss_fct = CrossEntropyLoss()
1407
+ loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
1408
+
1409
+ if not return_dict:
1410
+ output = (logits,) + transformer_outputs[2:]
1411
+ return ((loss,) + output) if loss is not None else output
1412
+
1413
+ return TokenClassifierOutput(
1414
+ loss=loss,
1415
+ logits=logits,
1416
+ hidden_states=transformer_outputs.hidden_states,
1417
+ attentions=transformer_outputs.attentions,
1418
+ )
1419
+
1420
+
1421
+ @add_start_docstrings(
1422
+ """
1423
+ The JAIS Model transformer with a span classification head on top for extractive question-answering tasks like
1424
+ SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`).
1425
+ """,
1426
+ JAIS_START_DOCSTRING,
1427
+ )
1428
+ class JAISForQuestionAnswering(JAISPreTrainedModel):
1429
+ _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"]
1430
+ _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias", r"lm_head.weight"]
1431
+
1432
+ def __init__(self, config):
1433
+ super().__init__(config)
1434
+ self.num_labels = config.num_labels
1435
+ self.transformer = JAISModel(config)
1436
+ self.qa_outputs = nn.Linear(config.hidden_size, 2)
1437
+ self.output_logits_scale = config.width_scale
1438
+
1439
+ # Model parallel
1440
+ self.model_parallel = False
1441
+ self.device_map = None
1442
+ self.gradient_checkpointing = False
1443
+
1444
+ # Initialize weights and apply final processing
1445
+ self.post_init()
1446
+
1447
+ def forward(
1448
+ self,
1449
+ input_ids: Optional[torch.LongTensor] = None,
1450
+ attention_mask: Optional[torch.FloatTensor] = None,
1451
+ token_type_ids: Optional[torch.LongTensor] = None,
1452
+ position_ids: Optional[torch.LongTensor] = None,
1453
+ head_mask: Optional[torch.FloatTensor] = None,
1454
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1455
+ start_positions: Optional[torch.LongTensor] = None,
1456
+ end_positions: Optional[torch.LongTensor] = None,
1457
+ output_attentions: Optional[bool] = None,
1458
+ output_hidden_states: Optional[bool] = None,
1459
+ return_dict: Optional[bool] = None,
1460
+ ) -> Union[Tuple, QuestionAnsweringModelOutput]:
1461
+ r"""
1462
+ start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1463
+ Labels for position (index) of the start of the labelled span for computing the token classification loss.
1464
+ Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
1465
+ are not taken into account for computing the loss.
1466
+ end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1467
+ Labels for position (index) of the end of the labelled span for computing the token classification loss.
1468
+ Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
1469
+ are not taken into account for computing the loss.
1470
+ """
1471
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1472
+
1473
+ outputs = self.transformer(
1474
+ input_ids,
1475
+ attention_mask=attention_mask,
1476
+ token_type_ids=token_type_ids,
1477
+ position_ids=position_ids,
1478
+ head_mask=head_mask,
1479
+ inputs_embeds=inputs_embeds,
1480
+ output_attentions=output_attentions,
1481
+ output_hidden_states=output_hidden_states,
1482
+ return_dict=return_dict,
1483
+ )
1484
+
1485
+ sequence_output = outputs[0]
1486
+
1487
+ logits = self.qa_outputs(sequence_output)
1488
+ logits *= torch.tensor(
1489
+ float(self.output_logits_scale), dtype=logits.dtype, device=logits.device
1490
+ )
1491
+ start_logits, end_logits = logits.split(1, dim=-1)
1492
+ start_logits = start_logits.squeeze(-1).contiguous()
1493
+ end_logits = end_logits.squeeze(-1).contiguous()
1494
+
1495
+ total_loss = None
1496
+ if start_positions is not None and end_positions is not None:
1497
+ # If we are on multi-GPU, split add a dimension
1498
+ if len(start_positions.size()) > 1:
1499
+ start_positions = start_positions.squeeze(-1).to(start_logits.device)
1500
+ if len(end_positions.size()) > 1:
1501
+ end_positions = end_positions.squeeze(-1).to(end_logits.device)
1502
+ # sometimes the start/end positions are outside our model inputs, we ignore these terms
1503
+ ignored_index = start_logits.size(1)
1504
+ start_positions = start_positions.clamp(0, ignored_index)
1505
+ end_positions = end_positions.clamp(0, ignored_index)
1506
+
1507
+ loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
1508
+ start_loss = loss_fct(start_logits, start_positions)
1509
+ end_loss = loss_fct(end_logits, end_positions)
1510
+ total_loss = (start_loss + end_loss) / 2
1511
+
1512
+ if not return_dict:
1513
+ output = (start_logits, end_logits) + outputs[2:]
1514
+ return ((total_loss,) + output) if total_loss is not None else output
1515
+
1516
+ return QuestionAnsweringModelOutput(
1517
+ loss=total_loss,
1518
+ start_logits=start_logits,
1519
+ end_logits=end_logits,
1520
+ hidden_states=outputs.hidden_states,
1521
+ attentions=outputs.attentions,
1522
+ )