thejagstudio commited on
Commit
3a02b0d
1 Parent(s): 2756b8d

Create modelling_RW.py

Browse files
Files changed (1) hide show
  1. modelling_RW.py +1095 -0
modelling_RW.py ADDED
@@ -0,0 +1,1095 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # port of models described in RW
2
+ # We use the bloom model as a starting point for these model.
3
+ # Please refer to the bloom models for usage instructions.
4
+
5
+ import math
6
+ import warnings
7
+ from typing import Optional, Tuple, Union
8
+
9
+ import torch
10
+ import torch.utils.checkpoint
11
+ from torch import nn
12
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, LayerNorm, MSELoss
13
+ from torch.nn import functional as F
14
+
15
+ from transformers.modeling_outputs import (
16
+ BaseModelOutputWithPastAndCrossAttentions,
17
+ CausalLMOutputWithCrossAttentions,
18
+ QuestionAnsweringModelOutput,
19
+ SequenceClassifierOutputWithPast,
20
+ TokenClassifierOutput,
21
+ )
22
+ from transformers.modeling_utils import PreTrainedModel
23
+ from transformers.utils import logging
24
+ from .configuration_RW import RWConfig
25
+
26
+ logger = logging.get_logger(__name__)
27
+
28
+ # NOTE(Hesslow): Unfortunately we did not fuse matmul and bias during training, this means that there's one additional quantization to bfloat16 between the operations.
29
+ # In order not to degrade the quality of our HF-port, we keep these characteristics in the final model.
30
+ class Linear(nn.Linear):
31
+ def forward(self, input: torch.Tensor) -> torch.Tensor:
32
+ ret = input @ self.weight.T
33
+ if self.bias is None:
34
+ return ret
35
+ else:
36
+ return ret + self.bias
37
+
38
+
39
+ from einops import rearrange
40
+
41
+ # rotary pos emb helpers (torch.jit.script does not seem to support staticmethod...)
42
+ def rotate_half(x):
43
+ x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
44
+ return torch.cat((-x2, x1), dim=x1.ndim - 1) # dim=-1 triggers a bug in torch < 1.8.0
45
+
46
+
47
+ class RotaryEmbedding(torch.nn.Module):
48
+ """Implementation of RotaryEmbedding from GPT-NeoX.
49
+ This implementation is design to operate on queries and keys that are compatible with
50
+ [batch_size, n_heads_per_partition, seq_len, head_dim] (e.g. MinGPTAttention format).
51
+ """
52
+
53
+ def __init__(
54
+ self,
55
+ head_dim: int,
56
+ base=10000,
57
+ ):
58
+ super().__init__()
59
+ inv_freq = 1.0 / (base ** (torch.arange(0, head_dim, 2).float() / head_dim))
60
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
61
+ self.head_dim = head_dim
62
+ self.seq_len_cached = None
63
+ self.batch_size_cached = None
64
+ self.cos_cached: torch.Tensor | None = None
65
+ self.sin_cached: torch.Tensor | None = None
66
+
67
+ def cos_sin(
68
+ self,
69
+ seq_len: int,
70
+ device="cuda",
71
+ dtype=torch.bfloat16,
72
+ ) -> torch.Tensor:
73
+ if seq_len != self.seq_len_cached:
74
+ self.seq_len_cached = seq_len
75
+ t = torch.arange(seq_len, device=device).type_as(self.inv_freq)
76
+ freqs = torch.einsum("i,j->ij", t, self.inv_freq)
77
+ emb = torch.cat((freqs, freqs), dim=-1).to(device)
78
+
79
+ if dtype in [torch.float16, torch.bfloat16]:
80
+ emb = emb.float()
81
+
82
+ self.cos_cached = emb.cos()[None, :, :]
83
+ self.sin_cached = emb.sin()[None, :, :]
84
+
85
+ self.cos_cached = self.cos_cached.type(dtype)
86
+ self.sin_cached = self.sin_cached.type(dtype)
87
+
88
+ return self.cos_cached, self.sin_cached
89
+
90
+ def forward(self, q, k):
91
+ batch, seq_len, head_dim = q.shape
92
+ cos, sin = self.cos_sin(seq_len, q.device, q.dtype)
93
+ return (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)
94
+
95
+
96
+ def _make_causal_mask(
97
+ input_ids_shape: torch.Size, device: torch.device, past_key_values_length: int
98
+ ) -> torch.BoolTensor:
99
+ batch_size, target_length = input_ids_shape
100
+ mask = torch.empty((target_length, target_length + past_key_values_length), dtype=torch.bool, device=device)
101
+ # ONNX doesn't support `torch.Tensor.triu` properly, thus we use this workaround
102
+ seq_ids = torch.arange(target_length, device=device)
103
+ mask[:, past_key_values_length:] = seq_ids[:, None] < seq_ids[None, :]
104
+
105
+ if past_key_values_length > 0:
106
+ mask[:, :past_key_values_length] = False
107
+
108
+ expanded_mask = mask[None, None, :, :].expand(batch_size, 1, target_length, target_length + past_key_values_length)
109
+ return expanded_mask
110
+
111
+
112
+ def _expand_mask(mask: torch.Tensor, tgt_length: int) -> torch.BoolTensor:
113
+ batch_size, src_length = mask.shape
114
+ tgt_length = tgt_length if tgt_length is not None else src_length
115
+
116
+ expanded_mask = ~(mask[:, None, None, :].to(torch.bool))
117
+ return expanded_mask.expand(batch_size, 1, tgt_length, src_length)
118
+
119
+
120
+ def build_alibi_tensor(attention_mask: torch.Tensor, num_heads: int, dtype: torch.dtype) -> torch.Tensor:
121
+ batch_size, seq_length = attention_mask.shape
122
+ closest_power_of_2 = 2 ** math.floor(math.log2(num_heads))
123
+ base = torch.tensor(
124
+ 2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
125
+ )
126
+ powers = torch.arange(1, 1 + closest_power_of_2, device=attention_mask.device, dtype=torch.int32)
127
+ slopes = torch.pow(base, powers)
128
+
129
+ if closest_power_of_2 != num_heads:
130
+ extra_base = torch.tensor(
131
+ 2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32
132
+ )
133
+ num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
134
+ extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, device=attention_mask.device, dtype=torch.int32)
135
+ slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)
136
+
137
+ # Note: alibi will added to the attention bias that will be applied to the query, key product of attention
138
+ # => therefore alibi will have to be of shape (batch_size, num_heads, query_length, key_length)
139
+ # => here we set (batch_size=1, num_heads=num_heads, query_length=1, key_length=max_length)
140
+ # => the query_length dimension will then be broadcasted correctly
141
+ # This is more or less identical to T5's relative position bias:
142
+ # https://github.com/huggingface/transformers/blob/f681437203baa7671de3174b0fa583c349d9d5e1/src/transformers/models/t5/modeling_t5.py#L527
143
+ arange_tensor = ((attention_mask.cumsum(dim=-1) - 1) * attention_mask)[:, None, :]
144
+ alibi = slopes[..., None].bfloat16() * arange_tensor
145
+ return alibi.reshape(batch_size * num_heads, 1, seq_length).to(dtype)
146
+
147
+
148
+ def dropout_add(x: torch.Tensor, residual: torch.Tensor, prob: float, training: bool) -> torch.Tensor:
149
+ out = F.dropout(x, p=prob, training=training)
150
+ out = residual + out
151
+ return out
152
+
153
+
154
+ class Attention(nn.Module):
155
+ def __init__(self, config: RWConfig):
156
+ super().__init__()
157
+
158
+ self.hidden_size = config.hidden_size
159
+ self.num_heads = config.n_head
160
+ self.head_dim = self.hidden_size // self.num_heads
161
+ self.split_size = self.hidden_size
162
+ self.hidden_dropout = config.hidden_dropout
163
+
164
+ if self.head_dim * self.num_heads != self.hidden_size:
165
+ raise ValueError(
166
+ f"`hidden_size` must be divisible by num_heads (got `hidden_size`: {self.hidden_size} and `num_heads`:"
167
+ f" {self.num_heads})."
168
+ )
169
+
170
+ self.maybe_rotary = RotaryEmbedding(config.head_dim) if config.rotary else lambda q, k: (q, k)
171
+
172
+ # Layer-wise attention scaling
173
+ self.inv_norm_factor = 1.0 / math.sqrt(self.head_dim)
174
+ self.beta = self.inv_norm_factor
175
+
176
+ self.query_key_value = Linear(
177
+ self.hidden_size,
178
+ 3 * self.hidden_size if not config.multi_query else (self.hidden_size + 2 * self.head_dim),
179
+ bias=config.bias,
180
+ )
181
+ self.multi_query = config.multi_query
182
+ self.dense = Linear(self.hidden_size, self.hidden_size, bias=config.bias)
183
+ self.attention_dropout = nn.Dropout(config.attention_dropout)
184
+ self.num_kv = config.n_head if not self.multi_query else 1
185
+
186
+ def _split_heads(self, fused_qkv: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
187
+ """
188
+ Split the last dimension into (num_heads, head_dim) without making any copies, results share same memory
189
+ storage as `fused_qkv`
190
+ Args:
191
+ fused_qkv (`torch.tensor`, *required*): [batch_size, seq_length, num_heads * 3 * head_dim]
192
+ Returns:
193
+ query: [batch_size, seq_length, num_heads, head_dim] key: [batch_size, seq_length, num_heads, head_dim]
194
+ value: [batch_size, seq_length, num_heads, head_dim]
195
+ """
196
+ if not self.multi_query:
197
+ batch_size, seq_length, three_times_hidden_size = fused_qkv.shape
198
+ fused_qkv = fused_qkv.view(batch_size, seq_length, self.num_heads, 3, self.head_dim)
199
+ return fused_qkv[..., 0, :], fused_qkv[..., 1, :], fused_qkv[..., 2, :]
200
+ else:
201
+ batch_size, seq_length, three_times_hidden_size = fused_qkv.shape
202
+ fused_qkv = fused_qkv.view(batch_size, seq_length, self.num_heads + 2, self.head_dim)
203
+ return fused_qkv[..., :-2, :], fused_qkv[..., [-2], :], fused_qkv[..., [-1], :]
204
+
205
+ def _merge_heads(self, x: torch.Tensor) -> torch.Tensor:
206
+ """
207
+ Merge heads together over the last dimenstion
208
+ Args:
209
+ x: (`torch.tensor`, *required*): [batch_size * num_heads, seq_length, head_dim]
210
+ Returns:
211
+ torch.tensor: [batch_size, seq_length, num_heads * head_dim]
212
+ """
213
+ # What we want to achieve is:
214
+ # batch_size * num_heads, seq_length, head_dim -> batch_size, seq_length, num_heads * head_dim
215
+ batch_size_and_num_heads, seq_length, _ = x.shape
216
+ batch_size = batch_size_and_num_heads // self.num_heads
217
+
218
+ # First view to decompose the batch size
219
+ # batch_size * num_heads, seq_length, head_dim -> batch_size, num_heads, seq_length, head_dim
220
+ x = x.view(batch_size, self.num_heads, seq_length, self.head_dim)
221
+
222
+ # batch_size, num_heads, seq_length, head_dim -> batch_size, seq_length, num_heads, head_dim
223
+ x = x.permute(0, 2, 1, 3)
224
+
225
+ # batch_size, seq_length, num_heads, head_dim -> batch_size, seq_length, num_heads * head_dim
226
+ return x.reshape(batch_size, seq_length, self.num_heads * self.head_dim)
227
+
228
+ def forward(
229
+ self,
230
+ hidden_states: torch.Tensor,
231
+ alibi: torch.Tensor,
232
+ attention_mask: torch.Tensor,
233
+ layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
234
+ head_mask: Optional[torch.Tensor] = None,
235
+ use_cache: bool = False,
236
+ output_attentions: bool = False,
237
+ ):
238
+ fused_qkv = self.query_key_value(hidden_states) # [batch_size, seq_length, 3 x hidden_size]
239
+
240
+ # 3 x [batch_size, seq_length, num_heads, head_dim]
241
+ (query_layer, key_layer, value_layer) = self._split_heads(fused_qkv)
242
+
243
+ batch_size, q_length, _, _ = query_layer.shape
244
+
245
+ query_layer = query_layer.transpose(1, 2).reshape(batch_size * self.num_heads, q_length, self.head_dim)
246
+ key_layer = key_layer.transpose(1, 2).reshape(
247
+ batch_size * self.num_kv,
248
+ q_length,
249
+ self.head_dim,
250
+ )
251
+ value_layer = value_layer.transpose(1, 2).reshape(batch_size * self.num_kv, q_length, self.head_dim)
252
+
253
+ query_layer, key_layer = self.maybe_rotary(query_layer, key_layer)
254
+
255
+ if layer_past is not None:
256
+ past_key, past_value = layer_past
257
+ # concatenate along seq_length dimension:
258
+ # - key: [batch_size * self.num_heads, head_dim, kv_length]
259
+ # - value: [batch_size * self.num_heads, kv_length, head_dim]
260
+ key_layer = torch.cat((past_key, key_layer), dim=1)
261
+ value_layer = torch.cat((past_value, value_layer), dim=1)
262
+
263
+ _, kv_length, _ = key_layer.shape
264
+
265
+ if use_cache is True:
266
+ present = (key_layer, value_layer)
267
+ else:
268
+ present = None
269
+
270
+ if alibi is None:
271
+ query_layer_ = query_layer.reshape(batch_size, self.num_heads, -1, self.head_dim)
272
+ key_layer_ = key_layer.reshape(batch_size, self.num_kv, -1, self.head_dim)
273
+ value_layer_ = value_layer.reshape(batch_size, self.num_kv, -1, self.head_dim)
274
+
275
+ attn_output = F.scaled_dot_product_attention(
276
+ query_layer_, key_layer_, value_layer_, None, 0.0, is_causal=True
277
+ )
278
+
279
+ x = attn_output.view(batch_size, self.num_heads, q_length, self.head_dim)
280
+ x = x.permute(0, 2, 1, 3)
281
+ attn_output = x.reshape(batch_size, q_length, self.num_heads * self.head_dim)
282
+
283
+ output_tensor = self.dense(attn_output)
284
+
285
+ outputs = (output_tensor, present)
286
+ assert not output_attentions # not supported.
287
+ return outputs
288
+ else:
289
+ attention_mask_float = (attention_mask * 1.0).masked_fill(attention_mask, -1e9).to(torch.bfloat16)
290
+ matmul_result = query_layer @ key_layer.transpose(-1, -2)
291
+
292
+ # change view to [batch_size, num_heads, q_length, kv_length]
293
+ attention_scores = matmul_result.view(batch_size, self.num_heads, q_length, kv_length)
294
+
295
+ # cast attention scores to fp32, compute scaled softmax and cast back to initial dtype - [batch_size, num_heads, q_length, kv_length]
296
+ input_dtype = attention_scores.dtype
297
+ # `float16` has a minimum value of -65504.0, whereas `bfloat16` and `float32` have a minimum value of `-3.4e+38`
298
+ if input_dtype == torch.float16 or input_dtype == torch.bfloat16:
299
+ attention_scores = attention_scores.to(torch.float32)
300
+ # attn_weights = torch.masked_fill(attention_scores, attention_mask, torch.finfo(attention_scores.dtype).min)
301
+ attention_probs = F.softmax(
302
+ (attention_scores + alibi.view(batch_size, self.num_heads, 1, -1)) * self.inv_norm_factor + attention_mask_float,
303
+ dim=-1,
304
+ dtype=hidden_states.dtype,
305
+ )
306
+ # [batch_size, num_heads, q_length, kv_length]
307
+ attention_probs = self.attention_dropout(attention_probs)
308
+
309
+ if head_mask is not None:
310
+ attention_probs = attention_probs * head_mask
311
+
312
+ # change view [batch_size x num_heads, q_length, kv_length]
313
+ attention_probs_reshaped = attention_probs.view(batch_size * self.num_heads, q_length, kv_length)
314
+
315
+ # matmul: [batch_size * num_heads, q_length, head_dim]
316
+ context_layer = attention_probs_reshaped @ value_layer
317
+
318
+ # change view [batch_size, num_heads, q_length, head_dim]
319
+ context_layer = self._merge_heads(context_layer)
320
+
321
+ output_tensor = self.dense(context_layer)
322
+
323
+ outputs = (output_tensor, present)
324
+ if output_attentions:
325
+ outputs += (attention_probs,)
326
+
327
+ return outputs
328
+
329
+
330
+ class MLP(nn.Module):
331
+ def __init__(self, config: RWConfig):
332
+ super().__init__()
333
+ hidden_size = config.hidden_size
334
+
335
+ self.dense_h_to_4h = Linear(hidden_size, 4 * hidden_size, bias=config.bias)
336
+ self.act = nn.GELU()
337
+ self.dense_4h_to_h = Linear(4 * hidden_size, hidden_size, bias=config.bias)
338
+ self.hidden_dropout = config.hidden_dropout
339
+
340
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
341
+ x = self.act(self.dense_h_to_4h(x))
342
+ x = self.dense_4h_to_h(x)
343
+ return x
344
+
345
+
346
+ class DecoderLayer(nn.Module):
347
+ def __init__(self, config: RWConfig):
348
+ super().__init__()
349
+ hidden_size = config.hidden_size
350
+
351
+ self.input_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
352
+ self.num_heads = config.n_head
353
+ self.self_attention = Attention(config)
354
+
355
+ if not config.parallel_attn:
356
+ # unused if parallel attn
357
+ self.post_attention_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
358
+
359
+ self.mlp = MLP(config)
360
+
361
+ self.apply_residual_connection_post_layernorm = config.apply_residual_connection_post_layernorm
362
+ self.hidden_dropout = config.hidden_dropout
363
+
364
+ self.config = config
365
+
366
+ def forward(
367
+ self,
368
+ hidden_states: torch.Tensor,
369
+ alibi: torch.Tensor,
370
+ attention_mask: torch.Tensor,
371
+ layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
372
+ head_mask: Optional[torch.Tensor] = None,
373
+ use_cache: bool = False,
374
+ output_attentions: bool = False,
375
+ ):
376
+
377
+ layernorm_output = self.input_layernorm(hidden_states)
378
+ residual = hidden_states
379
+
380
+ # Self attention.
381
+ attn_outputs = self.self_attention(
382
+ layernorm_output,
383
+ layer_past=layer_past,
384
+ attention_mask=attention_mask,
385
+ alibi=alibi,
386
+ head_mask=head_mask,
387
+ use_cache=use_cache,
388
+ output_attentions=output_attentions,
389
+ )
390
+
391
+ attention_output = attn_outputs[0]
392
+
393
+ if not self.config.parallel_attn:
394
+ residual = dropout_add(attention_output, residual, self.config.attention_dropout, training=self.training)
395
+ layernorm_output = self.post_attention_layernorm(residual)
396
+
397
+ outputs = attn_outputs[1:]
398
+
399
+ # MLP.
400
+ mlp_output = self.mlp(layernorm_output)
401
+
402
+ if self.config.parallel_attn:
403
+ mlp_output += attention_output
404
+
405
+ output = dropout_add(mlp_output, residual, self.config.hidden_dropout, training=self.training)
406
+
407
+ if use_cache:
408
+ outputs = (output,) + outputs
409
+ else:
410
+ outputs = (output,) + outputs[1:]
411
+
412
+ return outputs # hidden_states, present, attentions
413
+
414
+
415
+ class RWPreTrainedModel(PreTrainedModel):
416
+ _keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
417
+ """
418
+ An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
419
+ models.
420
+ """
421
+
422
+ config_class = RWConfig
423
+ base_model_prefix = "transformer"
424
+ supports_gradient_checkpointing = True
425
+ _no_split_modules = ["DecoderLayer"]
426
+
427
+ def __init__(self, *inputs, **kwargs):
428
+ super().__init__(*inputs, **kwargs)
429
+
430
+ def _init_weights(self, module: nn.Module):
431
+ """Initialize the weights."""
432
+ if isinstance(module, nn.Linear) or isinstance(module, Linear):
433
+ # Slightly different from the TF version which uses truncated_normal for initialization
434
+ # cf https://github.com/pytorch/pytorch/pull/5617
435
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
436
+ if module.bias is not None:
437
+ module.bias.data.zero_()
438
+ elif isinstance(module, nn.Embedding):
439
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
440
+ if module.padding_idx is not None:
441
+ module.weight.data[module.padding_idx].zero_()
442
+ elif isinstance(module, LayerNorm):
443
+ module.bias.data.zero_()
444
+ module.weight.data.fill_(1.0)
445
+
446
+ def _set_gradient_checkpointing(self, module: nn.Module, value: bool = False):
447
+ if isinstance(module, RWModel):
448
+ module.gradient_checkpointing = value
449
+
450
+ @staticmethod
451
+ def _convert_to_standard_cache(
452
+ past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]], batch_size: int
453
+ ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
454
+ """
455
+ Standardizes the format of the cache so as to match most implementations, i.e. to tuple(tuple([batch_size,
456
+ num_heads, ...]))
457
+ """
458
+ batch_size_times_num_heads, head_dim, seq_length = past_key_value[0][0].shape
459
+ num_heads = batch_size_times_num_heads // batch_size
460
+ # key: [batch_size * num_heads, head_dim, seq_length] -> [batch_size, num_heads, head_dim, seq_length]
461
+ # value: [batch_size * num_heads, seq_length, head_dim] -> [batch_size, num_heads, seq_length, head_dim]
462
+ return tuple(
463
+ (
464
+ layer_past[0].view(batch_size, num_heads, head_dim, seq_length),
465
+ layer_past[1].view(batch_size, num_heads, seq_length, head_dim),
466
+ )
467
+ for layer_past in past_key_value
468
+ )
469
+
470
+ @staticmethod
471
+ def _convert_to_rw_cache(
472
+ past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]]
473
+ ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]:
474
+ batch_size, num_heads, head_dim, seq_length = past_key_value[0][0].shape
475
+ batch_size_times_num_heads = batch_size * num_heads
476
+ # key: [batch_size, num_heads, head_dim, seq_length] -> [batch_size * num_heads, head_dim, seq_length]
477
+ # value: [batch_size, num_heads, seq_length, head_dim] -> [batch_size * num_heads, seq_length, head_dim]
478
+ return tuple(
479
+ (
480
+ layer_past[0].view(batch_size_times_num_heads, head_dim, seq_length),
481
+ layer_past[1].view(batch_size_times_num_heads, seq_length, head_dim),
482
+ )
483
+ for layer_past in past_key_value
484
+ )
485
+
486
+
487
+ class RWModel(RWPreTrainedModel):
488
+ def __init__(self, config: RWConfig):
489
+ super().__init__(config)
490
+
491
+ self.embed_dim = config.hidden_size
492
+ self.num_heads = config.n_head
493
+ self.alibi = config.alibi
494
+
495
+ # Embedding + LN Embedding
496
+ self.word_embeddings = nn.Embedding(config.vocab_size, self.embed_dim)
497
+
498
+ # Transformer blocks
499
+ self.h = nn.ModuleList([DecoderLayer(config) for _ in range(config.num_hidden_layers)])
500
+
501
+ # Final Layer Norm
502
+ self.ln_f = LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
503
+
504
+ self.gradient_checkpointing = False
505
+
506
+ # Initialize weights and apply final processing
507
+ self.post_init()
508
+
509
+ def get_input_embeddings(self):
510
+ return self.word_embeddings
511
+
512
+ def _prepare_attn_mask(
513
+ self, attention_mask: torch.Tensor, input_shape: Tuple[int, int], past_key_values_length: int
514
+ ) -> torch.BoolTensor:
515
+ # create causal mask
516
+ # [batch_size, seq_length] -> [batch_size, 1, tgt_length, src_length]
517
+ combined_attention_mask = None
518
+ device = attention_mask.device
519
+ _, src_length = input_shape
520
+
521
+ if src_length > 1:
522
+ combined_attention_mask = _make_causal_mask(
523
+ input_shape, device=device, past_key_values_length=past_key_values_length
524
+ )
525
+
526
+ # [batch_size, seq_length] -> [batch_size, 1, tgt_length, src_length]
527
+ expanded_attn_mask = _expand_mask(attention_mask, tgt_length=src_length)
528
+ combined_attention_mask = (
529
+ expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask | combined_attention_mask
530
+ )
531
+
532
+ return combined_attention_mask
533
+
534
+ def set_input_embeddings(self, new_embeddings: torch.Tensor):
535
+ self.word_embeddings = new_embeddings
536
+
537
+ def forward(
538
+ self,
539
+ input_ids: Optional[torch.LongTensor] = None,
540
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
541
+ attention_mask: Optional[torch.Tensor] = None,
542
+ head_mask: Optional[torch.LongTensor] = None,
543
+ inputs_embeds: Optional[torch.LongTensor] = None,
544
+ use_cache: Optional[bool] = None,
545
+ output_attentions: Optional[bool] = None,
546
+ output_hidden_states: Optional[bool] = None,
547
+ return_dict: Optional[bool] = None,
548
+ **deprecated_arguments,
549
+ ) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPastAndCrossAttentions]:
550
+ if deprecated_arguments.pop("position_ids", False) is not False:
551
+ # `position_ids` could have been `torch.Tensor` or `None` so defaulting pop to `False` allows to detect if users were passing explicitly `None`
552
+ warnings.warn(
553
+ "`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. You can safely ignore"
554
+ " passing `position_ids`.",
555
+ FutureWarning,
556
+ )
557
+ if len(deprecated_arguments) > 0:
558
+ raise ValueError(f"Got unexpected arguments: {deprecated_arguments}")
559
+
560
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
561
+ output_hidden_states = (
562
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
563
+ )
564
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
565
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
566
+
567
+ if input_ids is not None and inputs_embeds is not None:
568
+ raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
569
+ elif input_ids is not None:
570
+ batch_size, seq_length = input_ids.shape
571
+ elif inputs_embeds is not None:
572
+ batch_size, seq_length, _ = inputs_embeds.shape
573
+ else:
574
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
575
+
576
+ if past_key_values is None:
577
+ past_key_values = tuple([None] * len(self.h))
578
+
579
+ # Prepare head mask if needed
580
+ # 1.0 in head_mask indicate we keep the head
581
+ # attention_probs has shape batch_size x num_heads x N x N
582
+ # head_mask has shape n_layer x batch x num_heads x N x N
583
+ head_mask = self.get_head_mask(head_mask, self.config.n_layer)
584
+
585
+ if inputs_embeds is None:
586
+ inputs_embeds = self.word_embeddings(input_ids)
587
+
588
+ hidden_states = inputs_embeds
589
+
590
+ presents = () if use_cache else None
591
+ all_self_attentions = () if output_attentions else None
592
+ all_hidden_states = () if output_hidden_states else None
593
+
594
+ # Compute alibi tensor: check build_alibi_tensor documentation
595
+ seq_length_with_past = seq_length
596
+ past_key_values_length = 0
597
+ if past_key_values[0] is not None:
598
+ past_key_values_length = past_key_values[0][0].shape[2]
599
+ seq_length_with_past = seq_length_with_past + past_key_values_length
600
+ if attention_mask is None:
601
+ attention_mask = torch.ones((batch_size, seq_length_with_past), device=hidden_states.device)
602
+ else:
603
+ attention_mask = attention_mask.to(hidden_states.device)
604
+
605
+ if self.alibi:
606
+ alibi = build_alibi_tensor(attention_mask, self.num_heads, dtype=hidden_states.dtype)
607
+ else:
608
+ alibi = None
609
+
610
+ causal_mask = self._prepare_attn_mask(
611
+ attention_mask,
612
+ input_shape=(batch_size, seq_length),
613
+ past_key_values_length=past_key_values_length,
614
+ )
615
+
616
+ for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
617
+
618
+ if output_hidden_states:
619
+ all_hidden_states = all_hidden_states + (hidden_states,)
620
+
621
+ if self.gradient_checkpointing and self.training:
622
+
623
+ if use_cache:
624
+ logger.warning(
625
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
626
+ )
627
+ use_cache = False
628
+
629
+ def create_custom_forward(module):
630
+ def custom_forward(*inputs):
631
+ # None for past_key_value
632
+ return module(*inputs, use_cache=use_cache, output_attentions=output_attentions)
633
+
634
+ return custom_forward
635
+
636
+ outputs = torch.utils.checkpoint.checkpoint(
637
+ create_custom_forward(block),
638
+ hidden_states,
639
+ alibi,
640
+ causal_mask,
641
+ head_mask[i],
642
+ )
643
+ else:
644
+ outputs = block(
645
+ hidden_states,
646
+ layer_past=layer_past,
647
+ attention_mask=causal_mask,
648
+ head_mask=head_mask[i],
649
+ use_cache=use_cache,
650
+ output_attentions=output_attentions,
651
+ alibi=alibi,
652
+ )
653
+
654
+ hidden_states = outputs[0]
655
+ if use_cache is True:
656
+ presents = presents + (outputs[1],)
657
+
658
+ if output_attentions:
659
+ all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
660
+
661
+ # Add last hidden state
662
+ hidden_states = self.ln_f(hidden_states)
663
+
664
+ if output_hidden_states:
665
+ all_hidden_states = all_hidden_states + (hidden_states,)
666
+
667
+ if not return_dict:
668
+ return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
669
+
670
+ return BaseModelOutputWithPastAndCrossAttentions(
671
+ last_hidden_state=hidden_states,
672
+ past_key_values=presents,
673
+ hidden_states=all_hidden_states,
674
+ attentions=all_self_attentions,
675
+ )
676
+
677
+
678
+ class RWForCausalLM(RWPreTrainedModel):
679
+ _keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
680
+
681
+ def __init__(self, config: RWConfig):
682
+ super().__init__(config)
683
+ self.transformer = RWModel(config)
684
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
685
+
686
+ # Initialize weights and apply final processing
687
+ self.post_init()
688
+
689
+ def get_output_embeddings(self):
690
+ return self.lm_head
691
+
692
+ def set_output_embeddings(self, new_embeddings: torch.Tensor):
693
+ self.lm_head = new_embeddings
694
+
695
+ def prepare_inputs_for_generation(
696
+ self,
697
+ input_ids: torch.LongTensor,
698
+ past: Optional[torch.Tensor] = None,
699
+ attention_mask: Optional[torch.Tensor] = None,
700
+ **kwargs,
701
+ ) -> dict:
702
+ # only last token for input_ids if past is not None
703
+ if past:
704
+ input_ids = input_ids[:, -1].unsqueeze(-1)
705
+
706
+ # the cache may be in the stardard format (e.g. in contrastive search), convert to our's format if needed
707
+ if past[0][0].shape[0] == input_ids.shape[0]:
708
+ past = self._convert_to_rw_cache(past)
709
+
710
+ return {
711
+ "input_ids": input_ids,
712
+ "past_key_values": past,
713
+ "use_cache": kwargs.get("use_cache"),
714
+ "attention_mask": attention_mask,
715
+ }
716
+
717
+ def forward(
718
+ self,
719
+ input_ids: Optional[torch.LongTensor] = None,
720
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
721
+ attention_mask: Optional[torch.Tensor] = None,
722
+ head_mask: Optional[torch.Tensor] = None,
723
+ inputs_embeds: Optional[torch.Tensor] = None,
724
+ labels: Optional[torch.Tensor] = None,
725
+ use_cache: Optional[bool] = None,
726
+ output_attentions: Optional[bool] = None,
727
+ output_hidden_states: Optional[bool] = None,
728
+ return_dict: Optional[bool] = None,
729
+ **deprecated_arguments,
730
+ ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
731
+ r"""
732
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
733
+ Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
734
+ `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
735
+ are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
736
+ """
737
+ if deprecated_arguments.pop("position_ids", False) is not False:
738
+ # `position_ids` could have been `torch.Tensor` or `None` so defaulting pop to `False` allows to detect if users were passing explicitly `None`
739
+ warnings.warn(
740
+ "`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. You can safely ignore"
741
+ " passing `position_ids`.",
742
+ FutureWarning,
743
+ )
744
+ if len(deprecated_arguments) > 0:
745
+ raise ValueError(f"Got unexpected arguments: {deprecated_arguments}")
746
+
747
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
748
+
749
+ transformer_outputs = self.transformer(
750
+ input_ids,
751
+ past_key_values=past_key_values,
752
+ attention_mask=attention_mask,
753
+ head_mask=head_mask,
754
+ inputs_embeds=inputs_embeds,
755
+ use_cache=use_cache,
756
+ output_attentions=output_attentions,
757
+ output_hidden_states=output_hidden_states,
758
+ return_dict=return_dict,
759
+ )
760
+ hidden_states = transformer_outputs[0]
761
+
762
+ lm_logits = self.lm_head(hidden_states)
763
+
764
+ loss = None
765
+ if labels is not None:
766
+ # Shift so that tokens < n predict n
767
+ shift_logits = lm_logits[..., :-1, :].contiguous()
768
+ shift_labels = labels[..., 1:].contiguous()
769
+ batch_size, seq_length, vocab_size = shift_logits.shape
770
+ # Flatten the tokens
771
+ loss_fct = CrossEntropyLoss()
772
+ loss = loss_fct(
773
+ shift_logits.view(batch_size * seq_length, vocab_size), shift_labels.view(batch_size * seq_length)
774
+ )
775
+
776
+ if not return_dict:
777
+ output = (lm_logits,) + transformer_outputs[1:]
778
+ return ((loss,) + output) if loss is not None else output
779
+
780
+ return CausalLMOutputWithCrossAttentions(
781
+ loss=loss,
782
+ logits=lm_logits,
783
+ past_key_values=transformer_outputs.past_key_values,
784
+ hidden_states=transformer_outputs.hidden_states,
785
+ attentions=transformer_outputs.attentions,
786
+ )
787
+
788
+ def _reorder_cache(
789
+ self, past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor
790
+ ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]:
791
+ """
792
+ This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
793
+ [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
794
+ beam_idx at every generation step.
795
+ Output shares the same memory storage as `past`.
796
+ """
797
+ standardized_past = self._convert_to_standard_cache(past, batch_size=len(beam_idx))
798
+
799
+ # Get a copy of `beam_idx` on all the devices where we need those indices.
800
+ device_to_beam_idx = {
801
+ past_state.device: beam_idx.to(past_state.device) for layer_past in past for past_state in layer_past
802
+ }
803
+ reordered_past = tuple(
804
+ (
805
+ layer_past[0].index_select(0, device_to_beam_idx[layer_past[0].device]),
806
+ layer_past[1].index_select(0, device_to_beam_idx[layer_past[0].device]),
807
+ )
808
+ for layer_past in standardized_past
809
+ )
810
+ return self._convert_to_rw_cache(reordered_past)
811
+
812
+
813
+ class RWForSequenceClassification(RWPreTrainedModel):
814
+ _keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
815
+
816
+ def __init__(self, config: RWConfig):
817
+ super().__init__(config)
818
+ self.num_labels = config.num_labels
819
+ self.transformer = RWModel(config)
820
+ self.score = nn.Linear(config.hidden_size, config.num_labels, bias=False)
821
+
822
+ # Initialize weights and apply final processing
823
+ self.post_init()
824
+
825
+ def forward(
826
+ self,
827
+ input_ids: Optional[torch.LongTensor] = None,
828
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
829
+ attention_mask: Optional[torch.Tensor] = None,
830
+ head_mask: Optional[torch.Tensor] = None,
831
+ inputs_embeds: Optional[torch.Tensor] = None,
832
+ labels: Optional[torch.Tensor] = None,
833
+ use_cache: Optional[bool] = None,
834
+ output_attentions: Optional[bool] = None,
835
+ output_hidden_states: Optional[bool] = None,
836
+ return_dict: Optional[bool] = None,
837
+ **deprecated_arguments,
838
+ ) -> Union[Tuple[torch.Tensor], SequenceClassifierOutputWithPast]:
839
+ r"""
840
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
841
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
842
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
843
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
844
+ """
845
+ if deprecated_arguments.pop("position_ids", False) is not False:
846
+ # `position_ids` could have been `torch.Tensor` or `None` so defaulting pop to `False` allows to detect if users were passing explicitly `None`
847
+ warnings.warn(
848
+ "`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. You can safely ignore"
849
+ " passing `position_ids`.",
850
+ FutureWarning,
851
+ )
852
+ if len(deprecated_arguments) > 0:
853
+ raise ValueError(f"Got unexpected arguments: {deprecated_arguments}")
854
+
855
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
856
+
857
+ transformer_outputs = self.transformer(
858
+ input_ids,
859
+ past_key_values=past_key_values,
860
+ attention_mask=attention_mask,
861
+ head_mask=head_mask,
862
+ inputs_embeds=inputs_embeds,
863
+ use_cache=use_cache,
864
+ output_attentions=output_attentions,
865
+ output_hidden_states=output_hidden_states,
866
+ return_dict=return_dict,
867
+ )
868
+
869
+ hidden_states = transformer_outputs[0]
870
+ logits = self.score(hidden_states)
871
+
872
+ if input_ids is not None:
873
+ batch_size = input_ids.shape[0]
874
+ else:
875
+ batch_size = inputs_embeds.shape[0]
876
+
877
+ if self.config.pad_token_id is None and batch_size != 1:
878
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
879
+ if self.config.pad_token_id is None:
880
+ sequence_lengths = -1
881
+ else:
882
+ if input_ids is not None:
883
+ sequence_lengths = torch.ne(input_ids, self.config.pad_token_id).sum(dim=-1) - 1
884
+ else:
885
+ sequence_lengths = -1
886
+ logger.warning(
887
+ f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
888
+ "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
889
+ )
890
+
891
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
892
+
893
+ loss = None
894
+ if labels is not None:
895
+ if self.config.problem_type is None:
896
+ if self.num_labels == 1:
897
+ self.config.problem_type = "regression"
898
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
899
+ self.config.problem_type = "single_label_classification"
900
+ else:
901
+ self.config.problem_type = "multi_label_classification"
902
+
903
+ if self.config.problem_type == "regression":
904
+ loss_fct = MSELoss()
905
+ if self.num_labels == 1:
906
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
907
+ else:
908
+ loss = loss_fct(pooled_logits, labels)
909
+ elif self.config.problem_type == "single_label_classification":
910
+ loss_fct = CrossEntropyLoss()
911
+ loss = loss_fct(pooled_logits, labels)
912
+ elif self.config.problem_type == "multi_label_classification":
913
+ loss_fct = BCEWithLogitsLoss()
914
+ loss = loss_fct(pooled_logits, labels)
915
+ if not return_dict:
916
+ output = (pooled_logits,) + transformer_outputs[1:]
917
+ return ((loss,) + output) if loss is not None else output
918
+
919
+ return SequenceClassifierOutputWithPast(
920
+ loss=loss,
921
+ logits=pooled_logits,
922
+ past_key_values=transformer_outputs.past_key_values,
923
+ hidden_states=transformer_outputs.hidden_states,
924
+ attentions=transformer_outputs.attentions,
925
+ )
926
+
927
+
928
+ class RWForTokenClassification(RWPreTrainedModel):
929
+ _keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
930
+
931
+ def __init__(self, config: RWConfig):
932
+ super().__init__(config)
933
+ self.num_labels = config.num_labels
934
+
935
+ self.transformer = RWModel(config)
936
+ if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
937
+ classifier_dropout = config.classifier_dropout
938
+ elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
939
+ classifier_dropout = config.hidden_dropout
940
+ else:
941
+ classifier_dropout = 0.1
942
+ self.dropout = nn.Dropout(classifier_dropout)
943
+ self.classifier = nn.Linear(config.hidden_size, config.num_labels)
944
+
945
+ # Initialize weights and apply final processing
946
+ self.post_init()
947
+
948
+ def forward(
949
+ self,
950
+ input_ids: Optional[torch.LongTensor] = None,
951
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
952
+ attention_mask: Optional[torch.Tensor] = None,
953
+ head_mask: Optional[torch.Tensor] = None,
954
+ inputs_embeds: Optional[torch.Tensor] = None,
955
+ labels: Optional[torch.Tensor] = None,
956
+ use_cache: Optional[bool] = None,
957
+ output_attentions: Optional[bool] = None,
958
+ output_hidden_states: Optional[bool] = None,
959
+ return_dict: Optional[bool] = None,
960
+ **deprecated_arguments,
961
+ ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
962
+ r"""
963
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
964
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
965
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
966
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
967
+ """
968
+ if deprecated_arguments.pop("position_ids", False) is not False:
969
+ # `position_ids` could have been `torch.Tensor` or `None` so defaulting pop to `False` allows to detect if users were passing explicitly `None`
970
+ warnings.warn(
971
+ "`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. You can safely ignore"
972
+ " passing `position_ids`.",
973
+ FutureWarning,
974
+ )
975
+ if len(deprecated_arguments) > 0:
976
+ raise ValueError(f"Got unexpected arguments: {deprecated_arguments}")
977
+
978
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
979
+
980
+ transformer_outputs = self.transformer(
981
+ input_ids,
982
+ past_key_values=past_key_values,
983
+ attention_mask=attention_mask,
984
+ head_mask=head_mask,
985
+ inputs_embeds=inputs_embeds,
986
+ use_cache=use_cache,
987
+ output_attentions=output_attentions,
988
+ output_hidden_states=output_hidden_states,
989
+ return_dict=return_dict,
990
+ )
991
+
992
+ hidden_states = transformer_outputs[0]
993
+ hidden_states = self.dropout(hidden_states)
994
+ logits = self.classifier(hidden_states)
995
+
996
+ loss = None
997
+ if labels is not None:
998
+ batch_size, seq_length = labels.shape
999
+ loss_fct = CrossEntropyLoss()
1000
+ loss = loss_fct(logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length))
1001
+
1002
+ if not return_dict:
1003
+ output = (logits,) + transformer_outputs[2:]
1004
+ return ((loss,) + output) if loss is not None else output
1005
+
1006
+ return TokenClassifierOutput(
1007
+ loss=loss,
1008
+ logits=logits,
1009
+ hidden_states=transformer_outputs.hidden_states,
1010
+ attentions=transformer_outputs.attentions,
1011
+ )
1012
+
1013
+
1014
+ class RWForQuestionAnswering(RWPreTrainedModel):
1015
+ _keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
1016
+
1017
+ def __init__(self, config):
1018
+ super().__init__(config)
1019
+ self.transformer = RWModel(config)
1020
+ self.qa_outputs = nn.Linear(config.hidden_size, 2)
1021
+
1022
+ # Initialize weights and apply final processing
1023
+ self.post_init()
1024
+
1025
+ def forward(
1026
+ self,
1027
+ input_ids: Optional[torch.LongTensor] = None,
1028
+ attention_mask: Optional[torch.FloatTensor] = None,
1029
+ position_ids: Optional[torch.LongTensor] = None,
1030
+ head_mask: Optional[torch.FloatTensor] = None,
1031
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1032
+ start_positions: Optional[torch.LongTensor] = None,
1033
+ end_positions: Optional[torch.LongTensor] = None,
1034
+ output_attentions: Optional[bool] = None,
1035
+ output_hidden_states: Optional[bool] = None,
1036
+ return_dict: Optional[bool] = None,
1037
+ ) -> Union[Tuple, QuestionAnsweringModelOutput]:
1038
+ r"""
1039
+ start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1040
+ Labels for position (index) of the start of the labelled span for computing the token classification loss.
1041
+ Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
1042
+ are not taken into account for computing the loss.
1043
+ end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1044
+ Labels for position (index) of the end of the labelled span for computing the token classification loss.
1045
+ Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
1046
+ are not taken into account for computing the loss.
1047
+ """
1048
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1049
+
1050
+ outputs = self.transformer(
1051
+ input_ids,
1052
+ attention_mask=attention_mask,
1053
+ position_ids=position_ids,
1054
+ head_mask=head_mask,
1055
+ inputs_embeds=inputs_embeds,
1056
+ output_attentions=output_attentions,
1057
+ output_hidden_states=output_hidden_states,
1058
+ return_dict=return_dict,
1059
+ )
1060
+
1061
+ sequence_output = outputs[0]
1062
+
1063
+ logits = self.qa_outputs(sequence_output)
1064
+ start_logits, end_logits = logits.split(1, dim=-1)
1065
+ start_logits = start_logits.squeeze(-1).contiguous()
1066
+ end_logits = end_logits.squeeze(-1).contiguous()
1067
+
1068
+ total_loss = None
1069
+ if start_positions is not None and end_positions is not None:
1070
+ # If we are on multi-GPU, split add a dimension
1071
+ if len(start_positions.size()) > 1:
1072
+ start_positions = start_positions.squeeze(-1)
1073
+ if len(end_positions.size()) > 1:
1074
+ end_positions = end_positions.squeeze(-1)
1075
+ # sometimes the start/end positions are outside our model inputs, we ignore these terms
1076
+ ignored_index = start_logits.size(1)
1077
+ start_positions = start_positions.clamp(0, ignored_index)
1078
+ end_positions = end_positions.clamp(0, ignored_index)
1079
+
1080
+ loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
1081
+ start_loss = loss_fct(start_logits, start_positions)
1082
+ end_loss = loss_fct(end_logits, end_positions)
1083
+ total_loss = (start_loss + end_loss) / 2
1084
+
1085
+ if not return_dict:
1086
+ output = (start_logits, end_logits) + outputs[2:]
1087
+ return ((total_loss,) + output) if total_loss is not None else output
1088
+
1089
+ return QuestionAnsweringModelOutput(
1090
+ loss=total_loss,
1091
+ start_logits=start_logits,
1092
+ end_logits=end_logits,
1093
+ hidden_states=outputs.hidden_states,
1094
+ attentions=outputs.attentions,
1095
+ )