Key-value caching
Browse files- modeling_nort5.py +209 -97
modeling_nort5.py
CHANGED
@@ -1,19 +1,17 @@
|
|
1 |
-
from __future__ import absolute_import, division, print_function, unicode_literals
|
2 |
-
|
3 |
import math
|
4 |
from typing import List, Optional, Tuple, Union
|
5 |
|
6 |
import torch
|
7 |
import torch.nn as nn
|
8 |
import torch.nn.functional as F
|
9 |
-
from
|
10 |
from torch.utils import checkpoint
|
11 |
|
12 |
from configuration_nort5 import NorT5Config
|
13 |
from transformers.modeling_utils import PreTrainedModel
|
14 |
from transformers.activations import gelu_new
|
15 |
from transformers.modeling_outputs import (
|
16 |
-
Seq2SeqModelOutput, Seq2SeqLMOutput, BaseModelOutput
|
17 |
)
|
18 |
|
19 |
|
@@ -58,18 +56,37 @@ class Decoder(nn.Module):
|
|
58 |
layer.mlp.mlp[1].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
|
59 |
layer.mlp.mlp[-2].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
|
60 |
|
61 |
-
|
|
|
|
|
62 |
self_relative_embedding = self.self_relative_embedding()
|
63 |
cross_relative_embedding = self.cross_relative_embedding()
|
64 |
|
65 |
-
|
66 |
-
torch.
|
67 |
-
|
68 |
-
|
|
|
|
|
|
|
69 |
|
70 |
-
|
71 |
-
|
72 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
|
74 |
|
75 |
class MaskClassifier(nn.Module):
|
@@ -95,11 +112,11 @@ class MaskClassifier(nn.Module):
|
|
95 |
class EncoderLayer(nn.Module):
|
96 |
def __init__(self, config):
|
97 |
super().__init__()
|
98 |
-
self.attention = Attention(config)
|
99 |
self.mlp = FeedForward(config)
|
100 |
|
101 |
def forward(self, x, padding_mask, relative_embedding):
|
102 |
-
attention_output, attention_probs = self.attention(x, x, padding_mask, relative_embedding)
|
103 |
x = x + attention_output
|
104 |
x = x + self.mlp(x)
|
105 |
return x, attention_probs
|
@@ -108,15 +125,26 @@ class EncoderLayer(nn.Module):
|
|
108 |
class DecoderLayer(nn.Module):
|
109 |
def __init__(self, config):
|
110 |
super().__init__()
|
111 |
-
self.self_attention = Attention(config)
|
112 |
-
self.cross_attention = Attention(config)
|
113 |
self.mlp = FeedForward(config)
|
114 |
|
115 |
-
def forward(self, x, autoreg_mask, encoder_output, encoder_padding_mask, self_relative_embedding, cross_relative_embedding):
|
116 |
-
|
117 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
118 |
x = x + self.mlp(x)
|
119 |
-
|
|
|
120 |
|
121 |
|
122 |
class GeGLU(nn.Module):
|
@@ -152,24 +180,27 @@ class MaskedSoftmax(torch.autograd.Function):
|
|
152 |
@staticmethod
|
153 |
def forward(self, x, mask, dim):
|
154 |
self.dim = dim
|
155 |
-
|
|
|
156 |
x = torch.softmax(x, self.dim)
|
157 |
-
|
|
|
158 |
self.save_for_backward(x)
|
159 |
return x
|
160 |
|
161 |
@staticmethod
|
162 |
def backward(self, grad_output):
|
163 |
output, = self.saved_tensors
|
164 |
-
|
165 |
-
return
|
166 |
|
167 |
|
168 |
class Attention(nn.Module):
|
169 |
-
def __init__(self, config):
|
170 |
super().__init__()
|
171 |
|
172 |
self.config = config
|
|
|
173 |
|
174 |
if config.hidden_size % config.num_attention_heads != 0:
|
175 |
raise ValueError(f"The hidden size {config.hidden_size} is not a multiple of the number of attention heads {config.num_attention_heads}")
|
@@ -186,9 +217,9 @@ class Attention(nn.Module):
|
|
186 |
self.pre_layer_norm = nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=False)
|
187 |
self.post_layer_norm = nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=True)
|
188 |
|
189 |
-
position_indices = torch.arange(
|
190 |
-
- torch.arange(
|
191 |
-
position_indices = self.make_log_bucket_position(position_indices, config.position_bucket_size,
|
192 |
position_indices = config.position_bucket_size - 1 + position_indices
|
193 |
self.register_buffer("position_indices", position_indices, persistent=True)
|
194 |
|
@@ -215,59 +246,67 @@ class Attention(nn.Module):
|
|
215 |
self.in_proj_v.bias.data.zero_()
|
216 |
self.out_proj.bias.data.zero_()
|
217 |
|
218 |
-
def
|
219 |
key_len, batch_size, _ = kv.size()
|
220 |
query_len, _, _ = q.size()
|
221 |
|
222 |
-
if self.
|
223 |
-
|
224 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
225 |
position_indices = self.make_log_bucket_position(position_indices, self.config.position_bucket_size, 512)
|
226 |
position_indices = self.config.position_bucket_size - 1 + position_indices
|
227 |
self.register_buffer("position_indices", position_indices.to(q.device), persistent=True)
|
228 |
|
229 |
-
kv = self.pre_layer_norm(kv)
|
230 |
q = self.pre_layer_norm(q)
|
231 |
-
|
232 |
query = self.in_proj_q(q) # shape: [T, B, D]
|
233 |
-
key = self.in_proj_k(kv) # shape: [T, B, D]
|
234 |
-
value = self.in_proj_v(kv) # shape: [T, B, D]
|
235 |
-
|
236 |
-
query_pos = self.in_proj_q(self.dropout(relative_embedding)) # shape: [2T-1, 2D]
|
237 |
-
query_pos = F.embedding(self.position_indices[:query_len, :key_len], query_pos) # shape: [T, T, 2D]
|
238 |
-
query_pos = query_pos.view(query_len, key_len, self.num_heads, self.head_size)
|
239 |
-
|
240 |
-
key_pos = self.in_proj_k(self.dropout(relative_embedding)) # shape: [2T-1, 2D]
|
241 |
-
key_pos = F.embedding(self.position_indices[:query_len, :key_len], key_pos) # shape: [T, T, 2D]
|
242 |
-
key_pos = key_pos.view(query_len, key_len, self.num_heads, self.head_size)
|
243 |
-
|
244 |
query = query.reshape(query_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)
|
245 |
-
key = key.reshape(key_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)
|
246 |
-
value = value.view(key_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)
|
247 |
|
248 |
attention_scores = torch.bmm(query, key.transpose(1, 2) * self.scale)
|
249 |
-
|
250 |
-
|
251 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
252 |
attention_scores = attention_scores.view(batch_size, self.num_heads, query_len, key_len)
|
253 |
-
attention_scores.add_(
|
254 |
-
attention_scores.add_(
|
255 |
|
256 |
-
|
257 |
|
258 |
-
def compute_output(self, attention_probs, value):
|
259 |
attention_probs = self.dropout(attention_probs)
|
260 |
context = torch.bmm(attention_probs.flatten(0, 1), value) # shape: [B*H, Q, D]
|
261 |
context = context.transpose(0, 1).reshape(context.size(1), -1, self.hidden_size) # shape: [Q, B, H*D]
|
262 |
context = self.out_proj(context)
|
263 |
context = self.post_layer_norm(context)
|
264 |
context = self.dropout(context)
|
265 |
-
return context
|
266 |
|
267 |
-
|
268 |
-
attention_scores, value = self.compute_attention_scores(q, kv, relative_embedding)
|
269 |
-
attention_probs = MaskedSoftmax.apply(attention_scores, attention_mask, -1)
|
270 |
-
return self.compute_output(attention_probs, value), attention_probs.detach()
|
271 |
|
272 |
|
273 |
class WordEmbedding(nn.Module):
|
@@ -348,8 +387,8 @@ class NorT5Model(NorT5PreTrainedModel):
|
|
348 |
return self.get_encoder_output
|
349 |
|
350 |
def get_decoder(self):
|
351 |
-
return self.
|
352 |
-
|
353 |
def set_decoder_special_tokens(self, target_id):
|
354 |
target_id.masked_fill_(target_id == self.cls_token_id, self.bos_token_id)
|
355 |
target_id.masked_fill_(target_id == self.sep_token_id, self.eos_token_id)
|
@@ -359,12 +398,13 @@ class NorT5Model(NorT5PreTrainedModel):
|
|
359 |
shifted_input_ids = input_ids.new_zeros(input_ids.shape)
|
360 |
shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
|
361 |
shifted_input_ids[..., 0] = self.bos_token_id
|
|
|
362 |
|
363 |
return shifted_input_ids
|
364 |
|
365 |
def get_encoder_output(
|
366 |
self,
|
367 |
-
input_ids:
|
368 |
attention_mask: Optional[torch.Tensor] = None,
|
369 |
output_hidden_states: Optional[bool] = None,
|
370 |
output_attentions: Optional[bool] = None,
|
@@ -394,16 +434,28 @@ class NorT5Model(NorT5PreTrainedModel):
|
|
394 |
]
|
395 |
|
396 |
if not return_dict:
|
397 |
-
return
|
398 |
-
|
|
|
|
|
|
|
|
|
399 |
return BaseModelOutput(
|
400 |
last_hidden_state=last_layer,
|
401 |
-
hidden_states=contextualized_embeddings,
|
402 |
-
attentions=attention_probs
|
403 |
)
|
404 |
|
405 |
def get_decoder_output(
|
406 |
-
self,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
407 |
):
|
408 |
batch_size, seq_length, _ = encoder_output.shape
|
409 |
device = target_ids.device
|
@@ -414,11 +466,37 @@ class NorT5Model(NorT5PreTrainedModel):
|
|
414 |
attention_mask = ~attention_mask.bool()
|
415 |
attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
|
416 |
|
417 |
-
|
418 |
self.embedding(target_ids.t()),
|
419 |
encoder_output.transpose(0, 1),
|
420 |
-
attention_mask
|
421 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
422 |
|
423 |
def forward(
|
424 |
self,
|
@@ -426,28 +504,45 @@ class NorT5Model(NorT5PreTrainedModel):
|
|
426 |
attention_mask: Optional[torch.FloatTensor] = None,
|
427 |
decoder_input_ids: Optional[torch.LongTensor] = None,
|
428 |
decoder_attention_mask: Optional[torch.BoolTensor] = None,
|
429 |
-
|
|
|
|
|
|
|
|
|
|
|
430 |
):
|
431 |
|
432 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
433 |
|
434 |
decoder_input_ids = self.set_decoder_special_tokens(decoder_input_ids)
|
435 |
|
436 |
-
encoder_outputs
|
437 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
438 |
|
439 |
if not return_dict:
|
440 |
-
return
|
441 |
-
|
442 |
return Seq2SeqModelOutput(
|
443 |
-
last_hidden_state=decoder_outputs,
|
444 |
-
past_key_values=
|
445 |
-
decoder_hidden_states=
|
446 |
-
decoder_attentions=
|
447 |
-
cross_attentions=
|
448 |
-
encoder_last_hidden_state=encoder_outputs,
|
449 |
-
encoder_hidden_states=
|
450 |
-
encoder_attentions=
|
451 |
)
|
452 |
|
453 |
|
@@ -475,12 +570,19 @@ class NorT5ForConditionalGeneration(NorT5Model):
|
|
475 |
output_hidden_states: Optional[bool] = None,
|
476 |
return_dict: Optional[bool] = None,
|
477 |
):
|
478 |
-
|
479 |
-
use_cache = False
|
480 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
481 |
|
482 |
if encoder_outputs is None:
|
483 |
-
encoder_outputs = self.get_encoder_output(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
484 |
|
485 |
if labels is not None:
|
486 |
labels = self.set_decoder_special_tokens(labels)
|
@@ -490,24 +592,28 @@ class NorT5ForConditionalGeneration(NorT5Model):
|
|
490 |
elif decoder_input_ids is not None:
|
491 |
decoder_input_ids = self.set_decoder_special_tokens(decoder_input_ids)
|
492 |
|
493 |
-
decoder_outputs = self.get_decoder_output(
|
494 |
-
|
|
|
|
|
495 |
|
496 |
loss = None
|
497 |
if labels is not None:
|
498 |
-
|
|
|
499 |
loss = loss_fct(lm_logits.flatten(0, 1), labels.flatten())
|
500 |
|
501 |
if not return_dict:
|
502 |
-
output = (lm_logits,) + encoder_outputs
|
503 |
return ((loss,) + output) if loss is not None else output
|
504 |
|
505 |
return Seq2SeqLMOutput(
|
506 |
loss=loss,
|
507 |
logits=lm_logits,
|
508 |
-
|
509 |
-
|
510 |
-
|
|
|
511 |
encoder_last_hidden_state=encoder_outputs.last_hidden_state,
|
512 |
encoder_hidden_states=encoder_outputs.hidden_states,
|
513 |
encoder_attentions=encoder_outputs.attentions,
|
@@ -525,6 +631,9 @@ class NorT5ForConditionalGeneration(NorT5Model):
|
|
525 |
encoder_outputs=None,
|
526 |
**kwargs,
|
527 |
):
|
|
|
|
|
|
|
528 |
return {
|
529 |
"decoder_input_ids": input_ids,
|
530 |
"past_key_values": past_key_values,
|
@@ -553,9 +662,10 @@ class NorT5ForConditionalGeneration(NorT5Model):
|
|
553 |
reordered_layer_past_states = ()
|
554 |
for layer_past_state in layer_past_states:
|
555 |
# need to set correct `past` for each of the four key / value states
|
556 |
-
|
557 |
-
|
558 |
-
)
|
|
|
559 |
|
560 |
assert reordered_layer_past_states[0].shape == layer_past_states[0].shape
|
561 |
assert len(reordered_layer_past_states) == len(layer_past_states)
|
@@ -578,4 +688,6 @@ class NorT5Encoder(NorT5Model):
|
|
578 |
):
|
579 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
580 |
|
581 |
-
return self.get_encoder_output(
|
|
|
|
|
|
|
|
|
|
1 |
import math
|
2 |
from typing import List, Optional, Tuple, Union
|
3 |
|
4 |
import torch
|
5 |
import torch.nn as nn
|
6 |
import torch.nn.functional as F
|
7 |
+
from transformers.pytorch_utils import softmax_backward_data
|
8 |
from torch.utils import checkpoint
|
9 |
|
10 |
from configuration_nort5 import NorT5Config
|
11 |
from transformers.modeling_utils import PreTrainedModel
|
12 |
from transformers.activations import gelu_new
|
13 |
from transformers.modeling_outputs import (
|
14 |
+
Seq2SeqModelOutput, Seq2SeqLMOutput, BaseModelOutput, BaseModelOutputWithPastAndCrossAttentions
|
15 |
)
|
16 |
|
17 |
|
|
|
56 |
layer.mlp.mlp[1].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
|
57 |
layer.mlp.mlp[-2].weight.data *= math.sqrt(1.0 / (2.0 * (1 + i)))
|
58 |
|
59 |
+
self.activation_checkpointing = activation_checkpointing
|
60 |
+
|
61 |
+
def forward(self, x, encoder_output, encoder_padding_mask, past_key_values=None):
|
62 |
self_relative_embedding = self.self_relative_embedding()
|
63 |
cross_relative_embedding = self.cross_relative_embedding()
|
64 |
|
65 |
+
if past_key_values is not None:
|
66 |
+
autoreg_mask = torch.triu(
|
67 |
+
torch.full((x.size(0), x.size(0)), True, device=x.device),
|
68 |
+
diagonal=1
|
69 |
+
)
|
70 |
+
else:
|
71 |
+
autoreg_mask = None
|
72 |
|
73 |
+
# initialize past_key_values with `None` if past does not exist
|
74 |
+
if past_key_values is None:
|
75 |
+
past_key_values = [None] * len(self.layers)
|
76 |
+
|
77 |
+
hidden_states, self_attention_probs, cross_attention_probs, key_value_states = [x], [], [], []
|
78 |
+
for layer, past_key_value in zip(self.layers, past_key_values):
|
79 |
+
if self.activation_checkpointing:
|
80 |
+
hidden_state, self_attention_p, cross_attention_p, key_value_state = checkpoint.checkpoint(layer, hidden_states[-1], autoreg_mask, encoder_output, encoder_padding_mask, self_relative_embedding, cross_relative_embedding, past_key_value=None)
|
81 |
+
else:
|
82 |
+
hidden_state, self_attention_p, cross_attention_p, key_value_state = layer(hidden_states[-1], autoreg_mask, encoder_output, encoder_padding_mask, self_relative_embedding, cross_relative_embedding, past_key_value=past_key_value)
|
83 |
+
|
84 |
+
hidden_states.append(hidden_state)
|
85 |
+
self_attention_probs.append(self_attention_p)
|
86 |
+
cross_attention_probs.append(cross_attention_p)
|
87 |
+
key_value_states.append(key_value_state)
|
88 |
+
|
89 |
+
return hidden_states, self_attention_probs, cross_attention_probs, key_value_states
|
90 |
|
91 |
|
92 |
class MaskClassifier(nn.Module):
|
|
|
112 |
class EncoderLayer(nn.Module):
|
113 |
def __init__(self, config):
|
114 |
super().__init__()
|
115 |
+
self.attention = Attention(config, is_cross_attention=False)
|
116 |
self.mlp = FeedForward(config)
|
117 |
|
118 |
def forward(self, x, padding_mask, relative_embedding):
|
119 |
+
attention_output, attention_probs, _ = self.attention(x, x, padding_mask, relative_embedding)
|
120 |
x = x + attention_output
|
121 |
x = x + self.mlp(x)
|
122 |
return x, attention_probs
|
|
|
125 |
class DecoderLayer(nn.Module):
|
126 |
def __init__(self, config):
|
127 |
super().__init__()
|
128 |
+
self.self_attention = Attention(config, is_cross_attention=False)
|
129 |
+
self.cross_attention = Attention(config, is_cross_attention=True)
|
130 |
self.mlp = FeedForward(config)
|
131 |
|
132 |
+
def forward(self, x, autoreg_mask, encoder_output, encoder_padding_mask, self_relative_embedding, cross_relative_embedding, past_key_value=None):
|
133 |
+
query_offset = 0
|
134 |
+
if past_key_value is not None:
|
135 |
+
self_attn_past_key_value = past_key_value[:2]
|
136 |
+
cross_attn_past_key_value = past_key_value[2:]
|
137 |
+
query_offset = self_attn_past_key_value[0].size(1)
|
138 |
+
else:
|
139 |
+
self_attn_past_key_value, cross_attn_past_key_value = None, None
|
140 |
+
|
141 |
+
x_, self_attention_probs, self_key_value_state = self.self_attention(x, x, autoreg_mask, self_relative_embedding, past_key_value=self_attn_past_key_value, query_offset=query_offset)
|
142 |
+
x = x + x_
|
143 |
+
x_, cross_attention_probs, cross_key_value_state = self.cross_attention(x, encoder_output, encoder_padding_mask, cross_relative_embedding, past_key_value=cross_attn_past_key_value, query_offset=query_offset)
|
144 |
+
x = x + x_
|
145 |
x = x + self.mlp(x)
|
146 |
+
|
147 |
+
return x, self_attention_probs, cross_attention_probs, self_key_value_state + cross_key_value_state
|
148 |
|
149 |
|
150 |
class GeGLU(nn.Module):
|
|
|
180 |
@staticmethod
|
181 |
def forward(self, x, mask, dim):
|
182 |
self.dim = dim
|
183 |
+
if mask is not None:
|
184 |
+
x.masked_fill_(mask, float('-inf'))
|
185 |
x = torch.softmax(x, self.dim)
|
186 |
+
if mask is not None:
|
187 |
+
x.masked_fill_(mask, 0.0)
|
188 |
self.save_for_backward(x)
|
189 |
return x
|
190 |
|
191 |
@staticmethod
|
192 |
def backward(self, grad_output):
|
193 |
output, = self.saved_tensors
|
194 |
+
input_grad = softmax_backward_data(self, grad_output, output, self.dim, output)
|
195 |
+
return input_grad, None, None
|
196 |
|
197 |
|
198 |
class Attention(nn.Module):
|
199 |
+
def __init__(self, config, is_cross_attention=False):
|
200 |
super().__init__()
|
201 |
|
202 |
self.config = config
|
203 |
+
self.is_cross_attention = is_cross_attention
|
204 |
|
205 |
if config.hidden_size % config.num_attention_heads != 0:
|
206 |
raise ValueError(f"The hidden size {config.hidden_size} is not a multiple of the number of attention heads {config.num_attention_heads}")
|
|
|
217 |
self.pre_layer_norm = nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=False)
|
218 |
self.post_layer_norm = nn.LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=True)
|
219 |
|
220 |
+
position_indices = torch.arange(512, dtype=torch.long).unsqueeze(1) \
|
221 |
+
- torch.arange(512, dtype=torch.long).unsqueeze(0)
|
222 |
+
position_indices = self.make_log_bucket_position(position_indices, config.position_bucket_size, 512)
|
223 |
position_indices = config.position_bucket_size - 1 + position_indices
|
224 |
self.register_buffer("position_indices", position_indices, persistent=True)
|
225 |
|
|
|
246 |
self.in_proj_v.bias.data.zero_()
|
247 |
self.out_proj.bias.data.zero_()
|
248 |
|
249 |
+
def forward(self, q, kv, attention_mask, relative_embedding, past_key_value=None, query_offset=0):
|
250 |
key_len, batch_size, _ = kv.size()
|
251 |
query_len, _, _ = q.size()
|
252 |
|
253 |
+
if not self.is_cross_attention or past_key_value is None or past_key_value[0].size(1) != kv.size(0):
|
254 |
+
kv = self.pre_layer_norm(kv)
|
255 |
+
key = self.in_proj_k(kv) # shape: [T, B, D]
|
256 |
+
value = self.in_proj_v(kv) # shape: [T, B, D]
|
257 |
+
key = key.reshape(key_len, batch_size * self.num_heads, self.head_size).transpose(0, 1) # shape: [BxH, T, D]
|
258 |
+
value = value.view(key_len, batch_size * self.num_heads, self.head_size).transpose(0, 1) # shape: [BxH, T, D]
|
259 |
+
|
260 |
+
if past_key_value is not None:
|
261 |
+
if not self.is_cross_attention:
|
262 |
+
key = torch.cat([past_key_value[0], key], dim=1)
|
263 |
+
value = torch.cat([past_key_value[1], value], dim=1)
|
264 |
+
key_len = key.size(1)
|
265 |
+
elif past_key_value[0].size(1) == kv.size(0):
|
266 |
+
key = past_key_value[0]
|
267 |
+
value = past_key_value[1]
|
268 |
+
|
269 |
+
if self.position_indices.size(0) < max(query_len, key_len):
|
270 |
+
position_indices = torch.arange(max(query_len, key_len), dtype=torch.long).unsqueeze(1) \
|
271 |
+
- torch.arange(max(query_len, key_len), dtype=torch.long).unsqueeze(0)
|
272 |
position_indices = self.make_log_bucket_position(position_indices, self.config.position_bucket_size, 512)
|
273 |
position_indices = self.config.position_bucket_size - 1 + position_indices
|
274 |
self.register_buffer("position_indices", position_indices.to(q.device), persistent=True)
|
275 |
|
|
|
276 |
q = self.pre_layer_norm(q)
|
|
|
277 |
query = self.in_proj_q(q) # shape: [T, B, D]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
278 |
query = query.reshape(query_len, batch_size * self.num_heads, self.head_size).transpose(0, 1)
|
|
|
|
|
279 |
|
280 |
attention_scores = torch.bmm(query, key.transpose(1, 2) * self.scale)
|
281 |
+
|
282 |
+
query_pos = self.in_proj_q(self.dropout(relative_embedding)) # shape: [2T-1, D]
|
283 |
+
query_pos = query_pos.view(-1, self.num_heads, self.head_size) # shape: [2T-1, H, D]
|
284 |
+
key_pos = self.in_proj_k(self.dropout(relative_embedding)) # shape: [2T-1, D]
|
285 |
+
key_pos = key_pos.view(-1, self.num_heads, self.head_size) # shape: [2T-1, H, D]
|
286 |
+
|
287 |
+
query_ = query.view(batch_size, self.num_heads, query_len, self.head_size)
|
288 |
+
key_ = key.view(batch_size, self.num_heads, key_len, self.head_size)
|
289 |
+
|
290 |
+
attention_c_p = torch.einsum("bhqd,khd->bhqk", query_, key_pos.squeeze(1) * self.scale)
|
291 |
+
attention_p_c = torch.einsum("bhkd,qhd->bhqk", key_ * self.scale, query_pos.squeeze(1))
|
292 |
+
position_indices = self.position_indices[query_offset:query_offset+query_len, :key_len].expand(batch_size, self.num_heads, -1, -1)
|
293 |
+
attention_c_p = attention_c_p.gather(3, position_indices)
|
294 |
+
attention_p_c = attention_p_c.gather(2, position_indices)
|
295 |
+
|
296 |
attention_scores = attention_scores.view(batch_size, self.num_heads, query_len, key_len)
|
297 |
+
attention_scores.add_(attention_c_p)
|
298 |
+
attention_scores.add_(attention_p_c)
|
299 |
|
300 |
+
attention_probs = MaskedSoftmax.apply(attention_scores, attention_mask, -1)
|
301 |
|
|
|
302 |
attention_probs = self.dropout(attention_probs)
|
303 |
context = torch.bmm(attention_probs.flatten(0, 1), value) # shape: [B*H, Q, D]
|
304 |
context = context.transpose(0, 1).reshape(context.size(1), -1, self.hidden_size) # shape: [Q, B, H*D]
|
305 |
context = self.out_proj(context)
|
306 |
context = self.post_layer_norm(context)
|
307 |
context = self.dropout(context)
|
|
|
308 |
|
309 |
+
return context, attention_probs.detach(), (key.detach(), value.detach())
|
|
|
|
|
|
|
310 |
|
311 |
|
312 |
class WordEmbedding(nn.Module):
|
|
|
387 |
return self.get_encoder_output
|
388 |
|
389 |
def get_decoder(self):
|
390 |
+
return self.get_decoder_output
|
391 |
+
|
392 |
def set_decoder_special_tokens(self, target_id):
|
393 |
target_id.masked_fill_(target_id == self.cls_token_id, self.bos_token_id)
|
394 |
target_id.masked_fill_(target_id == self.sep_token_id, self.eos_token_id)
|
|
|
398 |
shifted_input_ids = input_ids.new_zeros(input_ids.shape)
|
399 |
shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
|
400 |
shifted_input_ids[..., 0] = self.bos_token_id
|
401 |
+
shifted_input_ids.masked_fill_(shifted_input_ids == -100, self.pad_token_id)
|
402 |
|
403 |
return shifted_input_ids
|
404 |
|
405 |
def get_encoder_output(
|
406 |
self,
|
407 |
+
input_ids: torch.Tensor = None,
|
408 |
attention_mask: Optional[torch.Tensor] = None,
|
409 |
output_hidden_states: Optional[bool] = None,
|
410 |
output_attentions: Optional[bool] = None,
|
|
|
434 |
]
|
435 |
|
436 |
if not return_dict:
|
437 |
+
return (
|
438 |
+
last_layer,
|
439 |
+
*([contextualized_embeddings] if output_hidden_states else []),
|
440 |
+
*([attention_probs] if output_attentions else [])
|
441 |
+
)
|
442 |
+
|
443 |
return BaseModelOutput(
|
444 |
last_hidden_state=last_layer,
|
445 |
+
hidden_states=contextualized_embeddings if output_hidden_states else None,
|
446 |
+
attentions=attention_probs if output_attentions else None
|
447 |
)
|
448 |
|
449 |
def get_decoder_output(
|
450 |
+
self,
|
451 |
+
target_ids: torch.Tensor = None,
|
452 |
+
encoder_output: torch.Tensor = None,
|
453 |
+
attention_mask: Optional[torch.Tensor] = None,
|
454 |
+
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
455 |
+
use_cache: Optional[bool] = None,
|
456 |
+
output_hidden_states: Optional[bool] = None,
|
457 |
+
output_attentions: Optional[bool] = None,
|
458 |
+
return_dict = False
|
459 |
):
|
460 |
batch_size, seq_length, _ = encoder_output.shape
|
461 |
device = target_ids.device
|
|
|
466 |
attention_mask = ~attention_mask.bool()
|
467 |
attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
|
468 |
|
469 |
+
hidden_states, self_attention_p, cross_attention_p, key_value_states = self.decoder(
|
470 |
self.embedding(target_ids.t()),
|
471 |
encoder_output.transpose(0, 1),
|
472 |
+
attention_mask,
|
473 |
+
past_key_values
|
474 |
+
)
|
475 |
+
|
476 |
+
hidden_states = [e.transpose(0, 1) for e in hidden_states]
|
477 |
+
last_layer = hidden_states[-1]
|
478 |
+
hidden_states = [hidden_states[0]] + [
|
479 |
+
hidden_states[i] - hidden_states[i - 1]
|
480 |
+
for i in range(1, len(hidden_states))
|
481 |
+
]
|
482 |
+
|
483 |
+
if not return_dict:
|
484 |
+
return (
|
485 |
+
last_layer,
|
486 |
+
*([key_value_states] if use_cache else []),
|
487 |
+
*([hidden_states] if output_hidden_states else []),
|
488 |
+
*([self_attention_p] if output_attentions else []),
|
489 |
+
*([cross_attention_p] if output_attentions else []),
|
490 |
+
)
|
491 |
+
|
492 |
+
return BaseModelOutputWithPastAndCrossAttentions(
|
493 |
+
last_hidden_state=last_layer,
|
494 |
+
past_key_values=key_value_states if use_cache else None,
|
495 |
+
hidden_states=hidden_states if output_hidden_states else None,
|
496 |
+
attentions=self_attention_p if output_attentions else None,
|
497 |
+
cross_attentions=cross_attention_p if output_attentions else None
|
498 |
+
)
|
499 |
+
|
500 |
|
501 |
def forward(
|
502 |
self,
|
|
|
504 |
attention_mask: Optional[torch.FloatTensor] = None,
|
505 |
decoder_input_ids: Optional[torch.LongTensor] = None,
|
506 |
decoder_attention_mask: Optional[torch.BoolTensor] = None,
|
507 |
+
encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
508 |
+
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
509 |
+
use_cache: Optional[bool] = None,
|
510 |
+
output_attentions: Optional[bool] = None,
|
511 |
+
output_hidden_states: Optional[bool] = None,
|
512 |
+
return_dict: Optional[bool] = None
|
513 |
):
|
514 |
|
515 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
516 |
|
517 |
decoder_input_ids = self.set_decoder_special_tokens(decoder_input_ids)
|
518 |
|
519 |
+
if encoder_outputs is None:
|
520 |
+
encoder_outputs = self.get_encoder_output(
|
521 |
+
input_ids, attention_mask, output_hidden_states, output_attentions, return_dict
|
522 |
+
)
|
523 |
+
elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
|
524 |
+
encoder_outputs = BaseModelOutput(
|
525 |
+
last_hidden_state=encoder_outputs[0],
|
526 |
+
hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
|
527 |
+
attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
|
528 |
+
)
|
529 |
+
|
530 |
+
decoder_outputs = self.get_decoder_output(
|
531 |
+
decoder_input_ids, encoder_outputs[0], attention_mask, past_key_values, use_cache, output_hidden_states, output_attentions, return_dict
|
532 |
+
)
|
533 |
|
534 |
if not return_dict:
|
535 |
+
return decoder_outputs + encoder_outputs
|
536 |
+
|
537 |
return Seq2SeqModelOutput(
|
538 |
+
last_hidden_state=decoder_outputs.last_hidden_state,
|
539 |
+
past_key_values=decoder_outputs.past_key_values,
|
540 |
+
decoder_hidden_states=decoder_outputs.hidden_states,
|
541 |
+
decoder_attentions=decoder_outputs.attentions,
|
542 |
+
cross_attentions=decoder_outputs.cross_attentions,
|
543 |
+
encoder_last_hidden_state=encoder_outputs.last_hidden_state,
|
544 |
+
encoder_hidden_states=encoder_outputs.hidden_states,
|
545 |
+
encoder_attentions=encoder_outputs.attentions,
|
546 |
)
|
547 |
|
548 |
|
|
|
570 |
output_hidden_states: Optional[bool] = None,
|
571 |
return_dict: Optional[bool] = None,
|
572 |
):
|
573 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
|
|
574 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
575 |
|
576 |
if encoder_outputs is None:
|
577 |
+
encoder_outputs = self.get_encoder_output(
|
578 |
+
input_ids, attention_mask, output_hidden_states, output_attentions, return_dict
|
579 |
+
)
|
580 |
+
elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
|
581 |
+
encoder_outputs = BaseModelOutput(
|
582 |
+
last_hidden_state=encoder_outputs[0],
|
583 |
+
hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
|
584 |
+
attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
|
585 |
+
)
|
586 |
|
587 |
if labels is not None:
|
588 |
labels = self.set_decoder_special_tokens(labels)
|
|
|
592 |
elif decoder_input_ids is not None:
|
593 |
decoder_input_ids = self.set_decoder_special_tokens(decoder_input_ids)
|
594 |
|
595 |
+
decoder_outputs = self.get_decoder_output(
|
596 |
+
decoder_input_ids, encoder_outputs[0], attention_mask, past_key_values, use_cache, output_hidden_states, output_attentions, return_dict
|
597 |
+
)
|
598 |
+
lm_logits = self.classifier(decoder_outputs[0])
|
599 |
|
600 |
loss = None
|
601 |
if labels is not None:
|
602 |
+
labels.masked_fill_(labels == self.pad_token_id, -100)
|
603 |
+
loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
|
604 |
loss = loss_fct(lm_logits.flatten(0, 1), labels.flatten())
|
605 |
|
606 |
if not return_dict:
|
607 |
+
output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs
|
608 |
return ((loss,) + output) if loss is not None else output
|
609 |
|
610 |
return Seq2SeqLMOutput(
|
611 |
loss=loss,
|
612 |
logits=lm_logits,
|
613 |
+
past_key_values=decoder_outputs.past_key_values,
|
614 |
+
decoder_hidden_states=decoder_outputs.hidden_states,
|
615 |
+
decoder_attentions=decoder_outputs.attentions,
|
616 |
+
cross_attentions=decoder_outputs.cross_attentions,
|
617 |
encoder_last_hidden_state=encoder_outputs.last_hidden_state,
|
618 |
encoder_hidden_states=encoder_outputs.hidden_states,
|
619 |
encoder_attentions=encoder_outputs.attentions,
|
|
|
631 |
encoder_outputs=None,
|
632 |
**kwargs,
|
633 |
):
|
634 |
+
if past_key_values is not None:
|
635 |
+
input_ids = input_ids[:, -1:]
|
636 |
+
|
637 |
return {
|
638 |
"decoder_input_ids": input_ids,
|
639 |
"past_key_values": past_key_values,
|
|
|
662 |
reordered_layer_past_states = ()
|
663 |
for layer_past_state in layer_past_states:
|
664 |
# need to set correct `past` for each of the four key / value states
|
665 |
+
layer_past_state = layer_past_state.unflatten(0, (-1, self.config.num_attention_heads))
|
666 |
+
layer_past_state = layer_past_state.index_select(0, beam_idx.to(layer_past_state.device))
|
667 |
+
layer_past_state = layer_past_state.flatten(0, 1)
|
668 |
+
reordered_layer_past_states = reordered_layer_past_states + (layer_past_state,)
|
669 |
|
670 |
assert reordered_layer_past_states[0].shape == layer_past_states[0].shape
|
671 |
assert len(reordered_layer_past_states) == len(layer_past_states)
|
|
|
688 |
):
|
689 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
690 |
|
691 |
+
return self.get_encoder_output(
|
692 |
+
input_ids, attention_mask, output_hidden_states, output_attentions, return_dict=return_dict
|
693 |
+
)
|