neo-nlp-dev
commited on
Commit
•
06859d4
1
Parent(s):
c2502f1
updating lola modeling class with auxiliary loss
Browse files- modeling_lola_gpt2.py +163 -217
modeling_lola_gpt2.py
CHANGED
@@ -7,6 +7,11 @@
|
|
7 |
import warnings
|
8 |
from typing import Optional, Tuple, Union
|
9 |
|
|
|
|
|
|
|
|
|
|
|
10 |
from .configuration_lola_gpt2 import LOLAConfig
|
11 |
import torch
|
12 |
import torch.utils.checkpoint
|
@@ -16,6 +21,7 @@ from torch.nn import CrossEntropyLoss
|
|
16 |
|
17 |
from transformers.modeling_outputs import (
|
18 |
BaseModelOutputWithPastAndCrossAttentions,
|
|
|
19 |
SequenceClassifierOutputWithPast,
|
20 |
QuestionAnsweringModelOutput
|
21 |
)
|
@@ -27,11 +33,68 @@ from transformers.utils import (
|
|
27 |
from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
|
28 |
|
29 |
from transformers.models.gpt2.modeling_gpt2 import GPT2Attention, GPT2MLP, GPT2Block, GPT2PreTrainedModel
|
30 |
-
from transformers.models.gpt2.modeling_gpt2 import GPT2LMHeadModel
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
|
32 |
|
33 |
logger = logging.get_logger(__name__)
|
34 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
# LOLA
|
36 |
class LOLAModel(GPT2PreTrainedModel):
|
37 |
|
@@ -39,7 +102,9 @@ class LOLAModel(GPT2PreTrainedModel):
|
|
39 |
|
40 |
def __init__(self, config):
|
41 |
super().__init__(config)
|
42 |
-
|
|
|
|
|
43 |
self.embed_dim = config.hidden_size
|
44 |
|
45 |
self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
|
@@ -63,9 +128,9 @@ class LOLAModel(GPT2PreTrainedModel):
|
|
63 |
def parallelize(self, device_map=None):
|
64 |
# Check validity of device_map
|
65 |
warnings.warn(
|
66 |
-
"
|
67 |
-
" model with
|
68 |
-
"
|
69 |
" ...}",
|
70 |
FutureWarning,
|
71 |
)
|
@@ -89,7 +154,7 @@ class LOLAModel(GPT2PreTrainedModel):
|
|
89 |
|
90 |
def deparallelize(self):
|
91 |
warnings.warn(
|
92 |
-
"Like
|
93 |
FutureWarning,
|
94 |
)
|
95 |
self.model_parallel = False
|
@@ -219,7 +284,7 @@ class LOLAModel(GPT2PreTrainedModel):
|
|
219 |
if self.gradient_checkpointing and self.training:
|
220 |
if use_cache:
|
221 |
logger.warning_once(
|
222 |
-
"
|
223 |
)
|
224 |
use_cache = False
|
225 |
|
@@ -227,6 +292,7 @@ class LOLAModel(GPT2PreTrainedModel):
|
|
227 |
all_self_attentions = () if output_attentions else None
|
228 |
all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
|
229 |
all_hidden_states = () if output_hidden_states else None
|
|
|
230 |
for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
|
231 |
# Model parallel
|
232 |
if self.model_parallel:
|
@@ -269,11 +335,14 @@ class LOLAModel(GPT2PreTrainedModel):
|
|
269 |
hidden_states = outputs[0]
|
270 |
if use_cache is True:
|
271 |
presents = presents + (outputs[1],)
|
|
|
|
|
|
|
|
|
|
|
272 |
|
273 |
if output_attentions:
|
274 |
all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
|
275 |
-
if self.config.add_cross_attention:
|
276 |
-
all_cross_attentions = all_cross_attentions + (outputs[3 if use_cache else 2],)
|
277 |
|
278 |
# Model Parallel: If it's the last layer for that device, put things on the next device
|
279 |
if self.model_parallel:
|
@@ -284,23 +353,27 @@ class LOLAModel(GPT2PreTrainedModel):
|
|
284 |
hidden_states = self.ln_f(hidden_states)
|
285 |
|
286 |
hidden_states = hidden_states.view(output_shape)
|
|
|
|
|
|
|
|
|
|
|
287 |
# Add last hidden state
|
288 |
if output_hidden_states:
|
289 |
all_hidden_states = all_hidden_states + (hidden_states,)
|
290 |
-
|
291 |
if not return_dict:
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
|
296 |
-
)
|
297 |
|
298 |
-
return
|
299 |
last_hidden_state=hidden_states,
|
300 |
past_key_values=presents,
|
301 |
hidden_states=all_hidden_states,
|
302 |
attentions=all_self_attentions,
|
303 |
-
|
|
|
304 |
)
|
305 |
|
306 |
class LOLABlock(nn.Module):
|
@@ -312,7 +385,6 @@ class LOLABlock(nn.Module):
|
|
312 |
self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
|
313 |
self.attn = GPT2Attention(config, layer_idx=layer_idx)
|
314 |
self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
|
315 |
-
|
316 |
self.moe = LOLAMOE(
|
317 |
hidden_size,
|
318 |
inner_dim,
|
@@ -336,7 +408,7 @@ class LOLABlock(nn.Module):
|
|
336 |
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
337 |
use_cache: Optional[bool] = False,
|
338 |
output_attentions: Optional[bool] = False,
|
339 |
-
) ->
|
340 |
residual = hidden_states
|
341 |
hidden_states = self.ln_1(hidden_states)
|
342 |
attn_outputs = self.attn(
|
@@ -347,45 +419,21 @@ class LOLABlock(nn.Module):
|
|
347 |
use_cache=use_cache,
|
348 |
output_attentions=output_attentions,
|
349 |
)
|
350 |
-
attn_output = attn_outputs[0]
|
351 |
outputs = attn_outputs[1:]
|
352 |
-
# residual connection
|
353 |
hidden_states = attn_output + residual
|
354 |
|
355 |
-
if encoder_hidden_states is not None:
|
356 |
-
# add one self-attention block for cross-attention
|
357 |
-
if not hasattr(self, "crossattention"):
|
358 |
-
raise ValueError(
|
359 |
-
f"If `encoder_hidden_states` are passed, {self} has to be instantiated with "
|
360 |
-
"cross-attention layers by setting `config.add_cross_attention=True`"
|
361 |
-
)
|
362 |
-
residual = hidden_states
|
363 |
-
hidden_states = self.ln_cross_attn(hidden_states)
|
364 |
-
cross_attn_outputs = self.crossattention(
|
365 |
-
hidden_states,
|
366 |
-
attention_mask=attention_mask,
|
367 |
-
head_mask=head_mask,
|
368 |
-
encoder_hidden_states=encoder_hidden_states,
|
369 |
-
encoder_attention_mask=encoder_attention_mask,
|
370 |
-
output_attentions=output_attentions,
|
371 |
-
)
|
372 |
-
attn_output = cross_attn_outputs[0]
|
373 |
-
# residual connection
|
374 |
-
hidden_states = residual + attn_output
|
375 |
-
outputs = outputs + cross_attn_outputs[2:] # add cross attentions if we output attention weights
|
376 |
-
|
377 |
residual = hidden_states
|
378 |
hidden_states = self.ln_2(hidden_states)
|
379 |
-
feed_forward_hidden_states,
|
380 |
-
# residual connection
|
381 |
hidden_states = residual + feed_forward_hidden_states
|
382 |
|
383 |
if use_cache:
|
384 |
-
outputs = (hidden_states,) + outputs
|
385 |
else:
|
386 |
-
outputs = (hidden_states,) + outputs
|
387 |
|
388 |
-
return outputs # hidden_states, present, (attentions,
|
389 |
|
390 |
class LOLAMOE(nn.Module):
|
391 |
def __init__(self,
|
@@ -404,50 +452,41 @@ class LOLAMOE(nn.Module):
|
|
404 |
self.experts = nn.ModuleList([GPT2MLP(inner_dim, config) for _ in range(self.num_experts)])
|
405 |
|
406 |
def forward(self, hidden_states):
|
407 |
-
# https://github.com/huggingface/transformers/blob/main/src/transformers/models/mixtral/modeling_mixtral.py#L816
|
408 |
-
# FIXME do it as in top1gating
|
409 |
-
# https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/moe/sharded_moe.py
|
410 |
-
|
411 |
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
412 |
hidden_states = hidden_states.view(-1, hidden_dim)
|
413 |
|
414 |
router_logits = self.gate(hidden_states)
|
415 |
-
|
|
|
|
|
|
|
|
|
416 |
|
417 |
-
#
|
418 |
-
|
419 |
-
|
420 |
-
# routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
|
421 |
-
# commenting the statement above for LOLA and removing the "/" operator to avoid getting weights as 1
|
422 |
-
routing_weights = routing_weights.sum(dim=-1, keepdim=True)
|
423 |
-
routing_weights = routing_weights.to(hidden_states.dtype)
|
424 |
|
|
|
|
|
|
|
|
|
425 |
final_hidden_states = torch.zeros(
|
426 |
(batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
|
427 |
)
|
428 |
-
expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
|
429 |
-
for expert_idx in range(self.num_experts):
|
430 |
-
expert_layer = self.experts[expert_idx]
|
431 |
-
idx, top_x = torch.where(expert_mask[expert_idx])
|
432 |
|
433 |
-
|
|
|
|
|
|
|
434 |
continue
|
|
|
|
|
|
|
|
|
435 |
|
436 |
-
|
437 |
-
|
438 |
-
|
439 |
-
|
440 |
-
# Index the correct hidden states and compute the expert hidden state for
|
441 |
-
# the current expert. We need to make sure to multiply the output hidden
|
442 |
-
# states by `routing_weights` on the corresponding tokens (top-1 and top-2)
|
443 |
-
current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)
|
444 |
-
current_hidden_states = expert_layer(current_state) * routing_weights[top_x_list, idx_list, None]
|
445 |
-
|
446 |
-
# However `index_add_` only support torch tensors for indexing so we'll use
|
447 |
-
# the `top_x` tensor here.
|
448 |
-
final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
|
449 |
-
final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
|
450 |
-
return final_hidden_states, router_logits
|
451 |
|
452 |
class LOLAAttention(GPT2Attention):
|
453 |
def __init__(self, config, is_cross_attention=False, layer_idx=None):
|
@@ -471,7 +510,7 @@ class LOLAAttention(GPT2Attention):
|
|
471 |
self.split_size = self.embed_dim
|
472 |
if self.head_dim * self.num_heads != self.embed_dim:
|
473 |
raise ValueError(
|
474 |
-
f"
|
475 |
f" {self.num_heads})."
|
476 |
)
|
477 |
|
@@ -512,156 +551,63 @@ class LOLALMHeadModel(GPT2LMHeadModel):
|
|
512 |
|
513 |
# Initialize weights and apply final processing
|
514 |
self.post_init()
|
515 |
-
|
516 |
-
|
517 |
-
class LOLADoubleHeadsModel(GPT2DoubleHeadsModel):
|
518 |
|
519 |
-
config_class = LOLAConfig
|
520 |
-
|
521 |
-
def __init__(self, config):
|
522 |
-
super(GPT2DoubleHeadsModel, self).__init__(config)
|
523 |
-
config.num_labels = 1
|
524 |
-
self.transformer = LOLAModel(config)
|
525 |
-
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
|
526 |
-
self.multiple_choice_head = SequenceSummary(config)
|
527 |
-
|
528 |
-
# Model parallel
|
529 |
-
self.model_parallel = False
|
530 |
-
self.device_map = None
|
531 |
-
|
532 |
-
# Initialize weights and apply final processing
|
533 |
-
self.post_init()
|
534 |
-
|
535 |
-
|
536 |
-
class LOLAForSequenceClassification(GPT2ForSequenceClassification):
|
537 |
-
|
538 |
-
config_class = LOLAConfig
|
539 |
-
|
540 |
-
def __init__(self, config):
|
541 |
-
super(GPT2ForSequenceClassification, self).__init__(config)
|
542 |
-
self.num_labels = config.num_labels
|
543 |
-
self.transformer = LOLAModel(config)
|
544 |
-
self.score = nn.Linear(config.n_embd, self.num_labels, bias=False)
|
545 |
-
|
546 |
-
# Model parallel
|
547 |
-
self.model_parallel = False
|
548 |
-
self.device_map = None
|
549 |
-
|
550 |
-
# Initialize weights and apply final processing
|
551 |
-
self.post_init()
|
552 |
-
|
553 |
-
class LOLAForTokenClassification(GPT2ForTokenClassification):
|
554 |
-
|
555 |
-
config_class = LOLAConfig
|
556 |
-
|
557 |
-
def __init__(self, config):
|
558 |
-
super(GPT2ForTokenClassification, self).__init__(config)
|
559 |
-
self.num_labels = config.num_labels
|
560 |
-
|
561 |
-
self.transformer = LOLAModel(config)
|
562 |
-
if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
|
563 |
-
classifier_dropout = config.classifier_dropout
|
564 |
-
elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
|
565 |
-
classifier_dropout = config.hidden_dropout
|
566 |
-
else:
|
567 |
-
classifier_dropout = 0.1
|
568 |
-
self.dropout = nn.Dropout(classifier_dropout)
|
569 |
-
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
|
570 |
-
|
571 |
-
# Model parallel
|
572 |
-
self.model_parallel = False
|
573 |
-
self.device_map = None
|
574 |
-
|
575 |
-
# Initialize weights and apply final processing
|
576 |
-
self.post_init()
|
577 |
-
|
578 |
-
class LOLAForQuestionAnswering(GPT2PreTrainedModel):
|
579 |
-
|
580 |
-
config_class = LOLAConfig
|
581 |
-
|
582 |
-
def __init__(self, config):
|
583 |
-
super().__init__(config)
|
584 |
-
self.num_labels = config.num_labels
|
585 |
-
self.transformer = LOLAModel(config)
|
586 |
-
self.qa_outputs = nn.Linear(config.hidden_size, 2)
|
587 |
-
|
588 |
-
# Model parallel
|
589 |
-
self.model_parallel = False
|
590 |
-
self.device_map = None
|
591 |
-
|
592 |
-
# Initialize weights and apply final processing
|
593 |
-
self.post_init()
|
594 |
-
|
595 |
def forward(
|
596 |
self,
|
597 |
-
input_ids
|
598 |
-
|
599 |
-
|
600 |
-
|
601 |
-
|
602 |
-
|
603 |
-
|
604 |
-
|
605 |
-
|
606 |
-
|
607 |
-
|
608 |
-
|
609 |
-
|
610 |
-
start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
611 |
-
Labels for position (index) of the start of the labelled span for computing the token classification loss.
|
612 |
-
Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
|
613 |
-
are not taken into account for computing the loss.
|
614 |
-
end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
615 |
-
Labels for position (index) of the end of the labelled span for computing the token classification loss.
|
616 |
-
Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
|
617 |
-
are not taken into account for computing the loss.
|
618 |
-
"""
|
619 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
620 |
|
621 |
-
|
622 |
input_ids,
|
|
|
623 |
attention_mask=attention_mask,
|
624 |
token_type_ids=token_type_ids,
|
625 |
position_ids=position_ids,
|
626 |
head_mask=head_mask,
|
627 |
inputs_embeds=inputs_embeds,
|
|
|
628 |
output_attentions=output_attentions,
|
629 |
output_hidden_states=output_hidden_states,
|
630 |
-
return_dict=
|
631 |
)
|
632 |
-
|
633 |
-
|
634 |
-
|
635 |
-
|
636 |
-
|
637 |
-
|
638 |
-
|
639 |
-
|
640 |
-
|
641 |
-
|
642 |
-
#
|
643 |
-
|
644 |
-
|
645 |
-
if
|
646 |
-
|
647 |
-
# sometimes the start/end positions are outside our model inputs, we ignore these terms
|
648 |
-
ignored_index = start_logits.size(1)
|
649 |
-
start_positions = start_positions.clamp(0, ignored_index)
|
650 |
-
end_positions = end_positions.clamp(0, ignored_index)
|
651 |
-
|
652 |
-
loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
|
653 |
-
start_loss = loss_fct(start_logits, start_positions)
|
654 |
-
end_loss = loss_fct(end_logits, end_positions)
|
655 |
-
total_loss = (start_loss + end_loss) / 2
|
656 |
|
657 |
if not return_dict:
|
658 |
-
output = (
|
659 |
-
return ((
|
660 |
-
|
661 |
-
return
|
662 |
-
loss=
|
663 |
-
|
664 |
-
|
665 |
-
|
666 |
-
|
|
|
|
|
667 |
)
|
|
|
7 |
import warnings
|
8 |
from typing import Optional, Tuple, Union
|
9 |
|
10 |
+
## Uncomment the below three and comment the other import for model conversion
|
11 |
+
#import sys
|
12 |
+
# sys.path.append(".")
|
13 |
+
# from configuration_lola_gpt2 import LOLAConfig
|
14 |
+
|
15 |
from .configuration_lola_gpt2 import LOLAConfig
|
16 |
import torch
|
17 |
import torch.utils.checkpoint
|
|
|
21 |
|
22 |
from transformers.modeling_outputs import (
|
23 |
BaseModelOutputWithPastAndCrossAttentions,
|
24 |
+
MoeCausalLMOutputWithPast,
|
25 |
SequenceClassifierOutputWithPast,
|
26 |
QuestionAnsweringModelOutput
|
27 |
)
|
|
|
33 |
from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
|
34 |
|
35 |
from transformers.models.gpt2.modeling_gpt2 import GPT2Attention, GPT2MLP, GPT2Block, GPT2PreTrainedModel
|
36 |
+
from transformers.models.gpt2.modeling_gpt2 import GPT2LMHeadModel
|
37 |
+
from dataclasses import dataclass
|
38 |
+
from typing import Optional, Tuple
|
39 |
+
|
40 |
+
import torch
|
41 |
+
from transformers.modeling_outputs import ModelOutput
|
42 |
+
import transformers
|
43 |
+
import importlib.util
|
44 |
|
45 |
|
46 |
logger = logging.get_logger(__name__)
|
47 |
|
48 |
+
expert_analysis_callback = lambda _: None
|
49 |
+
|
50 |
+
class LOLADependencyChecker:
|
51 |
+
def __init__(self):
|
52 |
+
self.expected_versions = {
|
53 |
+
"transformers": "4.38.2"
|
54 |
+
}
|
55 |
+
self.check_dependencies()
|
56 |
+
|
57 |
+
def check_dependencies(self):
|
58 |
+
# Check transformers version
|
59 |
+
self._check_version("transformers", transformers.__version__)
|
60 |
+
|
61 |
+
def _check_version(self, package_name, installed_version):
|
62 |
+
expected_version = self.expected_versions.get(package_name)
|
63 |
+
if installed_version != expected_version:
|
64 |
+
warnings.warn(
|
65 |
+
f"Warning: The installed {package_name} version ({installed_version}) "
|
66 |
+
f"differs from the expected version ({expected_version}). "
|
67 |
+
"This may lead to unexpected behavior.",
|
68 |
+
category=UserWarning
|
69 |
+
)
|
70 |
+
|
71 |
+
@dataclass
|
72 |
+
class MoeModelOutputWithPast(ModelOutput):
|
73 |
+
"""
|
74 |
+
Base class for model's outputs with potential hidden states and attentions, and includes auxiliary loss.
|
75 |
+
|
76 |
+
Args:
|
77 |
+
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
78 |
+
Sequence of hidden-states at the output of the last layer of the model.
|
79 |
+
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
80 |
+
Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
|
81 |
+
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed):
|
82 |
+
Hidden-states of the model at the output of each layer plus the initial embedding outputs.
|
83 |
+
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed):
|
84 |
+
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
|
85 |
+
router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_logits=True` is passed):
|
86 |
+
Router logits computed by MoE routers, used to compute the auxiliary loss for Mixture of Experts models.
|
87 |
+
aux_loss (`torch.FloatTensor`, *optional*):
|
88 |
+
The auxiliary loss computed from the MoE layers, used to encourage balanced expert utilization.
|
89 |
+
"""
|
90 |
+
|
91 |
+
last_hidden_state: torch.FloatTensor = None
|
92 |
+
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor, ...]]] = None
|
93 |
+
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
|
94 |
+
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
|
95 |
+
router_logits: Optional[Tuple[torch.FloatTensor, ...]] = None
|
96 |
+
aux_loss: Optional[torch.FloatTensor] = None
|
97 |
+
|
98 |
# LOLA
|
99 |
class LOLAModel(GPT2PreTrainedModel):
|
100 |
|
|
|
102 |
|
103 |
def __init__(self, config):
|
104 |
super().__init__(config)
|
105 |
+
# Checking dependencies version
|
106 |
+
LOLADependencyChecker()
|
107 |
+
|
108 |
self.embed_dim = config.hidden_size
|
109 |
|
110 |
self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
|
|
|
128 |
def parallelize(self, device_map=None):
|
129 |
# Check validity of device_map
|
130 |
warnings.warn(
|
131 |
+
"GPT2Model.parallelize is deprecated and will be removed in v5 of Transformers, you should load your"
|
132 |
+
" model with device_map='balanced' in the call to from_pretrained. You can also provide your own"
|
133 |
+
" device_map but it needs to be a dictionary module_name to device, so for instance {'h.0': 0, 'h.1': 1,"
|
134 |
" ...}",
|
135 |
FutureWarning,
|
136 |
)
|
|
|
154 |
|
155 |
def deparallelize(self):
|
156 |
warnings.warn(
|
157 |
+
"Like parallelize, deparallelize is deprecated and will be removed in v5 of Transformers.",
|
158 |
FutureWarning,
|
159 |
)
|
160 |
self.model_parallel = False
|
|
|
284 |
if self.gradient_checkpointing and self.training:
|
285 |
if use_cache:
|
286 |
logger.warning_once(
|
287 |
+
"use_cache=True is incompatible with gradient checkpointing. Setting use_cache=False..."
|
288 |
)
|
289 |
use_cache = False
|
290 |
|
|
|
292 |
all_self_attentions = () if output_attentions else None
|
293 |
all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
|
294 |
all_hidden_states = () if output_hidden_states else None
|
295 |
+
aux_losses = []
|
296 |
for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
|
297 |
# Model parallel
|
298 |
if self.model_parallel:
|
|
|
335 |
hidden_states = outputs[0]
|
336 |
if use_cache is True:
|
337 |
presents = presents + (outputs[1],)
|
338 |
+
|
339 |
+
if isinstance(block, LOLABlock):
|
340 |
+
# Collect auxiliary loss
|
341 |
+
aux_loss = outputs[-1]
|
342 |
+
aux_losses.append(aux_loss)
|
343 |
|
344 |
if output_attentions:
|
345 |
all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
|
|
|
|
|
346 |
|
347 |
# Model Parallel: If it's the last layer for that device, put things on the next device
|
348 |
if self.model_parallel:
|
|
|
353 |
hidden_states = self.ln_f(hidden_states)
|
354 |
|
355 |
hidden_states = hidden_states.view(output_shape)
|
356 |
+
# Aggregate auxiliary losses
|
357 |
+
if aux_losses:
|
358 |
+
total_aux_loss = torch.stack(aux_losses).sum()
|
359 |
+
else:
|
360 |
+
total_aux_loss = None
|
361 |
# Add last hidden state
|
362 |
if output_hidden_states:
|
363 |
all_hidden_states = all_hidden_states + (hidden_states,)
|
|
|
364 |
if not return_dict:
|
365 |
+
output = (hidden_states, presents, all_hidden_states, all_self_attentions)
|
366 |
+
if total_aux_loss is not None:
|
367 |
+
output += (total_aux_loss,)
|
368 |
+
return tuple(v for v in output if v is not None)
|
|
|
369 |
|
370 |
+
return MoeModelOutputWithPast(
|
371 |
last_hidden_state=hidden_states,
|
372 |
past_key_values=presents,
|
373 |
hidden_states=all_hidden_states,
|
374 |
attentions=all_self_attentions,
|
375 |
+
router_logits=None, # Include if router_logits are needed
|
376 |
+
aux_loss=total_aux_loss,
|
377 |
)
|
378 |
|
379 |
class LOLABlock(nn.Module):
|
|
|
385 |
self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
|
386 |
self.attn = GPT2Attention(config, layer_idx=layer_idx)
|
387 |
self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
|
|
|
388 |
self.moe = LOLAMOE(
|
389 |
hidden_size,
|
390 |
inner_dim,
|
|
|
408 |
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
409 |
use_cache: Optional[bool] = False,
|
410 |
output_attentions: Optional[bool] = False,
|
411 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]]:
|
412 |
residual = hidden_states
|
413 |
hidden_states = self.ln_1(hidden_states)
|
414 |
attn_outputs = self.attn(
|
|
|
419 |
use_cache=use_cache,
|
420 |
output_attentions=output_attentions,
|
421 |
)
|
422 |
+
attn_output = attn_outputs[0]
|
423 |
outputs = attn_outputs[1:]
|
|
|
424 |
hidden_states = attn_output + residual
|
425 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
426 |
residual = hidden_states
|
427 |
hidden_states = self.ln_2(hidden_states)
|
428 |
+
feed_forward_hidden_states, router_logits, aux_loss = self.moe(hidden_states)
|
|
|
429 |
hidden_states = residual + feed_forward_hidden_states
|
430 |
|
431 |
if use_cache:
|
432 |
+
outputs = (hidden_states,) + outputs + (aux_loss,)
|
433 |
else:
|
434 |
+
outputs = (hidden_states,) + outputs + (aux_loss,)
|
435 |
|
436 |
+
return outputs # hidden_states, present, (attentions), aux_loss
|
437 |
|
438 |
class LOLAMOE(nn.Module):
|
439 |
def __init__(self,
|
|
|
452 |
self.experts = nn.ModuleList([GPT2MLP(inner_dim, config) for _ in range(self.num_experts)])
|
453 |
|
454 |
def forward(self, hidden_states):
|
|
|
|
|
|
|
|
|
455 |
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
456 |
hidden_states = hidden_states.view(-1, hidden_dim)
|
457 |
|
458 |
router_logits = self.gate(hidden_states)
|
459 |
+
routing_probabilities = F.softmax(router_logits, dim=1)
|
460 |
+
routing_weights, selected_experts = torch.topk(routing_probabilities, self.top_k, dim=-1)
|
461 |
+
# Compute Expert Mask
|
462 |
+
expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts)
|
463 |
+
expert_mask = expert_mask.sum(dim=1) # Shape: [batch_size * seq_length, num_experts]
|
464 |
|
465 |
+
# Compute Tokens per Expert and Router Probabilities
|
466 |
+
token_fraction_per_expert = expert_mask.float().sum(dim=0) / expert_mask.float().sum()
|
467 |
+
mean_router_prob_per_expert = routing_probabilities.mean(dim=0)
|
|
|
|
|
|
|
|
|
468 |
|
469 |
+
# Calculate Auxiliary Loss
|
470 |
+
aux_loss = torch.sum(token_fraction_per_expert * mean_router_prob_per_expert) * self.num_experts
|
471 |
+
|
472 |
+
# Proceed with MoE computation as before
|
473 |
final_hidden_states = torch.zeros(
|
474 |
(batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
|
475 |
)
|
|
|
|
|
|
|
|
|
476 |
|
477 |
+
# Process tokens for each expert
|
478 |
+
for expert_idx in range(self.num_experts):
|
479 |
+
indices = (selected_experts == expert_idx).nonzero(as_tuple=True)[0]
|
480 |
+
if indices.numel() == 0:
|
481 |
continue
|
482 |
+
current_states = hidden_states[indices]
|
483 |
+
current_output = self.experts[expert_idx](current_states)
|
484 |
+
current_weights = routing_weights[indices, (selected_experts[indices] == expert_idx).nonzero(as_tuple=True)[1]]
|
485 |
+
final_hidden_states.index_add_(0, indices, current_output * current_weights.unsqueeze(-1))
|
486 |
|
487 |
+
final_hidden_states = final_hidden_states.view(batch_size, sequence_length, hidden_dim)
|
488 |
+
expert_analysis_callback(selected_experts)
|
489 |
+
return final_hidden_states, router_logits, aux_loss
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
490 |
|
491 |
class LOLAAttention(GPT2Attention):
|
492 |
def __init__(self, config, is_cross_attention=False, layer_idx=None):
|
|
|
510 |
self.split_size = self.embed_dim
|
511 |
if self.head_dim * self.num_heads != self.embed_dim:
|
512 |
raise ValueError(
|
513 |
+
f"embed_dim must be divisible by num_heads (got embed_dim: {self.embed_dim} and num_heads:"
|
514 |
f" {self.num_heads})."
|
515 |
)
|
516 |
|
|
|
551 |
|
552 |
# Initialize weights and apply final processing
|
553 |
self.post_init()
|
|
|
|
|
|
|
554 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
555 |
def forward(
|
556 |
self,
|
557 |
+
input_ids=None,
|
558 |
+
past_key_values=None,
|
559 |
+
attention_mask=None,
|
560 |
+
token_type_ids=None,
|
561 |
+
position_ids=None,
|
562 |
+
head_mask=None,
|
563 |
+
inputs_embeds=None,
|
564 |
+
labels=None,
|
565 |
+
use_cache=None,
|
566 |
+
output_attentions=None,
|
567 |
+
output_hidden_states=None,
|
568 |
+
return_dict=None,
|
569 |
+
) -> Union[Tuple, MoeCausalLMOutputWithPast]:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
570 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
571 |
|
572 |
+
transformer_outputs = self.transformer(
|
573 |
input_ids,
|
574 |
+
past_key_values=past_key_values,
|
575 |
attention_mask=attention_mask,
|
576 |
token_type_ids=token_type_ids,
|
577 |
position_ids=position_ids,
|
578 |
head_mask=head_mask,
|
579 |
inputs_embeds=inputs_embeds,
|
580 |
+
use_cache=use_cache,
|
581 |
output_attentions=output_attentions,
|
582 |
output_hidden_states=output_hidden_states,
|
583 |
+
return_dict=True, # Ensure we get a MoeModelOutputWithPast
|
584 |
)
|
585 |
+
hidden_states = transformer_outputs.last_hidden_state
|
586 |
+
lm_logits = self.lm_head(hidden_states)
|
587 |
+
|
588 |
+
aux_loss = transformer_outputs.aux_loss if hasattr(transformer_outputs, 'aux_loss') else None
|
589 |
+
|
590 |
+
loss = None
|
591 |
+
if labels is not None:
|
592 |
+
# Shift so that tokens < n predict n
|
593 |
+
shift_logits = lm_logits[..., :-1, :].contiguous()
|
594 |
+
shift_labels = labels[..., 1:].contiguous()
|
595 |
+
# Flatten the tokens
|
596 |
+
loss_fct = CrossEntropyLoss()
|
597 |
+
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
598 |
+
if aux_loss is not None:
|
599 |
+
loss += self.config.router_aux_loss_coef * aux_loss
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
600 |
|
601 |
if not return_dict:
|
602 |
+
output = (lm_logits,) + transformer_outputs[1:]
|
603 |
+
return ((loss,) + output) if loss is not None else output
|
604 |
+
|
605 |
+
return MoeCausalLMOutputWithPast(
|
606 |
+
loss=loss,
|
607 |
+
aux_loss=aux_loss,
|
608 |
+
logits=lm_logits,
|
609 |
+
past_key_values=transformer_outputs.past_key_values,
|
610 |
+
hidden_states=transformer_outputs.hidden_states,
|
611 |
+
attentions=transformer_outputs.attentions,
|
612 |
+
router_logits=transformer_outputs.router_logits,
|
613 |
)
|