try to simplify checkpointing
Browse files- modeling_bert.py +2 -247
modeling_bert.py
CHANGED
@@ -329,9 +329,7 @@ class BertPreTrainedModel(nn.Module):
|
|
329 |
"""
|
330 |
# Instantiate model.
|
331 |
model = cls(config, *inputs, **kwargs)
|
332 |
-
load_return = model.load_state_dict(
|
333 |
-
remap_state_dict(state_dict_from_pretrained(model_name), config), strict=False
|
334 |
-
)
|
335 |
logger.info(load_return)
|
336 |
return model
|
337 |
|
@@ -528,247 +526,4 @@ class BertForPreTraining(BertPreTrainedModel):
|
|
528 |
loss=total_loss,
|
529 |
prediction_logits=prediction_scores,
|
530 |
seq_relationship_logits=seq_relationship_score,
|
531 |
-
)
|
532 |
-
|
533 |
-
|
534 |
-
def remap_state_dict(state_dict, config: PretrainedConfig):
|
535 |
-
"""
|
536 |
-
Map the state_dict of a Huggingface BERT model to be flash_attn compatible.
|
537 |
-
"""
|
538 |
-
|
539 |
-
# LayerNorm
|
540 |
-
def key_mapping_ln_gamma_beta(key):
|
541 |
-
key = re.sub(r"LayerNorm.gamma$", "LayerNorm.weight", key)
|
542 |
-
key = re.sub(r"LayerNorm.beta$", "LayerNorm.bias", key)
|
543 |
-
return key
|
544 |
-
|
545 |
-
state_dict = OrderedDict((key_mapping_ln_gamma_beta(k), v) for k, v in state_dict.items())
|
546 |
-
|
547 |
-
# Layers
|
548 |
-
def key_mapping_layers(key):
|
549 |
-
return re.sub(r"^bert.encoder.layer.", "bert.encoder.layers.", key)
|
550 |
-
|
551 |
-
state_dict = OrderedDict((key_mapping_layers(k), v) for k, v in state_dict.items())
|
552 |
-
|
553 |
-
# LayerNorm
|
554 |
-
def key_mapping_ln(key):
|
555 |
-
key = re.sub(r"^bert.embeddings.LayerNorm.", "bert.emb_ln.", key)
|
556 |
-
key = re.sub(
|
557 |
-
r"^bert.encoder.layers.(\d+).attention.output.LayerNorm.(weight|bias)",
|
558 |
-
r"bert.encoder.layers.\1.norm1.\2",
|
559 |
-
key,
|
560 |
-
)
|
561 |
-
key = re.sub(
|
562 |
-
r"^bert.encoder.layers.(\d+).output.LayerNorm.(weight|bias)",
|
563 |
-
r"bert.encoder.layers.\1.norm2.\2",
|
564 |
-
key,
|
565 |
-
)
|
566 |
-
key = re.sub(
|
567 |
-
r"^cls.predictions.transform.LayerNorm.(weight|bias)",
|
568 |
-
r"cls.predictions.transform.layer_norm.\1",
|
569 |
-
key,
|
570 |
-
)
|
571 |
-
return key
|
572 |
-
|
573 |
-
state_dict = OrderedDict((key_mapping_ln(k), v) for k, v in state_dict.items())
|
574 |
-
|
575 |
-
# MLP
|
576 |
-
def key_mapping_mlp(key):
|
577 |
-
key = re.sub(
|
578 |
-
r"^bert.encoder.layers.(\d+).intermediate.dense.(weight|bias)",
|
579 |
-
r"bert.encoder.layers.\1.mlp.fc1.\2",
|
580 |
-
key,
|
581 |
-
)
|
582 |
-
key = re.sub(
|
583 |
-
r"^bert.encoder.layers.(\d+).output.dense.(weight|bias)",
|
584 |
-
r"bert.encoder.layers.\1.mlp.fc2.\2",
|
585 |
-
key,
|
586 |
-
)
|
587 |
-
return key
|
588 |
-
|
589 |
-
state_dict = OrderedDict((key_mapping_mlp(k), v) for k, v in state_dict.items())
|
590 |
-
|
591 |
-
# Attention
|
592 |
-
last_layer_subset = getattr(config, "last_layer_subset", False)
|
593 |
-
for d in range(config.num_hidden_layers):
|
594 |
-
Wq = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.query.weight")
|
595 |
-
Wk = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.key.weight")
|
596 |
-
Wv = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.value.weight")
|
597 |
-
bq = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.query.bias")
|
598 |
-
bk = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.key.bias")
|
599 |
-
bv = state_dict.pop(f"bert.encoder.layers.{d}.attention.self.value.bias")
|
600 |
-
if not (last_layer_subset and d == config.num_hidden_layers - 1):
|
601 |
-
state_dict[f"bert.encoder.layers.{d}.mixer.Wqkv.weight"] = torch.cat(
|
602 |
-
[Wq, Wk, Wv], dim=0
|
603 |
-
)
|
604 |
-
state_dict[f"bert.encoder.layers.{d}.mixer.Wqkv.bias"] = torch.cat([bq, bk, bv], dim=0)
|
605 |
-
else:
|
606 |
-
state_dict[f"bert.encoder.layers.{d}.mixer.Wq.weight"] = Wq
|
607 |
-
state_dict[f"bert.encoder.layers.{d}.mixer.Wkv.weight"] = torch.cat([Wk, Wv], dim=0)
|
608 |
-
state_dict[f"bert.encoder.layers.{d}.mixer.Wq.bias"] = bq
|
609 |
-
state_dict[f"bert.encoder.layers.{d}.mixer.Wkv.bias"] = torch.cat([bk, bv], dim=0)
|
610 |
-
|
611 |
-
def key_mapping_attn(key):
|
612 |
-
return re.sub(
|
613 |
-
r"^bert.encoder.layers.(\d+).attention.output.dense.(weight|bias)",
|
614 |
-
r"bert.encoder.layers.\1.mixer.out_proj.\2",
|
615 |
-
key,
|
616 |
-
)
|
617 |
-
|
618 |
-
state_dict = OrderedDict((key_mapping_attn(k), v) for k, v in state_dict.items())
|
619 |
-
|
620 |
-
def key_mapping_decoder_bias(key):
|
621 |
-
return re.sub(r"^cls.predictions.bias", "cls.predictions.decoder.bias", key)
|
622 |
-
|
623 |
-
state_dict = OrderedDict((key_mapping_decoder_bias(k), v) for k, v in state_dict.items())
|
624 |
-
|
625 |
-
# Word embedding
|
626 |
-
pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
|
627 |
-
if pad_vocab_size_multiple > 1:
|
628 |
-
word_embeddings = state_dict["bert.embeddings.word_embeddings.weight"]
|
629 |
-
state_dict["bert.embeddings.word_embeddings.weight"] = F.pad(
|
630 |
-
word_embeddings, (0, 0, 0, config.vocab_size - word_embeddings.shape[0])
|
631 |
-
)
|
632 |
-
decoder_weight = state_dict["cls.predictions.decoder.weight"]
|
633 |
-
state_dict["cls.predictions.decoder.weight"] = F.pad(
|
634 |
-
decoder_weight, (0, 0, 0, config.vocab_size - decoder_weight.shape[0])
|
635 |
-
)
|
636 |
-
# If the vocab was padded, we want to set the decoder bias for those padded indices to be
|
637 |
-
# strongly negative (i.e. the decoder shouldn't predict those indices).
|
638 |
-
# TD [2022-05-09]: I don't think it affects the MLPerf training.
|
639 |
-
decoder_bias = state_dict["cls.predictions.decoder.bias"]
|
640 |
-
state_dict["cls.predictions.decoder.bias"] = F.pad(
|
641 |
-
decoder_bias, (0, config.vocab_size - decoder_bias.shape[0]), value=-100.0
|
642 |
-
)
|
643 |
-
|
644 |
-
return state_dict
|
645 |
-
|
646 |
-
|
647 |
-
def inv_remap_state_dict(state_dict, config: PretrainedConfig):
|
648 |
-
"""
|
649 |
-
Map the state_dict of a flash_attn model to be Huggingface BERT compatible.
|
650 |
-
|
651 |
-
This function is meant to be the inverse of remap_state_dict.
|
652 |
-
"""
|
653 |
-
# Word embedding
|
654 |
-
pad_vocab_size_multiple = getattr(config, "pad_vocab_size_multiple", 1)
|
655 |
-
if pad_vocab_size_multiple > 1:
|
656 |
-
word_embeddings = state_dict["bert.embeddings.word_embeddings.weight"]
|
657 |
-
decoder_weight = state_dict["cls.predictions.decoder.weight"]
|
658 |
-
decoder_bias = state_dict["cls.predictions.decoder.bias"]
|
659 |
-
# unpad embeddings
|
660 |
-
state_dict["bert.embeddings.word_embeddings.weight"] = word_embeddings[
|
661 |
-
: config.orig_vocab_size, :
|
662 |
-
]
|
663 |
-
state_dict["cls.predictions.decoder.weight"] = decoder_weight[: config.orig_vocab_size, :]
|
664 |
-
state_dict["cls.predictions.decoder.bias"] = decoder_bias[: config.orig_vocab_size]
|
665 |
-
|
666 |
-
for d in range(config.num_hidden_layers):
|
667 |
-
last_layer_subset = getattr(config, "last_layer_subset", False)
|
668 |
-
if not last_layer_subset or d != (config.num_hidden_layers - 1):
|
669 |
-
Wqkv_weights = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wqkv.weight")
|
670 |
-
Wqkv_biases = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wqkv.bias")
|
671 |
-
state_dict[f"bert.encoder.layers.{d}.attention.self.query.weight"] = Wqkv_weights[
|
672 |
-
: Wqkv_weights.shape[0] // 3, :
|
673 |
-
]
|
674 |
-
state_dict[f"bert.encoder.layers.{d}.attention.self.key.weight"] = Wqkv_weights[
|
675 |
-
Wqkv_weights.shape[0] // 3 : 2 * Wqkv_weights.shape[0] // 3, :
|
676 |
-
]
|
677 |
-
state_dict[f"bert.encoder.layers.{d}.attention.self.value.weight"] = Wqkv_weights[
|
678 |
-
2 * Wqkv_weights.shape[0] // 3 :, :
|
679 |
-
]
|
680 |
-
state_dict[f"bert.encoder.layers.{d}.attention.self.query.bias"] = Wqkv_biases[
|
681 |
-
: Wqkv_biases.shape[0] // 3
|
682 |
-
]
|
683 |
-
state_dict[f"bert.encoder.layers.{d}.attention.self.key.bias"] = Wqkv_biases[
|
684 |
-
Wqkv_biases.shape[0] // 3 : 2 * Wqkv_biases.shape[0] // 3
|
685 |
-
]
|
686 |
-
state_dict[f"bert.encoder.layers.{d}.attention.self.value.bias"] = Wqkv_biases[
|
687 |
-
2 * Wqkv_biases.shape[0] // 3 :
|
688 |
-
]
|
689 |
-
else:
|
690 |
-
Wq_weight = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wq.weight")
|
691 |
-
Wkv_weights = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wkv.weight")
|
692 |
-
Wq_bias = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wq.bias")
|
693 |
-
Wkv_biases = state_dict.pop(f"bert.encoder.layers.{d}.mixer.Wkv.bias")
|
694 |
-
state_dict[f"bert.encoder.layers.{d}.attention.self.query.weight"] = Wq_weight
|
695 |
-
state_dict[f"bert.encoder.layers.{d}.attention.self.key.weight"] = Wkv_weights[
|
696 |
-
: Wkv_weights.shape[0] // 2, :
|
697 |
-
]
|
698 |
-
state_dict[f"bert.encoder.layers.{d}.attention.self.value.weight"] = Wkv_weights[
|
699 |
-
Wkv_weights.shape[0] // 2 :, :
|
700 |
-
]
|
701 |
-
state_dict[f"bert.encoder.layers.{d}.attention.self.query.bias"] = Wq_bias
|
702 |
-
state_dict[f"bert.encoder.layers.{d}.attention.self.key.bias"] = Wkv_biases[
|
703 |
-
: Wkv_biases.shape[0] // 2
|
704 |
-
]
|
705 |
-
state_dict[f"bert.encoder.layers.{d}.attention.self.value.bias"] = Wkv_biases[
|
706 |
-
Wkv_biases.shape[0] // 2 :
|
707 |
-
]
|
708 |
-
|
709 |
-
def inv_key_mapping_ln(key):
|
710 |
-
key = re.sub(r"bert.emb_ln.", "bert.embeddings.LayerNorm.", key)
|
711 |
-
key = re.sub(
|
712 |
-
r"bert.encoder.layers.(\d+).norm1.(weight|bias)",
|
713 |
-
r"bert.encoder.layers.\1.attention.output.LayerNorm.\2",
|
714 |
-
key,
|
715 |
-
)
|
716 |
-
key = re.sub(
|
717 |
-
r"bert.encoder.layers.(\d+).norm2.(weight|bias)",
|
718 |
-
r"bert.encoder.layers.\1.output.LayerNorm.\2",
|
719 |
-
key,
|
720 |
-
)
|
721 |
-
key = re.sub(
|
722 |
-
r"cls.predictions.transform.layer_norm.(weight|bias)",
|
723 |
-
r"cls.predictions.transform.LayerNorm.\1",
|
724 |
-
key,
|
725 |
-
)
|
726 |
-
return key
|
727 |
-
|
728 |
-
def inv_key_mapping_ln_gamma_beta(key):
|
729 |
-
key = re.sub(r"LayerNorm.weight$", "LayerNorm.gamma", key)
|
730 |
-
key = re.sub(r"LayerNorm.bias$", "LayerNorm.beta", key)
|
731 |
-
return key
|
732 |
-
|
733 |
-
def inv_key_mapping_layers(key):
|
734 |
-
return re.sub(r"bert.encoder.layers.", "bert.encoder.layer.", key)
|
735 |
-
|
736 |
-
def inv_key_mapping_mlp(key):
|
737 |
-
key = re.sub(
|
738 |
-
r"bert.encoder.layer.(\d+).mlp.fc1.(weight|bias)",
|
739 |
-
r"bert.encoder.layer.\1.intermediate.dense.\2",
|
740 |
-
key,
|
741 |
-
)
|
742 |
-
key = re.sub(
|
743 |
-
r"bert.encoder.layer.(\d+).mlp.fc2.(weight|bias)",
|
744 |
-
r"bert.encoder.layer.\1.output.dense.\2",
|
745 |
-
key,
|
746 |
-
)
|
747 |
-
return key
|
748 |
-
|
749 |
-
def inv_key_mapping_attn(key):
|
750 |
-
return re.sub(
|
751 |
-
r"bert.encoder.layer.(\d+).mixer.out_proj.(weight|bias)",
|
752 |
-
r"bert.encoder.layer.\1.attention.output.dense.\2",
|
753 |
-
key,
|
754 |
-
)
|
755 |
-
|
756 |
-
def inv_key_mapping_decoder_bias(key):
|
757 |
-
return re.sub(r"cls.predictions.decoder.bias", "cls.predictions.bias", key)
|
758 |
-
|
759 |
-
state_dict = OrderedDict((inv_key_mapping_ln(key), value) for key, value in state_dict.items())
|
760 |
-
state_dict = OrderedDict(
|
761 |
-
(inv_key_mapping_ln_gamma_beta(key), value) for key, value in state_dict.items()
|
762 |
-
)
|
763 |
-
state_dict = OrderedDict(
|
764 |
-
(inv_key_mapping_layers(key), value) for key, value in state_dict.items()
|
765 |
-
)
|
766 |
-
state_dict = OrderedDict((inv_key_mapping_mlp(key), value) for key, value in state_dict.items())
|
767 |
-
state_dict = OrderedDict(
|
768 |
-
(inv_key_mapping_attn(key), value) for key, value in state_dict.items()
|
769 |
-
)
|
770 |
-
state_dict = OrderedDict(
|
771 |
-
(inv_key_mapping_decoder_bias(key), value) for key, value in state_dict.items()
|
772 |
-
)
|
773 |
-
|
774 |
-
return state_dict
|
|
|
329 |
"""
|
330 |
# Instantiate model.
|
331 |
model = cls(config, *inputs, **kwargs)
|
332 |
+
load_return = model.load_state_dict(state_dict_from_pretrained(model_name), strict=False)
|
|
|
|
|
333 |
logger.info(load_return)
|
334 |
return model
|
335 |
|
|
|
526 |
loss=total_loss,
|
527 |
prediction_logits=prediction_scores,
|
528 |
seq_relationship_logits=seq_relationship_score,
|
529 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|