Update modelling_walsh.py
Browse files- Added support for inference cache.
- Refactor common code in attention
- Removed unused code (fragments from another project)
- modelling_walsh.py +369 -296
modelling_walsh.py
CHANGED
@@ -1,5 +1,5 @@
|
|
1 |
# See: https://huggingface.co/docs/transformers/custom_models
|
2 |
-
from typing import Optional, Tuple, Union
|
3 |
import math
|
4 |
import copy
|
5 |
import sys
|
@@ -9,7 +9,7 @@ import torch
|
|
9 |
from torch import nn, Tensor
|
10 |
import torch.nn.init as init
|
11 |
from torch.nn import functional as F
|
12 |
-
from transformers.modeling_outputs import CausalLMOutput
|
13 |
from transformers import (
|
14 |
PreTrainedModel,
|
15 |
PretrainedConfig,
|
@@ -18,6 +18,10 @@ from transformers import (
|
|
18 |
AutoModelForCausalLM,
|
19 |
)
|
20 |
|
|
|
|
|
|
|
|
|
21 |
from transformers.utils import (
|
22 |
is_flash_attn_2_available,
|
23 |
is_flash_attn_greater_or_equal_2_10,
|
@@ -26,6 +30,8 @@ from transformers.utils import (
|
|
26 |
if is_flash_attn_2_available():
|
27 |
from flash_attn import flash_attn_qkvpacked_func, flash_attn_func
|
28 |
|
|
|
|
|
29 |
# The model type string to bind.
|
30 |
model_type = "walsh-causal-v1"
|
31 |
|
@@ -78,6 +84,10 @@ class Config(PretrainedConfig):
|
|
78 |
layer_args=dict(),
|
79 |
embedding_args=dict(),
|
80 |
output_proj_args=dict(),
|
|
|
|
|
|
|
|
|
81 |
|
82 |
**kwargs,
|
83 |
):
|
@@ -113,6 +123,10 @@ class Config(PretrainedConfig):
|
|
113 |
self.layer_args = layer_args
|
114 |
self.embedding_args = embedding_args
|
115 |
self.output_proj_args = output_proj_args
|
|
|
|
|
|
|
|
|
116 |
|
117 |
super().__init__(**kwargs)
|
118 |
|
@@ -204,6 +218,8 @@ class HFCausalModel(PreTrainedModel):
|
|
204 |
_no_split_modules = ["DeepNetLayer"]
|
205 |
_supports_flash_attn_2 = True
|
206 |
_supports_sdpa = True
|
|
|
|
|
207 |
|
208 |
def __init__(self, config):
|
209 |
super().__init__(config)
|
@@ -221,40 +237,143 @@ class HFCausalModel(PreTrainedModel):
|
|
221 |
token_type_ids: Optional[torch.LongTensor] = None,
|
222 |
position_ids: Optional[torch.LongTensor] = None,
|
223 |
labels: Optional[torch.LongTensor] = None,
|
|
|
|
|
224 |
output_attentions: Optional[bool] = None,
|
225 |
output_hidden_states: Optional[bool] = None,
|
226 |
return_dict: Optional[bool] = None,
|
227 |
**kwargs,
|
228 |
) -> (Tensor, dict[str, Tensor]):
|
229 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
230 |
if self.gradient_checkpointing and self.training:
|
|
|
|
|
|
|
|
|
|
|
231 |
gradient_checkpointing_func = self._gradient_checkpointing_func
|
232 |
else:
|
233 |
gradient_checkpointing_func = None
|
|
|
234 |
|
235 |
-
|
236 |
input_ids=input_ids,
|
237 |
-
|
|
|
238 |
gradient_checkpointing_func=gradient_checkpointing_func,
|
|
|
|
|
|
|
239 |
)
|
|
|
|
|
|
|
240 |
|
241 |
# Compute loss.
|
242 |
if labels is not None:
|
243 |
loss = self.loss_function(logits=logits, labels=labels, input_ids=input_ids)
|
244 |
else:
|
245 |
loss = None
|
|
|
|
|
|
|
|
|
|
|
246 |
|
247 |
-
return
|
248 |
-
|
249 |
-
|
250 |
-
|
251 |
-
|
252 |
-
|
253 |
-
|
254 |
-
|
255 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
256 |
return model_inputs
|
257 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
258 |
def _make_embedding(self, config):
|
259 |
embedding_cls = get_dynamic_class(config.embdding_cls)
|
260 |
return embedding_cls(config.vocab_size, self.d_model, config.pad_index, **config.embedding_args)
|
@@ -278,7 +397,7 @@ class HFCausalModel(PreTrainedModel):
|
|
278 |
norm_cls = get_dynamic_class(config.norm_cls)
|
279 |
return norm_cls(self.d_model)
|
280 |
|
281 |
-
def _make_self_attention(self, config):
|
282 |
attention_cls = get_dynamic_class(config.attention_cls)
|
283 |
# Map HF _attn_implementation to attn_type
|
284 |
match config._attn_implementation:
|
@@ -299,28 +418,32 @@ class HFCausalModel(PreTrainedModel):
|
|
299 |
d_model=self.d_model,
|
300 |
num_heads=config.num_attention_heads,
|
301 |
attn_type=attn_type,
|
|
|
|
|
302 |
**config.attention_args,
|
303 |
)
|
304 |
|
305 |
-
def _make_feedforward(self, config):
|
306 |
feedforward_cls = get_dynamic_class(config.feedforward_cls)
|
307 |
return feedforward_cls(
|
308 |
d_model=self.d_model,
|
309 |
feedforward_dim=config.dim_feedforward,
|
310 |
dropout=config.dropout,
|
311 |
activation=self._make_activation(config),
|
|
|
312 |
**config.feedforward_args,
|
313 |
)
|
314 |
|
315 |
-
def _make_layer(self, config):
|
316 |
layer_cls = get_dynamic_class(config.layer_cls)
|
317 |
return layer_cls(
|
318 |
d_model=self.d_model,
|
319 |
dropout=self._make_dropout(config),
|
320 |
-
attention=self._make_self_attention(config),
|
321 |
-
feedforward=self._make_feedforward(config),
|
322 |
norm1=self._make_norm(config),
|
323 |
norm2=self._make_norm(config),
|
|
|
324 |
**config.layer_args,
|
325 |
)
|
326 |
|
@@ -328,7 +451,7 @@ class HFCausalModel(PreTrainedModel):
|
|
328 |
layer_stack_cls = get_dynamic_class(config.layer_stack_cls)
|
329 |
return layer_stack_cls(
|
330 |
layers=nn.ModuleList([
|
331 |
-
self._make_layer(config) for
|
332 |
]),
|
333 |
**config.layer_stack_args,
|
334 |
)
|
@@ -364,43 +487,35 @@ class Transformer(nn.Module):
|
|
364 |
self.sqrt_d_model = d_model**0.5
|
365 |
self.reset_parameters()
|
366 |
|
367 |
-
def forward(
|
368 |
-
|
369 |
-
|
370 |
-
|
371 |
-
|
372 |
-
|
373 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
374 |
)
|
375 |
|
376 |
-
# Translate output
|
377 |
-
logits = self.output_projection(
|
378 |
-
|
|
|
379 |
|
380 |
def reset_parameters(self):
|
381 |
init.xavier_uniform_(self.output_projection.weight)
|
382 |
init.constant_(self.output_projection.bias, 0.)
|
383 |
init.normal_(self.embedding.weight, std=self.d_model**-0.5)
|
384 |
|
385 |
-
# A vanilla positional encoder
|
386 |
-
class PositionalEncoder(nn.Module):
|
387 |
-
def __init__(self, d_embed, max_seq):
|
388 |
-
super().__init__()
|
389 |
-
self.d_embed = d_embed
|
390 |
-
self.max_seq = max_seq
|
391 |
-
|
392 |
-
weight = torch.zeros(max_seq, d_embed)
|
393 |
-
position = torch.arange(0, max_seq, dtype=torch.float).unsqueeze(1)
|
394 |
-
div_term = torch.exp(torch.arange(0, d_embed, 2).float() * (-math.log(10000.0) / d_embed))
|
395 |
-
weight[:, 0::2] = torch.sin(position * div_term)
|
396 |
-
weight[:, 1::2] = torch.cos(position * div_term)
|
397 |
-
weight = weight.unsqueeze(0)
|
398 |
-
self.register_buffer('weight', weight)
|
399 |
-
|
400 |
-
def forward(self, x):
|
401 |
-
seq_len = x.size(-2)
|
402 |
-
return x + self.weight[:, :seq_len]
|
403 |
-
|
404 |
# Converts a torch array of integers into their equivalent binary codes.
|
405 |
def binary_tensor(x, bits):
|
406 |
mask = 2**torch.arange(bits).to(x.device, x.dtype)
|
@@ -472,7 +587,7 @@ class RSWalshPositionalEncoder(nn.Module):
|
|
472 |
# walsh = (hadamard_walsh_matrix(k)[:bits,:d_embed] -0.5) * self.gain
|
473 |
self.register_buffer('walsh', walsh, persistent=False)
|
474 |
|
475 |
-
def forward(self, x):
|
476 |
seq_len = x.size(-2)
|
477 |
|
478 |
# Get sequence of binary codes...
|
@@ -486,6 +601,12 @@ class RSWalshPositionalEncoder(nn.Module):
|
|
486 |
shift = torch.randint(self.max_seq - seq_len + 1, (1,)).item()
|
487 |
seq = self.binary_code[shift:seq_len + shift,:]
|
488 |
|
|
|
|
|
|
|
|
|
|
|
|
|
489 |
# Disable shifting when not training. This does not appear to change the evaluation loss, but
|
490 |
# it does makes predictions easier to analyse when the attention weights are not shifting with each step.
|
491 |
else:
|
@@ -508,25 +629,58 @@ class TransformerLayerStack(nn.Module):
|
|
508 |
super().__init__()
|
509 |
self.layers = layers
|
510 |
|
511 |
-
def forward(
|
512 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
513 |
for layer in self.layers:
|
514 |
if gradient_checkpointing_func is not None:
|
515 |
-
|
516 |
layer.__call__,
|
517 |
-
|
518 |
-
|
519 |
-
|
|
|
|
|
520 |
)
|
521 |
else:
|
522 |
-
|
523 |
-
|
524 |
-
|
|
|
|
|
|
|
525 |
|
526 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
527 |
|
528 |
# DeepNet: Scaling Transformers to 1,000 Layers
|
529 |
# https://arxiv.org/abs/2203.00555
|
|
|
530 |
class DeepnetLayer(nn.Module):
|
531 |
def __init__(
|
532 |
self,
|
@@ -536,6 +690,7 @@ class DeepnetLayer(nn.Module):
|
|
536 |
norm1,
|
537 |
norm2,
|
538 |
dropout,
|
|
|
539 |
alpha=1.0,
|
540 |
):
|
541 |
super().__init__()
|
@@ -547,27 +702,45 @@ class DeepnetLayer(nn.Module):
|
|
547 |
self.dropout = dropout
|
548 |
# Deepnet alpha
|
549 |
self.alpha = alpha
|
|
|
550 |
|
551 |
-
def forward(
|
|
|
|
|
|
|
|
|
|
|
|
|
552 |
# Keep input as residual
|
553 |
-
residual =
|
554 |
|
555 |
# Compute attention
|
556 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
557 |
|
558 |
# Add attention with residual and normalize.
|
559 |
-
|
560 |
|
561 |
# Keep output as next residual.
|
562 |
-
residual =
|
563 |
|
564 |
# Pass through feedforward network.
|
565 |
-
|
566 |
|
567 |
# Combine residual and ff output, then normalize again.
|
568 |
-
|
569 |
|
570 |
-
return
|
|
|
|
|
|
|
|
|
571 |
|
572 |
# A vanilla MLP transfomer layer.
|
573 |
class FeedforwardLayer(nn.Module):
|
@@ -576,6 +749,7 @@ class FeedforwardLayer(nn.Module):
|
|
576 |
d_model: int,
|
577 |
feedforward_dim: int,
|
578 |
dropout,
|
|
|
579 |
activation=nn.ReLU(),
|
580 |
beta=1.0,
|
581 |
bias=True,
|
@@ -598,41 +772,6 @@ class FeedforwardLayer(nn.Module):
|
|
598 |
init.constant_(self.linear1.bias, 0.)
|
599 |
init.constant_(self.linear2.bias, 0.)
|
600 |
|
601 |
-
# GLU Variants Improve Transformer
|
602 |
-
# https://arxiv.org/pdf/2002.05202v1.pdf
|
603 |
-
class SwiGLUFeedforwardLayer(nn.Module):
|
604 |
-
def __init__(
|
605 |
-
self,
|
606 |
-
d_model,
|
607 |
-
d_feedforward,
|
608 |
-
beta=1.0,
|
609 |
-
dropout=0.1
|
610 |
-
):
|
611 |
-
super().__init__()
|
612 |
-
self.d_model = d_model
|
613 |
-
self.d_feedforward = d_feedforward
|
614 |
-
self.beta = 1.0
|
615 |
-
|
616 |
-
self.linear1 = nn.Linear(self.d_model, self.d_feedforward * 2, bias=False)
|
617 |
-
self.linear2 = nn.Linear(self.d_feedforward, self.d_model, bias=False)
|
618 |
-
self.dropout = nn.Dropout(dropout)
|
619 |
-
self.reset_parameters()
|
620 |
-
|
621 |
-
def forward(self, x):
|
622 |
-
x, gate = self.linear1(x).chunk(2, dim=-1)
|
623 |
-
x = x * F.silu(gate)
|
624 |
-
x = self.dropout(x)
|
625 |
-
x = self.linear2(x)
|
626 |
-
return x
|
627 |
-
|
628 |
-
def reset_parameters(self):
|
629 |
-
# Deepnet initialization
|
630 |
-
# https://arxiv.org/pdf/2203.00555.pdf
|
631 |
-
w, g = self.linear1.weight.chunk(2, dim=0)
|
632 |
-
init.xavier_uniform_(w, gain=self.beta)
|
633 |
-
init.xavier_uniform_(g, gain=self.beta)
|
634 |
-
init.xavier_uniform_(self.linear2.weight, gain=self.beta)
|
635 |
-
|
636 |
class CausalSelfAttention(nn.Module):
|
637 |
def __init__(
|
638 |
self,
|
@@ -643,6 +782,8 @@ class CausalSelfAttention(nn.Module):
|
|
643 |
# torch: Use pytorch "scaled_dot_product_attention()"; faster; generally good compatibility; does not support returning attn weights.
|
644 |
# flash2: Use Flash-Attention2 implementation; fastest; limited to int16 and bfloat16 types; least memory usage.
|
645 |
attn_type,
|
|
|
|
|
646 |
beta=1.0,
|
647 |
dropout=0.1,
|
648 |
):
|
@@ -651,6 +792,8 @@ class CausalSelfAttention(nn.Module):
|
|
651 |
self.num_heads = num_heads
|
652 |
self.beta = beta
|
653 |
self.attn_type = attn_type
|
|
|
|
|
654 |
|
655 |
assert d_model % num_heads == 0, "d_model must be evenly divisible by num_heads"
|
656 |
|
@@ -681,29 +824,56 @@ class CausalSelfAttention(nn.Module):
|
|
681 |
init.constant_(self.in_proj.bias, 0.)
|
682 |
init.constant_(self.output_linear.bias, 0.)
|
683 |
|
684 |
-
|
685 |
-
|
686 |
-
return proj.chunk(chunks=3, dim=-1)
|
687 |
-
|
688 |
-
def forward(self, qkv, need_weights):
|
689 |
-
if self.attn_type == "flash2":
|
690 |
-
return self.flash2_forward(qkv)
|
691 |
-
|
692 |
-
# qkv: (batch_size, seq_len, d_embed)
|
693 |
batch_size, seq_len, d_embed = qkv.shape
|
|
|
|
|
694 |
|
695 |
-
# Feed the inputs through the K, Q, V matrices.
|
696 |
-
query, key, value = self.project_input(qkv)
|
697 |
-
|
698 |
# Split projections into multiple heads and swap position of sequence / heads dimension
|
699 |
query = query.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
700 |
key = key.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
701 |
value = value.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
702 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
703 |
# Default to returning empty attention weights.
|
704 |
-
|
|
|
|
|
705 |
|
706 |
-
if
|
707 |
# This context manager can be used to force which implementation to use.
|
708 |
#with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
|
709 |
attended_values = F.scaled_dot_product_attention(
|
@@ -712,7 +882,7 @@ class CausalSelfAttention(nn.Module):
|
|
712 |
value,
|
713 |
attn_mask=None,
|
714 |
dropout_p=self.dropout.p if self.training else 0.0,
|
715 |
-
is_causal=
|
716 |
scale=self.dot_product_scale
|
717 |
)
|
718 |
# "native" scaled-dot-product attention implementation.
|
@@ -721,44 +891,57 @@ class CausalSelfAttention(nn.Module):
|
|
721 |
scores = torch.matmul(query, key.transpose(-2, -1)) * self.dot_product_scale
|
722 |
|
723 |
# Mask future positions from the past
|
724 |
-
|
725 |
-
|
726 |
-
torch.
|
727 |
-
|
728 |
-
|
729 |
-
|
730 |
-
|
|
|
731 |
|
732 |
# Calculate the attention weights; avoid NANs that might emerge from zeros in softmax's denominator
|
733 |
-
|
734 |
del scores
|
735 |
|
736 |
# Use the attention weights to get a weighted combination of value vectors
|
737 |
-
attended_values = torch.matmul(
|
738 |
-
if not
|
739 |
-
del
|
740 |
-
|
741 |
|
742 |
# Concatenate attention heads and project to original embedding size using the output linear layer
|
743 |
attended_values = attended_values.transpose(1, 2).contiguous().view(batch_size, seq_len, d_embed)
|
744 |
|
745 |
# Project the concatenated output through the output matrix.
|
746 |
attended_values = self.output_linear(attended_values)
|
747 |
-
return
|
748 |
-
|
749 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
750 |
batch_size, seq_len, d_embed = qkv.shape
|
751 |
|
752 |
# Feed the inputs through the K, Q, V matrices.
|
753 |
# query : (batch_size, seq_len, d_model)
|
754 |
# qkv : (batch_size, seq_len, 3, num_heads, d_kq)
|
|
|
|
|
|
|
|
|
755 |
qkv = self.in_proj(qkv).unflatten(
|
756 |
-1,
|
757 |
(3, self.num_heads, self.d_head)
|
758 |
)
|
759 |
-
|
760 |
attended_values = flash_attn_qkvpacked_func(
|
761 |
-
|
762 |
dropout_p=self.dropout.p if self.training else 0.0,
|
763 |
softmax_scale=self.dot_product_scale,
|
764 |
causal=True,
|
@@ -770,180 +953,70 @@ class CausalSelfAttention(nn.Module):
|
|
770 |
|
771 |
# Project the concatenated output through the output matrix.
|
772 |
attended_values = self.output_linear(attended_values)
|
773 |
-
return
|
774 |
-
|
775 |
-
|
776 |
-
|
777 |
-
|
778 |
-
def alibi_biases(query_len, key_len, device='cpu'):
|
779 |
-
x = torch.arange(key_len, device=device)[None, :]
|
780 |
-
y = torch.arange(query_len, device=device)[:, None]
|
781 |
-
return x - y
|
782 |
|
783 |
-
|
784 |
-
|
|
|
785 |
self,
|
786 |
-
|
787 |
-
|
788 |
-
beta=1.0,
|
789 |
-
dropout=0.1,
|
790 |
-
# values:
|
791 |
-
# native: Use local impementation; slowest option; good for debugging; useful when experimenting with non-standard stuff.
|
792 |
-
# torch: Use pytorch "scaled_dot_product_attention()"; faster; generally good compatibility; does not support returning attn weights.
|
793 |
-
# flash2: Use Flash-Attention2 implementation; fastest; limited to int16 and bfloat16 types; can't train Alibi weights; least memory usage.
|
794 |
-
# Note: You can perform initial training with "torch," then switch to "flash2," after the Alibi weights have settled.
|
795 |
-
window_size=None,
|
796 |
-
attn_type="native",
|
797 |
-
freeze_alibi=True,
|
798 |
):
|
799 |
-
super().__init__()
|
800 |
-
self.d_model = d_model
|
801 |
-
self.num_heads = num_heads
|
802 |
-
self.beta = beta
|
803 |
-
self.attn_type = attn_type
|
804 |
-
|
805 |
-
assert d_model % num_heads == 0, "d_model must be evenly divisible by num_heads"
|
806 |
-
|
807 |
-
# The dimension of each head.
|
808 |
-
self.d_head = d_model // num_heads
|
809 |
-
|
810 |
-
# We scale the attention scores by the inverse-square-root of the head dimension
|
811 |
-
# this shifts the temerature of softmax.
|
812 |
-
self.dot_product_scale = 1.0 / math.sqrt(self.d_head)
|
813 |
-
|
814 |
-
self.in_proj = nn.Parameter(torch.empty(3 * self.d_model, self.d_model))
|
815 |
-
self.output_linear = nn.Linear(self.d_model, self.d_model, bias=False)
|
816 |
-
|
817 |
-
if window_size is not None:
|
818 |
-
self.window_size=(window_size, -1)
|
819 |
-
else:
|
820 |
-
self.window_size = (-1, -1)
|
821 |
-
|
822 |
-
self.dropout = nn.Dropout(dropout)
|
823 |
-
|
824 |
-
# This generates the original slope distribution from the paper.
|
825 |
-
# Observations with trainable slopes suggest that the high half of the slopes shift
|
826 |
-
# towards / past 1.0 and the low half approach zero or even go slightly negative.
|
827 |
-
# alibi_slopes = 1.0 / torch.logspace(1, 8, self.num_heads, base=2, dtype=torch.float)
|
828 |
-
|
829 |
-
# These appear to work better, as initial values, in practice.
|
830 |
-
alibi_slopes = 1.0 / torch.logspace(0, 7, self.num_heads, base=2, dtype=torch.float)
|
831 |
-
|
832 |
-
# If not trainable, it can improve performance somewhat if the low half are set to zero. Apparently
|
833 |
-
# making roughly half of the slopes position-agnostic is somehow closer to optimal?
|
834 |
-
# alibi_slopes.masked_fill_(torch.where(torch.arange(0, self.num_heads) >= (self.num_heads / 2), True, False), 0)
|
835 |
-
|
836 |
-
self.alibi_slopes = nn.Parameter(alibi_slopes)
|
837 |
-
|
838 |
-
# Optionally, allow/disallow training of ALiBi slopes.
|
839 |
-
self.alibi_slopes.requires_grad = (not freeze_alibi)
|
840 |
-
self.reset_parameters()
|
841 |
-
|
842 |
-
def extra_repr(self) -> str:
|
843 |
-
return f'd_model={self.d_model}, num_heads={self.num_heads}, beta={self.beta}, attn_type={self.attn_type}, window_size={self.window_size}, dropout={self.dropout}'
|
844 |
-
|
845 |
-
def reset_parameters(self):
|
846 |
-
# Deepnet initialization
|
847 |
-
# https://arxiv.org/pdf/2203.00555.pdf
|
848 |
-
|
849 |
-
q, k, v = self.in_proj.chunk(3)
|
850 |
-
init.xavier_uniform_(q, gain=1.0)
|
851 |
-
init.xavier_uniform_(k, gain=1.0)
|
852 |
-
init.xavier_uniform_(v, gain=self.beta)
|
853 |
-
init.xavier_uniform_(self.output_linear.weight, gain=self.beta)
|
854 |
-
|
855 |
-
def project_input(self, qkv):
|
856 |
-
proj = F.linear(qkv, self.in_proj)
|
857 |
-
return proj.chunk(chunks=3, dim=-1)
|
858 |
-
|
859 |
-
def forward(self, qkv, need_weights):
|
860 |
-
if self.attn_type == "flash2":
|
861 |
-
return self.flash2_forward(qkv)
|
862 |
-
|
863 |
-
# qkv: (batch_size, seq_len, d_embed)
|
864 |
-
batch_size, seq_len, d_embed = qkv.shape
|
865 |
-
|
866 |
-
# Feed the inputs through the K, Q, V matrices.
|
867 |
-
query, key, value = self.project_input(qkv)
|
868 |
-
|
869 |
-
# Split projections into multiple heads and swap position of sequence / heads dimension
|
870 |
-
query = query.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
871 |
-
key = key.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
872 |
-
value = value.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
873 |
-
|
874 |
-
# Apply Alibi relative positional biases.
|
875 |
-
attn_bias = alibi_biases(seq_len, seq_len, device=query.device) * self.alibi_slopes.view(-1, 1, 1)
|
876 |
-
|
877 |
-
# Mask future positions from the past
|
878 |
-
causal_mask = torch.tril(torch.ones(seq_len, seq_len, dtype=torch.bool, device=qkv.device), diagonal=0)
|
879 |
-
attn_bias.masked_fill_(causal_mask.logical_not(), float('-inf'))
|
880 |
-
del causal_mask
|
881 |
-
|
882 |
-
# Default to returning empty attention weights.
|
883 |
-
attention_weights = None
|
884 |
-
|
885 |
-
if self.attn_type == "torch":
|
886 |
-
# This context manager can be used to force which implementation to use.
|
887 |
-
#with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
|
888 |
-
attended_values = F.scaled_dot_product_attention(
|
889 |
-
query,
|
890 |
-
key,
|
891 |
-
value,
|
892 |
-
attn_mask=attn_bias.to(dtype=query.dtype),
|
893 |
-
dropout_p=self.dropout.p if self.training else 0.0,
|
894 |
-
is_causal=False,
|
895 |
-
scale=self.dot_product_scale
|
896 |
-
)
|
897 |
-
# "native" scaled-dot-product attention implementation.
|
898 |
-
else:
|
899 |
-
# Compute attention scores
|
900 |
-
scores = torch.matmul(query, key.transpose(-2, -1)) * self.dot_product_scale
|
901 |
-
|
902 |
-
# Adjust scores with attn_mask
|
903 |
-
scores += attn_bias
|
904 |
-
|
905 |
-
# Calculate the attention weights; avoid NANs that might emerge from zeros in softmax's denominator
|
906 |
-
attention_weights = self.dropout(torch.softmax(scores, dim=-1).clamp(min=1e-10))
|
907 |
-
|
908 |
-
# Use the attention weights to get a weighted combination of value vectors
|
909 |
-
attended_values = torch.matmul(attention_weights, value)
|
910 |
-
if not need_weights:
|
911 |
-
attention_weights = None
|
912 |
-
|
913 |
-
# Concatenate attention heads and project to original embedding size using the output linear layer
|
914 |
-
attended_values = attended_values.transpose(1, 2).contiguous().view(batch_size, seq_len, d_embed)
|
915 |
-
|
916 |
-
# Project the concatenated output through the output matrix.
|
917 |
-
attended_values = self.output_linear(attended_values)
|
918 |
-
return attended_values, attention_weights
|
919 |
-
|
920 |
-
def flash2_forward(self, qkv):
|
921 |
batch_size, seq_len, d_embed = qkv.shape
|
922 |
|
923 |
# Feed the inputs through the K, Q, V matrices.
|
924 |
-
|
925 |
-
|
926 |
-
|
927 |
-
|
928 |
-
|
929 |
-
)
|
930 |
-
|
931 |
-
|
932 |
-
)
|
933 |
-
|
934 |
-
|
935 |
-
|
|
|
|
|
|
|
936 |
dropout_p=self.dropout.p if self.training else 0.0,
|
937 |
softmax_scale=self.dot_product_scale,
|
938 |
causal=True,
|
939 |
-
|
940 |
-
alibi_slopes=self.alibi_slopes.float(),
|
941 |
-
).to(dtype=qkv.dtype)
|
942 |
# attended_values: (batch_size, seqlen, nheads, headdim)
|
943 |
-
|
944 |
# Concatentate heads back into d_embed
|
945 |
attended_values = attended_values.view(batch_size, seq_len, d_embed)
|
946 |
|
947 |
# Project the concatenated output through the output matrix.
|
948 |
attended_values = self.output_linear(attended_values)
|
949 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
# See: https://huggingface.co/docs/transformers/custom_models
|
2 |
+
from typing import Optional, Tuple, Union, List
|
3 |
import math
|
4 |
import copy
|
5 |
import sys
|
|
|
9 |
from torch import nn, Tensor
|
10 |
import torch.nn.init as init
|
11 |
from torch.nn import functional as F
|
12 |
+
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutput, CausalLMOutputWithPast
|
13 |
from transformers import (
|
14 |
PreTrainedModel,
|
15 |
PretrainedConfig,
|
|
|
18 |
AutoModelForCausalLM,
|
19 |
)
|
20 |
|
21 |
+
from transformers.utils import logging
|
22 |
+
|
23 |
+
from transformers.cache_utils import Cache, DynamicCache
|
24 |
+
|
25 |
from transformers.utils import (
|
26 |
is_flash_attn_2_available,
|
27 |
is_flash_attn_greater_or_equal_2_10,
|
|
|
30 |
if is_flash_attn_2_available():
|
31 |
from flash_attn import flash_attn_qkvpacked_func, flash_attn_func
|
32 |
|
33 |
+
logger = logging.get_logger(__name__)
|
34 |
+
|
35 |
# The model type string to bind.
|
36 |
model_type = "walsh-causal-v1"
|
37 |
|
|
|
84 |
layer_args=dict(),
|
85 |
embedding_args=dict(),
|
86 |
output_proj_args=dict(),
|
87 |
+
|
88 |
+
output_attentions=False,
|
89 |
+
output_hidden_states=False,
|
90 |
+
use_cache=True,
|
91 |
|
92 |
**kwargs,
|
93 |
):
|
|
|
123 |
self.layer_args = layer_args
|
124 |
self.embedding_args = embedding_args
|
125 |
self.output_proj_args = output_proj_args
|
126 |
+
|
127 |
+
self.output_attentions = output_attentions
|
128 |
+
self.output_hidden_states = output_hidden_states
|
129 |
+
self.use_cache = use_cache
|
130 |
|
131 |
super().__init__(**kwargs)
|
132 |
|
|
|
218 |
_no_split_modules = ["DeepNetLayer"]
|
219 |
_supports_flash_attn_2 = True
|
220 |
_supports_sdpa = True
|
221 |
+
_supports_cache_class = True
|
222 |
+
_skip_keys_device_placement = "past_key_values"
|
223 |
|
224 |
def __init__(self, config):
|
225 |
super().__init__(config)
|
|
|
237 |
token_type_ids: Optional[torch.LongTensor] = None,
|
238 |
position_ids: Optional[torch.LongTensor] = None,
|
239 |
labels: Optional[torch.LongTensor] = None,
|
240 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
241 |
+
use_cache: Optional[bool] = None,
|
242 |
output_attentions: Optional[bool] = None,
|
243 |
output_hidden_states: Optional[bool] = None,
|
244 |
return_dict: Optional[bool] = None,
|
245 |
**kwargs,
|
246 |
) -> (Tensor, dict[str, Tensor]):
|
247 |
|
248 |
+
batch_size, seq_len = input_ids.shape
|
249 |
+
|
250 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
251 |
+
output_hidden_states = (
|
252 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
253 |
+
)
|
254 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
255 |
+
|
256 |
+
if use_cache:
|
257 |
+
# If legacy cache, convert to DynamicCache
|
258 |
+
use_legacy_cache = not isinstance(past_key_values, Cache)
|
259 |
+
if use_legacy_cache:
|
260 |
+
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
|
261 |
+
|
262 |
+
|
263 |
if self.gradient_checkpointing and self.training:
|
264 |
+
if use_cache:
|
265 |
+
logger.warning_once(
|
266 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
267 |
+
)
|
268 |
+
use_cache = False
|
269 |
gradient_checkpointing_func = self._gradient_checkpointing_func
|
270 |
else:
|
271 |
gradient_checkpointing_func = None
|
272 |
+
|
273 |
|
274 |
+
outputs = self.transformer_head(
|
275 |
input_ids=input_ids,
|
276 |
+
position_ids=position_ids,
|
277 |
+
output_attentions=output_attentions,
|
278 |
gradient_checkpointing_func=gradient_checkpointing_func,
|
279 |
+
past_key_values=past_key_values,
|
280 |
+
use_cache=use_cache,
|
281 |
+
output_hidden_states=output_hidden_states,
|
282 |
)
|
283 |
+
|
284 |
+
logits = outputs["logits"].float()
|
285 |
+
attentions = outputs["attentions"]
|
286 |
|
287 |
# Compute loss.
|
288 |
if labels is not None:
|
289 |
loss = self.loss_function(logits=logits, labels=labels, input_ids=input_ids)
|
290 |
else:
|
291 |
loss = None
|
292 |
+
|
293 |
+
# Convert back to legacy cache, if that's what we received
|
294 |
+
new_cache = outputs["past_key_values"]
|
295 |
+
if use_cache and new_cache is not None and use_legacy_cache:
|
296 |
+
new_cache = new_cache.to_legacy_cache()
|
297 |
|
298 |
+
return CausalLMOutputWithPast(
|
299 |
+
loss=loss,
|
300 |
+
logits=logits,
|
301 |
+
past_key_values=new_cache,
|
302 |
+
hidden_states=outputs["hidden_states"],
|
303 |
+
attentions=outputs["attentions"],
|
304 |
+
)
|
305 |
+
|
306 |
+
# Implementation from Huggingface Transformers,
|
307 |
+
# https://github.com/huggingface/transformers/blob/main/src/transformers/models/mistral/modeling_mistral.py
|
308 |
+
# Note: We do not implement attention mask at present, so some of this code is not applicable
|
309 |
+
# TODO: Reenable attention mask support for batch inference..
|
310 |
+
def prepare_inputs_for_generation(
|
311 |
+
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
|
312 |
+
):
|
313 |
+
# Omit tokens covered by past_key_values
|
314 |
+
if past_key_values is not None:
|
315 |
+
if isinstance(past_key_values, Cache):
|
316 |
+
cache_length = past_key_values.get_seq_length()
|
317 |
+
past_length = past_key_values.seen_tokens
|
318 |
+
max_cache_length = past_key_values.get_max_length()
|
319 |
+
else:
|
320 |
+
cache_length = past_length = past_key_values[0][0].shape[2]
|
321 |
+
max_cache_length = None
|
322 |
+
|
323 |
+
# Keep only the unprocessed tokens:
|
324 |
+
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
|
325 |
+
# some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
|
326 |
+
# input)
|
327 |
+
if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
|
328 |
+
input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
|
329 |
+
# 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
|
330 |
+
# input_ids based on the past_length.
|
331 |
+
elif past_length < input_ids.shape[1]:
|
332 |
+
input_ids = input_ids[:, past_length:]
|
333 |
+
# 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
|
334 |
+
|
335 |
+
# If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
|
336 |
+
if (
|
337 |
+
max_cache_length is not None
|
338 |
+
and attention_mask is not None
|
339 |
+
and cache_length + input_ids.shape[1] > max_cache_length
|
340 |
+
):
|
341 |
+
attention_mask = attention_mask[:, -max_cache_length:]
|
342 |
+
|
343 |
+
position_ids = kwargs.get("position_ids", None)
|
344 |
+
if attention_mask is not None and position_ids is None:
|
345 |
+
# create position_ids on the fly for batch generation
|
346 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
347 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
348 |
+
if past_key_values:
|
349 |
+
position_ids = position_ids[:, -input_ids.shape[1] :]
|
350 |
+
|
351 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
352 |
+
# NOTE: Injecting positional embeddings is not yet supported.
|
353 |
+
if inputs_embeds is not None and past_key_values is None:
|
354 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
355 |
+
else:
|
356 |
+
model_inputs = {"input_ids": input_ids}
|
357 |
+
|
358 |
+
model_inputs.update(
|
359 |
+
{
|
360 |
+
"position_ids": position_ids,
|
361 |
+
"past_key_values": past_key_values,
|
362 |
+
"use_cache": kwargs.get("use_cache"),
|
363 |
+
"attention_mask": attention_mask,
|
364 |
+
}
|
365 |
+
)
|
366 |
return model_inputs
|
367 |
|
368 |
+
@staticmethod
|
369 |
+
def _reorder_cache(past_key_values, beam_idx):
|
370 |
+
reordered_past = ()
|
371 |
+
for layer_past in past_key_values:
|
372 |
+
reordered_past += (
|
373 |
+
tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
|
374 |
+
)
|
375 |
+
return reordered_past
|
376 |
+
|
377 |
def _make_embedding(self, config):
|
378 |
embedding_cls = get_dynamic_class(config.embdding_cls)
|
379 |
return embedding_cls(config.vocab_size, self.d_model, config.pad_index, **config.embedding_args)
|
|
|
397 |
norm_cls = get_dynamic_class(config.norm_cls)
|
398 |
return norm_cls(self.d_model)
|
399 |
|
400 |
+
def _make_self_attention(self, layer_idx, config):
|
401 |
attention_cls = get_dynamic_class(config.attention_cls)
|
402 |
# Map HF _attn_implementation to attn_type
|
403 |
match config._attn_implementation:
|
|
|
418 |
d_model=self.d_model,
|
419 |
num_heads=config.num_attention_heads,
|
420 |
attn_type=attn_type,
|
421 |
+
layer_idx=layer_idx,
|
422 |
+
config=config,
|
423 |
**config.attention_args,
|
424 |
)
|
425 |
|
426 |
+
def _make_feedforward(self, layer_idx, config):
|
427 |
feedforward_cls = get_dynamic_class(config.feedforward_cls)
|
428 |
return feedforward_cls(
|
429 |
d_model=self.d_model,
|
430 |
feedforward_dim=config.dim_feedforward,
|
431 |
dropout=config.dropout,
|
432 |
activation=self._make_activation(config),
|
433 |
+
layer_idx=layer_idx,
|
434 |
**config.feedforward_args,
|
435 |
)
|
436 |
|
437 |
+
def _make_layer(self, layer_idx, config):
|
438 |
layer_cls = get_dynamic_class(config.layer_cls)
|
439 |
return layer_cls(
|
440 |
d_model=self.d_model,
|
441 |
dropout=self._make_dropout(config),
|
442 |
+
attention=self._make_self_attention(layer_idx, config),
|
443 |
+
feedforward=self._make_feedforward(layer_idx, config),
|
444 |
norm1=self._make_norm(config),
|
445 |
norm2=self._make_norm(config),
|
446 |
+
layer_idx=layer_idx,
|
447 |
**config.layer_args,
|
448 |
)
|
449 |
|
|
|
451 |
layer_stack_cls = get_dynamic_class(config.layer_stack_cls)
|
452 |
return layer_stack_cls(
|
453 |
layers=nn.ModuleList([
|
454 |
+
self._make_layer(layer_idx, config) for layer_idx in range(config.num_hidden_layers)
|
455 |
]),
|
456 |
**config.layer_stack_args,
|
457 |
)
|
|
|
487 |
self.sqrt_d_model = d_model**0.5
|
488 |
self.reset_parameters()
|
489 |
|
490 |
+
def forward(
|
491 |
+
self,
|
492 |
+
input_ids,
|
493 |
+
position_ids,
|
494 |
+
output_attentions,
|
495 |
+
gradient_checkpointing_func,
|
496 |
+
past_key_values,
|
497 |
+
use_cache,
|
498 |
+
output_hidden_states,
|
499 |
+
):
|
500 |
+
outputs = self.layer_stack(
|
501 |
+
self.positional_encoder(self.embedding(input_ids) * self.sqrt_d_model, position_ids),
|
502 |
+
output_attentions=output_attentions,
|
503 |
+
gradient_checkpointing_func=gradient_checkpointing_func,
|
504 |
+
past_key_values=past_key_values,
|
505 |
+
use_cache=use_cache,
|
506 |
+
output_hidden_states=output_hidden_states,
|
507 |
)
|
508 |
|
509 |
+
# Translate output states to logits.
|
510 |
+
outputs["logits"] = self.output_projection(outputs["last_hidden_state"])
|
511 |
+
del outputs["last_hidden_state"]
|
512 |
+
return outputs
|
513 |
|
514 |
def reset_parameters(self):
|
515 |
init.xavier_uniform_(self.output_projection.weight)
|
516 |
init.constant_(self.output_projection.bias, 0.)
|
517 |
init.normal_(self.embedding.weight, std=self.d_model**-0.5)
|
518 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
519 |
# Converts a torch array of integers into their equivalent binary codes.
|
520 |
def binary_tensor(x, bits):
|
521 |
mask = 2**torch.arange(bits).to(x.device, x.dtype)
|
|
|
587 |
# walsh = (hadamard_walsh_matrix(k)[:bits,:d_embed] -0.5) * self.gain
|
588 |
self.register_buffer('walsh', walsh, persistent=False)
|
589 |
|
590 |
+
def forward(self, x, position_ids=None):
|
591 |
seq_len = x.size(-2)
|
592 |
|
593 |
# Get sequence of binary codes...
|
|
|
601 |
shift = torch.randint(self.max_seq - seq_len + 1, (1,)).item()
|
602 |
seq = self.binary_code[shift:seq_len + shift,:]
|
603 |
|
604 |
+
# When the cache is used for generation, after the first call, we are only passed a single token at a time,
|
605 |
+
# with the remaining tokens being in the cache. We need to make sure that the newly injected tokens have the
|
606 |
+
# correct relative position by indexing the codes with the position_ids.
|
607 |
+
elif position_ids != None:
|
608 |
+
seq = self.binary_code[position_ids, :]
|
609 |
+
|
610 |
# Disable shifting when not training. This does not appear to change the evaluation loss, but
|
611 |
# it does makes predictions easier to analyse when the attention weights are not shifting with each step.
|
612 |
else:
|
|
|
629 |
super().__init__()
|
630 |
self.layers = layers
|
631 |
|
632 |
+
def forward(
|
633 |
+
self,
|
634 |
+
hidden_states,
|
635 |
+
output_attentions,
|
636 |
+
past_key_values,
|
637 |
+
use_cache,
|
638 |
+
output_hidden_states,
|
639 |
+
gradient_checkpointing_func=None,
|
640 |
+
):
|
641 |
+
present_key_value = None
|
642 |
+
all_attentions = [] if output_attentions else None
|
643 |
+
all_hidden_states = [hidden_states] if output_hidden_states else None
|
644 |
+
|
645 |
for layer in self.layers:
|
646 |
if gradient_checkpointing_func is not None:
|
647 |
+
layer_outputs = gradient_checkpointing_func(
|
648 |
layer.__call__,
|
649 |
+
hidden_states,
|
650 |
+
output_attentions,
|
651 |
+
past_key_values,
|
652 |
+
use_cache,
|
653 |
+
use_reentrant=False,
|
654 |
)
|
655 |
else:
|
656 |
+
layer_outputs = layer(
|
657 |
+
hidden_states,
|
658 |
+
output_attentions,
|
659 |
+
past_key_values,
|
660 |
+
use_cache,
|
661 |
+
)
|
662 |
|
663 |
+
hidden_states = layer_outputs["hidden_states"]
|
664 |
+
|
665 |
+
if output_hidden_states:
|
666 |
+
all_hidden_states.append(hidden_states)
|
667 |
+
|
668 |
+
if use_cache:
|
669 |
+
present_key_value = layer_outputs["past_key_values"]
|
670 |
+
|
671 |
+
if output_attentions:
|
672 |
+
all_attentions.append(layer_outputs["attentions"])
|
673 |
+
|
674 |
+
return dict(
|
675 |
+
last_hidden_state=hidden_states,
|
676 |
+
past_key_values=present_key_value,
|
677 |
+
hidden_states=hidden_states,
|
678 |
+
attentions=all_attentions,
|
679 |
+
)
|
680 |
|
681 |
# DeepNet: Scaling Transformers to 1,000 Layers
|
682 |
# https://arxiv.org/abs/2203.00555
|
683 |
+
# Note: This is a type of Pre-Layer-Norm Transformer layer.
|
684 |
class DeepnetLayer(nn.Module):
|
685 |
def __init__(
|
686 |
self,
|
|
|
690 |
norm1,
|
691 |
norm2,
|
692 |
dropout,
|
693 |
+
layer_idx,
|
694 |
alpha=1.0,
|
695 |
):
|
696 |
super().__init__()
|
|
|
702 |
self.dropout = dropout
|
703 |
# Deepnet alpha
|
704 |
self.alpha = alpha
|
705 |
+
self.layer_idx = layer_idx
|
706 |
|
707 |
+
def forward(
|
708 |
+
self,
|
709 |
+
hidden_states,
|
710 |
+
output_attentions,
|
711 |
+
past_key_values,
|
712 |
+
use_cache,
|
713 |
+
):
|
714 |
# Keep input as residual
|
715 |
+
residual = hidden_states * self.alpha
|
716 |
|
717 |
# Compute attention
|
718 |
+
attn_outputs = self.attention(
|
719 |
+
hidden_states,
|
720 |
+
past_key_values=past_key_values,
|
721 |
+
use_cache=use_cache,
|
722 |
+
output_attentions=output_attentions
|
723 |
+
)
|
724 |
+
|
725 |
+
hidden_states = attn_outputs["hidden_states"]
|
726 |
|
727 |
# Add attention with residual and normalize.
|
728 |
+
hidden_states = self.norm1(residual + self.dropout(hidden_states))
|
729 |
|
730 |
# Keep output as next residual.
|
731 |
+
residual = hidden_states * self.alpha
|
732 |
|
733 |
# Pass through feedforward network.
|
734 |
+
hidden_states = self.feedforward(hidden_states)
|
735 |
|
736 |
# Combine residual and ff output, then normalize again.
|
737 |
+
hidden_states = self.norm2(residual + self.dropout(hidden_states))
|
738 |
|
739 |
+
return dict(
|
740 |
+
hidden_states=hidden_states,
|
741 |
+
attentions=attn_outputs["attentions"],
|
742 |
+
past_key_values=attn_outputs["past_key_values"]
|
743 |
+
)
|
744 |
|
745 |
# A vanilla MLP transfomer layer.
|
746 |
class FeedforwardLayer(nn.Module):
|
|
|
749 |
d_model: int,
|
750 |
feedforward_dim: int,
|
751 |
dropout,
|
752 |
+
layer_idx,
|
753 |
activation=nn.ReLU(),
|
754 |
beta=1.0,
|
755 |
bias=True,
|
|
|
772 |
init.constant_(self.linear1.bias, 0.)
|
773 |
init.constant_(self.linear2.bias, 0.)
|
774 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
775 |
class CausalSelfAttention(nn.Module):
|
776 |
def __init__(
|
777 |
self,
|
|
|
782 |
# torch: Use pytorch "scaled_dot_product_attention()"; faster; generally good compatibility; does not support returning attn weights.
|
783 |
# flash2: Use Flash-Attention2 implementation; fastest; limited to int16 and bfloat16 types; least memory usage.
|
784 |
attn_type,
|
785 |
+
layer_idx,
|
786 |
+
config,
|
787 |
beta=1.0,
|
788 |
dropout=0.1,
|
789 |
):
|
|
|
792 |
self.num_heads = num_heads
|
793 |
self.beta = beta
|
794 |
self.attn_type = attn_type
|
795 |
+
self.layer_idx = layer_idx
|
796 |
+
self.config = config
|
797 |
|
798 |
assert d_model % num_heads == 0, "d_model must be evenly divisible by num_heads"
|
799 |
|
|
|
824 |
init.constant_(self.in_proj.bias, 0.)
|
825 |
init.constant_(self.output_linear.bias, 0.)
|
826 |
|
827 |
+
# Project QKV input through input matrices, reshape to (batch_size, n_heads, seq_len, d_model), and apply cache.
|
828 |
+
def project_input(self, qkv, past_key_values):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
829 |
batch_size, seq_len, d_embed = qkv.shape
|
830 |
+
proj = self.in_proj(qkv)
|
831 |
+
query, key, value = proj.chunk(chunks=3, dim=-1)
|
832 |
|
|
|
|
|
|
|
833 |
# Split projections into multiple heads and swap position of sequence / heads dimension
|
834 |
query = query.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
835 |
key = key.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
836 |
value = value.view(batch_size, seq_len, self.num_heads, self.d_head).transpose(1, 2)
|
837 |
|
838 |
+
# Update the cache values.
|
839 |
+
if past_key_values is not None:
|
840 |
+
key, value = past_key_values.update(key, value, self.layer_idx)
|
841 |
+
return query, key, value
|
842 |
+
|
843 |
+
def forward(
|
844 |
+
self,
|
845 |
+
qkv,
|
846 |
+
output_attentions,
|
847 |
+
past_key_values,
|
848 |
+
use_cache,
|
849 |
+
):
|
850 |
+
attn_type = self.attn_type
|
851 |
+
if output_attentions and attn_type != "native":
|
852 |
+
logger.warning_once(
|
853 |
+
"CausalSelfAttention(output_attentions=True) and attn_type is not 'native': "
|
854 |
+
"Forcing native attention."
|
855 |
+
)
|
856 |
+
attn_type = "native"
|
857 |
+
|
858 |
+
if attn_type == "flash2":
|
859 |
+
if use_cache is None or use_cache == False:
|
860 |
+
return self.flash2_forward(qkv)
|
861 |
+
else:
|
862 |
+
return self.flash2_forward_cached(qkv, past_key_values)
|
863 |
+
|
864 |
+
# qkv: (batch_size, seq_len, d_embed)
|
865 |
+
batch_size, seq_len, d_embed = qkv.shape
|
866 |
+
|
867 |
+
# Feed the inputs through the K, Q, V matrices.
|
868 |
+
query, key, value = self.project_input(qkv, past_key_values)
|
869 |
+
kv_seq_len = key.shape[-2]
|
870 |
+
|
871 |
# Default to returning empty attention weights.
|
872 |
+
attentions = None
|
873 |
+
|
874 |
+
# https://github.com/pytorch/pytorch/issues/112577
|
875 |
|
876 |
+
if attn_type == "torch":
|
877 |
# This context manager can be used to force which implementation to use.
|
878 |
#with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
|
879 |
attended_values = F.scaled_dot_product_attention(
|
|
|
882 |
value,
|
883 |
attn_mask=None,
|
884 |
dropout_p=self.dropout.p if self.training else 0.0,
|
885 |
+
is_causal=(seq_len > 1),
|
886 |
scale=self.dot_product_scale
|
887 |
)
|
888 |
# "native" scaled-dot-product attention implementation.
|
|
|
891 |
scores = torch.matmul(query, key.transpose(-2, -1)) * self.dot_product_scale
|
892 |
|
893 |
# Mask future positions from the past
|
894 |
+
if seq_len > 1:
|
895 |
+
scores.masked_fill_(
|
896 |
+
torch.tril(
|
897 |
+
torch.ones(seq_len, kv_seq_len, dtype=torch.bool, device=qkv.device),
|
898 |
+
diagonal=0,
|
899 |
+
).logical_not(),
|
900 |
+
float('-inf'),
|
901 |
+
)
|
902 |
|
903 |
# Calculate the attention weights; avoid NANs that might emerge from zeros in softmax's denominator
|
904 |
+
attentions = self.dropout(torch.softmax(scores, dim=-1).clamp(min=1e-10))
|
905 |
del scores
|
906 |
|
907 |
# Use the attention weights to get a weighted combination of value vectors
|
908 |
+
attended_values = torch.matmul(attentions, value)
|
909 |
+
if not output_attentions:
|
910 |
+
del attentions
|
911 |
+
attentions = None
|
912 |
|
913 |
# Concatenate attention heads and project to original embedding size using the output linear layer
|
914 |
attended_values = attended_values.transpose(1, 2).contiguous().view(batch_size, seq_len, d_embed)
|
915 |
|
916 |
# Project the concatenated output through the output matrix.
|
917 |
attended_values = self.output_linear(attended_values)
|
918 |
+
return dict(
|
919 |
+
hidden_states=attended_values,
|
920 |
+
attentions=attentions,
|
921 |
+
past_key_values=past_key_values
|
922 |
+
)
|
923 |
+
|
924 |
+
# No cache support, but faster
|
925 |
+
def flash2_forward(
|
926 |
+
self,
|
927 |
+
qkv,
|
928 |
+
):
|
929 |
batch_size, seq_len, d_embed = qkv.shape
|
930 |
|
931 |
# Feed the inputs through the K, Q, V matrices.
|
932 |
# query : (batch_size, seq_len, d_model)
|
933 |
# qkv : (batch_size, seq_len, 3, num_heads, d_kq)
|
934 |
+
# Feed the inputs through the K, Q, V matrices.
|
935 |
+
# query : (batch_size, seq_len, d_model)
|
936 |
+
# qkv : (batch_size, seq_len, 3, num_heads, d_kq)
|
937 |
+
|
938 |
qkv = self.in_proj(qkv).unflatten(
|
939 |
-1,
|
940 |
(3, self.num_heads, self.d_head)
|
941 |
)
|
942 |
+
|
943 |
attended_values = flash_attn_qkvpacked_func(
|
944 |
+
self._downcast_to_float16(qkv)[0],
|
945 |
dropout_p=self.dropout.p if self.training else 0.0,
|
946 |
softmax_scale=self.dot_product_scale,
|
947 |
causal=True,
|
|
|
953 |
|
954 |
# Project the concatenated output through the output matrix.
|
955 |
attended_values = self.output_linear(attended_values)
|
956 |
+
return dict(
|
957 |
+
hidden_states=attended_values,
|
958 |
+
attentions=None,
|
959 |
+
past_key_values=None
|
960 |
+
)
|
|
|
|
|
|
|
|
|
961 |
|
962 |
+
# See https://github.com/huggingface/transformers/blob/main/src/transformers/cache_utils.py
|
963 |
+
#https://huggingface.co/docs/transformers/internal/generation_utils
|
964 |
+
def flash2_forward_cached(
|
965 |
self,
|
966 |
+
qkv,
|
967 |
+
past_key_values,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
968 |
):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
969 |
batch_size, seq_len, d_embed = qkv.shape
|
970 |
|
971 |
# Feed the inputs through the K, Q, V matrices.
|
972 |
+
query, key, value = self.project_input(qkv, past_key_values)
|
973 |
+
query, key, value = self._downcast_to_float16(query, key, value)
|
974 |
+
|
975 |
+
# Expected inputs to flash2:
|
976 |
+
# q: (batch_size, seqlen, nheads, headdim)
|
977 |
+
# k: (batch_size, seqlen, nheads_k, headdim)
|
978 |
+
# v: (batch_size, seqlen, nheads_k, headdim)
|
979 |
+
query = query.transpose(1, 2)
|
980 |
+
key = key.transpose(1, 2)
|
981 |
+
value = value.transpose(1, 2)
|
982 |
+
|
983 |
+
attended_values = flash_attn_func(
|
984 |
+
q=query,
|
985 |
+
k=key,
|
986 |
+
v=value,
|
987 |
dropout_p=self.dropout.p if self.training else 0.0,
|
988 |
softmax_scale=self.dot_product_scale,
|
989 |
causal=True,
|
990 |
+
)
|
|
|
|
|
991 |
# attended_values: (batch_size, seqlen, nheads, headdim)
|
992 |
+
|
993 |
# Concatentate heads back into d_embed
|
994 |
attended_values = attended_values.view(batch_size, seq_len, d_embed)
|
995 |
|
996 |
# Project the concatenated output through the output matrix.
|
997 |
attended_values = self.output_linear(attended_values)
|
998 |
+
return dict(
|
999 |
+
hidden_states=attended_values,
|
1000 |
+
attentions=None,
|
1001 |
+
past_key_values=past_key_values
|
1002 |
+
)
|
1003 |
+
|
1004 |
+
def _downcast_to_float16(self, *args):
|
1005 |
+
if args[0].dtype != torch.float32:
|
1006 |
+
return args
|
1007 |
+
|
1008 |
+
if torch.is_autocast_enabled():
|
1009 |
+
target_dtype = torch.get_autocast_gpu_dtype()
|
1010 |
+
# Handle the case where the model is quantized
|
1011 |
+
elif hasattr(self.config, "_pre_quantization_dtype"):
|
1012 |
+
target_dtype = self.config._pre_quantization_dtype
|
1013 |
+
else:
|
1014 |
+
target_dtype = self.output_linear.weight.dtype
|
1015 |
+
|
1016 |
+
logger.warning_once(
|
1017 |
+
f"The input hidden states seems to be silently casted in float32, this might be related to"
|
1018 |
+
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
|
1019 |
+
f" {target_dtype}."
|
1020 |
+
)
|
1021 |
+
|
1022 |
+
return (arg.to(target_dtype) for arg in args)
|