add flash-attn support
Browse files- modeling_internlm2.py +151 -20
modeling_internlm2.py
CHANGED
@@ -25,6 +25,7 @@ import torch
|
|
25 |
import torch.utils.checkpoint
|
26 |
from einops import rearrange
|
27 |
from torch import nn
|
|
|
28 |
from transformers.activations import ACT2FN
|
29 |
from transformers.modeling_outputs import BaseModelOutputWithPast
|
30 |
from transformers.modeling_utils import PreTrainedModel
|
@@ -42,6 +43,30 @@ logger = logging.get_logger(__name__)
|
|
42 |
|
43 |
_CONFIG_FOR_DOC = 'InternLM2Config'
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
|
46 |
# Copied from transformers.models.bart.modeling_bart._make_causal_mask
|
47 |
def _make_causal_mask(input_ids_shape: torch.Size,
|
@@ -264,21 +289,21 @@ class InternLM2MLP(nn.Module):
|
|
264 |
bias=False,
|
265 |
lora_r=256,
|
266 |
lora_alpha=256,
|
267 |
-
lora_len=
|
268 |
self.w3 = PLoRA(
|
269 |
self.hidden_size,
|
270 |
self.intermediate_size,
|
271 |
bias=False,
|
272 |
lora_r=256,
|
273 |
lora_alpha=256,
|
274 |
-
lora_len=
|
275 |
self.w2 = PLoRA(
|
276 |
self.intermediate_size,
|
277 |
self.hidden_size,
|
278 |
bias=False,
|
279 |
lora_r=256,
|
280 |
lora_alpha=256,
|
281 |
-
lora_len=
|
282 |
|
283 |
self.act_fn = ACT2FN[config.hidden_act]
|
284 |
|
@@ -332,7 +357,7 @@ class InternLM2Attention(nn.Module):
|
|
332 |
bias=config.bias,
|
333 |
lora_r=256,
|
334 |
lora_alpha=256,
|
335 |
-
lora_len=
|
336 |
|
337 |
self.wo = PLoRA(
|
338 |
self.num_heads * self.head_dim,
|
@@ -340,7 +365,7 @@ class InternLM2Attention(nn.Module):
|
|
340 |
bias=config.bias,
|
341 |
lora_r=256,
|
342 |
lora_alpha=256,
|
343 |
-
lora_len=
|
344 |
self._init_rope()
|
345 |
|
346 |
def _init_rope(self):
|
@@ -498,7 +523,7 @@ class InternLM2FlashAttention2(InternLM2Attention):
|
|
498 |
qkv_states = rearrange(
|
499 |
qkv_states,
|
500 |
'b q (h gs d) -> b q h gs d',
|
501 |
-
gs=
|
502 |
d=self.head_dim,
|
503 |
q=q_len,
|
504 |
)
|
@@ -507,6 +532,10 @@ class InternLM2FlashAttention2(InternLM2Attention):
|
|
507 |
query_states = rearrange(query_states, 'b q h gs d -> b q (h gs) d')
|
508 |
key_states = qkv_states[..., -2, :]
|
509 |
value_states = qkv_states[..., -1, :]
|
|
|
|
|
|
|
|
|
510 |
|
511 |
kv_seq_len = key_states.shape[-2]
|
512 |
if past_key_value is not None:
|
@@ -523,12 +552,12 @@ class InternLM2FlashAttention2(InternLM2Attention):
|
|
523 |
value_states = torch.cat([past_key_value[1], value_states], dim=2)
|
524 |
|
525 |
past_key_value = (key_states, value_states) if use_cache else None
|
526 |
-
|
527 |
query_states = query_states.transpose(1, 2)
|
528 |
key_states = key_states.transpose(1, 2)
|
529 |
value_states = value_states.transpose(1, 2)
|
530 |
|
531 |
-
dropout_rate = 0.0 if not self.training else self.
|
532 |
|
533 |
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
|
534 |
# therefore the input hidden states gets silently casted in float32. Hence, we need
|
@@ -569,17 +598,110 @@ class InternLM2FlashAttention2(InternLM2Attention):
|
|
569 |
attn_weights = None
|
570 |
|
571 |
return attn_output, attn_weights, past_key_value
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
572 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
573 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
574 |
class InternLM2DecoderLayer(nn.Module):
|
575 |
|
576 |
def __init__(self, config: InternLM2Config):
|
577 |
super().__init__()
|
578 |
self.hidden_size = config.hidden_size
|
579 |
-
self.attention = (
|
580 |
-
|
581 |
-
if not getattr(config, '_flash_attn_2_enabled', False) else
|
582 |
-
InternLM2FlashAttention2(config=config))
|
583 |
self.feed_forward = InternLM2MLP(config)
|
584 |
self.attention_norm = InternLM2RMSNorm(
|
585 |
config.hidden_size, eps=config.rms_norm_eps)
|
@@ -773,6 +895,8 @@ class InternLM2Model(InternLM2PreTrainedModel):
|
|
773 |
|
774 |
def __init__(self, config: InternLM2Config):
|
775 |
super().__init__(config)
|
|
|
|
|
776 |
self.padding_idx = config.pad_token_id
|
777 |
self.vocab_size = config.vocab_size
|
778 |
|
@@ -843,6 +967,9 @@ class InternLM2Model(InternLM2PreTrainedModel):
|
|
843 |
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
844 |
|
845 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
|
|
|
|
|
|
846 |
|
847 |
# retrieve input_ids and inputs_embeds
|
848 |
if input_ids is not None and inputs_embeds is not None:
|
@@ -876,14 +1003,18 @@ class InternLM2Model(InternLM2PreTrainedModel):
|
|
876 |
inputs_embeds = self.tok_embeddings(input_ids)
|
877 |
im_mask = torch.zeros(inputs_embeds.shape[:2]).to(
|
878 |
inputs_embeds.device).bool()
|
879 |
-
|
880 |
-
if
|
881 |
-
|
882 |
-
|
883 |
-
|
884 |
-
|
885 |
-
|
886 |
-
|
|
|
|
|
|
|
|
|
887 |
|
888 |
# embed positions
|
889 |
hidden_states = inputs_embeds
|
|
|
25 |
import torch.utils.checkpoint
|
26 |
from einops import rearrange
|
27 |
from torch import nn
|
28 |
+
import torch.nn.functional as F
|
29 |
from transformers.activations import ACT2FN
|
30 |
from transformers.modeling_outputs import BaseModelOutputWithPast
|
31 |
from transformers.modeling_utils import PreTrainedModel
|
|
|
43 |
|
44 |
_CONFIG_FOR_DOC = 'InternLM2Config'
|
45 |
|
46 |
+
flash_attn_func, flash_attn_varlen_func = None, None
|
47 |
+
pad_input, index_first_axis, unpad_input = None, None, None
|
48 |
+
def _import_flash_attn():
|
49 |
+
global flash_attn_func, flash_attn_varlen_func
|
50 |
+
global pad_input, index_first_axis, unpad_input
|
51 |
+
try:
|
52 |
+
from flash_attn import flash_attn_func as _flash_attn_func, flash_attn_varlen_func as _flash_attn_varlen_func
|
53 |
+
from flash_attn.bert_padding import pad_input as _pad_input, index_first_axis as _index_first_axis, unpad_input as _unpad_input
|
54 |
+
flash_attn_func, flash_attn_varlen_func = _flash_attn_func, _flash_attn_varlen_func
|
55 |
+
pad_input, index_first_axis, unpad_input = _pad_input, _index_first_axis, _unpad_input
|
56 |
+
except ImportError:
|
57 |
+
raise ImportError("flash_attn is not installed.")
|
58 |
+
|
59 |
+
# Copied from transformers.models.llama.modeling_llama._get_unpad_data
|
60 |
+
def _get_unpad_data(attention_mask):
|
61 |
+
seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
|
62 |
+
indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
|
63 |
+
max_seqlen_in_batch = seqlens_in_batch.max().item()
|
64 |
+
cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
|
65 |
+
return (
|
66 |
+
indices,
|
67 |
+
cu_seqlens,
|
68 |
+
max_seqlen_in_batch,
|
69 |
+
)
|
70 |
|
71 |
# Copied from transformers.models.bart.modeling_bart._make_causal_mask
|
72 |
def _make_causal_mask(input_ids_shape: torch.Size,
|
|
|
289 |
bias=False,
|
290 |
lora_r=256,
|
291 |
lora_alpha=256,
|
292 |
+
lora_len=1225)
|
293 |
self.w3 = PLoRA(
|
294 |
self.hidden_size,
|
295 |
self.intermediate_size,
|
296 |
bias=False,
|
297 |
lora_r=256,
|
298 |
lora_alpha=256,
|
299 |
+
lora_len=1225)
|
300 |
self.w2 = PLoRA(
|
301 |
self.intermediate_size,
|
302 |
self.hidden_size,
|
303 |
bias=False,
|
304 |
lora_r=256,
|
305 |
lora_alpha=256,
|
306 |
+
lora_len=1225)
|
307 |
|
308 |
self.act_fn = ACT2FN[config.hidden_act]
|
309 |
|
|
|
357 |
bias=config.bias,
|
358 |
lora_r=256,
|
359 |
lora_alpha=256,
|
360 |
+
lora_len=1225)
|
361 |
|
362 |
self.wo = PLoRA(
|
363 |
self.num_heads * self.head_dim,
|
|
|
365 |
bias=config.bias,
|
366 |
lora_r=256,
|
367 |
lora_alpha=256,
|
368 |
+
lora_len=1225)
|
369 |
self._init_rope()
|
370 |
|
371 |
def _init_rope(self):
|
|
|
523 |
qkv_states = rearrange(
|
524 |
qkv_states,
|
525 |
'b q (h gs d) -> b q h gs d',
|
526 |
+
gs=2 + self.num_key_value_groups,
|
527 |
d=self.head_dim,
|
528 |
q=q_len,
|
529 |
)
|
|
|
532 |
query_states = rearrange(query_states, 'b q h gs d -> b q (h gs) d')
|
533 |
key_states = qkv_states[..., -2, :]
|
534 |
value_states = qkv_states[..., -1, :]
|
535 |
+
|
536 |
+
query_states = query_states.transpose(1, 2)
|
537 |
+
key_states = key_states.transpose(1, 2)
|
538 |
+
value_states = value_states.transpose(1, 2)
|
539 |
|
540 |
kv_seq_len = key_states.shape[-2]
|
541 |
if past_key_value is not None:
|
|
|
552 |
value_states = torch.cat([past_key_value[1], value_states], dim=2)
|
553 |
|
554 |
past_key_value = (key_states, value_states) if use_cache else None
|
555 |
+
|
556 |
query_states = query_states.transpose(1, 2)
|
557 |
key_states = key_states.transpose(1, 2)
|
558 |
value_states = value_states.transpose(1, 2)
|
559 |
|
560 |
+
dropout_rate = 0.0 if not self.training else getattr(self, "dropout_rate", 0.0)
|
561 |
|
562 |
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
|
563 |
# therefore the input hidden states gets silently casted in float32. Hence, we need
|
|
|
598 |
attn_weights = None
|
599 |
|
600 |
return attn_output, attn_weights, past_key_value
|
601 |
+
|
602 |
+
def _flash_attention_forward(
|
603 |
+
self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
|
604 |
+
):
|
605 |
+
"""
|
606 |
+
Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
|
607 |
+
first unpad the input, then computes the attention scores and pad the final attention scores.
|
608 |
+
Args:
|
609 |
+
query_states (`torch.Tensor`):
|
610 |
+
Input query states to be passed to Flash Attention API
|
611 |
+
key_states (`torch.Tensor`):
|
612 |
+
Input key states to be passed to Flash Attention API
|
613 |
+
value_states (`torch.Tensor`):
|
614 |
+
Input value states to be passed to Flash Attention API
|
615 |
+
attention_mask (`torch.Tensor`):
|
616 |
+
The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
|
617 |
+
position of padding tokens and 1 for the position of non-padding tokens.
|
618 |
+
dropout (`int`, *optional*):
|
619 |
+
Attention dropout
|
620 |
+
softmax_scale (`float`, *optional*):
|
621 |
+
The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
|
622 |
+
"""
|
623 |
+
# Contains at least one padding token in the sequence
|
624 |
+
causal = self.is_causal and query_length != 1
|
625 |
+
if attention_mask is not None:
|
626 |
+
batch_size = query_states.shape[0]
|
627 |
+
query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._unpad_input(
|
628 |
+
query_states, key_states, value_states, attention_mask, query_length
|
629 |
+
)
|
630 |
|
631 |
+
cu_seqlens_q, cu_seqlens_k = cu_seq_lens
|
632 |
+
max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
|
633 |
+
|
634 |
+
attn_output_unpad = flash_attn_varlen_func(
|
635 |
+
query_states,
|
636 |
+
key_states,
|
637 |
+
value_states,
|
638 |
+
cu_seqlens_q=cu_seqlens_q,
|
639 |
+
cu_seqlens_k=cu_seqlens_k,
|
640 |
+
max_seqlen_q=max_seqlen_in_batch_q,
|
641 |
+
max_seqlen_k=max_seqlen_in_batch_k,
|
642 |
+
dropout_p=dropout,
|
643 |
+
softmax_scale=softmax_scale,
|
644 |
+
causal=causal,
|
645 |
+
)
|
646 |
+
|
647 |
+
attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
|
648 |
+
else:
|
649 |
+
attn_output = flash_attn_func(
|
650 |
+
query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
|
651 |
+
)
|
652 |
|
653 |
+
return attn_output
|
654 |
+
|
655 |
+
def _unpad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
|
656 |
+
indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
|
657 |
+
batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
|
658 |
+
|
659 |
+
key_layer = index_first_axis(
|
660 |
+
key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
|
661 |
+
)
|
662 |
+
value_layer = index_first_axis(
|
663 |
+
value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
|
664 |
+
)
|
665 |
+
|
666 |
+
if query_length == kv_seq_len:
|
667 |
+
query_layer = index_first_axis(
|
668 |
+
query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
|
669 |
+
)
|
670 |
+
cu_seqlens_q = cu_seqlens_k
|
671 |
+
max_seqlen_in_batch_q = max_seqlen_in_batch_k
|
672 |
+
indices_q = indices_k
|
673 |
+
elif query_length == 1:
|
674 |
+
max_seqlen_in_batch_q = 1
|
675 |
+
cu_seqlens_q = torch.arange(
|
676 |
+
batch_size + 1, dtype=torch.int32, device=query_layer.device
|
677 |
+
) # There is a memcpy here, that is very bad.
|
678 |
+
indices_q = cu_seqlens_q[:-1]
|
679 |
+
query_layer = query_layer.squeeze(1)
|
680 |
+
else:
|
681 |
+
# The -q_len: slice assumes left padding.
|
682 |
+
attention_mask = attention_mask[:, -query_length:]
|
683 |
+
query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
|
684 |
+
|
685 |
+
return (
|
686 |
+
query_layer,
|
687 |
+
key_layer,
|
688 |
+
value_layer,
|
689 |
+
indices_q.to(torch.int64),
|
690 |
+
(cu_seqlens_q, cu_seqlens_k),
|
691 |
+
(max_seqlen_in_batch_q, max_seqlen_in_batch_k),
|
692 |
+
)
|
693 |
+
|
694 |
+
INTERNLM2_ATTENTION_CLASSES = {
|
695 |
+
"eager": InternLM2Attention,
|
696 |
+
"flash_attention_2": InternLM2FlashAttention2,
|
697 |
+
}
|
698 |
class InternLM2DecoderLayer(nn.Module):
|
699 |
|
700 |
def __init__(self, config: InternLM2Config):
|
701 |
super().__init__()
|
702 |
self.hidden_size = config.hidden_size
|
703 |
+
self.attention = INTERNLM2_ATTENTION_CLASSES[config.attn_implementation](config=config)
|
704 |
+
|
|
|
|
|
705 |
self.feed_forward = InternLM2MLP(config)
|
706 |
self.attention_norm = InternLM2RMSNorm(
|
707 |
config.hidden_size, eps=config.rms_norm_eps)
|
|
|
895 |
|
896 |
def __init__(self, config: InternLM2Config):
|
897 |
super().__init__(config)
|
898 |
+
print(f"Attention Implementation: {self.config.attn_implementation}")
|
899 |
+
|
900 |
self.padding_idx = config.pad_token_id
|
901 |
self.vocab_size = config.vocab_size
|
902 |
|
|
|
967 |
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
968 |
|
969 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
970 |
+
|
971 |
+
if self.config.attn_implementation == "flash_attention_2":
|
972 |
+
_import_flash_attn()
|
973 |
|
974 |
# retrieve input_ids and inputs_embeds
|
975 |
if input_ids is not None and inputs_embeds is not None:
|
|
|
1003 |
inputs_embeds = self.tok_embeddings(input_ids)
|
1004 |
im_mask = torch.zeros(inputs_embeds.shape[:2]).to(
|
1005 |
inputs_embeds.device).bool()
|
1006 |
+
|
1007 |
+
if self.config.attn_implementation == "flash_attention_2":
|
1008 |
+
# 2d mask is passed through the layers
|
1009 |
+
attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
|
1010 |
+
else:
|
1011 |
+
if attention_mask is None:
|
1012 |
+
attention_mask = torch.ones((batch_size, seq_length_with_past),
|
1013 |
+
dtype=torch.bool,
|
1014 |
+
device=inputs_embeds.device)
|
1015 |
+
attention_mask = self._prepare_decoder_attention_mask(
|
1016 |
+
attention_mask, (batch_size, seq_length), inputs_embeds,
|
1017 |
+
past_key_values_length)
|
1018 |
|
1019 |
# embed positions
|
1020 |
hidden_states = inputs_embeds
|