jupyterjazz
commited on
feat: merge with recent changes
Browse filesSigned-off-by: jupyterjazz <saba.sturua@jina.ai>
- block.py +1 -1
- configuration_xlm_roberta.py +2 -0
- embedding.py +2 -2
- mha.py +2 -2
- mlp.py +2 -2
- modeling_lora.py +23 -19
- modeling_xlm_roberta.py +15 -21
- rotary.py +44 -21
block.py
CHANGED
@@ -233,7 +233,7 @@ class Block(nn.Module):
|
|
233 |
is_rms_norm=isinstance(self.norm1, RMSNorm),
|
234 |
)
|
235 |
if not isinstance(self.mlp, nn.Identity):
|
236 |
-
mlp_out = self.mlp(hidden_states,
|
237 |
if self.return_residual: # mlp out is actually a pair here
|
238 |
mlp_out, hidden_states = mlp_out
|
239 |
if not self.fused_dropout_add_ln:
|
|
|
233 |
is_rms_norm=isinstance(self.norm1, RMSNorm),
|
234 |
)
|
235 |
if not isinstance(self.mlp, nn.Identity):
|
236 |
+
mlp_out = self.mlp(hidden_states, task_type=mixer_kwargs.get('task_type'))
|
237 |
if self.return_residual: # mlp out is actually a pair here
|
238 |
mlp_out, hidden_states = mlp_out
|
239 |
if not self.fused_dropout_add_ln:
|
configuration_xlm_roberta.py
CHANGED
@@ -23,6 +23,7 @@ class XLMRobertaFlashConfig(PretrainedConfig):
|
|
23 |
use_cache=True,
|
24 |
classifier_dropout=None,
|
25 |
lora_adaptations=None,
|
|
|
26 |
lora_rank=4,
|
27 |
lora_dropout_p=0.0,
|
28 |
lora_alpha=1,
|
@@ -55,6 +56,7 @@ class XLMRobertaFlashConfig(PretrainedConfig):
|
|
55 |
self.classifier_dropout = classifier_dropout
|
56 |
self.load_trained_adapters = load_trained_adapters
|
57 |
self.lora_adaptations = lora_adaptations
|
|
|
58 |
self.lora_rank = lora_rank
|
59 |
self.lora_dropout_p = lora_dropout_p
|
60 |
self.lora_alpha = lora_alpha
|
|
|
23 |
use_cache=True,
|
24 |
classifier_dropout=None,
|
25 |
lora_adaptations=None,
|
26 |
+
lora_prompts=None,
|
27 |
lora_rank=4,
|
28 |
lora_dropout_p=0.0,
|
29 |
lora_alpha=1,
|
|
|
56 |
self.classifier_dropout = classifier_dropout
|
57 |
self.load_trained_adapters = load_trained_adapters
|
58 |
self.lora_adaptations = lora_adaptations
|
59 |
+
self.lora_prompts = lora_prompts
|
60 |
self.lora_rank = lora_rank
|
61 |
self.lora_dropout_p = lora_dropout_p
|
62 |
self.lora_alpha = lora_alpha
|
embedding.py
CHANGED
@@ -40,14 +40,14 @@ class XLMRobertaEmbeddings(nn.Module):
|
|
40 |
if self.type_vocab_size > 0:
|
41 |
self.token_type_embeddings = nn.Embedding(type_vocab_size, embed_dim, **factory_kwargs)
|
42 |
|
43 |
-
def forward(self, input_ids, position_ids=None, token_type_ids=None,
|
44 |
"""
|
45 |
input_ids: (batch, seqlen)
|
46 |
position_ids: (batch, seqlen)
|
47 |
token_type_ids: (batch, seqlen)
|
48 |
"""
|
49 |
batch_size, seqlen = input_ids.shape
|
50 |
-
lora_kwargs = {'
|
51 |
embeddings = self.word_embeddings(input_ids, **lora_kwargs)
|
52 |
if self.max_position_embeddings > 0:
|
53 |
if position_ids is None:
|
|
|
40 |
if self.type_vocab_size > 0:
|
41 |
self.token_type_embeddings = nn.Embedding(type_vocab_size, embed_dim, **factory_kwargs)
|
42 |
|
43 |
+
def forward(self, input_ids, position_ids=None, token_type_ids=None, task_type=None):
|
44 |
"""
|
45 |
input_ids: (batch, seqlen)
|
46 |
position_ids: (batch, seqlen)
|
47 |
token_type_ids: (batch, seqlen)
|
48 |
"""
|
49 |
batch_size, seqlen = input_ids.shape
|
50 |
+
lora_kwargs = {'task_type': task_type} if task_type is not None else {}
|
51 |
embeddings = self.word_embeddings(input_ids, **lora_kwargs)
|
52 |
if self.max_position_embeddings > 0:
|
53 |
if position_ids is None:
|
mha.py
CHANGED
@@ -590,7 +590,7 @@ class MHA(nn.Module):
|
|
590 |
max_seqlen=None,
|
591 |
mixer_subset=None,
|
592 |
inference_params=None,
|
593 |
-
|
594 |
**kwargs,
|
595 |
):
|
596 |
"""
|
@@ -645,7 +645,7 @@ class MHA(nn.Module):
|
|
645 |
batch, seqlen = x.shape[:2]
|
646 |
if not self.cross_attn and self.num_heads_kv == self.num_heads:
|
647 |
assert x_kv is None and mixer_subset is None
|
648 |
-
lora_kwargs = {'
|
649 |
if not self.return_residual:
|
650 |
qkv = self.Wqkv(x, **lora_kwargs)
|
651 |
else:
|
|
|
590 |
max_seqlen=None,
|
591 |
mixer_subset=None,
|
592 |
inference_params=None,
|
593 |
+
task_type=None,
|
594 |
**kwargs,
|
595 |
):
|
596 |
"""
|
|
|
645 |
batch, seqlen = x.shape[:2]
|
646 |
if not self.cross_attn and self.num_heads_kv == self.num_heads:
|
647 |
assert x_kv is None and mixer_subset is None
|
648 |
+
lora_kwargs = {'task_type': task_type} if task_type is not None else {}
|
649 |
if not self.return_residual:
|
650 |
qkv = self.Wqkv(x, **lora_kwargs)
|
651 |
else:
|
mlp.py
CHANGED
@@ -47,8 +47,8 @@ class Mlp(nn.Module):
|
|
47 |
self.activation = activation
|
48 |
self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2, **factory_kwargs)
|
49 |
|
50 |
-
def forward(self, x,
|
51 |
-
lora_kwargs = {'
|
52 |
y = self.fc1(x, **lora_kwargs)
|
53 |
y = self.activation(y)
|
54 |
y = self.fc2(y, **lora_kwargs)
|
|
|
47 |
self.activation = activation
|
48 |
self.fc2 = nn.Linear(hidden_features, out_features, bias=bias2, **factory_kwargs)
|
49 |
|
50 |
+
def forward(self, x, task_type=None):
|
51 |
+
lora_kwargs = {'task_type': task_type} if task_type is not None else {}
|
52 |
y = self.fc1(x, **lora_kwargs)
|
53 |
y = self.activation(y)
|
54 |
y = self.fc2(y, **lora_kwargs)
|
modeling_lora.py
CHANGED
@@ -15,9 +15,6 @@ from transformers import PretrainedConfig
|
|
15 |
from .modeling_xlm_roberta import XLMRobertaFlashConfig, XLMRobertaModel, XLMRobertaPreTrainedModel
|
16 |
|
17 |
|
18 |
-
LORA_NO_UPDATE = '__lora_no_update__'
|
19 |
-
|
20 |
-
|
21 |
def initialized_weights(
|
22 |
shape: Tuple[int], num_adaptations: int, init: str = "kaiming"
|
23 |
) -> torch.Tensor:
|
@@ -179,8 +176,8 @@ class LoRAParametrization(nn.Module):
|
|
179 |
),
|
180 |
)
|
181 |
|
182 |
-
def new_forward(self, input,
|
183 |
-
task_idx = adaptation_map[
|
184 |
if task_idx is not None:
|
185 |
weights = self.parametrizations.weight[0].lora_forward(self.weight, current_task=task_idx)
|
186 |
else:
|
@@ -207,8 +204,8 @@ class LoRAParametrization(nn.Module):
|
|
207 |
),
|
208 |
)
|
209 |
|
210 |
-
def new_forward(self, input,
|
211 |
-
task_idx = adaptation_map[
|
212 |
if task_idx is not None:
|
213 |
weights = self.parametrizations.weight[0].lora_forward(self.weight, current_task=task_idx)
|
214 |
else:
|
@@ -244,6 +241,16 @@ class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
|
|
244 |
raise ValueError(
|
245 |
f'`lora_adaptations` must be a list and contain at least one element'
|
246 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
247 |
self._adaptation_map = {
|
248 |
name: idx for idx, name in enumerate(self._lora_adaptations)
|
249 |
}
|
@@ -335,25 +342,22 @@ class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
|
|
335 |
def encode(
|
336 |
self,
|
337 |
*args,
|
338 |
-
|
339 |
**kwargs,
|
340 |
) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
|
341 |
"""
|
342 |
Computes sentence embeddings
|
343 |
|
344 |
-
|
345 |
-
Specifies the task for which the encoding is intended.
|
346 |
-
|
347 |
-
|
348 |
-
existing adapter configuration. If `task` is explicitly set to `None`, all LoRA
|
349 |
-
adapters are disabled, and the model reverts to its original, general-purpose weights.
|
350 |
-
If `task` is set to a specific LoRA adaptation, that adaptation is activated.
|
351 |
"""
|
352 |
-
if
|
353 |
raise ValueError(
|
354 |
-
f"Unsupported task '{
|
355 |
f"Supported tasks are: {', '.join(self.config.lora_adaptations)}."
|
356 |
-
f"Alternatively, don't pass the `
|
357 |
)
|
358 |
|
359 |
-
return self.roberta.encode(*args, **kwargs)
|
|
|
15 |
from .modeling_xlm_roberta import XLMRobertaFlashConfig, XLMRobertaModel, XLMRobertaPreTrainedModel
|
16 |
|
17 |
|
|
|
|
|
|
|
18 |
def initialized_weights(
|
19 |
shape: Tuple[int], num_adaptations: int, init: str = "kaiming"
|
20 |
) -> torch.Tensor:
|
|
|
176 |
),
|
177 |
)
|
178 |
|
179 |
+
def new_forward(self, input, task_type, residual=False):
|
180 |
+
task_idx = adaptation_map[task_type] if task_type else None
|
181 |
if task_idx is not None:
|
182 |
weights = self.parametrizations.weight[0].lora_forward(self.weight, current_task=task_idx)
|
183 |
else:
|
|
|
204 |
),
|
205 |
)
|
206 |
|
207 |
+
def new_forward(self, input, task_type):
|
208 |
+
task_idx = adaptation_map[task_type] if task_type else None
|
209 |
if task_idx is not None:
|
210 |
weights = self.parametrizations.weight[0].lora_forward(self.weight, current_task=task_idx)
|
211 |
else:
|
|
|
241 |
raise ValueError(
|
242 |
f'`lora_adaptations` must be a list and contain at least one element'
|
243 |
)
|
244 |
+
self._lora_prompts = config.lora_prompts
|
245 |
+
if (
|
246 |
+
not isinstance(self._lora_prompts, dict)
|
247 |
+
or len(self._lora_prompts) != len(self._lora_adaptations)
|
248 |
+
or not all([v in self._lora_adaptations for v in self._lora_prompts.keys()])
|
249 |
+
):
|
250 |
+
raise ValueError(
|
251 |
+
f'`lora_prompts` must be a dict and contain the same number of elements '
|
252 |
+
f'as `lora_adaptations` with all keys in `lora_prompts` present in `lora_adaptations`.'
|
253 |
+
)
|
254 |
self._adaptation_map = {
|
255 |
name: idx for idx, name in enumerate(self._lora_adaptations)
|
256 |
}
|
|
|
342 |
def encode(
|
343 |
self,
|
344 |
*args,
|
345 |
+
task_type: Optional[str] = None,
|
346 |
**kwargs,
|
347 |
) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
|
348 |
"""
|
349 |
Computes sentence embeddings
|
350 |
|
351 |
+
task_type(`str`, *optional*, defaults to `None`):
|
352 |
+
Specifies the task for which the encoding is intended. If `task_type` is not provide,
|
353 |
+
all LoRA adapters are disabled, and the model reverts to its original,
|
354 |
+
general-purpose weights.
|
|
|
|
|
|
|
355 |
"""
|
356 |
+
if task_type and task_type not in self._lora_adaptations:
|
357 |
raise ValueError(
|
358 |
+
f"Unsupported task '{task_type}'. "
|
359 |
f"Supported tasks are: {', '.join(self.config.lora_adaptations)}."
|
360 |
+
f"Alternatively, don't pass the `task_type` argument to disable LoRA."
|
361 |
)
|
362 |
|
363 |
+
return self.roberta.encode(*args, task_type=task_type, **kwargs)
|
modeling_xlm_roberta.py
CHANGED
@@ -21,7 +21,7 @@ import torch.nn.functional as F
|
|
21 |
import torch.utils.checkpoint
|
22 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
23 |
from einops import rearrange
|
24 |
-
from transformers import PretrainedConfig
|
25 |
from transformers.modeling_utils import PreTrainedModel
|
26 |
from transformers.modeling_outputs import MaskedLMOutput,SequenceClassifierOutput
|
27 |
from transformers.models.xlm_roberta.modeling_xlm_roberta import XLMRobertaLMHead
|
@@ -204,7 +204,7 @@ class XLMRobertaEncoder(nn.Module):
|
|
204 |
def gradient_checkpointing(self, value):
|
205 |
self._grad_checkpointing = value
|
206 |
|
207 |
-
def forward(self, hidden_states, key_padding_mask=None, subset_mask=None,
|
208 |
"""If subset_mask is not None, we only want output for the subset of the sequence.
|
209 |
This means that we only compute the last layer output for these tokens.
|
210 |
subset_mask: (batch, seqlen), dtype=torch.bool
|
@@ -215,7 +215,7 @@ class XLMRobertaEncoder(nn.Module):
|
|
215 |
if key_padding_mask is not None
|
216 |
else None
|
217 |
)
|
218 |
-
mixer_kwargs['
|
219 |
for layer in self.layers:
|
220 |
if self._grad_checkpointing:
|
221 |
hidden_states = torch.utils.checkpoint.checkpoint(
|
@@ -233,7 +233,7 @@ class XLMRobertaEncoder(nn.Module):
|
|
233 |
hidden_states, indices, cu_seqlens, max_seqlen_in_batch = unpad_input(
|
234 |
hidden_states, key_padding_mask
|
235 |
)
|
236 |
-
mixer_kwargs = {"cu_seqlens": cu_seqlens, "max_seqlen": max_seqlen_in_batch, "
|
237 |
if subset_mask is None:
|
238 |
for layer in self.layers:
|
239 |
if self._grad_checkpointing:
|
@@ -310,10 +310,10 @@ class XLMRobertaPooler(nn.Module):
|
|
310 |
self.dense = linear_cls(config.hidden_size, config.hidden_size)
|
311 |
self.activation = nn.Tanh()
|
312 |
|
313 |
-
def forward(self, hidden_states, pool=True,
|
314 |
# We "pool" the model by simply taking the hidden state corresponding
|
315 |
# to the first token.
|
316 |
-
lora_kwargs = {'
|
317 |
|
318 |
first_token_tensor = hidden_states[:, 0] if pool else hidden_states
|
319 |
pooled_output = self.dense(first_token_tensor, **lora_kwargs)
|
@@ -443,7 +443,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
443 |
self.pooler = XLMRobertaPooler(config) if add_pooling_layer else None
|
444 |
|
445 |
self.apply(partial(_init_weights, initializer_range=config.initializer_range))
|
446 |
-
|
447 |
|
448 |
@torch.inference_mode()
|
449 |
def encode(
|
@@ -457,7 +457,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
457 |
device: Optional[torch.device] = None,
|
458 |
normalize_embeddings: bool = False,
|
459 |
truncate_dim: Optional[int] = None,
|
460 |
-
|
461 |
**tokenizer_kwargs,
|
462 |
) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
|
463 |
"""
|
@@ -496,12 +496,6 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
496 |
If convert_to_tensor, a stacked tensor is returned.
|
497 |
If convert_to_numpy, a numpy matrix is returned.
|
498 |
"""
|
499 |
-
from transformers import AutoTokenizer
|
500 |
-
|
501 |
-
self.tokenizer = AutoTokenizer.from_pretrained(
|
502 |
-
self.name_or_path, trust_remote_code=True
|
503 |
-
)
|
504 |
-
|
505 |
is_training = self.training
|
506 |
self.eval()
|
507 |
|
@@ -548,7 +542,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
548 |
)
|
549 |
else:
|
550 |
range_iter = range(0, len(sentences), batch_size)
|
551 |
-
lora_kwargs = {'
|
552 |
for i in range_iter:
|
553 |
encoded_input = self.tokenizer(
|
554 |
sentences[i : i + batch_size],
|
@@ -643,7 +637,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
643 |
layer output for these tokens.
|
644 |
masked_tokens_mask: (batch, seqlen), dtype=torch.bool
|
645 |
"""
|
646 |
-
|
647 |
if kwargs:
|
648 |
for key, value in kwargs.items():
|
649 |
if value is not None:
|
@@ -657,7 +651,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
657 |
)
|
658 |
|
659 |
hidden_states = self.embeddings(
|
660 |
-
input_ids, position_ids=position_ids, token_type_ids=token_type_ids,
|
661 |
)
|
662 |
# TD [2022-12:18]: Don't need to force residual in fp32
|
663 |
# BERT puts embedding LayerNorm before embedding dropout.
|
@@ -681,12 +675,12 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
681 |
subset_mask = None
|
682 |
|
683 |
sequence_output = self.encoder(
|
684 |
-
hidden_states, key_padding_mask=attention_mask, subset_mask=subset_mask,
|
685 |
)
|
686 |
|
687 |
if masked_tokens_mask is None:
|
688 |
pooled_output = (
|
689 |
-
self.pooler(sequence_output,
|
690 |
)
|
691 |
else:
|
692 |
# TD [2022-03-01]: the indexing here is very tricky.
|
@@ -700,7 +694,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
700 |
pool_input = sequence_output[first_col_mask[subset_mask]]
|
701 |
sequence_output = sequence_output[masked_tokens_mask[subset_mask]]
|
702 |
pooled_output = (
|
703 |
-
self.pooler(pool_input, pool=False,
|
704 |
)
|
705 |
|
706 |
if not return_dict:
|
@@ -1282,4 +1276,4 @@ class XLMRobertaForSequenceClassification(XLMRobertaPreTrainedModel):
|
|
1282 |
logits=logits,
|
1283 |
hidden_states=outputs.hidden_states,
|
1284 |
attentions=outputs.attentions,
|
1285 |
-
)
|
|
|
21 |
import torch.utils.checkpoint
|
22 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
23 |
from einops import rearrange
|
24 |
+
from transformers import PretrainedConfig, AutoTokenizer
|
25 |
from transformers.modeling_utils import PreTrainedModel
|
26 |
from transformers.modeling_outputs import MaskedLMOutput,SequenceClassifierOutput
|
27 |
from transformers.models.xlm_roberta.modeling_xlm_roberta import XLMRobertaLMHead
|
|
|
204 |
def gradient_checkpointing(self, value):
|
205 |
self._grad_checkpointing = value
|
206 |
|
207 |
+
def forward(self, hidden_states, key_padding_mask=None, subset_mask=None, task_type=None):
|
208 |
"""If subset_mask is not None, we only want output for the subset of the sequence.
|
209 |
This means that we only compute the last layer output for these tokens.
|
210 |
subset_mask: (batch, seqlen), dtype=torch.bool
|
|
|
215 |
if key_padding_mask is not None
|
216 |
else None
|
217 |
)
|
218 |
+
mixer_kwargs['task_type'] = task_type
|
219 |
for layer in self.layers:
|
220 |
if self._grad_checkpointing:
|
221 |
hidden_states = torch.utils.checkpoint.checkpoint(
|
|
|
233 |
hidden_states, indices, cu_seqlens, max_seqlen_in_batch = unpad_input(
|
234 |
hidden_states, key_padding_mask
|
235 |
)
|
236 |
+
mixer_kwargs = {"cu_seqlens": cu_seqlens, "max_seqlen": max_seqlen_in_batch, "task_type": task_type}
|
237 |
if subset_mask is None:
|
238 |
for layer in self.layers:
|
239 |
if self._grad_checkpointing:
|
|
|
310 |
self.dense = linear_cls(config.hidden_size, config.hidden_size)
|
311 |
self.activation = nn.Tanh()
|
312 |
|
313 |
+
def forward(self, hidden_states, pool=True, task_type=None):
|
314 |
# We "pool" the model by simply taking the hidden state corresponding
|
315 |
# to the first token.
|
316 |
+
lora_kwargs = {'task_type': task_type} if task_type is not None else {}
|
317 |
|
318 |
first_token_tensor = hidden_states[:, 0] if pool else hidden_states
|
319 |
pooled_output = self.dense(first_token_tensor, **lora_kwargs)
|
|
|
443 |
self.pooler = XLMRobertaPooler(config) if add_pooling_layer else None
|
444 |
|
445 |
self.apply(partial(_init_weights, initializer_range=config.initializer_range))
|
446 |
+
self.tokenizer = AutoTokenizer.from_pretrained(self.name_or_path, trust_remote_code=True)
|
447 |
|
448 |
@torch.inference_mode()
|
449 |
def encode(
|
|
|
457 |
device: Optional[torch.device] = None,
|
458 |
normalize_embeddings: bool = False,
|
459 |
truncate_dim: Optional[int] = None,
|
460 |
+
task_type: Optional[str] = None,
|
461 |
**tokenizer_kwargs,
|
462 |
) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
|
463 |
"""
|
|
|
496 |
If convert_to_tensor, a stacked tensor is returned.
|
497 |
If convert_to_numpy, a numpy matrix is returned.
|
498 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
499 |
is_training = self.training
|
500 |
self.eval()
|
501 |
|
|
|
542 |
)
|
543 |
else:
|
544 |
range_iter = range(0, len(sentences), batch_size)
|
545 |
+
lora_kwargs = {'task_type': task_type} if task_type is not None else {}
|
546 |
for i in range_iter:
|
547 |
encoded_input = self.tokenizer(
|
548 |
sentences[i : i + batch_size],
|
|
|
637 |
layer output for these tokens.
|
638 |
masked_tokens_mask: (batch, seqlen), dtype=torch.bool
|
639 |
"""
|
640 |
+
task_type = kwargs.pop('task_type', None)
|
641 |
if kwargs:
|
642 |
for key, value in kwargs.items():
|
643 |
if value is not None:
|
|
|
651 |
)
|
652 |
|
653 |
hidden_states = self.embeddings(
|
654 |
+
input_ids, position_ids=position_ids, token_type_ids=token_type_ids, task_type=task_type
|
655 |
)
|
656 |
# TD [2022-12:18]: Don't need to force residual in fp32
|
657 |
# BERT puts embedding LayerNorm before embedding dropout.
|
|
|
675 |
subset_mask = None
|
676 |
|
677 |
sequence_output = self.encoder(
|
678 |
+
hidden_states, key_padding_mask=attention_mask, subset_mask=subset_mask, task_type=task_type
|
679 |
)
|
680 |
|
681 |
if masked_tokens_mask is None:
|
682 |
pooled_output = (
|
683 |
+
self.pooler(sequence_output, task_type=task_type) if self.pooler is not None else None
|
684 |
)
|
685 |
else:
|
686 |
# TD [2022-03-01]: the indexing here is very tricky.
|
|
|
694 |
pool_input = sequence_output[first_col_mask[subset_mask]]
|
695 |
sequence_output = sequence_output[masked_tokens_mask[subset_mask]]
|
696 |
pooled_output = (
|
697 |
+
self.pooler(pool_input, pool=False, task_type=task_type) if self.pooler is not None else None
|
698 |
)
|
699 |
|
700 |
if not return_dict:
|
|
|
1276 |
logits=logits,
|
1277 |
hidden_states=outputs.hidden_states,
|
1278 |
attentions=outputs.attentions,
|
1279 |
+
)
|
rotary.py
CHANGED
@@ -6,11 +6,13 @@ from typing import Optional, Tuple, Union
|
|
6 |
|
7 |
import torch
|
8 |
from einops import rearrange, repeat
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
|
|
|
|
14 |
|
15 |
|
16 |
def rotate_half(x, interleaved=False):
|
@@ -29,6 +31,10 @@ def apply_rotary_emb_torch(x, cos, sin, interleaved=False):
|
|
29 |
"""
|
30 |
ro_dim = cos.shape[-1] * 2
|
31 |
assert ro_dim <= x.shape[-1]
|
|
|
|
|
|
|
|
|
32 |
cos = repeat(cos, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
|
33 |
sin = repeat(sin, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
|
34 |
return torch.cat(
|
@@ -60,6 +66,7 @@ class ApplyRotaryEmb(torch.autograd.Function):
|
|
60 |
interleaved=interleaved,
|
61 |
inplace=inplace,
|
62 |
)
|
|
|
63 |
if isinstance(seqlen_offsets, int):
|
64 |
ctx.save_for_backward(cos, sin, cu_seqlens) # Can't save int with save_for_backward
|
65 |
ctx.seqlen_offsets = seqlen_offsets
|
@@ -82,6 +89,7 @@ class ApplyRotaryEmb(torch.autograd.Function):
|
|
82 |
# "[CUDA]: invalid device context", and cloning makes it work. Idk why. Triton 2.1.0 works.
|
83 |
if not ctx.interleaved and not ctx.inplace:
|
84 |
do = do.clone()
|
|
|
85 |
dx = apply_rotary(
|
86 |
do,
|
87 |
cos,
|
@@ -150,21 +158,37 @@ class ApplyRotaryEmbQKV_(torch.autograd.Function):
|
|
150 |
# batch, seqlen, three, nheads, headdim = qkv.shape
|
151 |
assert qkv.shape[-3] == 3
|
152 |
if cos_k is None and sin_k is None and qkv.is_contiguous():
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
qk,
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
168 |
else:
|
169 |
cos_k = cos if cos_k is None else cos_k
|
170 |
sin_k = sin if sin_k is None else sin_k
|
@@ -228,7 +252,6 @@ class ApplyRotaryEmbQKV_(torch.autograd.Function):
|
|
228 |
sin_k = sin if sin_k is None else sin_k
|
229 |
dq, dk = dqkv[..., 0, :, :], dqkv[..., 1, :, :]
|
230 |
apply_rotary(
|
231 |
-
|
232 |
dq,
|
233 |
cos,
|
234 |
sin,
|
|
|
6 |
|
7 |
import torch
|
8 |
from einops import rearrange, repeat
|
9 |
+
|
10 |
+
if torch.cuda.is_available():
|
11 |
+
try:
|
12 |
+
from flash_attn.ops.triton.rotary import apply_rotary
|
13 |
+
except ImportError:
|
14 |
+
def apply_rotary(*args, **kwargs):
|
15 |
+
raise RuntimeError('RoPE requires flash-attention to be installed')
|
16 |
|
17 |
|
18 |
def rotate_half(x, interleaved=False):
|
|
|
31 |
"""
|
32 |
ro_dim = cos.shape[-1] * 2
|
33 |
assert ro_dim <= x.shape[-1]
|
34 |
+
cos, sin = (
|
35 |
+
cos[:x.shape[1]],
|
36 |
+
sin[:x.shape[1]],
|
37 |
+
)
|
38 |
cos = repeat(cos, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
|
39 |
sin = repeat(sin, "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
|
40 |
return torch.cat(
|
|
|
66 |
interleaved=interleaved,
|
67 |
inplace=inplace,
|
68 |
)
|
69 |
+
|
70 |
if isinstance(seqlen_offsets, int):
|
71 |
ctx.save_for_backward(cos, sin, cu_seqlens) # Can't save int with save_for_backward
|
72 |
ctx.seqlen_offsets = seqlen_offsets
|
|
|
89 |
# "[CUDA]: invalid device context", and cloning makes it work. Idk why. Triton 2.1.0 works.
|
90 |
if not ctx.interleaved and not ctx.inplace:
|
91 |
do = do.clone()
|
92 |
+
|
93 |
dx = apply_rotary(
|
94 |
do,
|
95 |
cos,
|
|
|
158 |
# batch, seqlen, three, nheads, headdim = qkv.shape
|
159 |
assert qkv.shape[-3] == 3
|
160 |
if cos_k is None and sin_k is None and qkv.is_contiguous():
|
161 |
+
|
162 |
+
if torch.cuda.is_available():
|
163 |
+
# Call 1 kernel instead of 2 kernels
|
164 |
+
# We need qkv to be contiguous so that when we reshape to combine (3, nheads)
|
165 |
+
# dimensions, we get the same tensor
|
166 |
+
qk = rearrange(qkv[..., :2, :, :], "... t h d -> ... (t h) d")
|
167 |
+
# qk = qkv[:, :, :2].reshape(batch, seqlen, -1, headdim)
|
168 |
+
apply_rotary(
|
169 |
+
qk,
|
170 |
+
cos,
|
171 |
+
sin,
|
172 |
+
seqlen_offsets=seqlen_offsets,
|
173 |
+
interleaved=interleaved,
|
174 |
+
inplace=True,
|
175 |
+
cu_seqlens=cu_seqlens,
|
176 |
+
max_seqlen=max_seqlen,
|
177 |
+
)
|
178 |
+
else:
|
179 |
+
q_rot = apply_rotary_emb_torch(
|
180 |
+
qkv[:, :, 0],
|
181 |
+
cos,
|
182 |
+
sin,
|
183 |
+
interleaved=interleaved,
|
184 |
+
)
|
185 |
+
k_rot = apply_rotary_emb_torch(
|
186 |
+
qkv[:, :, 1],
|
187 |
+
cos,
|
188 |
+
sin,
|
189 |
+
interleaved=interleaved,
|
190 |
+
)
|
191 |
+
qkv = torch.stack((q_rot, k_rot, qkv[:, :, 2]), dim=2)
|
192 |
else:
|
193 |
cos_k = cos if cos_k is None else cos_k
|
194 |
sin_k = sin if sin_k is None else sin_k
|
|
|
252 |
sin_k = sin if sin_k is None else sin_k
|
253 |
dq, dk = dqkv[..., 0, :, :], dqkv[..., 1, :, :]
|
254 |
apply_rotary(
|
|
|
255 |
dq,
|
256 |
cos,
|
257 |
sin,
|