jupyterjazz
commited on
Commit
•
e860caa
1
Parent(s):
6a92924
make-lora-stateless (#25)
Browse files- draft (6cc0f517fef666dc2c5c0f0fe7c92935f7b7d7bb)
- poc (eefe43c08d4952b1e2a3c17d6e20090a72a52474)
- refactor: finalize impl (509511d7e5d7b46de5b143e5fc908784eada3f07)
- refactor: modify encode (3eb20d0a7451e244a9c2ee80e7bb762d17a9f9d1)
- fix: 0 is not none (ae40cb93420908fdd3f370e6af3cb0cbe9cbf90d)
- fix: remove prints (acffa62b444442f4933a3023f19d09148de9e2fb)
- refactor: kwargs comprehension (4e13c908bafe0f54becc8c278b3752a37c207c50)
- fix: residual is kept in kwargs (d9d83063f0fcb5bfa22cd4946f23baa1421b9930)
- feat: merge with recent changes (493416fa0cabc0de4deba3c317c58043f9998d49)
- block.py +1 -1
- embedding.py +4 -3
- mha.py +11 -3
- mlp.py +4 -3
- modeling_lora.py +51 -91
- modeling_xlm_roberta.py +15 -11
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:
|
embedding.py
CHANGED
@@ -40,14 +40,15 @@ 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 |
-
|
|
|
51 |
if self.max_position_embeddings > 0:
|
52 |
if position_ids is None:
|
53 |
position_ids = create_position_ids_from_input_ids(input_ids, padding_idx=self.word_embeddings.padding_idx).to(input_ids.device)
|
@@ -57,6 +58,6 @@ class XLMRobertaEmbeddings(nn.Module):
|
|
57 |
if self.type_vocab_size > 0:
|
58 |
if token_type_ids is None:
|
59 |
token_type_ids = torch.zeros(seqlen, dtype=torch.long, device=input_ids.device)
|
60 |
-
token_type_embeddings = self.token_type_embeddings(token_type_ids)
|
61 |
embeddings = embeddings + token_type_embeddings
|
62 |
return embeddings
|
|
|
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:
|
54 |
position_ids = create_position_ids_from_input_ids(input_ids, padding_idx=self.word_embeddings.padding_idx).to(input_ids.device)
|
|
|
58 |
if self.type_vocab_size > 0:
|
59 |
if token_type_ids is None:
|
60 |
token_type_ids = torch.zeros(seqlen, dtype=torch.long, device=input_ids.device)
|
61 |
+
token_type_embeddings = self.token_type_embeddings(token_type_ids, **lora_kwargs)
|
62 |
embeddings = embeddings + token_type_embeddings
|
63 |
return embeddings
|
mha.py
CHANGED
@@ -450,6 +450,7 @@ class MHA(nn.Module):
|
|
450 |
|
451 |
if fused_bias_fc and FusedDense is None:
|
452 |
raise ImportError("fused_dense is not installed")
|
|
|
453 |
linear_cls = nn.Linear if not fused_bias_fc else FusedDense
|
454 |
linear_resid_cls = (
|
455 |
LinearResidual if not fused_bias_fc else partial(FusedDense, return_residual=True)
|
@@ -589,6 +590,7 @@ class MHA(nn.Module):
|
|
589 |
max_seqlen=None,
|
590 |
mixer_subset=None,
|
591 |
inference_params=None,
|
|
|
592 |
**kwargs,
|
593 |
):
|
594 |
"""
|
@@ -643,10 +645,14 @@ class MHA(nn.Module):
|
|
643 |
batch, seqlen = x.shape[:2]
|
644 |
if not self.cross_attn and self.num_heads_kv == self.num_heads:
|
645 |
assert x_kv is None and mixer_subset is None
|
|
|
646 |
if not self.return_residual:
|
647 |
-
qkv = self.Wqkv(x)
|
648 |
else:
|
649 |
-
|
|
|
|
|
|
|
650 |
if self.dwconv:
|
651 |
qkv = rearrange(
|
652 |
self.dwconv_qkv(rearrange(qkv, "b s d -> b d s"))[..., :-2], "b d s -> b s d"
|
@@ -731,5 +737,7 @@ class MHA(nn.Module):
|
|
731 |
context = self._update_kvcache_attention(q, kv, inference_params)
|
732 |
else:
|
733 |
context = self._apply_rotary_update_kvcache_attention(q, kv, inference_params)
|
734 |
-
|
|
|
|
|
735 |
return out if not self.return_residual else (out, x)
|
|
|
450 |
|
451 |
if fused_bias_fc and FusedDense is None:
|
452 |
raise ImportError("fused_dense is not installed")
|
453 |
+
|
454 |
linear_cls = nn.Linear if not fused_bias_fc else FusedDense
|
455 |
linear_resid_cls = (
|
456 |
LinearResidual if not fused_bias_fc else partial(FusedDense, return_residual=True)
|
|
|
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:
|
652 |
+
if lora_kwargs:
|
653 |
+
lora_kwargs['residual'] = True
|
654 |
+
qkv, x = self.Wqkv(x, **lora_kwargs)
|
655 |
+
|
656 |
if self.dwconv:
|
657 |
qkv = rearrange(
|
658 |
self.dwconv_qkv(rearrange(qkv, "b s d -> b d s"))[..., :-2], "b d s -> b s d"
|
|
|
737 |
context = self._update_kvcache_attention(q, kv, inference_params)
|
738 |
else:
|
739 |
context = self._apply_rotary_update_kvcache_attention(q, kv, inference_params)
|
740 |
+
|
741 |
+
lora_kwargs.pop('residual', None)
|
742 |
+
out = self.out_proj(rearrange(context, "... h d -> ... (h d)"), **lora_kwargs)
|
743 |
return out if not self.return_residual else (out, x)
|
mlp.py
CHANGED
@@ -47,10 +47,11 @@ 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 |
-
|
|
|
52 |
y = self.activation(y)
|
53 |
-
y = self.fc2(y)
|
54 |
return y if not self.return_residual else (y, x)
|
55 |
|
56 |
|
|
|
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)
|
55 |
return y if not self.return_residual else (y, x)
|
56 |
|
57 |
|
modeling_lora.py
CHANGED
@@ -9,6 +9,7 @@ import torch
|
|
9 |
import torch.nn.utils.parametrize as parametrize
|
10 |
from torch import nn
|
11 |
from torch.nn import Parameter
|
|
|
12 |
from transformers import PretrainedConfig
|
13 |
|
14 |
from .modeling_xlm_roberta import XLMRobertaFlashConfig, XLMRobertaModel, XLMRobertaPreTrainedModel
|
@@ -88,22 +89,19 @@ class LoRAParametrization(nn.Module):
|
|
88 |
torch.ones(self.swap((1, fan_in)), dtype=self.lora_A.dtype),
|
89 |
persistent=False,
|
90 |
)
|
91 |
-
self.forward_fn = lambda x: x
|
92 |
-
self.current_task = None
|
93 |
|
94 |
def _dropout(self, A):
|
95 |
# to mimic the original implementation: A @ dropout(x), we do (A * dropout(ones)) @ x
|
96 |
return A * self.lora_dropout(self.lora_dropout_mask)
|
97 |
|
98 |
-
def lora_forward(self, X):
|
99 |
-
assert self.current_task is not None
|
100 |
return (
|
101 |
X
|
102 |
+ torch.matmul(
|
103 |
*self.swap(
|
104 |
(
|
105 |
-
self.lora_B[
|
106 |
-
self.dropout_fn(self.lora_A[
|
107 |
)
|
108 |
)
|
109 |
).view(X.shape)
|
@@ -111,19 +109,7 @@ class LoRAParametrization(nn.Module):
|
|
111 |
)
|
112 |
|
113 |
def forward(self, X):
|
114 |
-
return
|
115 |
-
|
116 |
-
@property
|
117 |
-
def current_task(self):
|
118 |
-
return self._current_task
|
119 |
-
|
120 |
-
@current_task.setter
|
121 |
-
def current_task(self, task: Union[None, int]):
|
122 |
-
self._current_task = task
|
123 |
-
if task is None:
|
124 |
-
self.forward_fn = lambda x: x
|
125 |
-
else:
|
126 |
-
self.forward_fn = self.lora_forward
|
127 |
|
128 |
@classmethod
|
129 |
def from_linear(
|
@@ -175,6 +161,7 @@ class LoRAParametrization(nn.Module):
|
|
175 |
rank: int,
|
176 |
dropout_p: float,
|
177 |
alpha: float,
|
|
|
178 |
):
|
179 |
if isinstance(layer, nn.Linear):
|
180 |
parametrize.register_parametrization(
|
@@ -188,6 +175,22 @@ class LoRAParametrization(nn.Module):
|
|
188 |
alpha=alpha,
|
189 |
),
|
190 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
191 |
elif isinstance(layer, nn.Embedding):
|
192 |
parametrize.register_parametrization(
|
193 |
layer,
|
@@ -201,10 +204,20 @@ class LoRAParametrization(nn.Module):
|
|
201 |
),
|
202 |
)
|
203 |
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
208 |
|
209 |
|
210 |
class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
|
@@ -251,9 +264,7 @@ class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
|
|
251 |
alpha=self._alpha,
|
252 |
)
|
253 |
self.main_params_trainable = config.lora_main_params_trainable
|
254 |
-
|
255 |
-
# By default, disable LoRA until it's specified which adapter/task to use
|
256 |
-
self.current_task = None
|
257 |
|
258 |
@property
|
259 |
def main_params_trainable(self):
|
@@ -307,51 +318,11 @@ class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
|
|
307 |
rank=rank,
|
308 |
dropout_p=dropout_p,
|
309 |
alpha=alpha,
|
|
|
310 |
)
|
311 |
)
|
312 |
|
313 |
-
|
314 |
-
def current_task(self):
|
315 |
-
"""Which LoRA is currently selected
|
316 |
-
:return: Integer or None (when LoRA is disabled)
|
317 |
-
"""
|
318 |
-
return self._task_idx
|
319 |
-
|
320 |
-
@current_task.setter
|
321 |
-
def current_task(self, task_name: Union[None, str]):
|
322 |
-
"""Set the LoRA that is to be used.
|
323 |
-
The LoRA is specified by `task_idx`, which may be an integer >= 0,
|
324 |
-
indexing the available LoRAs. If it is None, no LoRA is used.
|
325 |
-
:param task_name: Which LoRA to use
|
326 |
-
:return:
|
327 |
-
"""
|
328 |
-
if task_name and task_name not in self._lora_adaptations:
|
329 |
-
raise ValueError(
|
330 |
-
f"Unsupported task '{task_name}'. "
|
331 |
-
f"Supported tasks are: {', '.join(self.config.lora_adaptations)}."
|
332 |
-
f"Alternatively, set `task` to `None` if you want to disable LoRA."
|
333 |
-
)
|
334 |
-
task_idx = self._adaptation_map[task_name] if task_name else None
|
335 |
-
if self._task_idx != task_idx:
|
336 |
-
# In this case, we need to update the LoRAs everywhere
|
337 |
-
self._task_idx = task_idx
|
338 |
-
self.apply(
|
339 |
-
partial(LoRAParametrization.select_task_for_layer, task_idx=task_idx)
|
340 |
-
)
|
341 |
-
|
342 |
-
def forward(self, *args, task_type: Union[str, None] = None, **kwargs):
|
343 |
-
if task_type:
|
344 |
-
self.current_task = task_type
|
345 |
-
else:
|
346 |
-
input_ids = kwargs["input_ids"]
|
347 |
-
input_text = self.roberta.tokenizer.decode(input_ids[0], skip_special_tokens=True)
|
348 |
-
for task_name, prompt in self._lora_prompts.items():
|
349 |
-
if input_text.startswith(prompt):
|
350 |
-
self.current_task = task_name
|
351 |
-
break
|
352 |
-
else:
|
353 |
-
self.current_task = None # No task-specific adapter is found, just use the general-purpose weights
|
354 |
-
|
355 |
return self.roberta(*args, **kwargs)
|
356 |
|
357 |
def parameters(self, recurse: bool = True) -> Iterator[Parameter]:
|
@@ -371,33 +342,22 @@ class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
|
|
371 |
def encode(
|
372 |
self,
|
373 |
*args,
|
374 |
-
task_type:
|
375 |
**kwargs,
|
376 |
) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
|
377 |
"""
|
378 |
Computes sentence embeddings
|
379 |
|
380 |
-
|
381 |
-
Specifies the task for which the encoding is intended.
|
382 |
-
|
383 |
-
|
384 |
-
general-purpose weights. If `task` is set to a specific LoRA adaptation, that adaptation
|
385 |
-
is activated.
|
386 |
"""
|
387 |
-
if task_type:
|
388 |
-
|
389 |
-
|
390 |
-
|
391 |
-
|
392 |
-
|
393 |
-
self.current_task = task_name
|
394 |
-
break
|
395 |
-
else:
|
396 |
-
warnings.warn(
|
397 |
-
f"Task-specific embeddings are disabled. To enable, specify the `task` "
|
398 |
-
f"argument with one of the supported tasks: {', '.join(self.config.lora_adaptations)}",
|
399 |
-
category=UserWarning,
|
400 |
-
)
|
401 |
-
self.current_task = None # No task-specific adapter is found, just use the general-purpose weights
|
402 |
|
403 |
-
return self.roberta.encode(*args, **kwargs)
|
|
|
9 |
import torch.nn.utils.parametrize as parametrize
|
10 |
from torch import nn
|
11 |
from torch.nn import Parameter
|
12 |
+
from torch.nn import functional as F
|
13 |
from transformers import PretrainedConfig
|
14 |
|
15 |
from .modeling_xlm_roberta import XLMRobertaFlashConfig, XLMRobertaModel, XLMRobertaPreTrainedModel
|
|
|
89 |
torch.ones(self.swap((1, fan_in)), dtype=self.lora_A.dtype),
|
90 |
persistent=False,
|
91 |
)
|
|
|
|
|
92 |
|
93 |
def _dropout(self, A):
|
94 |
# to mimic the original implementation: A @ dropout(x), we do (A * dropout(ones)) @ x
|
95 |
return A * self.lora_dropout(self.lora_dropout_mask)
|
96 |
|
97 |
+
def lora_forward(self, X, current_task):
|
|
|
98 |
return (
|
99 |
X
|
100 |
+ torch.matmul(
|
101 |
*self.swap(
|
102 |
(
|
103 |
+
self.lora_B[current_task],
|
104 |
+
self.dropout_fn(self.lora_A[current_task]),
|
105 |
)
|
106 |
)
|
107 |
).view(X.shape)
|
|
|
109 |
)
|
110 |
|
111 |
def forward(self, X):
|
112 |
+
return X
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
113 |
|
114 |
@classmethod
|
115 |
def from_linear(
|
|
|
161 |
rank: int,
|
162 |
dropout_p: float,
|
163 |
alpha: float,
|
164 |
+
adaptation_map: dict,
|
165 |
):
|
166 |
if isinstance(layer, nn.Linear):
|
167 |
parametrize.register_parametrization(
|
|
|
175 |
alpha=alpha,
|
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:
|
184 |
+
weights = self.weight
|
185 |
+
|
186 |
+
out = F.linear(input, weights, self.bias)
|
187 |
+
|
188 |
+
if residual:
|
189 |
+
return out, input
|
190 |
+
return out
|
191 |
+
|
192 |
+
layer.forward = new_forward.__get__(layer, layer.__class__)
|
193 |
+
|
194 |
elif isinstance(layer, nn.Embedding):
|
195 |
parametrize.register_parametrization(
|
196 |
layer,
|
|
|
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:
|
212 |
+
weights = self.weight
|
213 |
+
|
214 |
+
out = F.embedding(
|
215 |
+
input, weights, self.padding_idx, self.max_norm,
|
216 |
+
self.norm_type, self.scale_grad_by_freq, self.sparse)
|
217 |
+
|
218 |
+
return out
|
219 |
+
|
220 |
+
layer.forward = new_forward.__get__(layer, layer.__class__)
|
221 |
|
222 |
|
223 |
class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
|
|
|
264 |
alpha=self._alpha,
|
265 |
)
|
266 |
self.main_params_trainable = config.lora_main_params_trainable
|
267 |
+
|
|
|
|
|
268 |
|
269 |
@property
|
270 |
def main_params_trainable(self):
|
|
|
318 |
rank=rank,
|
319 |
dropout_p=dropout_p,
|
320 |
alpha=alpha,
|
321 |
+
adaptation_map=self._adaptation_map,
|
322 |
)
|
323 |
)
|
324 |
|
325 |
+
def forward(self, *args, **kwargs):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
326 |
return self.roberta(*args, **kwargs)
|
327 |
|
328 |
def parameters(self, recurse: bool = True) -> Iterator[Parameter]:
|
|
|
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
@@ -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
|
@@ -215,6 +215,7 @@ class XLMRobertaEncoder(nn.Module):
|
|
215 |
if key_padding_mask is not None
|
216 |
else None
|
217 |
)
|
|
|
218 |
for layer in self.layers:
|
219 |
if self._grad_checkpointing:
|
220 |
hidden_states = torch.utils.checkpoint.checkpoint(
|
@@ -232,7 +233,7 @@ class XLMRobertaEncoder(nn.Module):
|
|
232 |
hidden_states, indices, cu_seqlens, max_seqlen_in_batch = unpad_input(
|
233 |
hidden_states, key_padding_mask
|
234 |
)
|
235 |
-
mixer_kwargs = {"cu_seqlens": cu_seqlens, "max_seqlen": max_seqlen_in_batch}
|
236 |
if subset_mask is None:
|
237 |
for layer in self.layers:
|
238 |
if self._grad_checkpointing:
|
@@ -309,11 +310,13 @@ class XLMRobertaPooler(nn.Module):
|
|
309 |
self.dense = linear_cls(config.hidden_size, config.hidden_size)
|
310 |
self.activation = nn.Tanh()
|
311 |
|
312 |
-
def forward(self, hidden_states, pool=True):
|
313 |
# We "pool" the model by simply taking the hidden state corresponding
|
314 |
# to the first token.
|
|
|
|
|
315 |
first_token_tensor = hidden_states[:, 0] if pool else hidden_states
|
316 |
-
pooled_output = self.dense(first_token_tensor)
|
317 |
pooled_output = self.activation(pooled_output)
|
318 |
return pooled_output
|
319 |
|
@@ -454,6 +457,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
454 |
device: Optional[torch.device] = None,
|
455 |
normalize_embeddings: bool = False,
|
456 |
truncate_dim: Optional[int] = None,
|
|
|
457 |
**tokenizer_kwargs,
|
458 |
) -> Union[List[torch.Tensor], np.ndarray, torch.Tensor]:
|
459 |
"""
|
@@ -538,14 +542,14 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
538 |
)
|
539 |
else:
|
540 |
range_iter = range(0, len(sentences), batch_size)
|
541 |
-
|
542 |
for i in range_iter:
|
543 |
encoded_input = self.tokenizer(
|
544 |
sentences[i : i + batch_size],
|
545 |
return_tensors='pt',
|
546 |
**tokenizer_kwargs,
|
547 |
).to(self.device)
|
548 |
-
token_embs = self.forward(**encoded_input)[0]
|
549 |
|
550 |
# Accumulate in fp32 to avoid overflow
|
551 |
token_embs = token_embs.float()
|
@@ -633,7 +637,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
633 |
layer output for these tokens.
|
634 |
masked_tokens_mask: (batch, seqlen), dtype=torch.bool
|
635 |
"""
|
636 |
-
|
637 |
if kwargs:
|
638 |
for key, value in kwargs.items():
|
639 |
if value is not None:
|
@@ -647,7 +651,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
647 |
)
|
648 |
|
649 |
hidden_states = self.embeddings(
|
650 |
-
input_ids, position_ids=position_ids, token_type_ids=token_type_ids
|
651 |
)
|
652 |
# TD [2022-12:18]: Don't need to force residual in fp32
|
653 |
# BERT puts embedding LayerNorm before embedding dropout.
|
@@ -671,12 +675,12 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
671 |
subset_mask = None
|
672 |
|
673 |
sequence_output = self.encoder(
|
674 |
-
hidden_states, key_padding_mask=attention_mask, subset_mask=subset_mask
|
675 |
)
|
676 |
|
677 |
if masked_tokens_mask is None:
|
678 |
pooled_output = (
|
679 |
-
self.pooler(sequence_output) if self.pooler is not None else None
|
680 |
)
|
681 |
else:
|
682 |
# TD [2022-03-01]: the indexing here is very tricky.
|
@@ -690,7 +694,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
690 |
pool_input = sequence_output[first_col_mask[subset_mask]]
|
691 |
sequence_output = sequence_output[masked_tokens_mask[subset_mask]]
|
692 |
pooled_output = (
|
693 |
-
self.pooler(pool_input, pool=False) if self.pooler is not None else None
|
694 |
)
|
695 |
|
696 |
if not return_dict:
|
|
|
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)
|
320 |
pooled_output = self.activation(pooled_output)
|
321 |
return pooled_output
|
322 |
|
|
|
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 |
"""
|
|
|
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],
|
549 |
return_tensors='pt',
|
550 |
**tokenizer_kwargs,
|
551 |
).to(self.device)
|
552 |
+
token_embs = self.forward(**encoded_input, **lora_kwargs)[0]
|
553 |
|
554 |
# Accumulate in fp32 to avoid overflow
|
555 |
token_embs = token_embs.float()
|
|
|
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:
|