jupyterjazz
commited on
Commit
•
509511d
1
Parent(s):
eefe43c
refactor: finalize impl
Browse filesSigned-off-by: jupyterjazz <[email protected]>
- block.py +1 -1
- embedding.py +6 -3
- mha.py +12 -6
- mlp.py +6 -3
- modeling_lora.py +1 -56
- modeling_xlm_roberta.py +13 -8
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=mixer_kwargs.get('task'))
|
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 +40,17 @@ 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 +60,6 @@ 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=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 |
+
if task is not None:
|
52 |
+
lora_kwargs['task'] = task
|
53 |
+
embeddings = self.word_embeddings(input_ids, **lora_kwargs)
|
54 |
if self.max_position_embeddings > 0:
|
55 |
if position_ids is None:
|
56 |
position_ids = create_position_ids_from_input_ids(input_ids, padding_idx=self.word_embeddings.padding_idx).to(input_ids.device)
|
|
|
60 |
if self.type_vocab_size > 0:
|
61 |
if token_type_ids is None:
|
62 |
token_type_ids = torch.zeros(seqlen, dtype=torch.long, device=input_ids.device)
|
63 |
+
token_type_embeddings = self.token_type_embeddings(token_type_ids, **lora_kwargs)
|
64 |
embeddings = embeddings + token_type_embeddings
|
65 |
return embeddings
|
mha.py
CHANGED
@@ -340,9 +340,8 @@ class CrossAttention(nn.Module):
|
|
340 |
class LinearResidual(nn.Linear):
|
341 |
"""Wrap nn.Linear to return the residual as well. For compatibility with FusedDense."""
|
342 |
|
343 |
-
def forward(self, input: torch.Tensor
|
344 |
-
|
345 |
-
return super().forward(input, task=task), input
|
346 |
|
347 |
|
348 |
def _update_kv_cache(kv, inference_params, layer_idx):
|
@@ -591,6 +590,7 @@ class MHA(nn.Module):
|
|
591 |
max_seqlen=None,
|
592 |
mixer_subset=None,
|
593 |
inference_params=None,
|
|
|
594 |
**kwargs,
|
595 |
):
|
596 |
"""
|
@@ -645,10 +645,15 @@ 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 |
if not self.return_residual:
|
649 |
-
qkv = self.Wqkv(x)
|
650 |
else:
|
651 |
-
qkv, x = self.Wqkv(x,
|
652 |
|
653 |
if self.dwconv:
|
654 |
qkv = rearrange(
|
@@ -734,5 +739,6 @@ class MHA(nn.Module):
|
|
734 |
context = self._update_kvcache_attention(q, kv, inference_params)
|
735 |
else:
|
736 |
context = self._apply_rotary_update_kvcache_attention(q, kv, inference_params)
|
737 |
-
|
|
|
738 |
return out if not self.return_residual else (out, x)
|
|
|
340 |
class LinearResidual(nn.Linear):
|
341 |
"""Wrap nn.Linear to return the residual as well. For compatibility with FusedDense."""
|
342 |
|
343 |
+
def forward(self, input: torch.Tensor) -> torch.Tensor:
|
344 |
+
return super().forward(input), input
|
|
|
345 |
|
346 |
|
347 |
def _update_kv_cache(kv, inference_params, layer_idx):
|
|
|
590 |
max_seqlen=None,
|
591 |
mixer_subset=None,
|
592 |
inference_params=None,
|
593 |
+
task=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 = {}
|
649 |
+
if task:
|
650 |
+
lora_kwargs['task'] = task
|
651 |
+
lora_kwargs['residual'] = self.return_residual
|
652 |
+
|
653 |
if not self.return_residual:
|
654 |
+
qkv = self.Wqkv(x, **lora_kwargs)
|
655 |
else:
|
656 |
+
qkv, x = self.Wqkv(x, **lora_kwargs)
|
657 |
|
658 |
if self.dwconv:
|
659 |
qkv = rearrange(
|
|
|
739 |
context = self._update_kvcache_attention(q, kv, inference_params)
|
740 |
else:
|
741 |
context = self._apply_rotary_update_kvcache_attention(q, kv, inference_params)
|
742 |
+
lora_kwargs.pop('residual', None)
|
743 |
+
out = self.out_proj(rearrange(context, "... h d -> ... (h d)"), **lora_kwargs)
|
744 |
return out if not self.return_residual else (out, x)
|
mlp.py
CHANGED
@@ -47,10 +47,13 @@ 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):
|
51 |
+
lora_kwargs = {}
|
52 |
+
if task:
|
53 |
+
lora_kwargs['task'] = task
|
54 |
+
y = self.fc1(x, **lora_kwargs)
|
55 |
y = self.activation(y)
|
56 |
+
y = self.fc2(y, **lora_kwargs)
|
57 |
return y if not self.return_residual else (y, x)
|
58 |
|
59 |
|
modeling_lora.py
CHANGED
@@ -92,8 +92,6 @@ class LoRAParametrization(nn.Module):
|
|
92 |
torch.ones(self.swap((1, fan_in)), dtype=self.lora_A.dtype),
|
93 |
persistent=False,
|
94 |
)
|
95 |
-
self.forward_fn = lambda x: x
|
96 |
-
self.current_task = None
|
97 |
|
98 |
def _dropout(self, A):
|
99 |
# to mimic the original implementation: A @ dropout(x), we do (A * dropout(ones)) @ x
|
@@ -116,18 +114,6 @@ class LoRAParametrization(nn.Module):
|
|
116 |
def forward(self, X):
|
117 |
return X
|
118 |
|
119 |
-
@property
|
120 |
-
def current_task(self):
|
121 |
-
return self._current_task
|
122 |
-
|
123 |
-
@current_task.setter
|
124 |
-
def current_task(self, task: Union[None, int]):
|
125 |
-
self._current_task = task
|
126 |
-
if task is None:
|
127 |
-
self.forward_fn = lambda x: x
|
128 |
-
else:
|
129 |
-
self.forward_fn = self.lora_forward
|
130 |
-
|
131 |
@classmethod
|
132 |
def from_linear(
|
133 |
cls,
|
@@ -239,12 +225,6 @@ class LoRAParametrization(nn.Module):
|
|
239 |
layer.forward = new_forward.__get__(layer, layer.__class__)
|
240 |
|
241 |
|
242 |
-
@staticmethod
|
243 |
-
def select_task_for_layer(layer: nn.Module, task_idx: Optional[int] = None):
|
244 |
-
if isinstance(layer, LoRAParametrization):
|
245 |
-
layer.current_task = task_idx
|
246 |
-
|
247 |
-
|
248 |
class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
|
249 |
def __init__(
|
250 |
self,
|
@@ -279,9 +259,6 @@ class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
|
|
279 |
alpha=self._alpha,
|
280 |
)
|
281 |
self.main_params_trainable = config.lora_main_params_trainable
|
282 |
-
self._task_idx = None
|
283 |
-
# By default, disable LoRA until it's specified which adapter/task to use
|
284 |
-
self.current_task = None
|
285 |
|
286 |
|
287 |
@property
|
@@ -340,39 +317,7 @@ class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
|
|
340 |
)
|
341 |
)
|
342 |
|
343 |
-
|
344 |
-
def current_task(self):
|
345 |
-
"""Which LoRA is currently selected
|
346 |
-
:return: Integer or None (when LoRA is disabled)
|
347 |
-
"""
|
348 |
-
return self._task_idx
|
349 |
-
|
350 |
-
@current_task.setter
|
351 |
-
def current_task(self, task_name: Union[None, str]):
|
352 |
-
"""Set the LoRA that is to be used.
|
353 |
-
The LoRA is specified by `task_idx`, which may be an integer >= 0,
|
354 |
-
indexing the available LoRAs. If it is None, no LoRA is used.
|
355 |
-
:param task_name: Which LoRA to use
|
356 |
-
:return:
|
357 |
-
"""
|
358 |
-
if task_name and task_name not in self._lora_adaptations:
|
359 |
-
raise ValueError(
|
360 |
-
f"Unsupported task '{task_name}'. "
|
361 |
-
f"Supported tasks are: {', '.join(self.config.lora_adaptations)}."
|
362 |
-
f"Alternatively, set `task` to `None` if you want to disable LoRA."
|
363 |
-
)
|
364 |
-
task_idx = self._adaptation_map[task_name] if task_name else None
|
365 |
-
# if self._task_idx != task_idx:
|
366 |
-
# # In this case, we need to update the LoRAs everywhere
|
367 |
-
# self._task_idx = task_idx
|
368 |
-
# self.apply(
|
369 |
-
# partial(LoRAParametrization.select_task_for_layer, task_idx=task_idx)
|
370 |
-
# )
|
371 |
-
|
372 |
-
def forward(self, *args, task: Union[str, None] = LORA_NO_UPDATE, **kwargs):
|
373 |
-
if task != LORA_NO_UPDATE:
|
374 |
-
self.current_task = task
|
375 |
-
|
376 |
return self.roberta(*args, **kwargs)
|
377 |
|
378 |
def parameters(self, recurse: bool = True) -> Iterator[Parameter]:
|
|
|
92 |
torch.ones(self.swap((1, fan_in)), dtype=self.lora_A.dtype),
|
93 |
persistent=False,
|
94 |
)
|
|
|
|
|
95 |
|
96 |
def _dropout(self, A):
|
97 |
# to mimic the original implementation: A @ dropout(x), we do (A * dropout(ones)) @ x
|
|
|
114 |
def forward(self, X):
|
115 |
return X
|
116 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
117 |
@classmethod
|
118 |
def from_linear(
|
119 |
cls,
|
|
|
225 |
layer.forward = new_forward.__get__(layer, layer.__class__)
|
226 |
|
227 |
|
|
|
|
|
|
|
|
|
|
|
|
|
228 |
class XLMRobertaLoRA(XLMRobertaPreTrainedModel):
|
229 |
def __init__(
|
230 |
self,
|
|
|
259 |
alpha=self._alpha,
|
260 |
)
|
261 |
self.main_params_trainable = config.lora_main_params_trainable
|
|
|
|
|
|
|
262 |
|
263 |
|
264 |
@property
|
|
|
317 |
)
|
318 |
)
|
319 |
|
320 |
+
def forward(self, *args, **kwargs):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
321 |
return self.roberta(*args, **kwargs)
|
322 |
|
323 |
def parameters(self, recurse: bool = True) -> Iterator[Parameter]:
|
modeling_xlm_roberta.py
CHANGED
@@ -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 +310,15 @@ 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 |
|
@@ -639,7 +644,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
639 |
layer output for these tokens.
|
640 |
masked_tokens_mask: (batch, seqlen), dtype=torch.bool
|
641 |
"""
|
642 |
-
|
643 |
if kwargs:
|
644 |
for key, value in kwargs.items():
|
645 |
if value is not None:
|
@@ -653,7 +658,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
653 |
)
|
654 |
|
655 |
hidden_states = self.embeddings(
|
656 |
-
input_ids, position_ids=position_ids, token_type_ids=token_type_ids
|
657 |
)
|
658 |
# TD [2022-12:18]: Don't need to force residual in fp32
|
659 |
# BERT puts embedding LayerNorm before embedding dropout.
|
@@ -677,12 +682,12 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
677 |
subset_mask = None
|
678 |
|
679 |
sequence_output = self.encoder(
|
680 |
-
hidden_states, key_padding_mask=attention_mask, subset_mask=subset_mask
|
681 |
)
|
682 |
|
683 |
if masked_tokens_mask is None:
|
684 |
pooled_output = (
|
685 |
-
self.pooler(sequence_output) if self.pooler is not None else None
|
686 |
)
|
687 |
else:
|
688 |
# TD [2022-03-01]: the indexing here is very tricky.
|
@@ -696,7 +701,7 @@ class XLMRobertaModel(XLMRobertaPreTrainedModel):
|
|
696 |
pool_input = sequence_output[first_col_mask[subset_mask]]
|
697 |
sequence_output = sequence_output[masked_tokens_mask[subset_mask]]
|
698 |
pooled_output = (
|
699 |
-
self.pooler(pool_input, pool=False) if self.pooler is not None else None
|
700 |
)
|
701 |
|
702 |
if not return_dict:
|
|
|
215 |
if key_padding_mask is not None
|
216 |
else None
|
217 |
)
|
218 |
+
mixer_kwargs['task'] = task
|
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": task}
|
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=None):
|
314 |
# We "pool" the model by simply taking the hidden state corresponding
|
315 |
# to the first token.
|
316 |
+
lora_kwargs = {}
|
317 |
+
if task:
|
318 |
+
lora_kwargs['task'] = task
|
319 |
+
|
320 |
first_token_tensor = hidden_states[:, 0] if pool else hidden_states
|
321 |
+
pooled_output = self.dense(first_token_tensor, **lora_kwargs)
|
322 |
pooled_output = self.activation(pooled_output)
|
323 |
return pooled_output
|
324 |
|
|
|
644 |
layer output for these tokens.
|
645 |
masked_tokens_mask: (batch, seqlen), dtype=torch.bool
|
646 |
"""
|
647 |
+
task = kwargs.pop('task', None)
|
648 |
if kwargs:
|
649 |
for key, value in kwargs.items():
|
650 |
if value is not None:
|
|
|
658 |
)
|
659 |
|
660 |
hidden_states = self.embeddings(
|
661 |
+
input_ids, position_ids=position_ids, token_type_ids=token_type_ids, task=task
|
662 |
)
|
663 |
# TD [2022-12:18]: Don't need to force residual in fp32
|
664 |
# BERT puts embedding LayerNorm before embedding dropout.
|
|
|
682 |
subset_mask = None
|
683 |
|
684 |
sequence_output = self.encoder(
|
685 |
+
hidden_states, key_padding_mask=attention_mask, subset_mask=subset_mask, task=task
|
686 |
)
|
687 |
|
688 |
if masked_tokens_mask is None:
|
689 |
pooled_output = (
|
690 |
+
self.pooler(sequence_output, task=task) if self.pooler is not None else None
|
691 |
)
|
692 |
else:
|
693 |
# TD [2022-03-01]: the indexing here is very tricky.
|
|
|
701 |
pool_input = sequence_output[first_col_mask[subset_mask]]
|
702 |
sequence_output = sequence_output[masked_tokens_mask[subset_mask]]
|
703 |
pooled_output = (
|
704 |
+
self.pooler(pool_input, pool=False, task=task) if self.pooler is not None else None
|
705 |
)
|
706 |
|
707 |
if not return_dict:
|