Crystalcareai
commited on
Commit
•
292a484
1
Parent(s):
c28e2ee
Update modeling_gemmoe.py
Browse files- modeling_gemmoe.py +583 -605
modeling_gemmoe.py
CHANGED
@@ -28,6 +28,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
|
28 |
from transformers.activations import ACT2FN
|
29 |
from transformers.cache_utils import Cache, DynamicCache, StaticCache
|
30 |
from transformers.modeling_attn_mask_utils import (
|
|
|
31 |
_prepare_4d_causal_attention_mask,
|
32 |
)
|
33 |
from transformers.modeling_outputs import SequenceClassifierOutputWithPast, MoeModelOutputWithPast, MoeCausalLMOutputWithPast
|
@@ -176,52 +177,71 @@ ALL_LAYERNORM_LAYERS.append(GemmoeRMSNorm)
|
|
176 |
class GemmoeRotaryEmbedding(nn.Module):
|
177 |
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
178 |
super().__init__()
|
|
|
179 |
self.dim = dim
|
180 |
self.max_position_embeddings = max_position_embeddings
|
181 |
self.base = base
|
182 |
-
self.
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
self.cos_cached[:seq_len],
|
206 |
-
self.sin_cached[:seq_len],
|
207 |
-
)
|
208 |
-
|
209 |
def rotate_half(x):
|
210 |
"""Rotates half the hidden dims of the input."""
|
211 |
x1 = x[..., : x.shape[-1] // 2]
|
212 |
x2 = x[..., x.shape[-1] // 2 :]
|
213 |
return torch.cat((-x2, x1), dim=-1)
|
214 |
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
|
219 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
220 |
q_embed = (q * cos) + (rotate_half(q) * sin)
|
221 |
k_embed = (k * cos) + (rotate_half(k) * sin)
|
222 |
return q_embed, k_embed
|
223 |
|
224 |
-
|
|
|
|
|
|
|
225 |
"""
|
226 |
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
227 |
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
@@ -233,14 +253,9 @@ def repeat_kv(self, hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
|
233 |
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
234 |
|
235 |
class GemmoeAttention(nn.Module):
|
236 |
-
"""
|
237 |
-
Multi-headed attention module for Gemmoe model.
|
238 |
-
|
239 |
-
Args:
|
240 |
-
config (GemmoeConfig): The configuration object for the Gemmoe model.
|
241 |
-
layer_idx (Optional[int]): The index of the layer. Default is None.
|
242 |
-
"""
|
243 |
|
|
|
244 |
def __init__(self, config: GemmoeConfig, layer_idx: Optional[int] = None):
|
245 |
super().__init__()
|
246 |
self.config = config
|
@@ -251,6 +266,7 @@ class GemmoeAttention(nn.Module):
|
|
251 |
"lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
|
252 |
"when creating this class."
|
253 |
)
|
|
|
254 |
self.attention_dropout = config.attention_dropout
|
255 |
self.hidden_size = config.hidden_size
|
256 |
self.num_heads = config.num_attention_heads
|
@@ -266,15 +282,16 @@ class GemmoeAttention(nn.Module):
|
|
266 |
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
|
267 |
f" and `num_heads`: {self.num_heads})."
|
268 |
)
|
|
|
269 |
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
|
270 |
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
|
271 |
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
|
272 |
-
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=
|
273 |
self.rotary_emb = GemmoeRotaryEmbedding(
|
274 |
-
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
|
279 |
def forward(
|
280 |
self,
|
@@ -287,25 +304,6 @@ class GemmoeAttention(nn.Module):
|
|
287 |
cache_position: Optional[torch.LongTensor] = None,
|
288 |
**kwargs,
|
289 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
290 |
-
"""
|
291 |
-
Forward pass of the attention module.
|
292 |
-
|
293 |
-
Args:
|
294 |
-
hidden_states (torch.Tensor): The input hidden states.
|
295 |
-
attention_mask (Optional[torch.Tensor]): The attention mask. Default is None.
|
296 |
-
position_ids (Optional[torch.LongTensor]): The position IDs. Default is None.
|
297 |
-
past_key_value (Optional[Cache]): The past key-value cache. Default is None.
|
298 |
-
output_attentions (bool): Whether to output the attention weights. Default is False.
|
299 |
-
use_cache (bool): Whether to use caching. Default is False.
|
300 |
-
cache_position (Optional[torch.LongTensor]): The cache position. Default is None.
|
301 |
-
**kwargs: Additional keyword arguments.
|
302 |
-
|
303 |
-
Returns:
|
304 |
-
Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
305 |
-
- The output hidden states.
|
306 |
-
- The attention weights (if `output_attentions=True`).
|
307 |
-
- The past key-value cache (if `use_cache=True`).
|
308 |
-
"""
|
309 |
bsz, q_len, _ = hidden_states.size()
|
310 |
|
311 |
query_states = self.q_proj(hidden_states)
|
@@ -317,17 +315,16 @@ class GemmoeAttention(nn.Module):
|
|
317 |
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
318 |
|
319 |
past_key_value = getattr(self, "past_key_value", past_key_value)
|
320 |
-
|
321 |
cos, sin = self.rotary_emb(value_states, position_ids, seq_len=None)
|
322 |
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, None)
|
323 |
|
324 |
if past_key_value is not None:
|
325 |
-
# sin and cos are specific to RoPE models;
|
326 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
327 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
328 |
|
329 |
-
|
330 |
-
|
331 |
|
332 |
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
333 |
|
@@ -341,7 +338,6 @@ class GemmoeAttention(nn.Module):
|
|
341 |
# upcast attention to fp32
|
342 |
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
343 |
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
|
344 |
-
|
345 |
attn_output = torch.matmul(attn_weights, value_states)
|
346 |
|
347 |
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
@@ -351,8 +347,8 @@ class GemmoeAttention(nn.Module):
|
|
351 |
)
|
352 |
|
353 |
attn_output = attn_output.transpose(1, 2).contiguous()
|
354 |
-
attn_output = attn_output.view(bsz, q_len, -1)
|
355 |
|
|
|
356 |
attn_output = self.o_proj(attn_output)
|
357 |
|
358 |
if not output_attentions:
|
@@ -360,17 +356,24 @@ class GemmoeAttention(nn.Module):
|
|
360 |
|
361 |
return attn_output, attn_weights, past_key_value
|
362 |
|
|
|
|
|
363 |
class GemmoeFlashAttention2(GemmoeAttention):
|
364 |
"""
|
365 |
Gemmoe flash attention module. This module inherits from `GemmoeAttention` as the weights of the module stays
|
366 |
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
|
367 |
flash attention and deal with padding tokens in case the input contains any of them.
|
368 |
"""
|
|
|
369 |
def __init__(self, *args, **kwargs):
|
370 |
super().__init__(*args, **kwargs)
|
371 |
-
|
|
|
|
|
|
|
372 |
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
|
373 |
|
|
|
374 |
def forward(
|
375 |
self,
|
376 |
hidden_states: torch.Tensor,
|
@@ -401,8 +404,9 @@ class GemmoeFlashAttention2(GemmoeAttention):
|
|
401 |
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, None)
|
402 |
|
403 |
past_key_value = getattr(self, "past_key_value", past_key_value)
|
|
|
404 |
if past_key_value is not None:
|
405 |
-
# sin and cos are specific to RoPE models;
|
406 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
407 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
408 |
|
@@ -419,6 +423,7 @@ class GemmoeFlashAttention2(GemmoeAttention):
|
|
419 |
# cast them back in the correct dtype just to be sure everything works as expected.
|
420 |
# This might slowdown training & inference so it is recommended to not cast the LayerNorms
|
421 |
# in fp32. (GemmoeRMSNorm handles it correctly)
|
|
|
422 |
input_dtype = query_states.dtype
|
423 |
if input_dtype == torch.float32:
|
424 |
if torch.is_autocast_enabled():
|
@@ -434,6 +439,7 @@ class GemmoeFlashAttention2(GemmoeAttention):
|
|
434 |
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
|
435 |
f" {target_dtype}."
|
436 |
)
|
|
|
437 |
query_states = query_states.to(target_dtype)
|
438 |
key_states = key_states.to(target_dtype)
|
439 |
value_states = value_states.to(target_dtype)
|
@@ -467,7 +473,7 @@ class GemmoeFlashAttention2(GemmoeAttention):
|
|
467 |
attention_mask (`torch.Tensor`):
|
468 |
The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
|
469 |
position of padding tokens and 1 for the position of non-padding tokens.
|
470 |
-
dropout (`
|
471 |
Attention dropout
|
472 |
softmax_scale (`float`, *optional*):
|
473 |
The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
|
@@ -484,6 +490,7 @@ class GemmoeFlashAttention2(GemmoeAttention):
|
|
484 |
query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
|
485 |
query_states, key_states, value_states, attention_mask, query_length
|
486 |
)
|
|
|
487 |
cu_seqlens_q, cu_seqlens_k = cu_seq_lens
|
488 |
max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
|
489 |
|
@@ -499,6 +506,7 @@ class GemmoeFlashAttention2(GemmoeAttention):
|
|
499 |
softmax_scale=softmax_scale,
|
500 |
causal=causal,
|
501 |
)
|
|
|
502 |
attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
|
503 |
else:
|
504 |
attn_output = flash_attn_func(
|
@@ -509,15 +517,14 @@ class GemmoeFlashAttention2(GemmoeAttention):
|
|
509 |
|
510 |
def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
|
511 |
indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
|
512 |
-
|
513 |
batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
|
|
|
514 |
key_layer = index_first_axis(
|
515 |
key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
|
516 |
)
|
517 |
value_layer = index_first_axis(
|
518 |
value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
|
519 |
)
|
520 |
-
|
521 |
if query_length == kv_seq_len:
|
522 |
query_layer = index_first_axis(
|
523 |
query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
|
@@ -546,24 +553,16 @@ class GemmoeFlashAttention2(GemmoeAttention):
|
|
546 |
(max_seqlen_in_batch_q, max_seqlen_in_batch_k),
|
547 |
)
|
548 |
|
|
|
|
|
549 |
class GemmoeSdpaAttention(GemmoeAttention):
|
550 |
"""
|
551 |
Gemmoe attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
|
552 |
-
GemmoeAttention as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
|
553 |
SDPA API.
|
554 |
"""
|
555 |
|
556 |
-
|
557 |
-
"""
|
558 |
-
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
559 |
-
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
560 |
-
"""
|
561 |
-
batch, num_key_value_heads, slen, head_dim = x.shape
|
562 |
-
if n_rep == 1:
|
563 |
-
return x
|
564 |
-
x = x[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
565 |
-
return x.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
566 |
-
|
567 |
def forward(
|
568 |
self,
|
569 |
hidden_states: torch.Tensor,
|
@@ -576,11 +575,10 @@ class GemmoeSdpaAttention(GemmoeAttention):
|
|
576 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
577 |
if output_attentions:
|
578 |
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
|
579 |
-
|
580 |
-
|
581 |
-
|
582 |
-
|
583 |
-
|
584 |
return super().forward(
|
585 |
hidden_states=hidden_states,
|
586 |
attention_mask=attention_mask,
|
@@ -590,7 +588,7 @@ class GemmoeSdpaAttention(GemmoeAttention):
|
|
590 |
use_cache=use_cache,
|
591 |
cache_position=cache_position,
|
592 |
)
|
593 |
-
|
594 |
bsz, q_len, _ = hidden_states.size()
|
595 |
|
596 |
query_states = self.q_proj(hidden_states)
|
@@ -605,23 +603,19 @@ class GemmoeSdpaAttention(GemmoeAttention):
|
|
605 |
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, None)
|
606 |
|
607 |
past_key_value = getattr(self, "past_key_value", past_key_value)
|
|
|
608 |
if past_key_value is not None:
|
609 |
-
# sin and cos are specific to RoPE models;
|
610 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
611 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
612 |
|
613 |
-
key_states =
|
614 |
-
value_states =
|
615 |
|
616 |
causal_mask = attention_mask
|
617 |
if attention_mask is not None and cache_position is not None:
|
618 |
causal_mask = causal_mask[:, :, cache_position, : key_states.shape[-2]]
|
619 |
|
620 |
-
# Ensure query, key, and value states have the same dtype
|
621 |
-
common_dtype = query_states.dtype
|
622 |
-
key_states = key_states.to(dtype=common_dtype)
|
623 |
-
value_states = value_states.to(dtype=common_dtype)
|
624 |
-
|
625 |
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
|
626 |
# Reference: https://github.com/pytorch/pytorch/issues/112577.
|
627 |
if query_states.device.type == "cuda" and causal_mask is not None:
|
@@ -629,10 +623,6 @@ class GemmoeSdpaAttention(GemmoeAttention):
|
|
629 |
key_states = key_states.contiguous()
|
630 |
value_states = value_states.contiguous()
|
631 |
|
632 |
-
# Cast causal_mask to the same dtype as query_states
|
633 |
-
if causal_mask is not None:
|
634 |
-
causal_mask = causal_mask.to(dtype=query_states.dtype)
|
635 |
-
|
636 |
attn_output = torch.nn.functional.scaled_dot_product_attention(
|
637 |
query_states,
|
638 |
key_states,
|
@@ -643,167 +633,113 @@ class GemmoeSdpaAttention(GemmoeAttention):
|
|
643 |
|
644 |
attn_output = attn_output.transpose(1, 2).contiguous()
|
645 |
attn_output = attn_output.view(bsz, q_len, -1)
|
|
|
646 |
attn_output = self.o_proj(attn_output)
|
647 |
|
648 |
return attn_output, None, past_key_value
|
649 |
|
650 |
-
GEMMOE_ATTENTION_CLASSES = {
|
651 |
-
"eager": GemmoeAttention,
|
652 |
-
"flash_attention_2": GemmoeFlashAttention2,
|
653 |
-
"sdpa": GemmoeSdpaAttention,
|
654 |
-
}
|
655 |
-
|
656 |
-
class GemmoeMLP(nn.Module):
|
657 |
-
def __init__(self, config, hidden_size=None, intermediate_size=None):
|
658 |
-
super().__init__()
|
659 |
-
self.config = config
|
660 |
-
self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
|
661 |
-
self.intermediate_size = config.intermediate_size if intermediate_size is None else intermediate_size
|
662 |
|
663 |
-
|
664 |
-
|
665 |
-
|
666 |
-
|
667 |
-
|
668 |
-
def forward(self, x):
|
669 |
-
if self.config.pretraining_tp > 1:
|
670 |
-
slice = self.intermediate_size // self.config.pretraining_tp
|
671 |
-
gate_proj_slices = self.gate_proj.weight.split(slice, dim=0)
|
672 |
-
up_proj_slices = self.up_proj.weight.split(slice, dim=0)
|
673 |
-
down_proj_slices = self.down_proj.weight.split(slice, dim=1)
|
674 |
-
|
675 |
-
gate_proj = torch.cat(
|
676 |
-
[F.linear(x, gate_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1
|
677 |
-
)
|
678 |
-
up_proj = torch.cat([F.linear(x, up_proj_slices[i]) for i in range(self.config.pretraining_tp)], dim=-1)
|
679 |
-
|
680 |
-
intermediate_states = (self.act_fn(gate_proj) * up_proj).split(slice, dim=2)
|
681 |
-
down_proj = [
|
682 |
-
F.linear(intermediate_states[i], down_proj_slices[i]) for i in range(self.config.pretraining_tp)
|
683 |
-
]
|
684 |
-
down_proj = sum(down_proj)
|
685 |
-
else:
|
686 |
-
down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
|
687 |
|
688 |
-
|
689 |
-
|
690 |
-
class MoEGate(nn.Module):
|
691 |
-
def __init__(self, config):
|
692 |
super().__init__()
|
693 |
-
self.
|
694 |
-
self.
|
695 |
-
self.n_routed_experts = config.n_routed_experts
|
696 |
-
|
697 |
-
self.scoring_func = config.scoring_func
|
698 |
-
self.alpha = config.aux_loss_alpha
|
699 |
-
self.seq_aux = config.seq_aux
|
700 |
|
701 |
-
|
702 |
-
self.
|
703 |
-
self.
|
704 |
-
self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim)))
|
705 |
-
self.reset_parameters()
|
706 |
|
707 |
-
|
708 |
-
import torch.nn.init as init
|
709 |
-
init.kaiming_uniform_(self.weight, a=math.sqrt(5))
|
710 |
|
711 |
def forward(self, hidden_states):
|
712 |
-
|
713 |
-
|
714 |
-
|
715 |
-
logits = F.linear(hidden_states, self.weight, None)
|
716 |
-
if self.scoring_func == 'softmax':
|
717 |
-
scores = logits.softmax(dim=-1)
|
718 |
-
else:
|
719 |
-
raise NotImplementedError(f'insupportable scoring function for MoE gating: {self.scoring_func}')
|
720 |
-
|
721 |
-
### select top-k experts
|
722 |
-
topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False)
|
723 |
-
|
724 |
-
### norm gate to sum 1
|
725 |
-
if self.top_k > 1 and self.norm_topk_prob:
|
726 |
-
denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
|
727 |
-
topk_weight = topk_weight / denominator
|
728 |
-
|
729 |
-
### expert-level computation auxiliary loss
|
730 |
-
if self.training and self.alpha > 0.0:
|
731 |
-
scores_for_aux = scores
|
732 |
-
aux_topk = self.top_k
|
733 |
-
# always compute aux loss based on the naive greedy topk method
|
734 |
-
topk_idx_for_aux_loss = topk_idx.view(bsz, -1)
|
735 |
-
if self.seq_aux:
|
736 |
-
scores_for_seq_aux = scores_for_aux.view(bsz, seq_len, -1)
|
737 |
-
ce = torch.zeros(bsz, self.n_routed_experts, device=hidden_states.device)
|
738 |
-
ce.scatter_add_(1, topk_idx_for_aux_loss, torch.ones(bsz, seq_len * aux_topk, device=hidden_states.device)).div_(seq_len * aux_topk / self.n_routed_experts)
|
739 |
-
aux_loss = (ce * scores_for_seq_aux.mean(dim=1)).sum(dim=1).mean() * self.alpha
|
740 |
-
else:
|
741 |
-
mask_ce = F.one_hot(topk_idx_for_aux_loss.view(-1), num_classes=self.n_routed_experts)
|
742 |
-
ce = mask_ce.float().mean(0)
|
743 |
-
Pi = scores_for_aux.mean(0)
|
744 |
-
fi = ce * self.n_routed_experts
|
745 |
-
aux_loss = (Pi * fi).sum() * self.alpha
|
746 |
-
else:
|
747 |
-
aux_loss = None
|
748 |
-
return topk_idx, topk_weight, aux_loss
|
749 |
|
|
|
|
|
|
|
|
|
|
|
|
|
750 |
|
751 |
class GemmoeSparseMoeBlock(nn.Module):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
752 |
def __init__(self, config):
|
753 |
super().__init__()
|
754 |
self.hidden_dim = config.hidden_size
|
755 |
self.ffn_dim = config.intermediate_size
|
756 |
self.num_experts = config.num_local_experts
|
757 |
-
self.top_k =
|
758 |
|
759 |
-
|
|
|
760 |
|
761 |
-
self.experts = nn.ModuleList([
|
762 |
|
763 |
-
def forward(self, hidden_states: torch.Tensor) ->
|
|
|
764 |
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
765 |
hidden_states = hidden_states.view(-1, hidden_dim)
|
|
|
|
|
766 |
|
767 |
-
|
768 |
-
|
|
|
769 |
# we cast back to the input dtype
|
770 |
-
|
771 |
|
772 |
-
|
|
|
|
|
773 |
|
774 |
-
|
|
|
|
|
775 |
|
776 |
-
|
777 |
-
for
|
778 |
-
|
779 |
-
|
780 |
-
y[flat_topk_idx == i] = expert_output.to(y.dtype) # Cast expert_output to the same dtype as y
|
781 |
|
782 |
-
|
|
|
783 |
|
784 |
-
|
785 |
-
|
786 |
-
|
787 |
-
class AddAuxiliaryLoss(torch.autograd.Function):
|
788 |
-
"""
|
789 |
-
The trick function of adding auxiliary (aux) loss,
|
790 |
-
which includes the gradient of the aux loss during backpropagation.
|
791 |
-
"""
|
792 |
-
@staticmethod
|
793 |
-
def forward(ctx, x, loss):
|
794 |
-
assert loss.numel() == 1
|
795 |
-
ctx.dtype = loss.dtype
|
796 |
-
ctx.required_aux_loss = loss.requires_grad
|
797 |
-
return x
|
798 |
|
799 |
-
|
800 |
-
|
801 |
-
|
802 |
-
|
803 |
-
|
804 |
-
|
|
|
|
|
|
|
|
|
|
|
805 |
|
806 |
|
|
|
807 |
class GemmoeDecoderLayer(nn.Module):
|
808 |
def __init__(self, config: GemmoeConfig, layer_idx: int):
|
809 |
super().__init__()
|
@@ -824,10 +760,31 @@ class GemmoeDecoderLayer(nn.Module):
|
|
824 |
output_attentions: Optional[bool] = False,
|
825 |
output_router_logits: Optional[bool] = False,
|
826 |
use_cache: Optional[bool] = False,
|
827 |
-
cache_position: Optional[torch.LongTensor] = None,
|
828 |
**kwargs,
|
829 |
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
830 |
residual = hidden_states
|
|
|
831 |
hidden_states = self.input_layernorm(hidden_states)
|
832 |
|
833 |
# Self Attention
|
@@ -838,20 +795,15 @@ class GemmoeDecoderLayer(nn.Module):
|
|
838 |
past_key_value=past_key_value,
|
839 |
output_attentions=output_attentions,
|
840 |
use_cache=use_cache,
|
841 |
-
cache_position=cache_position,
|
842 |
-
**kwargs,
|
843 |
)
|
844 |
hidden_states = residual + hidden_states
|
845 |
|
846 |
# Fully Connected
|
847 |
residual = hidden_states
|
848 |
hidden_states = self.post_attention_layernorm(hidden_states)
|
849 |
-
hidden_states,
|
850 |
hidden_states = residual + hidden_states
|
851 |
|
852 |
-
if aux_loss is not None:
|
853 |
-
hidden_states = AddAuxiliaryLoss.apply(hidden_states, aux_loss)
|
854 |
-
|
855 |
outputs = (hidden_states,)
|
856 |
|
857 |
if output_attentions:
|
@@ -860,331 +812,364 @@ class GemmoeDecoderLayer(nn.Module):
|
|
860 |
if use_cache:
|
861 |
outputs += (present_key_value,)
|
862 |
|
|
|
|
|
|
|
863 |
return outputs
|
864 |
|
|
|
865 |
GEMMOE_START_DOCSTRING = r"""
|
866 |
-
This model inherits from [PreTrainedModel]. Check the superclass documentation for the generic methods the
|
867 |
-
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
868 |
-
etc.)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
869 |
"""
|
870 |
|
|
|
871 |
@add_start_docstrings(
|
872 |
-
"The bare Gemmoe Model outputting raw hidden-states without any specific head on top.",
|
873 |
-
GEMMOE_START_DOCSTRING,
|
874 |
)
|
875 |
|
876 |
class GemmoePreTrainedModel(PreTrainedModel):
|
877 |
-
|
878 |
-
|
879 |
-
|
880 |
-
|
881 |
-
|
882 |
-
|
883 |
-
|
884 |
-
|
885 |
-
|
886 |
-
|
887 |
-
|
888 |
-
|
889 |
-
|
890 |
-
|
891 |
-
|
892 |
-
|
893 |
-
|
894 |
-
|
895 |
-
|
896 |
-
|
897 |
-
|
898 |
-
|
899 |
-
|
900 |
-
|
901 |
-
|
902 |
-
|
903 |
-
|
904 |
-
|
905 |
-
|
906 |
-
|
907 |
-
|
908 |
-
|
909 |
-
|
910 |
-
|
911 |
-
|
912 |
-
|
913 |
-
|
914 |
-
|
915 |
-
|
916 |
-
|
|
|
|
|
917 |
|
918 |
GEMMOE_INPUTS_DOCSTRING = r"""
|
919 |
-
Args:
|
920 |
-
input_ids (torch.LongTensor of shape (batch_size, sequence_length)):
|
921 |
-
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
922 |
-
it.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
923 |
"""
|
924 |
|
|
|
925 |
@add_start_docstrings(
|
926 |
-
"The bare Gemmoe Model outputting raw hidden-states without any specific head on top.",
|
927 |
-
GEMMOE_START_DOCSTRING,
|
928 |
)
|
929 |
|
930 |
class GemmoeModel(GemmoePreTrainedModel):
|
931 |
-
|
932 |
-
|
933 |
-
config: GemmoeConfig
|
934 |
-
"""
|
935 |
-
|
936 |
-
|
937 |
-
def __init__(self, config: GemmoeConfig):
|
938 |
-
super().__init__(config)
|
939 |
-
self.padding_idx = config.pad_token_id
|
940 |
-
self.vocab_size = config.vocab_size
|
941 |
-
|
942 |
-
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
943 |
-
self.layers = nn.ModuleList(
|
944 |
-
[GemmoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
945 |
-
)
|
946 |
-
|
947 |
-
self.norm = GemmoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
948 |
-
|
949 |
-
self.gradient_checkpointing = False
|
950 |
-
|
951 |
-
# Register a causal mask to separate causal and padding mask creation. Merging happens in the attention class.
|
952 |
-
# NOTE: This is not friendly with TorchScript, ONNX, ExportedProgram serialization for very large `max_position_embeddings`.
|
953 |
-
causal_mask = torch.full(
|
954 |
-
(config.max_position_embeddings, config.max_position_embeddings), fill_value=True, dtype=torch.bool
|
955 |
-
)
|
956 |
-
self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
|
957 |
-
|
958 |
-
# Initialize weights and apply final processing
|
959 |
-
self.post_init()
|
960 |
-
|
961 |
-
def get_input_embeddings(self):
|
962 |
-
return self.embed_tokens
|
963 |
-
|
964 |
-
def set_input_embeddings(self, value):
|
965 |
-
self.embed_tokens = value
|
966 |
-
|
967 |
-
@add_start_docstrings_to_model_forward(GEMMOE_INPUTS_DOCSTRING)
|
968 |
-
@replace_return_docstrings(output_type=MoeModelOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
969 |
-
def forward(
|
970 |
-
self,
|
971 |
-
input_ids: torch.LongTensor = None,
|
972 |
-
attention_mask: Optional[torch.Tensor] = None,
|
973 |
-
position_ids: Optional[torch.LongTensor] = None,
|
974 |
-
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
975 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
976 |
-
use_cache: Optional[bool] = None,
|
977 |
-
output_attentions: Optional[bool] = None,
|
978 |
-
output_hidden_states: Optional[bool] = None,
|
979 |
-
output_router_logits: Optional[bool] = None,
|
980 |
-
return_dict: Optional[bool] = None,
|
981 |
-
cache_position: Optional[torch.LongTensor] = None,
|
982 |
-
) -> Union[Tuple, MoeModelOutputWithPast]:
|
983 |
-
"""
|
984 |
-
Forward pass of the sequence classification model.
|
985 |
-
|
986 |
-
Args:
|
987 |
-
input_ids: Input token IDs.
|
988 |
-
attention_mask: Attention mask.
|
989 |
-
position_ids: Position IDs.
|
990 |
-
past_key_values: Past key-value pairs.
|
991 |
-
inputs_embeds: Input embeddings.
|
992 |
-
labels: Labels for sequence classification.
|
993 |
-
use_cache: Whether to use cache.
|
994 |
-
output_attentions: Whether to output attentions.
|
995 |
-
output_hidden_states: Whether to output hidden states.
|
996 |
-
return_dict: Whether to return a dictionary or tuple.
|
997 |
-
|
998 |
-
Returns:
|
999 |
-
Output of the sequence classification model.
|
1000 |
-
"""
|
1001 |
-
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
1002 |
-
output_hidden_states = (
|
1003 |
-
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
1004 |
-
)
|
1005 |
-
output_router_logits = (
|
1006 |
-
output_router_logits if output_router_logits is not None else self.config.output_router_logits
|
1007 |
-
)
|
1008 |
-
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
1009 |
-
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1010 |
-
|
1011 |
-
if (input_ids is None) ^ (inputs_embeds is not None):
|
1012 |
-
raise ValueError(
|
1013 |
-
"You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
|
1014 |
-
)
|
1015 |
-
|
1016 |
-
if self.gradient_checkpointing and self.training and use_cache:
|
1017 |
-
logger.warning_once(
|
1018 |
-
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
|
1019 |
-
)
|
1020 |
-
use_cache = False
|
1021 |
-
|
1022 |
-
if inputs_embeds is None:
|
1023 |
-
inputs_embeds = self.embed_tokens(input_ids)
|
1024 |
-
|
1025 |
-
past_seen_tokens = 0
|
1026 |
-
if use_cache: # kept for BC (cache positions)
|
1027 |
-
if not isinstance(past_key_values, StaticCache):
|
1028 |
-
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
|
1029 |
-
past_seen_tokens = past_key_values.get_seq_length()
|
1030 |
-
|
1031 |
-
if cache_position is None:
|
1032 |
-
cache_position = torch.arange(
|
1033 |
-
past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
|
1034 |
-
)
|
1035 |
-
|
1036 |
-
if position_ids is None:
|
1037 |
-
position_ids = cache_position.unsqueeze(0)
|
1038 |
-
|
1039 |
-
causal_mask = self._update_causal_mask(attention_mask, inputs_embeds)
|
1040 |
-
|
1041 |
-
hidden_states = inputs_embeds
|
1042 |
-
|
1043 |
-
# Normalize
|
1044 |
-
scale_factor = torch.tensor(math_sqrt(self.config.hidden_size), dtype=hidden_states.dtype)
|
1045 |
-
hidden_states = hidden_states * scale_factor
|
1046 |
-
# Decoder layers
|
1047 |
-
all_hidden_states = () if output_hidden_states else None
|
1048 |
-
all_self_attns = () if output_attentions else None
|
1049 |
-
all_router_logits = () if output_router_logits else None
|
1050 |
-
next_decoder_cache = None
|
1051 |
-
|
1052 |
-
for decoder_layer in self.layers:
|
1053 |
-
if output_hidden_states:
|
1054 |
-
all_hidden_states += (hidden_states,)
|
1055 |
-
|
1056 |
-
if self.gradient_checkpointing and self.training:
|
1057 |
-
layer_outputs = self._gradient_checkpointing_func(
|
1058 |
-
decoder_layer.__call__,
|
1059 |
-
hidden_states,
|
1060 |
-
causal_mask,
|
1061 |
-
position_ids,
|
1062 |
-
past_key_values,
|
1063 |
-
output_attentions,
|
1064 |
-
output_router_logits,
|
1065 |
-
use_cache,
|
1066 |
-
cache_position,
|
1067 |
-
)
|
1068 |
-
else:
|
1069 |
-
layer_outputs = decoder_layer(
|
1070 |
-
hidden_states,
|
1071 |
-
attention_mask=causal_mask,
|
1072 |
-
position_ids=position_ids,
|
1073 |
-
past_key_value=past_key_values,
|
1074 |
-
output_attentions=output_attentions,
|
1075 |
-
output_router_logits=output_router_logits,
|
1076 |
-
use_cache=use_cache,
|
1077 |
-
cache_position=cache_position,
|
1078 |
-
)
|
1079 |
-
|
1080 |
-
hidden_states = layer_outputs[0]
|
1081 |
-
if use_cache:
|
1082 |
-
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
|
1083 |
-
if output_attentions:
|
1084 |
-
all_self_attns += (layer_outputs[1],)
|
1085 |
-
if output_router_logits:
|
1086 |
-
all_router_logits += (layer_outputs[-1],)
|
1087 |
-
|
1088 |
-
hidden_states = self.norm(hidden_states)
|
1089 |
-
|
1090 |
-
# Add hidden states from the last decoder layer
|
1091 |
-
if output_hidden_states:
|
1092 |
-
all_hidden_states += (hidden_states,)
|
1093 |
-
|
1094 |
-
next_cache = None
|
1095 |
-
if use_cache:
|
1096 |
-
next_cache = (
|
1097 |
-
next_decoder_cache.to_legacy_cache() if isinstance(next_decoder_cache, Cache) else next_decoder_cache
|
1098 |
-
)
|
1099 |
-
|
1100 |
-
if not return_dict:
|
1101 |
-
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits] if v is not None)
|
1102 |
-
|
1103 |
-
return MoeModelOutputWithPast(
|
1104 |
-
last_hidden_state=hidden_states,
|
1105 |
-
past_key_values=next_cache,
|
1106 |
-
hidden_states=all_hidden_states,
|
1107 |
-
attentions=all_self_attns,
|
1108 |
-
router_logits=all_router_logits
|
1109 |
-
)
|
1110 |
-
|
1111 |
-
def _update_causal_mask(self, attention_mask, input_tensor):
|
1112 |
-
"""
|
1113 |
-
Update the causal mask based on the attention mask and input tensor.
|
1114 |
-
|
1115 |
-
Args:
|
1116 |
-
attention_mask (torch.Tensor): The attention mask.
|
1117 |
-
input_tensor (torch.Tensor): The input tensor.
|
1118 |
-
|
1119 |
-
Returns:
|
1120 |
-
torch.Tensor: The updated causal mask.
|
1121 |
-
"""
|
1122 |
-
|
1123 |
-
if self.config._attn_implementation == "flash_attention_2":
|
1124 |
-
if attention_mask is not None and 0.0 in attention_mask:
|
1125 |
-
return attention_mask
|
1126 |
-
return None
|
1127 |
-
|
1128 |
-
batch_size, seq_length = input_tensor.shape[:2]
|
1129 |
-
dtype = input_tensor.dtype
|
1130 |
-
device = input_tensor.device
|
1131 |
-
|
1132 |
-
# support going beyond cached `max_position_embedding`
|
1133 |
-
if seq_length > self.causal_mask.shape[-1]:
|
1134 |
-
logger.info(f"Resizing causal mask buffer from {self.causal_mask.shape[-1]} to {2 * self.causal_mask.shape[-1]}")
|
1135 |
-
causal_mask = torch.full((2 * self.causal_mask.shape[-1], 2 * self.causal_mask.shape[-1]), fill_value=1)
|
1136 |
-
self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
|
1137 |
-
|
1138 |
-
# We use the current dtype to avoid any overflows
|
1139 |
-
min_dtype = torch.finfo(dtype).min
|
1140 |
-
causal_mask = self.causal_mask[None, None, :, :].repeat(batch_size, 1, 1, 1).to(dtype) * min_dtype
|
1141 |
-
causal_mask = causal_mask.to(dtype=dtype, device=device)
|
1142 |
-
|
1143 |
-
if attention_mask is not None and attention_mask.dim() == 2:
|
1144 |
-
mask_length = attention_mask.shape[-1]
|
1145 |
-
padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
|
1146 |
-
causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
|
1147 |
-
|
1148 |
-
if self.config._attn_implementation == "sdpa" and attention_mask is not None:
|
1149 |
-
# TODO: For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
|
1150 |
-
is_tracing = (
|
1151 |
-
torch.jit.is_tracing()
|
1152 |
-
or isinstance(input_tensor, torch.fx.Proxy)
|
1153 |
-
or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
|
1154 |
-
)
|
1155 |
-
|
1156 |
-
if not is_tracing and torch.any(attention_mask != 1):
|
1157 |
-
# Attend to all tokens in masked rows from the causal_mask, for example the relevant first rows when
|
1158 |
-
# using left padding. This is required by
|
1159 |
-
# F.scaled_dot_product_attention memory-efficient attention path.
|
1160 |
-
# Details: https://github.com/pytorch/pytorch/issues/110213
|
1161 |
-
causal_mask = causal_mask.mul(~torch.all(causal_mask == min_dtype, dim=-1, keepdim=True)).to(dtype)
|
1162 |
-
|
1163 |
-
return causal_mask
|
1164 |
-
|
1165 |
-
class GemmoeForCausalLM(GemmoePreTrainedModel):
|
1166 |
-
r"""
|
1167 |
-
The Gemmoe Model transformer with a language modeling head on top for causal language modeling (CLM).
|
1168 |
|
1169 |
Args:
|
1170 |
-
config
|
|
|
|
|
|
|
|
|
|
|
|
|
1171 |
|
1172 |
-
|
1173 |
-
|
1174 |
-
|
|
|
|
|
|
|
1175 |
|
1176 |
-
|
1177 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
1178 |
|
1179 |
-
|
1180 |
-
|
1181 |
|
1182 |
-
|
1183 |
-
|
1184 |
-
|
1185 |
-
|
1186 |
-
|
1187 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1188 |
_tied_weights_keys = ["lm_head.weight"]
|
1189 |
|
1190 |
def __init__(self, config):
|
@@ -1193,9 +1178,8 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
|
|
1193 |
self.vocab_size = config.vocab_size
|
1194 |
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
1195 |
self.router_aux_loss_coef = config.router_aux_loss_coef
|
1196 |
-
self.num_experts =
|
1197 |
self.num_experts_per_tok = config.num_experts_per_tok
|
1198 |
-
|
1199 |
# Initialize weights and apply final processing
|
1200 |
self.post_init()
|
1201 |
|
@@ -1219,6 +1203,7 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
|
|
1219 |
|
1220 |
@add_start_docstrings_to_model_forward(GEMMOE_INPUTS_DOCSTRING)
|
1221 |
@replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
|
|
1222 |
def forward(
|
1223 |
self,
|
1224 |
input_ids: torch.LongTensor = None,
|
@@ -1232,7 +1217,6 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
|
|
1232 |
output_hidden_states: Optional[bool] = None,
|
1233 |
output_router_logits: Optional[bool] = None,
|
1234 |
return_dict: Optional[bool] = None,
|
1235 |
-
cache_position: Optional[torch.LongTensor] = None,
|
1236 |
) -> Union[Tuple, MoeCausalLMOutputWithPast]:
|
1237 |
r"""
|
1238 |
Args:
|
@@ -1248,26 +1232,29 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
|
|
1248 |
```python
|
1249 |
>>> from transformers import AutoTokenizer, GemmoeForCausalLM
|
1250 |
|
1251 |
-
>>> model = GemmoeForCausalLM.from_pretrained("
|
1252 |
-
>>> tokenizer = AutoTokenizer.from_pretrained("
|
1253 |
|
1254 |
-
>>> prompt = "
|
1255 |
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
1256 |
|
1257 |
>>> # Generate
|
1258 |
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
1259 |
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
1260 |
-
"
|
1261 |
```"""
|
|
|
1262 |
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
1263 |
output_router_logits = (
|
1264 |
-
output_router_logits if output_router_logits is not None else
|
1265 |
)
|
|
|
1266 |
output_hidden_states = (
|
1267 |
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
1268 |
)
|
1269 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1270 |
|
|
|
1271 |
outputs = self.model(
|
1272 |
input_ids=input_ids,
|
1273 |
attention_mask=attention_mask,
|
@@ -1279,42 +1266,39 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
|
|
1279 |
output_hidden_states=output_hidden_states,
|
1280 |
output_router_logits=output_router_logits,
|
1281 |
return_dict=return_dict,
|
1282 |
-
cache_position=cache_position,
|
1283 |
)
|
1284 |
|
1285 |
hidden_states = outputs[0]
|
1286 |
-
|
1287 |
-
# Ensure hidden_states and lm_head have compatible dtypes
|
1288 |
-
hidden_states = hidden_states.to(dtype=self.lm_head.weight.dtype)
|
1289 |
-
|
1290 |
logits = self.lm_head(hidden_states)
|
|
|
1291 |
|
1292 |
loss = None
|
1293 |
if labels is not None:
|
|
|
1294 |
shift_logits = logits[..., :-1, :].contiguous()
|
1295 |
shift_labels = labels[..., 1:].contiguous()
|
|
|
1296 |
loss_fct = CrossEntropyLoss()
|
1297 |
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
1298 |
shift_labels = shift_labels.view(-1)
|
|
|
1299 |
shift_labels = shift_labels.to(shift_logits.device)
|
1300 |
loss = loss_fct(shift_logits, shift_labels)
|
1301 |
|
1302 |
aux_loss = None
|
1303 |
if output_router_logits:
|
1304 |
-
|
1305 |
-
|
1306 |
-
|
1307 |
-
|
1308 |
-
|
1309 |
-
|
1310 |
-
|
1311 |
-
)
|
1312 |
-
if labels is not None:
|
1313 |
-
loss += self.router_aux_loss_coef * aux_loss.to(loss.device)
|
1314 |
|
1315 |
if not return_dict:
|
1316 |
output = (logits,) + outputs[1:]
|
1317 |
-
if
|
1318 |
output = (aux_loss,) + output
|
1319 |
return (loss,) + output if loss is not None else output
|
1320 |
|
@@ -1329,9 +1313,15 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
|
|
1329 |
)
|
1330 |
|
1331 |
def prepare_inputs_for_generation(
|
1332 |
-
self,
|
|
|
|
|
|
|
|
|
|
|
|
|
1333 |
):
|
1334 |
-
|
1335 |
if past_key_values is not None:
|
1336 |
if isinstance(past_key_values, Cache):
|
1337 |
cache_length = past_key_values.get_seq_length()
|
@@ -1341,11 +1331,19 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
|
|
1341 |
cache_length = past_length = past_key_values[0][0].shape[2]
|
1342 |
max_cache_length = None
|
1343 |
|
|
|
|
|
|
|
|
|
1344 |
if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
|
1345 |
input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
|
|
|
|
|
1346 |
elif past_length < input_ids.shape[1]:
|
1347 |
input_ids = input_ids[:, past_length:]
|
1348 |
-
|
|
|
|
|
1349 |
if (
|
1350 |
max_cache_length is not None
|
1351 |
and attention_mask is not None
|
@@ -1355,37 +1353,27 @@ class GemmoeForCausalLM(GemmoePreTrainedModel):
|
|
1355 |
|
1356 |
position_ids = kwargs.get("position_ids", None)
|
1357 |
if attention_mask is not None and position_ids is None:
|
|
|
1358 |
position_ids = attention_mask.long().cumsum(-1) - 1
|
1359 |
position_ids.masked_fill_(attention_mask == 0, 1)
|
1360 |
if past_key_values:
|
1361 |
position_ids = position_ids[:, -input_ids.shape[1] :]
|
1362 |
|
1363 |
-
if
|
1364 |
-
cache_position = kwargs.get("cache_position", None)
|
1365 |
-
if cache_position is None:
|
1366 |
-
past_length = 0
|
1367 |
-
else:
|
1368 |
-
past_length = cache_position[-1] + 1
|
1369 |
-
input_ids = input_ids[:, -1].unsqueeze(-1)
|
1370 |
-
position_ids = position_ids[:, -1].unsqueeze(-1)
|
1371 |
-
|
1372 |
-
cache_position = torch.arange(past_length, past_length + position_ids.shape[-1], device=position_ids.device)
|
1373 |
-
|
1374 |
if inputs_embeds is not None and past_key_values is None:
|
1375 |
model_inputs = {"inputs_embeds": inputs_embeds}
|
1376 |
else:
|
1377 |
-
model_inputs = {"input_ids": input_ids
|
1378 |
|
1379 |
model_inputs.update(
|
1380 |
{
|
1381 |
-
"position_ids": position_ids
|
1382 |
-
"cache_position": cache_position,
|
1383 |
"past_key_values": past_key_values,
|
1384 |
"use_cache": kwargs.get("use_cache"),
|
1385 |
"attention_mask": attention_mask,
|
|
|
1386 |
}
|
1387 |
)
|
1388 |
-
|
1389 |
return model_inputs
|
1390 |
|
1391 |
@staticmethod
|
@@ -1418,6 +1406,7 @@ class GemmoeForSequenceClassification(GemmoePreTrainedModel):
|
|
1418 |
self.num_labels = config.num_labels
|
1419 |
self.model = GemmoeModel(config)
|
1420 |
self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
|
|
|
1421 |
# Initialize weights and apply final processing
|
1422 |
self.post_init()
|
1423 |
|
@@ -1428,7 +1417,6 @@ class GemmoeForSequenceClassification(GemmoePreTrainedModel):
|
|
1428 |
self.model.embed_tokens = value
|
1429 |
|
1430 |
@add_start_docstrings_to_model_forward(GEMMOE_INPUTS_DOCSTRING)
|
1431 |
-
@replace_return_docstrings(output_type=SequenceClassifierOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
1432 |
def forward(
|
1433 |
self,
|
1434 |
input_ids: torch.LongTensor = None,
|
@@ -1442,25 +1430,14 @@ class GemmoeForSequenceClassification(GemmoePreTrainedModel):
|
|
1442 |
output_hidden_states: Optional[bool] = None,
|
1443 |
return_dict: Optional[bool] = None,
|
1444 |
) -> Union[Tuple, SequenceClassifierOutputWithPast]:
|
1445 |
-
"""
|
1446 |
-
|
1447 |
-
|
1448 |
-
|
1449 |
-
|
1450 |
-
attention_mask (torch.Tensor, optional): Attention mask.
|
1451 |
-
position_ids (torch.LongTensor, optional): Position IDs.
|
1452 |
-
past_key_values (List[torch.FloatTensor], optional): Past key-value pairs.
|
1453 |
-
inputs_embeds (torch.FloatTensor, optional): Input embeddings.
|
1454 |
-
labels (torch.LongTensor, optional): Labels for sequence classification.
|
1455 |
-
use_cache (bool, optional): Whether to use cache.
|
1456 |
-
output_attentions (bool, optional): Whether to output attentions.
|
1457 |
-
output_hidden_states (bool, optional): Whether to output hidden states.
|
1458 |
-
return_dict (bool, optional): Whether to return a dictionary or tuple.
|
1459 |
-
|
1460 |
-
Returns:
|
1461 |
-
Union[Tuple, SequenceClassifierOutputWithPast]: Output of the sequence classification model.
|
1462 |
"""
|
1463 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
|
|
1464 |
transformer_outputs = self.model(
|
1465 |
input_ids,
|
1466 |
attention_mask=attention_mask,
|
@@ -1486,8 +1463,10 @@ class GemmoeForSequenceClassification(GemmoePreTrainedModel):
|
|
1486 |
sequence_lengths = -1
|
1487 |
else:
|
1488 |
if input_ids is not None:
|
1489 |
-
|
1490 |
-
sequence_lengths =
|
|
|
|
|
1491 |
else:
|
1492 |
sequence_lengths = -1
|
1493 |
|
@@ -1516,7 +1495,6 @@ class GemmoeForSequenceClassification(GemmoePreTrainedModel):
|
|
1516 |
elif self.config.problem_type == "multi_label_classification":
|
1517 |
loss_fct = BCEWithLogitsLoss()
|
1518 |
loss = loss_fct(pooled_logits, labels)
|
1519 |
-
|
1520 |
if not return_dict:
|
1521 |
output = (pooled_logits,) + transformer_outputs[1:]
|
1522 |
return ((loss,) + output) if loss is not None else output
|
@@ -1527,4 +1505,4 @@ class GemmoeForSequenceClassification(GemmoePreTrainedModel):
|
|
1527 |
past_key_values=transformer_outputs.past_key_values,
|
1528 |
hidden_states=transformer_outputs.hidden_states,
|
1529 |
attentions=transformer_outputs.attentions,
|
1530 |
-
)
|
|
|
28 |
from transformers.activations import ACT2FN
|
29 |
from transformers.cache_utils import Cache, DynamicCache, StaticCache
|
30 |
from transformers.modeling_attn_mask_utils import (
|
31 |
+
AttentionMaskConverter,
|
32 |
_prepare_4d_causal_attention_mask,
|
33 |
)
|
34 |
from transformers.modeling_outputs import SequenceClassifierOutputWithPast, MoeModelOutputWithPast, MoeCausalLMOutputWithPast
|
|
|
177 |
class GemmoeRotaryEmbedding(nn.Module):
|
178 |
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
179 |
super().__init__()
|
180 |
+
|
181 |
self.dim = dim
|
182 |
self.max_position_embeddings = max_position_embeddings
|
183 |
self.base = base
|
184 |
+
self.register_buffer("inv_freq", None, persistent=False)
|
185 |
+
|
186 |
+
@torch.no_grad()
|
187 |
+
def forward(self, x, position_ids, seq_len=None):
|
188 |
+
# x: [bs, num_attention_heads, seq_len, head_size]
|
189 |
+
if self.inv_freq is None:
|
190 |
+
self.inv_freq = 1.0 / (
|
191 |
+
self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
|
192 |
+
)
|
193 |
+
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
|
194 |
+
position_ids_expanded = position_ids[:, None, :].float()
|
195 |
+
# Force float32 since bfloat16 loses precision on long contexts
|
196 |
+
# See https://github.com/huggingface/transformers/pull/29285
|
197 |
+
device_type = x.device.type
|
198 |
+
device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
|
199 |
+
with torch.autocast(device_type=device_type, enabled=False):
|
200 |
+
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
201 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
202 |
+
cos = emb.cos()
|
203 |
+
sin = emb.sin()
|
204 |
+
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
205 |
+
|
206 |
+
# Copied from transformers.models.llama.modeling_llama.rotate_half
|
|
|
|
|
|
|
|
|
207 |
def rotate_half(x):
|
208 |
"""Rotates half the hidden dims of the input."""
|
209 |
x1 = x[..., : x.shape[-1] // 2]
|
210 |
x2 = x[..., x.shape[-1] // 2 :]
|
211 |
return torch.cat((-x2, x1), dim=-1)
|
212 |
|
213 |
+
|
214 |
+
# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
|
215 |
+
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
216 |
+
"""Applies Rotary Position Embedding to the query and key tensors.
|
217 |
+
|
218 |
+
Args:
|
219 |
+
q (`torch.Tensor`): The query tensor.
|
220 |
+
k (`torch.Tensor`): The key tensor.
|
221 |
+
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
222 |
+
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
223 |
+
position_ids (`torch.Tensor`, *optional*):
|
224 |
+
Deprecated and unused.
|
225 |
+
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
226 |
+
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
227 |
+
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
228 |
+
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
229 |
+
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
230 |
+
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
231 |
+
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
232 |
+
Returns:
|
233 |
+
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
234 |
+
"""
|
235 |
+
cos = cos.unsqueeze(unsqueeze_dim)
|
236 |
+
sin = sin.unsqueeze(unsqueeze_dim)
|
237 |
q_embed = (q * cos) + (rotate_half(q) * sin)
|
238 |
k_embed = (k * cos) + (rotate_half(k) * sin)
|
239 |
return q_embed, k_embed
|
240 |
|
241 |
+
|
242 |
+
|
243 |
+
# Copied from transformers.models.llama.modeling_llama.repeat_kv
|
244 |
+
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
245 |
"""
|
246 |
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
247 |
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
|
|
253 |
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
254 |
|
255 |
class GemmoeAttention(nn.Module):
|
256 |
+
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
|
|
|
|
|
|
|
|
|
|
|
|
257 |
|
258 |
+
# Ignore copy
|
259 |
def __init__(self, config: GemmoeConfig, layer_idx: Optional[int] = None):
|
260 |
super().__init__()
|
261 |
self.config = config
|
|
|
266 |
"lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
|
267 |
"when creating this class."
|
268 |
)
|
269 |
+
|
270 |
self.attention_dropout = config.attention_dropout
|
271 |
self.hidden_size = config.hidden_size
|
272 |
self.num_heads = config.num_attention_heads
|
|
|
282 |
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
|
283 |
f" and `num_heads`: {self.num_heads})."
|
284 |
)
|
285 |
+
|
286 |
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
|
287 |
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
|
288 |
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
|
289 |
+
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias)
|
290 |
self.rotary_emb = GemmoeRotaryEmbedding(
|
291 |
+
self.head_dim,
|
292 |
+
max_position_embeddings=self.max_position_embeddings,
|
293 |
+
base=self.rope_theta,
|
294 |
+
)
|
295 |
|
296 |
def forward(
|
297 |
self,
|
|
|
304 |
cache_position: Optional[torch.LongTensor] = None,
|
305 |
**kwargs,
|
306 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
307 |
bsz, q_len, _ = hidden_states.size()
|
308 |
|
309 |
query_states = self.q_proj(hidden_states)
|
|
|
315 |
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
316 |
|
317 |
past_key_value = getattr(self, "past_key_value", past_key_value)
|
|
|
318 |
cos, sin = self.rotary_emb(value_states, position_ids, seq_len=None)
|
319 |
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, None)
|
320 |
|
321 |
if past_key_value is not None:
|
322 |
+
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
323 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
324 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
325 |
|
326 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
327 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
328 |
|
329 |
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
330 |
|
|
|
338 |
# upcast attention to fp32
|
339 |
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
340 |
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
|
|
|
341 |
attn_output = torch.matmul(attn_weights, value_states)
|
342 |
|
343 |
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
|
|
347 |
)
|
348 |
|
349 |
attn_output = attn_output.transpose(1, 2).contiguous()
|
|
|
350 |
|
351 |
+
attn_output = attn_output.view(bsz, q_len, -1)
|
352 |
attn_output = self.o_proj(attn_output)
|
353 |
|
354 |
if not output_attentions:
|
|
|
356 |
|
357 |
return attn_output, attn_weights, past_key_value
|
358 |
|
359 |
+
|
360 |
+
# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2 with Llama->Gemmoe
|
361 |
class GemmoeFlashAttention2(GemmoeAttention):
|
362 |
"""
|
363 |
Gemmoe flash attention module. This module inherits from `GemmoeAttention` as the weights of the module stays
|
364 |
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
|
365 |
flash attention and deal with padding tokens in case the input contains any of them.
|
366 |
"""
|
367 |
+
|
368 |
def __init__(self, *args, **kwargs):
|
369 |
super().__init__(*args, **kwargs)
|
370 |
+
|
371 |
+
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
|
372 |
+
# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
|
373 |
+
# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
|
374 |
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
|
375 |
|
376 |
+
# Ignore copy
|
377 |
def forward(
|
378 |
self,
|
379 |
hidden_states: torch.Tensor,
|
|
|
404 |
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, None)
|
405 |
|
406 |
past_key_value = getattr(self, "past_key_value", past_key_value)
|
407 |
+
|
408 |
if past_key_value is not None:
|
409 |
+
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
410 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
411 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
412 |
|
|
|
423 |
# cast them back in the correct dtype just to be sure everything works as expected.
|
424 |
# This might slowdown training & inference so it is recommended to not cast the LayerNorms
|
425 |
# in fp32. (GemmoeRMSNorm handles it correctly)
|
426 |
+
|
427 |
input_dtype = query_states.dtype
|
428 |
if input_dtype == torch.float32:
|
429 |
if torch.is_autocast_enabled():
|
|
|
439 |
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
|
440 |
f" {target_dtype}."
|
441 |
)
|
442 |
+
|
443 |
query_states = query_states.to(target_dtype)
|
444 |
key_states = key_states.to(target_dtype)
|
445 |
value_states = value_states.to(target_dtype)
|
|
|
473 |
attention_mask (`torch.Tensor`):
|
474 |
The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
|
475 |
position of padding tokens and 1 for the position of non-padding tokens.
|
476 |
+
dropout (`float`):
|
477 |
Attention dropout
|
478 |
softmax_scale (`float`, *optional*):
|
479 |
The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
|
|
|
490 |
query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
|
491 |
query_states, key_states, value_states, attention_mask, query_length
|
492 |
)
|
493 |
+
|
494 |
cu_seqlens_q, cu_seqlens_k = cu_seq_lens
|
495 |
max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
|
496 |
|
|
|
506 |
softmax_scale=softmax_scale,
|
507 |
causal=causal,
|
508 |
)
|
509 |
+
|
510 |
attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
|
511 |
else:
|
512 |
attn_output = flash_attn_func(
|
|
|
517 |
|
518 |
def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
|
519 |
indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
|
|
|
520 |
batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
|
521 |
+
|
522 |
key_layer = index_first_axis(
|
523 |
key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
|
524 |
)
|
525 |
value_layer = index_first_axis(
|
526 |
value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
|
527 |
)
|
|
|
528 |
if query_length == kv_seq_len:
|
529 |
query_layer = index_first_axis(
|
530 |
query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
|
|
|
553 |
(max_seqlen_in_batch_q, max_seqlen_in_batch_k),
|
554 |
)
|
555 |
|
556 |
+
|
557 |
+
# Copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Gemmoe
|
558 |
class GemmoeSdpaAttention(GemmoeAttention):
|
559 |
"""
|
560 |
Gemmoe attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
|
561 |
+
`GemmoeAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
|
562 |
SDPA API.
|
563 |
"""
|
564 |
|
565 |
+
# Ignore copy
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
566 |
def forward(
|
567 |
self,
|
568 |
hidden_states: torch.Tensor,
|
|
|
575 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
576 |
if output_attentions:
|
577 |
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
|
578 |
+
logger.warning_once(
|
579 |
+
"GemmoeModel is using GemmoeSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
|
580 |
+
'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
|
581 |
+
)
|
|
|
582 |
return super().forward(
|
583 |
hidden_states=hidden_states,
|
584 |
attention_mask=attention_mask,
|
|
|
588 |
use_cache=use_cache,
|
589 |
cache_position=cache_position,
|
590 |
)
|
591 |
+
|
592 |
bsz, q_len, _ = hidden_states.size()
|
593 |
|
594 |
query_states = self.q_proj(hidden_states)
|
|
|
603 |
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, None)
|
604 |
|
605 |
past_key_value = getattr(self, "past_key_value", past_key_value)
|
606 |
+
|
607 |
if past_key_value is not None:
|
608 |
+
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
609 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
610 |
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
611 |
|
612 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
613 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
614 |
|
615 |
causal_mask = attention_mask
|
616 |
if attention_mask is not None and cache_position is not None:
|
617 |
causal_mask = causal_mask[:, :, cache_position, : key_states.shape[-2]]
|
618 |
|
|
|
|
|
|
|
|
|
|
|
619 |
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
|
620 |
# Reference: https://github.com/pytorch/pytorch/issues/112577.
|
621 |
if query_states.device.type == "cuda" and causal_mask is not None:
|
|
|
623 |
key_states = key_states.contiguous()
|
624 |
value_states = value_states.contiguous()
|
625 |
|
|
|
|
|
|
|
|
|
626 |
attn_output = torch.nn.functional.scaled_dot_product_attention(
|
627 |
query_states,
|
628 |
key_states,
|
|
|
633 |
|
634 |
attn_output = attn_output.transpose(1, 2).contiguous()
|
635 |
attn_output = attn_output.view(bsz, q_len, -1)
|
636 |
+
|
637 |
attn_output = self.o_proj(attn_output)
|
638 |
|
639 |
return attn_output, None, past_key_value
|
640 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
641 |
|
642 |
+
GEMMOE_ATTENTION_CLASSES = {
|
643 |
+
"eager": GemmoeAttention,
|
644 |
+
"flash_attention_2": GemmoeFlashAttention2,
|
645 |
+
"sdpa": GemmoeSdpaAttention,
|
646 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
647 |
|
648 |
+
class GemmoeBlockSparseTop2MLP(nn.Module):
|
649 |
+
def __init__(self, config: GemmoeConfig):
|
|
|
|
|
650 |
super().__init__()
|
651 |
+
self.ffn_dim = config.intermediate_size
|
652 |
+
self.hidden_dim = config.hidden_size
|
|
|
|
|
|
|
|
|
|
|
653 |
|
654 |
+
self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
|
655 |
+
self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)
|
656 |
+
self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
|
|
|
|
|
657 |
|
658 |
+
self.act_fn = approx_gelu
|
|
|
|
|
659 |
|
660 |
def forward(self, hidden_states):
|
661 |
+
current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)
|
662 |
+
current_hidden_states = self.w2(current_hidden_states)
|
663 |
+
return current_hidden_states
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
664 |
|
665 |
+
class GemmoeBlockSparseTop2MLP(GemmoeBlockSparseTop2MLP):
|
666 |
+
def __init__(self, *args, **kwargs):
|
667 |
+
logger.warning_once(
|
668 |
+
"GemmoeBLockSparseTop2MLP is deprecated by GemmoeBlockSparseTop2MLP and will be removed in v4.40."
|
669 |
+
)
|
670 |
+
super().__init__(*args, **kwargs)
|
671 |
|
672 |
class GemmoeSparseMoeBlock(nn.Module):
|
673 |
+
"""
|
674 |
+
This implementation is
|
675 |
+
strictly equivalent to standard MoE with full capacity (no
|
676 |
+
dropped tokens). It's faster since it formulates MoE operations
|
677 |
+
in terms of block-sparse operations to accomodate imbalanced
|
678 |
+
assignments of tokens to experts, whereas standard MoE either
|
679 |
+
(1) drop tokens at the cost of reduced performance or (2) set
|
680 |
+
capacity factor to number of experts and thus waste computation
|
681 |
+
and memory on padding.
|
682 |
+
"""
|
683 |
+
|
684 |
def __init__(self, config):
|
685 |
super().__init__()
|
686 |
self.hidden_dim = config.hidden_size
|
687 |
self.ffn_dim = config.intermediate_size
|
688 |
self.num_experts = config.num_local_experts
|
689 |
+
self.top_k = config.num_experts_per_tok
|
690 |
|
691 |
+
# gating
|
692 |
+
self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
|
693 |
|
694 |
+
self.experts = nn.ModuleList([GemmoeBlockSparseTop2MLP(config) for _ in range(self.num_experts)])
|
695 |
|
696 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
697 |
+
""" """
|
698 |
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
699 |
hidden_states = hidden_states.view(-1, hidden_dim)
|
700 |
+
# router_logits: (batch * sequence_length, n_experts)
|
701 |
+
router_logits = self.gate(hidden_states)
|
702 |
|
703 |
+
routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
|
704 |
+
routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
|
705 |
+
routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
|
706 |
# we cast back to the input dtype
|
707 |
+
routing_weights = routing_weights.to(hidden_states.dtype)
|
708 |
|
709 |
+
final_hidden_states = torch.zeros(
|
710 |
+
(batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
|
711 |
+
)
|
712 |
|
713 |
+
# One hot encode the selected experts to create an expert mask
|
714 |
+
# this will be used to easily index which expert is going to be sollicitated
|
715 |
+
expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
|
716 |
|
717 |
+
# Loop over all available experts in the model and perform the computation on each expert
|
718 |
+
for expert_idx in range(self.num_experts):
|
719 |
+
expert_layer = self.experts[expert_idx]
|
720 |
+
idx, top_x = torch.where(expert_mask[expert_idx])
|
|
|
721 |
|
722 |
+
if top_x.shape[0] == 0:
|
723 |
+
continue
|
724 |
|
725 |
+
# in torch it is faster to index using lists than torch tensors
|
726 |
+
top_x_list = top_x.tolist()
|
727 |
+
idx_list = idx.tolist()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
728 |
|
729 |
+
# Index the correct hidden states and compute the expert hidden state for
|
730 |
+
# the current expert. We need to make sure to multiply the output hidden
|
731 |
+
# states by `routing_weights` on the corresponding tokens (top-1 and top-2)
|
732 |
+
current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)
|
733 |
+
current_hidden_states = expert_layer(current_state) * routing_weights[top_x_list, idx_list, None]
|
734 |
+
|
735 |
+
# However `index_add_` only support torch tensors for indexing so we'll use
|
736 |
+
# the `top_x` tensor here.
|
737 |
+
final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
|
738 |
+
final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
|
739 |
+
return final_hidden_states, router_logits
|
740 |
|
741 |
|
742 |
+
# Copied from transformers.models.llama.modeling_llama.LlamaDecoderLayer with LLAMA->GEMMOE,Llama->Gemmoe
|
743 |
class GemmoeDecoderLayer(nn.Module):
|
744 |
def __init__(self, config: GemmoeConfig, layer_idx: int):
|
745 |
super().__init__()
|
|
|
760 |
output_attentions: Optional[bool] = False,
|
761 |
output_router_logits: Optional[bool] = False,
|
762 |
use_cache: Optional[bool] = False,
|
|
|
763 |
**kwargs,
|
764 |
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
765 |
+
if "padding_mask" in kwargs:
|
766 |
+
warnings.warn(
|
767 |
+
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
|
768 |
+
)
|
769 |
+
"""
|
770 |
+
Args:
|
771 |
+
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
772 |
+
attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
|
773 |
+
`(batch, sequence_length)` where padding elements are indicated by 0.
|
774 |
+
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
|
775 |
+
output_attentions (`bool`, *optional*):
|
776 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
777 |
+
returned tensors for more detail.
|
778 |
+
output_router_logits (`bool`, *optional*):
|
779 |
+
Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
|
780 |
+
should not be returned during inference.
|
781 |
+
use_cache (`bool`, *optional*):
|
782 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
|
783 |
+
(see `past_key_values`).
|
784 |
+
"""
|
785 |
+
|
786 |
residual = hidden_states
|
787 |
+
|
788 |
hidden_states = self.input_layernorm(hidden_states)
|
789 |
|
790 |
# Self Attention
|
|
|
795 |
past_key_value=past_key_value,
|
796 |
output_attentions=output_attentions,
|
797 |
use_cache=use_cache,
|
|
|
|
|
798 |
)
|
799 |
hidden_states = residual + hidden_states
|
800 |
|
801 |
# Fully Connected
|
802 |
residual = hidden_states
|
803 |
hidden_states = self.post_attention_layernorm(hidden_states)
|
804 |
+
hidden_states, router_logits = self.block_sparse_moe(hidden_states)
|
805 |
hidden_states = residual + hidden_states
|
806 |
|
|
|
|
|
|
|
807 |
outputs = (hidden_states,)
|
808 |
|
809 |
if output_attentions:
|
|
|
812 |
if use_cache:
|
813 |
outputs += (present_key_value,)
|
814 |
|
815 |
+
if output_router_logits:
|
816 |
+
outputs += (router_logits,)
|
817 |
+
|
818 |
return outputs
|
819 |
|
820 |
+
|
821 |
GEMMOE_START_DOCSTRING = r"""
|
822 |
+
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
823 |
+
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
824 |
+
etc.)
|
825 |
+
|
826 |
+
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
827 |
+
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
828 |
+
and behavior.
|
829 |
+
|
830 |
+
Parameters:
|
831 |
+
config ([`GemmoeConfig`]):
|
832 |
+
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
833 |
+
load the weights associated with the model, only the configuration. Check out the
|
834 |
+
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
835 |
"""
|
836 |
|
837 |
+
|
838 |
@add_start_docstrings(
|
839 |
+
"The bare Gemmoe Model outputting raw hidden-states without any specific head on top.",
|
840 |
+
GEMMOE_START_DOCSTRING,
|
841 |
)
|
842 |
|
843 |
class GemmoePreTrainedModel(PreTrainedModel):
|
844 |
+
config_class = GemmoeConfig
|
845 |
+
base_model_prefix = "model"
|
846 |
+
supports_gradient_checkpointing = True
|
847 |
+
_keep_in_fp32_modules = ["inv_freq", "rotary_emb", "cos_cached", "sin_cached"]
|
848 |
+
_no_split_modules = ["GemmoeDecoderLayer"]
|
849 |
+
_skip_keys_device_placement = ["past_key_values", "causal_mask"]
|
850 |
+
_supports_flash_attn_2 = True
|
851 |
+
_supports_sdpa = True
|
852 |
+
_supports_cache_class = True
|
853 |
+
|
854 |
+
def _init_weights(self, module):
|
855 |
+
std = self.config.initializer_range
|
856 |
+
if isinstance(module, nn.Linear):
|
857 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
858 |
+
if module.bias is not None:
|
859 |
+
module.bias.data.zero_()
|
860 |
+
elif isinstance(module, nn.Embedding):
|
861 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
862 |
+
if module.padding_idx is not None:
|
863 |
+
module.weight.data[module.padding_idx].zero_()
|
864 |
+
|
865 |
+
def _setup_cache(self, cache_cls, max_batch_size, max_cache_len: Optional[int] = None):
|
866 |
+
if self.config._attn_implementation == "flash_attention_2" and cache_cls == StaticCache:
|
867 |
+
raise ValueError(
|
868 |
+
"`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` "
|
869 |
+
"make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformers"
|
870 |
+
)
|
871 |
+
|
872 |
+
if max_cache_len > self.model.causal_mask.shape[-1] or self.device != self.model.causal_mask.device:
|
873 |
+
causal_mask = torch.full((max_cache_len, max_cache_len), fill_value=1, device=self.device)
|
874 |
+
self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
|
875 |
+
|
876 |
+
for layer in self.model.layers:
|
877 |
+
weights = layer.self_attn.o_proj.weight
|
878 |
+
layer.self_attn.past_key_value = cache_cls(
|
879 |
+
self.config, max_batch_size, max_cache_len, device=weights.device, dtype=weights.dtype
|
880 |
+
)
|
881 |
+
|
882 |
+
def _reset_cache(self):
|
883 |
+
for layer in self.model.layers:
|
884 |
+
layer.self_attn.past_key_value = None
|
885 |
+
|
886 |
|
887 |
GEMMOE_INPUTS_DOCSTRING = r"""
|
888 |
+
Args:
|
889 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
890 |
+
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
891 |
+
it.
|
892 |
+
|
893 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
894 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
895 |
+
|
896 |
+
[What are input IDs?](../glossary#input-ids)
|
897 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
898 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
899 |
+
|
900 |
+
- 1 for tokens that are **not masked**,
|
901 |
+
- 0 for tokens that are **masked**.
|
902 |
+
|
903 |
+
[What are attention masks?](../glossary#attention-mask)
|
904 |
+
|
905 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
906 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
907 |
+
|
908 |
+
If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
|
909 |
+
`past_key_values`).
|
910 |
+
|
911 |
+
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
912 |
+
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
913 |
+
information on the default strategy.
|
914 |
+
|
915 |
+
- 1 indicates the head is **not masked**,
|
916 |
+
- 0 indicates the head is **masked**.
|
917 |
+
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
918 |
+
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
919 |
+
config.n_positions - 1]`.
|
920 |
+
|
921 |
+
[What are position IDs?](../glossary#position-ids)
|
922 |
+
past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
|
923 |
+
Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
924 |
+
blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
|
925 |
+
returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
|
926 |
+
|
927 |
+
Two formats are allowed:
|
928 |
+
- a [`~cache_utils.Cache`] instance;
|
929 |
+
- Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
|
930 |
+
shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
|
931 |
+
cache format.
|
932 |
+
|
933 |
+
The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
|
934 |
+
legacy cache format will be returned.
|
935 |
+
|
936 |
+
If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
|
937 |
+
have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
|
938 |
+
of shape `(batch_size, sequence_length)`.
|
939 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
940 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
941 |
+
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
942 |
+
model's internal embedding lookup matrix.
|
943 |
+
use_cache (`bool`, *optional*):
|
944 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
945 |
+
`past_key_values`).
|
946 |
+
output_attentions (`bool`, *optional*):
|
947 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
948 |
+
tensors for more detail.
|
949 |
+
output_hidden_states (`bool`, *optional*):
|
950 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
951 |
+
more detail.
|
952 |
+
return_dict (`bool`, *optional*):
|
953 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
954 |
+
cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
|
955 |
+
Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
|
956 |
+
this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
|
957 |
+
the complete sequence length.
|
958 |
"""
|
959 |
|
960 |
+
|
961 |
@add_start_docstrings(
|
962 |
+
"The bare Gemmoe Model outputting raw hidden-states without any specific head on top.",
|
963 |
+
GEMMOE_START_DOCSTRING,
|
964 |
)
|
965 |
|
966 |
class GemmoeModel(GemmoePreTrainedModel):
|
967 |
+
"""
|
968 |
+
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`GemmoeDecoderLayer`]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
969 |
|
970 |
Args:
|
971 |
+
config: GemmoeConfig
|
972 |
+
"""
|
973 |
+
|
974 |
+
def __init__(self, config: GemmoeConfig):
|
975 |
+
super().__init__(config)
|
976 |
+
self.padding_idx = config.pad_token_id
|
977 |
+
self.vocab_size = config.vocab_size
|
978 |
|
979 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
980 |
+
self.layers = nn.ModuleList(
|
981 |
+
[GemmoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
982 |
+
)
|
983 |
+
self.norm = GemmoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
984 |
+
self.gradient_checkpointing = False
|
985 |
|
986 |
+
# Register a causal mask to separate causal and padding mask creation. Merging happens in the attention class.
|
987 |
+
# NOTE: This is not friendly with TorchScript, ONNX, ExportedProgram serialization for very large `max_position_embeddings`.
|
988 |
+
causal_mask = torch.full(
|
989 |
+
(config.max_position_embeddings, config.max_position_embeddings), fill_value=True, dtype=torch.bool
|
990 |
+
)
|
991 |
+
self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
|
992 |
+
# Initialize weights and apply final processing
|
993 |
+
self.post_init()
|
994 |
|
995 |
+
def get_input_embeddings(self):
|
996 |
+
return self.embed_tokens
|
997 |
|
998 |
+
def set_input_embeddings(self, value):
|
999 |
+
self.embed_tokens = value
|
1000 |
+
|
1001 |
+
@add_start_docstrings_to_model_forward(GEMMOE_INPUTS_DOCSTRING)
|
1002 |
+
# Ignore copy
|
1003 |
+
def forward(
|
1004 |
+
self,
|
1005 |
+
input_ids: torch.LongTensor = None,
|
1006 |
+
attention_mask: Optional[torch.Tensor] = None,
|
1007 |
+
position_ids: Optional[torch.LongTensor] = None,
|
1008 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
1009 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
1010 |
+
use_cache: Optional[bool] = None,
|
1011 |
+
output_attentions: Optional[bool] = None,
|
1012 |
+
output_hidden_states: Optional[bool] = None,
|
1013 |
+
return_dict: Optional[bool] = None,
|
1014 |
+
cache_position: Optional[torch.LongTensor] = None,
|
1015 |
+
) -> Union[Tuple, MoeModelOutputWithPast]:
|
1016 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
1017 |
+
output_hidden_states = (
|
1018 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
1019 |
+
)
|
1020 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
1021 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1022 |
+
|
1023 |
+
if (input_ids is None) ^ (inputs_embeds is not None):
|
1024 |
+
raise ValueError(
|
1025 |
+
"You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
|
1026 |
+
)
|
1027 |
+
|
1028 |
+
if self.gradient_checkpointing and self.training and use_cache:
|
1029 |
+
logger.warning_once(
|
1030 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
|
1031 |
+
)
|
1032 |
+
use_cache = False
|
1033 |
+
|
1034 |
+
if inputs_embeds is None:
|
1035 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
1036 |
+
|
1037 |
+
past_seen_tokens = 0
|
1038 |
+
if use_cache: # kept for BC (cache positions)
|
1039 |
+
if not isinstance(past_key_values, StaticCache):
|
1040 |
+
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
|
1041 |
+
past_seen_tokens = past_key_values.get_seq_length()
|
1042 |
+
|
1043 |
+
if cache_position is None:
|
1044 |
+
cache_position = torch.arange(
|
1045 |
+
past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
|
1046 |
+
)
|
1047 |
+
|
1048 |
+
if position_ids is None:
|
1049 |
+
position_ids = cache_position.unsqueeze(0)
|
1050 |
+
|
1051 |
+
causal_mask = self._update_causal_mask(attention_mask, inputs_embeds)
|
1052 |
+
|
1053 |
+
# embed positions
|
1054 |
+
hidden_states = inputs_embeds
|
1055 |
+
|
1056 |
+
# normalized
|
1057 |
+
hidden_states = hidden_states * (self.config.hidden_size**0.5)
|
1058 |
+
|
1059 |
+
# decoder layers
|
1060 |
+
all_hidden_states = () if output_hidden_states else None
|
1061 |
+
all_self_attns = () if output_attentions else None
|
1062 |
+
next_decoder_cache = None
|
1063 |
+
|
1064 |
+
for decoder_layer in self.layers:
|
1065 |
+
if output_hidden_states:
|
1066 |
+
all_hidden_states += (hidden_states,)
|
1067 |
+
|
1068 |
+
if self.gradient_checkpointing and self.training:
|
1069 |
+
layer_outputs = self._gradient_checkpointing_func(
|
1070 |
+
decoder_layer.__call__,
|
1071 |
+
hidden_states,
|
1072 |
+
causal_mask,
|
1073 |
+
position_ids,
|
1074 |
+
past_key_values,
|
1075 |
+
output_attentions,
|
1076 |
+
use_cache,
|
1077 |
+
cache_position,
|
1078 |
+
)
|
1079 |
+
else:
|
1080 |
+
layer_outputs = decoder_layer(
|
1081 |
+
hidden_states,
|
1082 |
+
attention_mask=causal_mask,
|
1083 |
+
position_ids=position_ids,
|
1084 |
+
past_key_value=past_key_values,
|
1085 |
+
output_attentions=output_attentions,
|
1086 |
+
use_cache=use_cache,
|
1087 |
+
cache_position=cache_position,
|
1088 |
+
)
|
1089 |
+
|
1090 |
+
hidden_states = layer_outputs[0]
|
1091 |
+
|
1092 |
+
if use_cache:
|
1093 |
+
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
|
1094 |
+
|
1095 |
+
if output_attentions:
|
1096 |
+
all_self_attns += (layer_outputs[1],)
|
1097 |
+
|
1098 |
+
hidden_states = self.norm(hidden_states)
|
1099 |
+
|
1100 |
+
# add hidden states from the last decoder layer
|
1101 |
+
if output_hidden_states:
|
1102 |
+
all_hidden_states += (hidden_states,)
|
1103 |
+
|
1104 |
+
next_cache = None
|
1105 |
+
if use_cache:
|
1106 |
+
next_cache = (
|
1107 |
+
next_decoder_cache.to_legacy_cache() if isinstance(next_decoder_cache, Cache) else next_decoder_cache
|
1108 |
+
)
|
1109 |
+
if not return_dict:
|
1110 |
+
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
|
1111 |
+
return MoeModelOutputWithPast(
|
1112 |
+
last_hidden_state=hidden_states,
|
1113 |
+
past_key_values=next_cache,
|
1114 |
+
hidden_states=all_hidden_states,
|
1115 |
+
attentions=all_self_attns,
|
1116 |
+
)
|
1117 |
+
|
1118 |
+
# TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
|
1119 |
+
# KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
|
1120 |
+
# (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
|
1121 |
+
# `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
|
1122 |
+
def _update_causal_mask(self, attention_mask, input_tensor):
|
1123 |
+
if self.config._attn_implementation == "flash_attention_2":
|
1124 |
+
if attention_mask is not None and 0.0 in attention_mask:
|
1125 |
+
return attention_mask
|
1126 |
+
return None
|
1127 |
+
|
1128 |
+
batch_size, seq_length = input_tensor.shape[:2]
|
1129 |
+
dtype = input_tensor.dtype
|
1130 |
+
device = input_tensor.device
|
1131 |
+
|
1132 |
+
# support going beyond cached `max_position_embedding`
|
1133 |
+
if seq_length > self.causal_mask.shape[-1]:
|
1134 |
+
causal_mask = torch.full((2 * self.causal_mask.shape[-1], 2 * self.causal_mask.shape[-1]), fill_value=1)
|
1135 |
+
self.register_buffer("causal_mask", torch.triu(causal_mask, diagonal=1), persistent=False)
|
1136 |
+
|
1137 |
+
# We use the current dtype to avoid any overflows
|
1138 |
+
min_dtype = torch.finfo(dtype).min
|
1139 |
+
|
1140 |
+
causal_mask = self.causal_mask[None, None, :, :].to(dtype=dtype, device=device) * min_dtype
|
1141 |
+
causal_mask = causal_mask.expand(batch_size, 1, -1, -1)
|
1142 |
+
if attention_mask is not None:
|
1143 |
+
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
|
1144 |
+
if attention_mask.dim() == 2:
|
1145 |
+
mask_length = attention_mask.shape[-1]
|
1146 |
+
padding_mask = causal_mask[..., :mask_length].eq(0.0) * attention_mask[:, None, None, :].eq(0.0)
|
1147 |
+
causal_mask[..., :mask_length] = causal_mask[..., :mask_length].masked_fill(padding_mask, min_dtype)
|
1148 |
+
elif attention_mask.dim() == 4:
|
1149 |
+
mask_shape = attention_mask.shape
|
1150 |
+
mask_slice = (attention_mask.eq(0.0)).to(dtype=dtype) * min_dtype
|
1151 |
+
causal_mask[: mask_shape[0], : mask_shape[1], : mask_shape[2], : mask_shape[3]] = mask_slice
|
1152 |
+
|
1153 |
+
if (
|
1154 |
+
self.config._attn_implementation == "sdpa"
|
1155 |
+
and attention_mask is not None
|
1156 |
+
and attention_mask.device.type == "cuda"
|
1157 |
+
):
|
1158 |
+
# TODO: For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400).
|
1159 |
+
is_tracing = (
|
1160 |
+
torch.jit.is_tracing()
|
1161 |
+
or isinstance(input_tensor, torch.fx.Proxy)
|
1162 |
+
or (hasattr(torch, "_dynamo") and torch._dynamo.is_compiling())
|
1163 |
+
)
|
1164 |
+
if not is_tracing and torch.any(attention_mask != 1):
|
1165 |
+
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
|
1166 |
+
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
|
1167 |
+
# Details: https://github.com/pytorch/pytorch/issues/110213
|
1168 |
+
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
|
1169 |
+
|
1170 |
+
return causal_mask
|
1171 |
+
|
1172 |
+
class GemmoeForCausalLM(GemmoePreTrainedModel):
|
1173 |
_tied_weights_keys = ["lm_head.weight"]
|
1174 |
|
1175 |
def __init__(self, config):
|
|
|
1178 |
self.vocab_size = config.vocab_size
|
1179 |
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
1180 |
self.router_aux_loss_coef = config.router_aux_loss_coef
|
1181 |
+
self.num_experts = config.num_local_experts
|
1182 |
self.num_experts_per_tok = config.num_experts_per_tok
|
|
|
1183 |
# Initialize weights and apply final processing
|
1184 |
self.post_init()
|
1185 |
|
|
|
1203 |
|
1204 |
@add_start_docstrings_to_model_forward(GEMMOE_INPUTS_DOCSTRING)
|
1205 |
@replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
1206 |
+
# Ignore copy
|
1207 |
def forward(
|
1208 |
self,
|
1209 |
input_ids: torch.LongTensor = None,
|
|
|
1217 |
output_hidden_states: Optional[bool] = None,
|
1218 |
output_router_logits: Optional[bool] = None,
|
1219 |
return_dict: Optional[bool] = None,
|
|
|
1220 |
) -> Union[Tuple, MoeCausalLMOutputWithPast]:
|
1221 |
r"""
|
1222 |
Args:
|
|
|
1232 |
```python
|
1233 |
>>> from transformers import AutoTokenizer, GemmoeForCausalLM
|
1234 |
|
1235 |
+
>>> model = GemmoeForCausalLM.from_pretrained("mistralai/Gemmoe-8x7B-v0.1")
|
1236 |
+
>>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Gemmoe-8x7B-v0.1")
|
1237 |
|
1238 |
+
>>> prompt = "Hey, are you conscious? Can you talk to me?"
|
1239 |
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
1240 |
|
1241 |
>>> # Generate
|
1242 |
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
1243 |
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
1244 |
+
"Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
|
1245 |
```"""
|
1246 |
+
|
1247 |
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
1248 |
output_router_logits = (
|
1249 |
+
output_router_logits if output_router_logits is not None else self.config.output_router_logits
|
1250 |
)
|
1251 |
+
|
1252 |
output_hidden_states = (
|
1253 |
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
1254 |
)
|
1255 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1256 |
|
1257 |
+
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
1258 |
outputs = self.model(
|
1259 |
input_ids=input_ids,
|
1260 |
attention_mask=attention_mask,
|
|
|
1266 |
output_hidden_states=output_hidden_states,
|
1267 |
output_router_logits=output_router_logits,
|
1268 |
return_dict=return_dict,
|
|
|
1269 |
)
|
1270 |
|
1271 |
hidden_states = outputs[0]
|
|
|
|
|
|
|
|
|
1272 |
logits = self.lm_head(hidden_states)
|
1273 |
+
logits = logits.float()
|
1274 |
|
1275 |
loss = None
|
1276 |
if labels is not None:
|
1277 |
+
# Shift so that tokens < n predict n
|
1278 |
shift_logits = logits[..., :-1, :].contiguous()
|
1279 |
shift_labels = labels[..., 1:].contiguous()
|
1280 |
+
# Flatten the tokens
|
1281 |
loss_fct = CrossEntropyLoss()
|
1282 |
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
1283 |
shift_labels = shift_labels.view(-1)
|
1284 |
+
# Enable model parallelism
|
1285 |
shift_labels = shift_labels.to(shift_logits.device)
|
1286 |
loss = loss_fct(shift_logits, shift_labels)
|
1287 |
|
1288 |
aux_loss = None
|
1289 |
if output_router_logits:
|
1290 |
+
aux_loss = load_balancing_loss_func(
|
1291 |
+
outputs.router_logits if return_dict else outputs[-1],
|
1292 |
+
self.num_experts,
|
1293 |
+
self.num_experts_per_tok,
|
1294 |
+
attention_mask,
|
1295 |
+
)
|
1296 |
+
if labels is not None:
|
1297 |
+
loss += self.router_aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device
|
|
|
|
|
1298 |
|
1299 |
if not return_dict:
|
1300 |
output = (logits,) + outputs[1:]
|
1301 |
+
if output_router_logits:
|
1302 |
output = (aux_loss,) + output
|
1303 |
return (loss,) + output if loss is not None else output
|
1304 |
|
|
|
1313 |
)
|
1314 |
|
1315 |
def prepare_inputs_for_generation(
|
1316 |
+
self,
|
1317 |
+
input_ids,
|
1318 |
+
past_key_values=None,
|
1319 |
+
attention_mask=None,
|
1320 |
+
inputs_embeds=None,
|
1321 |
+
output_router_logits=False,
|
1322 |
+
**kwargs,
|
1323 |
):
|
1324 |
+
# Omit tokens covered by past_key_values
|
1325 |
if past_key_values is not None:
|
1326 |
if isinstance(past_key_values, Cache):
|
1327 |
cache_length = past_key_values.get_seq_length()
|
|
|
1331 |
cache_length = past_length = past_key_values[0][0].shape[2]
|
1332 |
max_cache_length = None
|
1333 |
|
1334 |
+
# Keep only the unprocessed tokens:
|
1335 |
+
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
|
1336 |
+
# some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
|
1337 |
+
# input)
|
1338 |
if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
|
1339 |
input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
|
1340 |
+
# 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
|
1341 |
+
# input_ids based on the past_length.
|
1342 |
elif past_length < input_ids.shape[1]:
|
1343 |
input_ids = input_ids[:, past_length:]
|
1344 |
+
# 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
|
1345 |
+
|
1346 |
+
# If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
|
1347 |
if (
|
1348 |
max_cache_length is not None
|
1349 |
and attention_mask is not None
|
|
|
1353 |
|
1354 |
position_ids = kwargs.get("position_ids", None)
|
1355 |
if attention_mask is not None and position_ids is None:
|
1356 |
+
# create position_ids on the fly for batch generation
|
1357 |
position_ids = attention_mask.long().cumsum(-1) - 1
|
1358 |
position_ids.masked_fill_(attention_mask == 0, 1)
|
1359 |
if past_key_values:
|
1360 |
position_ids = position_ids[:, -input_ids.shape[1] :]
|
1361 |
|
1362 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1363 |
if inputs_embeds is not None and past_key_values is None:
|
1364 |
model_inputs = {"inputs_embeds": inputs_embeds}
|
1365 |
else:
|
1366 |
+
model_inputs = {"input_ids": input_ids}
|
1367 |
|
1368 |
model_inputs.update(
|
1369 |
{
|
1370 |
+
"position_ids": position_ids,
|
|
|
1371 |
"past_key_values": past_key_values,
|
1372 |
"use_cache": kwargs.get("use_cache"),
|
1373 |
"attention_mask": attention_mask,
|
1374 |
+
"output_router_logits": output_router_logits,
|
1375 |
}
|
1376 |
)
|
|
|
1377 |
return model_inputs
|
1378 |
|
1379 |
@staticmethod
|
|
|
1406 |
self.num_labels = config.num_labels
|
1407 |
self.model = GemmoeModel(config)
|
1408 |
self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
|
1409 |
+
|
1410 |
# Initialize weights and apply final processing
|
1411 |
self.post_init()
|
1412 |
|
|
|
1417 |
self.model.embed_tokens = value
|
1418 |
|
1419 |
@add_start_docstrings_to_model_forward(GEMMOE_INPUTS_DOCSTRING)
|
|
|
1420 |
def forward(
|
1421 |
self,
|
1422 |
input_ids: torch.LongTensor = None,
|
|
|
1430 |
output_hidden_states: Optional[bool] = None,
|
1431 |
return_dict: Optional[bool] = None,
|
1432 |
) -> Union[Tuple, SequenceClassifierOutputWithPast]:
|
1433 |
+
r"""
|
1434 |
+
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
1435 |
+
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
1436 |
+
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
1437 |
+
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1438 |
"""
|
1439 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1440 |
+
|
1441 |
transformer_outputs = self.model(
|
1442 |
input_ids,
|
1443 |
attention_mask=attention_mask,
|
|
|
1463 |
sequence_lengths = -1
|
1464 |
else:
|
1465 |
if input_ids is not None:
|
1466 |
+
# if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
|
1467 |
+
sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
|
1468 |
+
sequence_lengths = sequence_lengths % input_ids.shape[-1]
|
1469 |
+
sequence_lengths = sequence_lengths.to(logits.device)
|
1470 |
else:
|
1471 |
sequence_lengths = -1
|
1472 |
|
|
|
1495 |
elif self.config.problem_type == "multi_label_classification":
|
1496 |
loss_fct = BCEWithLogitsLoss()
|
1497 |
loss = loss_fct(pooled_logits, labels)
|
|
|
1498 |
if not return_dict:
|
1499 |
output = (pooled_logits,) + transformer_outputs[1:]
|
1500 |
return ((loss,) + output) if loss is not None else output
|
|
|
1505 |
past_key_values=transformer_outputs.past_key_values,
|
1506 |
hidden_states=transformer_outputs.hidden_states,
|
1507 |
attentions=transformer_outputs.attentions,
|
1508 |
+
)
|