Update modeling_motif.py
Browse files- modeling_motif.py +821 -101
modeling_motif.py
CHANGED
@@ -1,5 +1,5 @@
|
|
1 |
import math
|
2 |
-
from typing import List, Optional, Tuple, Union
|
3 |
|
4 |
import torch
|
5 |
import torch.utils.checkpoint
|
@@ -28,25 +28,14 @@ from .configuration_motif import MotifConfig
|
|
28 |
from dataclasses import dataclass
|
29 |
|
30 |
import torch.nn.functional as F
|
31 |
-
import time
|
32 |
-
|
33 |
-
logger = logging.get_logger(__name__)
|
34 |
-
|
35 |
-
if is_flash_attn_2_available():
|
36 |
-
from transformers.modeling_flash_attention_utils import _flash_attention_forward
|
37 |
-
|
38 |
-
|
39 |
-
_CONFIG_FOR_DOC = "MotifConfig"
|
40 |
|
41 |
from transformers.activations import ACT2CLS as _ACT2CLS
|
42 |
from transformers.activations import ClassInstantier
|
43 |
-
|
44 |
-
|
45 |
class PolyNorm(torch.nn.Module):
|
46 |
-
"""
|
47 |
A trainable activation function introduced in https://arxiv.org/html/2411.03884v1.
|
48 |
The code is copied from https://github.com/BryceZhuo/PolyCom?tab=readme-ov-file/README.md,
|
49 |
-
with the change `* torch.rsqrt` => `/ torch.sqrt`
|
50 |
"""
|
51 |
|
52 |
def __init__(self, eps=1e-6):
|
@@ -62,11 +51,52 @@ class PolyNorm(torch.nn.Module):
|
|
62 |
return self.weight[0] * self._norm(x ** 3) + self.weight[1] * self._norm(
|
63 |
x ** 2) + self.weight[2] * self._norm(x) + self.bias
|
64 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
65 |
|
66 |
-
CUSTOM_ACT2CLS = {"poly_norm": PolyNorm}
|
67 |
ACT2CLS = {**_ACT2CLS, **CUSTOM_ACT2CLS}
|
68 |
ACT2FN = ClassInstantier(ACT2CLS)
|
69 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
|
71 |
class MotifRMSNorm(nn.Module):
|
72 |
|
@@ -80,7 +110,8 @@ class MotifRMSNorm(nn.Module):
|
|
80 |
|
81 |
def forward(self, hidden_states):
|
82 |
input_dtype = hidden_states.dtype
|
83 |
-
|
|
|
84 |
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
85 |
return self.weight * hidden_states.to(input_dtype)
|
86 |
|
@@ -88,21 +119,24 @@ class MotifRMSNorm(nn.Module):
|
|
88 |
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
89 |
|
90 |
|
91 |
-
ALL_LAYERNORM_LAYERS.append(MotifRMSNorm)
|
92 |
|
93 |
|
94 |
class MotifRotaryEmbeddingWithCache(nn.Module):
|
95 |
"""
|
96 |
Rotary positional embedding module with caching for efficiency.
|
|
|
97 |
Args:
|
98 |
dim (int): Dimensionality of the embedding.
|
99 |
max_position_embeddings (int): Maximum sequence length for caching. Default is 2048.
|
100 |
base (int): Base for computing inverse frequency. Default is 10000.
|
101 |
device (torch.device, optional): Device for tensor storage.
|
|
|
102 |
Methods:
|
103 |
forward(x, seq_len=None):
|
104 |
Computes cosine and sine embeddings for input sequence length.
|
105 |
Automatically updates cache if `seq_len` exceeds cached length.
|
|
|
106 |
Attributes:
|
107 |
inv_freq (torch.Tensor): Inverse frequency tensor for position encoding.
|
108 |
cos_cached (torch.Tensor): Cached cosine embeddings.
|
@@ -138,8 +172,8 @@ class MotifRotaryEmbeddingWithCache(nn.Module):
|
|
138 |
self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
|
139 |
|
140 |
return (
|
141 |
-
self.cos_cached[
|
142 |
-
self.sin_cached[
|
143 |
)
|
144 |
|
145 |
|
@@ -156,6 +190,7 @@ class MotifRotaryEmbedding(nn.Module):
|
|
156 |
config: Optional[MotifConfig] = None,
|
157 |
):
|
158 |
super().__init__()
|
|
|
159 |
self.rope_kwargs = {}
|
160 |
if config is None:
|
161 |
logger.warning_once(
|
@@ -200,7 +235,7 @@ class MotifRotaryEmbedding(nn.Module):
|
|
200 |
device,
|
201 |
seq_len=seq_len,
|
202 |
**self.rope_kwargs)
|
203 |
-
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
204 |
self.max_seq_len_cached = seq_len
|
205 |
|
206 |
if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len: # reset
|
@@ -234,8 +269,10 @@ class MotifRotaryEmbedding(nn.Module):
|
|
234 |
def rotate_half(x):
|
235 |
"""
|
236 |
Rotates half of the dimensions of the input tensor using torch.roll and in-place negation.
|
|
|
237 |
Args:
|
238 |
x (torch.Tensor): The input tensor.
|
|
|
239 |
Returns:
|
240 |
torch.Tensor: A tensor where the latter half of the dimensions are negated
|
241 |
and moved before the first half.
|
@@ -247,27 +284,60 @@ def rotate_half(x):
|
|
247 |
return rotated_tensor
|
248 |
|
249 |
|
250 |
-
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
251 |
"""
|
252 |
Applies rotary position embeddings to the input tensors.
|
|
|
253 |
Args:
|
254 |
q (torch.Tensor): Query tensor of shape (B, NH, S, D_KV).
|
255 |
k (torch.Tensor): Key tensor of shape (B, NH, S, D_KV).
|
256 |
cos (torch.Tensor): Cosine values for rotary embedding.
|
257 |
sin (torch.Tensor): Sine values for rotary embedding.
|
258 |
-
unsqueeze_dim (int, optional): Dimension along which `cos` and `sin` are unsqueezed.
|
259 |
Defaults to 1.
|
|
|
|
|
|
|
|
|
260 |
Returns:
|
261 |
Tuple[torch.Tensor, torch.Tensor]: Returns transformed query and key tensors after applying rotary embeddings.
|
262 |
"""
|
263 |
-
|
264 |
-
|
265 |
-
|
266 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
267 |
|
268 |
|
269 |
class MotifMLP(nn.Module):
|
270 |
-
|
271 |
def __init__(self, config):
|
272 |
super().__init__()
|
273 |
self.hidden_size = config.hidden_size
|
@@ -277,33 +347,420 @@ class MotifMLP(nn.Module):
|
|
277 |
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
|
278 |
self.act_fn = ACT2FN[config.hidden_act]
|
279 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
280 |
def forward(self, hidden_state):
|
281 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
282 |
|
283 |
|
284 |
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
285 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
286 |
|
|
|
|
|
287 |
|
288 |
-
# @log_timing
|
289 |
class MotifAttention(nn.Module):
|
290 |
"""
|
291 |
Differential Attention (DiffAttention) module.
|
292 |
-
|
|
|
293 |
"DIFFERENTIAL TRANSFORMER" (https://arxiv.org/pdf/2410.05258).
|
|
|
294 |
Overview
|
295 |
Standard transformers often over-allocate attention to irrelevant context.
|
296 |
-
DiffAttention addresses this by computing attention as the difference between
|
297 |
-
two separate softmax attention maps, effectively canceling noise and promoting
|
298 |
sparse, structured attention patterns.
|
|
|
299 |
Reference Implementation
|
300 |
https://github.com/microsoft/unilm/tree/master/Diff-Transformer
|
|
|
301 |
Args
|
302 |
-
The differential attention mechanism computes attention as the difference of two softmax attention scores, weighted by a learnable scalar λ.
|
303 |
λ is re-parameterized as λ = exp(λ_q1 · λ_k1) − exp(λ_q2 · λ_k2) + λ_init.
|
304 |
- lambda_q1, lambda_q2 (nn.Parameter): Learnable vectors used to compute the first and second components of λ for query transformations.
|
305 |
- lambda_k1, lambda_k2 (nn.Parameter): Learnable vectors used to compute the first and second components of λ for key transformations.
|
306 |
- lambda_init (float): A constant used for initializing λ, typically set as λ_init = 0.8 − 0.6 × exp(−0.3 × (layer_index − 1)).
|
|
|
307 |
"""
|
308 |
|
309 |
def __init__(self, config: MotifConfig, layer_idx: Optional[int] = None):
|
@@ -383,7 +840,7 @@ class MotifAttention(nn.Module):
|
|
383 |
self.subln = MotifRMSNorm(2 * self.head_dim, eps=1e-5)
|
384 |
self.lambda_init = 0.8 - 0.6 * math.exp(-0.3 * (layer_idx - 1))
|
385 |
|
386 |
-
self.rotary_emb =
|
387 |
max_position_embeddings=self.max_position_embeddings,
|
388 |
base=self.rope_theta)
|
389 |
|
@@ -429,12 +886,12 @@ class MotifAttention(nn.Module):
|
|
429 |
cos, sin = (self.rotary_emb(value_states, q_len + past_key_value.get_usable_length(q_len, self.layer_idx))
|
430 |
if use_cache else position_embeddings)
|
431 |
|
432 |
-
query_states, key_states = apply_rotary_pos_emb(
|
433 |
-
|
434 |
-
|
435 |
-
|
436 |
-
|
437 |
-
|
438 |
|
439 |
if past_key_value is not None:
|
440 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
|
@@ -503,7 +960,6 @@ class MotifAttention(nn.Module):
|
|
503 |
return attn_output, attn_weights, past_key_value
|
504 |
|
505 |
|
506 |
-
# @log_timing
|
507 |
class MotifFlashAttention2(MotifAttention):
|
508 |
"""
|
509 |
Motif flash attention module, following Motif attention module. This module inherits from `MotifAttention`
|
@@ -517,7 +973,7 @@ class MotifFlashAttention2(MotifAttention):
|
|
517 |
def __init__(self, *args, **kwargs):
|
518 |
super().__init__(*args, **kwargs)
|
519 |
|
520 |
-
|
521 |
|
522 |
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
|
523 |
# 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.
|
@@ -525,6 +981,8 @@ class MotifFlashAttention2(MotifAttention):
|
|
525 |
|
526 |
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
|
527 |
|
|
|
|
|
528 |
def _reshape_heads(self, tensor, batch_size, seq_len):
|
529 |
"""2-way head split tensor reshape"""
|
530 |
return tensor.reshape(batch_size, seq_len, self.num_heads, 2, self.head_dim)
|
@@ -534,27 +992,59 @@ class MotifFlashAttention2(MotifAttention):
|
|
534 |
return tensor.reshape(batch_size, seq_len, self.num_heads, self.head_dim)
|
535 |
|
536 |
def _compute_attention(self, query_states, key_states, value_states, attention_mask, q_len, position_ids,
|
537 |
-
dropout_rate, sliding_window, batch_num):
|
538 |
"""Flash Attention 2 implements"""
|
|
|
539 |
scale_factor = 1.0 / math.sqrt(self.head_dim)
|
540 |
-
|
541 |
if not self._flash_attn_uses_top_left_mask:
|
542 |
causal = self.is_causal
|
543 |
else:
|
544 |
causal = self.is_causal and q_len != 1
|
|
|
|
|
|
|
545 |
|
546 |
-
|
|
|
|
|
|
|
547 |
|
548 |
-
|
549 |
-
|
550 |
-
|
551 |
-
|
552 |
-
|
553 |
-
|
554 |
-
|
555 |
-
|
556 |
-
|
557 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
558 |
|
559 |
def forward(
|
560 |
self,
|
@@ -588,12 +1078,12 @@ class MotifFlashAttention2(MotifAttention):
|
|
588 |
cos, sin = (self.rotary_emb(value_states, q_len + past_key_value.get_usable_length(q_len, self.layer_idx))
|
589 |
if use_cache else position_embeddings)
|
590 |
|
591 |
-
query_states, key_states = apply_rotary_pos_emb(
|
592 |
-
|
593 |
-
|
594 |
-
|
595 |
-
|
596 |
-
|
597 |
|
598 |
if past_key_value is not None:
|
599 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
|
@@ -604,6 +1094,28 @@ class MotifFlashAttention2(MotifAttention):
|
|
604 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
605 |
dropout_rate = 0.0 if not self.training else self.attention_dropout
|
606 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
607 |
q_len = query_states.shape[-2]
|
608 |
kv_seq_len = key_states.shape[-2]
|
609 |
|
@@ -613,7 +1125,7 @@ class MotifFlashAttention2(MotifAttention):
|
|
613 |
value_states = value_states.transpose(1, 2)
|
614 |
|
615 |
if (self.config.use_sliding_window and getattr(self.config, "sliding_window", None) is not None
|
616 |
-
and self.layer_idx >= self.config.max_window_layers):
|
617 |
sliding_window = self.config.sliding_window
|
618 |
else:
|
619 |
sliding_window = None
|
@@ -633,13 +1145,14 @@ class MotifFlashAttention2(MotifAttention):
|
|
633 |
k1, k2 = k1.contiguous(), k2.contiguous()
|
634 |
v1, v2 = v1.contiguous(), v2.contiguous()
|
635 |
|
636 |
-
|
637 |
-
self._compute_attention(q1, k1, v2, attention_mask, q_len, position_ids, dropout_rate, sliding_window, self.batch_num)
|
638 |
-
attn21, attn22 = self._compute_attention(q2, k2, v1, attention_mask, q_len, position_ids, dropout_rate, sliding_window, self.batch_num), \
|
639 |
-
self._compute_attention(q2, k2, v2, attention_mask, q_len, position_ids, dropout_rate, sliding_window, self.batch_num)
|
640 |
|
641 |
-
|
642 |
-
|
|
|
|
|
|
|
|
|
643 |
|
644 |
lambda_q1 = self.lambda_q1.unsqueeze(0).expand([bsz, self.lambda_q1.shape[0]]) # bsz, num_head
|
645 |
lambda_q2 = self.lambda_q2.unsqueeze(0).expand([bsz, self.lambda_q2.shape[0]]) # bsz, num_head
|
@@ -655,16 +1168,15 @@ class MotifFlashAttention2(MotifAttention):
|
|
655 |
attn_output = attn_output * (1 - self.lambda_init)
|
656 |
|
657 |
if attn_output.size() != (bsz, q_len, self.num_heads, self.head_dim * 2):
|
658 |
-
raise ValueError(f"`attn_output` should be of size {(bsz, self.num_heads,
|
659 |
f" {attn_output.size()}")
|
660 |
|
661 |
-
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
|
662 |
attn_output = self.o_proj(attn_output) * self.o_proj_alpha
|
663 |
|
664 |
-
return attn_output
|
665 |
|
666 |
|
667 |
-
# @log_timing
|
668 |
class MotifSdpaAttention(MotifAttention):
|
669 |
"""
|
670 |
Motif attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
|
@@ -758,16 +1270,17 @@ class MotifSdpaAttention(MotifAttention):
|
|
758 |
MOTIF_ATTENTION_CLASSES = {
|
759 |
"eager": MotifAttention,
|
760 |
"flash_attention_2": MotifFlashAttention2,
|
761 |
-
"sdpa":
|
762 |
}
|
763 |
|
764 |
|
765 |
class MotifDecoderLayer(nn.Module):
|
766 |
|
767 |
-
def __init__(self, config: MotifConfig, layer_idx: int):
|
768 |
super().__init__()
|
769 |
self.hidden_size = config.hidden_size
|
770 |
-
|
|
|
771 |
if config.sliding_window and config._attn_implementation != "flash_attention_2":
|
772 |
logger.warning_once(
|
773 |
f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
|
@@ -777,8 +1290,12 @@ class MotifDecoderLayer(nn.Module):
|
|
777 |
else:
|
778 |
self.self_attn = MOTIF_ATTENTION_CLASSES["eager"](config, layer_idx)
|
779 |
self.mlp = MotifMLP(config)
|
780 |
-
|
781 |
-
|
|
|
|
|
|
|
|
|
782 |
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
783 |
self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
784 |
|
@@ -847,7 +1364,13 @@ class MotifDecoderLayer(nn.Module):
|
|
847 |
residual = hidden_states
|
848 |
hidden_states = self.post_attention_layernorm(hidden_states) * self.post_attention_layernorm_alpha
|
849 |
|
850 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
851 |
|
852 |
hidden_states = residual + hidden_states
|
853 |
|
@@ -866,9 +1389,11 @@ MOTIF_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 |
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
870 |
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
871 |
and behavior.
|
|
|
872 |
Parameters:
|
873 |
config ([`MotifConfig`]):
|
874 |
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
@@ -918,23 +1443,26 @@ class MotifPreTrainedModel(PreTrainedModel):
|
|
918 |
module_std = module_std / math.sqrt(self.config.dim_model_base_lmh) ### lmhead.. 1
|
919 |
else:
|
920 |
module_std = module_std
|
921 |
-
|
922 |
-
|
|
|
923 |
if module.bias is not None:
|
924 |
module.bias.data.zero_()
|
925 |
|
926 |
elif isinstance(module, nn.Embedding):
|
927 |
-
|
|
|
|
|
928 |
if module.padding_idx is not None:
|
929 |
module.weight.data[module.padding_idx].zero_()
|
930 |
|
931 |
|
932 |
@dataclass
|
933 |
class MotifModelOutputWithPast(ModelOutput):
|
934 |
-
"""
|
935 |
-
This augments `BaseModelOutputWithPast` in `transformers.modeling_outputs` with new optional keys: `causal_mask`, `position_embeddings`.
|
936 |
The optional keys are currently used in the following ways:
|
937 |
-
- pass information to the token-wise last attention layers in multi-token training
|
938 |
"""
|
939 |
last_hidden_state: torch.FloatTensor = None
|
940 |
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
|
@@ -949,39 +1477,51 @@ MOTIF_INPUTS_DOCSTRING = r"""
|
|
949 |
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
950 |
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
951 |
it.
|
|
|
952 |
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
953 |
[`PreTrainedTokenizer.__call__`] for details.
|
|
|
954 |
[What are input IDs?](../glossary#input-ids)
|
955 |
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
956 |
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
|
|
957 |
- 1 for tokens that are **not masked**,
|
958 |
- 0 for tokens that are **masked**.
|
|
|
959 |
[What are attention masks?](../glossary#attention-mask)
|
|
|
960 |
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
961 |
[`PreTrainedTokenizer.__call__`] for details.
|
|
|
962 |
If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
|
963 |
`past_key_values`).
|
|
|
964 |
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
965 |
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
966 |
information on the default strategy.
|
|
|
967 |
- 1 indicates the head is **not masked**,
|
968 |
- 0 indicates the head is **masked**.
|
969 |
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
970 |
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
971 |
config.n_positions - 1]`.
|
|
|
972 |
[What are position IDs?](../glossary#position-ids)
|
973 |
past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
|
974 |
Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
975 |
blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
|
976 |
returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
|
|
|
977 |
Two formats are allowed:
|
978 |
- a [`~cache_utils.Cache`] instance, see our
|
979 |
[kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
|
980 |
- Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
|
981 |
shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
|
982 |
cache format.
|
|
|
983 |
The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
|
984 |
legacy cache format will be returned.
|
|
|
985 |
If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
|
986 |
have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
|
987 |
of shape `(batch_size, sequence_length)`.
|
@@ -1014,6 +1554,7 @@ MOTIF_INPUTS_DOCSTRING = r"""
|
|
1014 |
class MotifModel(MotifPreTrainedModel):
|
1015 |
"""
|
1016 |
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MotifDecoderLayer`]
|
|
|
1017 |
Args:
|
1018 |
config: MotifConfig
|
1019 |
"""
|
@@ -1025,14 +1566,19 @@ class MotifModel(MotifPreTrainedModel):
|
|
1025 |
self.multi_token_heads = config.multi_token_heads
|
1026 |
|
1027 |
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
|
|
|
|
1028 |
|
1029 |
num_hidden_layers = config.num_hidden_layers if self.multi_token_heads is None else config.num_hidden_layers - 1
|
1030 |
-
|
1031 |
-
|
1032 |
-
|
1033 |
-
|
|
|
|
|
|
|
1034 |
self._attn_implementation = config._attn_implementation
|
1035 |
-
RMSNorm = MotifRMSNorm
|
1036 |
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
1037 |
self.hidden_size = config.hidden_size
|
1038 |
self.num_heads = config.num_attention_heads
|
@@ -1046,6 +1592,34 @@ class MotifModel(MotifPreTrainedModel):
|
|
1046 |
self.gradient_checkpointing = False
|
1047 |
self.post_init()
|
1048 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1049 |
def get_input_embeddings(self):
|
1050 |
return self.embed_tokens
|
1051 |
|
@@ -1084,6 +1658,7 @@ class MotifModel(MotifPreTrainedModel):
|
|
1084 |
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...")
|
1085 |
use_cache = False
|
1086 |
|
|
|
1087 |
return_legacy_cache = False
|
1088 |
if use_cache and not isinstance(past_key_values, Cache):
|
1089 |
return_legacy_cache = True
|
@@ -1097,17 +1672,17 @@ class MotifModel(MotifPreTrainedModel):
|
|
1097 |
"(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)")
|
1098 |
|
1099 |
if inputs_embeds is None:
|
1100 |
-
inputs_embeds = self.embed_tokens(input_ids)
|
1101 |
|
1102 |
if cache_position is None:
|
1103 |
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
1104 |
cache_position = torch.arange(past_seen_tokens,
|
1105 |
past_seen_tokens + inputs_embeds.shape[1],
|
1106 |
device=inputs_embeds.device)
|
1107 |
-
position_ids = None
|
1108 |
if position_ids is None:
|
1109 |
position_ids = cache_position.unsqueeze(0)
|
1110 |
-
|
1111 |
causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_key_values,
|
1112 |
output_attentions)
|
1113 |
|
@@ -1150,6 +1725,10 @@ class MotifModel(MotifPreTrainedModel):
|
|
1150 |
)
|
1151 |
|
1152 |
hidden_states = layer_outputs[0]
|
|
|
|
|
|
|
|
|
1153 |
|
1154 |
if use_cache:
|
1155 |
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
|
@@ -1157,8 +1736,9 @@ class MotifModel(MotifPreTrainedModel):
|
|
1157 |
if output_attentions:
|
1158 |
all_self_attns += (layer_outputs[1], )
|
1159 |
|
1160 |
-
|
1161 |
-
|
|
|
1162 |
# add hidden states from the last decoder layer
|
1163 |
if output_hidden_states:
|
1164 |
all_hidden_states += (hidden_states, )
|
@@ -1190,6 +1770,8 @@ class MotifModel(MotifPreTrainedModel):
|
|
1190 |
output_attentions: bool,
|
1191 |
):
|
1192 |
if self.config._attn_implementation == "flash_attention_2":
|
|
|
|
|
1193 |
if attention_mask is not None and 0.0 in attention_mask:
|
1194 |
return attention_mask
|
1195 |
return None
|
@@ -1261,6 +1843,7 @@ class MotifModel(MotifPreTrainedModel):
|
|
1261 |
"""
|
1262 |
Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
|
1263 |
`(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
|
|
|
1264 |
Args:
|
1265 |
attention_mask (`torch.Tensor`):
|
1266 |
A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
|
@@ -1318,14 +1901,33 @@ class MotifForCausalLM(MotifPreTrainedModel, GenerationMixin):
|
|
1318 |
self.vocab_size = config.vocab_size
|
1319 |
self.multi_token_heads = config.multi_token_heads
|
1320 |
|
1321 |
-
self.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1322 |
|
1323 |
# Initialize weights and apply final processing
|
1324 |
self.post_init()
|
1325 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1326 |
if getattr(config, "tie_word_embeddings", True):
|
1327 |
logger.info('tie embeddings')
|
1328 |
self.tie_weights()
|
|
|
|
|
|
|
1329 |
|
1330 |
def get_input_embeddings(self):
|
1331 |
return self.model.embed_tokens
|
@@ -1345,7 +1947,101 @@ class MotifForCausalLM(MotifPreTrainedModel, GenerationMixin):
|
|
1345 |
def get_decoder(self):
|
1346 |
return self.model
|
1347 |
|
1348 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1349 |
@add_start_docstrings_to_model_forward(MOTIF_INPUTS_DOCSTRING)
|
1350 |
@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
1351 |
def forward(
|
@@ -1370,18 +2066,25 @@ class MotifForCausalLM(MotifPreTrainedModel, GenerationMixin):
|
|
1370 |
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
1371 |
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
1372 |
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
|
|
1373 |
num_logits_to_keep (`int`, *optional*):
|
1374 |
Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
|
1375 |
`input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
|
1376 |
token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
|
|
|
1377 |
Returns:
|
|
|
1378 |
Example:
|
|
|
1379 |
```python
|
1380 |
>>> from transformers import AutoTokenizer, MotifForCausalLM
|
1381 |
-
|
1382 |
-
>>>
|
|
|
|
|
1383 |
>>> prompt = "Hey, are you conscious? Can you talk to me?"
|
1384 |
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
|
|
1385 |
>>> # Generate
|
1386 |
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
1387 |
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
@@ -1394,6 +2097,8 @@ class MotifForCausalLM(MotifPreTrainedModel, GenerationMixin):
|
|
1394 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
1395 |
|
1396 |
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
|
|
|
|
1397 |
outputs: MotifModelOutputWithPast = self.model(
|
1398 |
input_ids=input_ids,
|
1399 |
attention_mask=attention_mask,
|
@@ -1405,16 +2110,31 @@ class MotifForCausalLM(MotifPreTrainedModel, GenerationMixin):
|
|
1405 |
output_hidden_states=output_hidden_states,
|
1406 |
return_dict=return_dict,
|
1407 |
cache_position=cache_position,
|
|
|
|
|
1408 |
)
|
1409 |
|
1410 |
hidden_states = outputs[0]
|
1411 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1412 |
# Only compute necessary logits, and do not upcast them to float if we are not computing the loss
|
|
|
1413 |
logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
|
1414 |
logits = logits.float()
|
1415 |
|
1416 |
loss = None
|
1417 |
if labels is not None:
|
|
|
1418 |
# Shift so that tokens < n predict n
|
1419 |
shift_logits = logits[..., :-1, :].contiguous()
|
1420 |
shift_labels = labels[..., 1:].contiguous()
|
@@ -1436,4 +2156,4 @@ class MotifForCausalLM(MotifPreTrainedModel, GenerationMixin):
|
|
1436 |
past_key_values=outputs.past_key_values,
|
1437 |
hidden_states=outputs.hidden_states,
|
1438 |
attentions=outputs.attentions,
|
1439 |
-
)
|
|
|
1 |
import math
|
2 |
+
from typing import List, Optional, Tuple, Union
|
3 |
|
4 |
import torch
|
5 |
import torch.utils.checkpoint
|
|
|
28 |
from dataclasses import dataclass
|
29 |
|
30 |
import torch.nn.functional as F
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
|
32 |
from transformers.activations import ACT2CLS as _ACT2CLS
|
33 |
from transformers.activations import ClassInstantier
|
|
|
|
|
34 |
class PolyNorm(torch.nn.Module):
|
35 |
+
"""
|
36 |
A trainable activation function introduced in https://arxiv.org/html/2411.03884v1.
|
37 |
The code is copied from https://github.com/BryceZhuo/PolyCom?tab=readme-ov-file/README.md,
|
38 |
+
with the change `* torch.rsqrt` => `/ torch.sqrt` for potential MAF incompatibility.
|
39 |
"""
|
40 |
|
41 |
def __init__(self, eps=1e-6):
|
|
|
51 |
return self.weight[0] * self._norm(x ** 3) + self.weight[1] * self._norm(
|
52 |
x ** 2) + self.weight[2] * self._norm(x) + self.bias
|
53 |
|
54 |
+
class PolyNorm_Test(torch.nn.Module):
|
55 |
+
"""
|
56 |
+
A trainable activation function introduced in https://arxiv.org/html/2411.03884v1.
|
57 |
+
The code is copied from https://github.com/BryceZhuo/PolyCom?tab=readme-ov-file/README.md,
|
58 |
+
with the change `* torch.rsqrt` => `/ torch.sqrt` for potential MAF incompatibility.
|
59 |
+
"""
|
60 |
+
|
61 |
+
def __init__(self, eps=1e-6):
|
62 |
+
super(PolyNorm_Test, self).__init__()
|
63 |
+
self.weight = torch.nn.Parameter(torch.ones(3) / 3)
|
64 |
+
self.bias = torch.nn.Parameter(torch.zeros(1))
|
65 |
+
self.eps = eps
|
66 |
+
|
67 |
+
def forward(self, x):
|
68 |
+
|
69 |
+
#return torch.nn.SiLU(x)
|
70 |
+
return moreh_ops.poly_norm(x, self.weight, self.bias)
|
71 |
+
|
72 |
+
|
73 |
+
CUSTOM_ACT2CLS = {"poly_norm": PolyNorm, "poly_norm_test": PolyNorm_Test}
|
74 |
|
|
|
75 |
ACT2CLS = {**_ACT2CLS, **CUSTOM_ACT2CLS}
|
76 |
ACT2FN = ClassInstantier(ACT2CLS)
|
77 |
|
78 |
+
logger = logging.get_logger(__name__)
|
79 |
+
|
80 |
+
if is_flash_attn_2_available():
|
81 |
+
from transformers.modeling_flash_attention_utils import _flash_attention_forward
|
82 |
+
|
83 |
+
try:
|
84 |
+
moreh_ops = torch.ops.moreh
|
85 |
+
MorehRMSNorm = moreh_ops.T5LayerNorm
|
86 |
+
ScaledDotProductAttention = moreh_ops.scaled_dot_product_attention
|
87 |
+
MorehFlashAttention = moreh_ops.flash_attention
|
88 |
+
logger.warning_once("Using moreh ops")
|
89 |
+
except AttributeError:
|
90 |
+
MorehRMSNorm = None
|
91 |
+
ScaledDotProductAttention = None
|
92 |
+
MorehFlashAttention = None
|
93 |
+
logger.warning_once("Failed to import moreh ops")
|
94 |
+
|
95 |
+
#_CHECKPOINT_FOR_DOC = "moreh/Motif-102B"
|
96 |
+
_CONFIG_FOR_DOC = "MotifConfig"
|
97 |
+
|
98 |
+
#from .moreh_moe import MorehMoeMLP, MorehMoeFusedMLP
|
99 |
+
|
100 |
|
101 |
class MotifRMSNorm(nn.Module):
|
102 |
|
|
|
110 |
|
111 |
def forward(self, hidden_states):
|
112 |
input_dtype = hidden_states.dtype
|
113 |
+
hidden_states = hidden_states.to(torch.float32)
|
114 |
+
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
115 |
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
116 |
return self.weight * hidden_states.to(input_dtype)
|
117 |
|
|
|
119 |
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
120 |
|
121 |
|
122 |
+
ALL_LAYERNORM_LAYERS.append(MotifRMSNorm if MorehRMSNorm is None else MorehRMSNorm)
|
123 |
|
124 |
|
125 |
class MotifRotaryEmbeddingWithCache(nn.Module):
|
126 |
"""
|
127 |
Rotary positional embedding module with caching for efficiency.
|
128 |
+
|
129 |
Args:
|
130 |
dim (int): Dimensionality of the embedding.
|
131 |
max_position_embeddings (int): Maximum sequence length for caching. Default is 2048.
|
132 |
base (int): Base for computing inverse frequency. Default is 10000.
|
133 |
device (torch.device, optional): Device for tensor storage.
|
134 |
+
|
135 |
Methods:
|
136 |
forward(x, seq_len=None):
|
137 |
Computes cosine and sine embeddings for input sequence length.
|
138 |
Automatically updates cache if `seq_len` exceeds cached length.
|
139 |
+
|
140 |
Attributes:
|
141 |
inv_freq (torch.Tensor): Inverse frequency tensor for position encoding.
|
142 |
cos_cached (torch.Tensor): Cached cosine embeddings.
|
|
|
172 |
self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
|
173 |
|
174 |
return (
|
175 |
+
self.cos_cached[ :seq_len].to(dtype=x.dtype),
|
176 |
+
self.sin_cached[ :seq_len].to(dtype=x.dtype),
|
177 |
)
|
178 |
|
179 |
|
|
|
190 |
config: Optional[MotifConfig] = None,
|
191 |
):
|
192 |
super().__init__()
|
193 |
+
# TODO (joao): remove the `if` below, only used for BC
|
194 |
self.rope_kwargs = {}
|
195 |
if config is None:
|
196 |
logger.warning_once(
|
|
|
235 |
device,
|
236 |
seq_len=seq_len,
|
237 |
**self.rope_kwargs)
|
238 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False) # TODO joao: may break with compilation
|
239 |
self.max_seq_len_cached = seq_len
|
240 |
|
241 |
if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len: # reset
|
|
|
269 |
def rotate_half(x):
|
270 |
"""
|
271 |
Rotates half of the dimensions of the input tensor using torch.roll and in-place negation.
|
272 |
+
|
273 |
Args:
|
274 |
x (torch.Tensor): The input tensor.
|
275 |
+
|
276 |
Returns:
|
277 |
torch.Tensor: A tensor where the latter half of the dimensions are negated
|
278 |
and moved before the first half.
|
|
|
284 |
return rotated_tensor
|
285 |
|
286 |
|
287 |
+
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1, fused_rope=False):
|
288 |
"""
|
289 |
Applies rotary position embeddings to the input tensors.
|
290 |
+
|
291 |
Args:
|
292 |
q (torch.Tensor): Query tensor of shape (B, NH, S, D_KV).
|
293 |
k (torch.Tensor): Key tensor of shape (B, NH, S, D_KV).
|
294 |
cos (torch.Tensor): Cosine values for rotary embedding.
|
295 |
sin (torch.Tensor): Sine values for rotary embedding.
|
296 |
+
unsqueeze_dim (int, optional): Dimension along which `cos` and `sin` are unsqueezed.
|
297 |
Defaults to 1.
|
298 |
+
fused_rope (bool, optional): If True, applies fused rotary embeddings using
|
299 |
+
`moreh_ops.apply_rotary_emb`. If False, computes rotary embeddings manually.
|
300 |
+
Defaults to False.
|
301 |
+
|
302 |
Returns:
|
303 |
Tuple[torch.Tensor, torch.Tensor]: Returns transformed query and key tensors after applying rotary embeddings.
|
304 |
"""
|
305 |
+
'''
|
306 |
+
# (B, NH, S, D_KV) -> (B, S, NH, D_KV)
|
307 |
+
cos = cos.unsqueeze(unsqueeze_dim)
|
308 |
+
sin = sin.unsqueeze(unsqueeze_dim)
|
309 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
310 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
311 |
+
'''
|
312 |
+
if not fused_rope:
|
313 |
+
device = q.device
|
314 |
+
return map(
|
315 |
+
lambda x: (x * cos[position_ids].unsqueeze(unsqueeze_dim).to(device)) +
|
316 |
+
(rotate_half(x) * sin[position_ids].unsqueeze(unsqueeze_dim).to(device)), (q, k))
|
317 |
+
else:
|
318 |
+
# (B, NH, S, D_KV) -> (B, S, NH, D_KV)
|
319 |
+
cos = cos[position_ids]
|
320 |
+
sin = sin[position_ids]
|
321 |
+
|
322 |
+
q = q.transpose(1, 2)
|
323 |
+
k = k.transpose(1, 2)
|
324 |
+
|
325 |
+
# Expand 'batch' dim
|
326 |
+
cos = cos.expand(q.shape[0], *cos.shape[1:])
|
327 |
+
sin = sin.expand(q.shape[0], *sin.shape[1:])
|
328 |
+
|
329 |
+
q_embed = moreh_ops.apply_rotary_emb(q, cos, sin, opcode=1)
|
330 |
+
k_embed = moreh_ops.apply_rotary_emb(k, cos, sin, opcode=1)
|
331 |
+
|
332 |
+
# (B, S, NH, D_KV) -> (B, NH, S, D_KV)
|
333 |
+
q_embed = q_embed.transpose(1, 2)
|
334 |
+
k_embed = k_embed.transpose(1, 2)
|
335 |
+
|
336 |
+
return q_embed, k_embed
|
337 |
|
338 |
|
339 |
class MotifMLP(nn.Module):
|
340 |
+
|
341 |
def __init__(self, config):
|
342 |
super().__init__()
|
343 |
self.hidden_size = config.hidden_size
|
|
|
347 |
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
|
348 |
self.act_fn = ACT2FN[config.hidden_act]
|
349 |
|
350 |
+
if config.wesar_weights:
|
351 |
+
self.gate_up_proj_alpha = nn.Parameter(torch.tensor(1) *config.gate_up_proj_alpha)
|
352 |
+
self.down_proj_alpha = nn.Parameter(torch.tensor(1) * config.down_proj_alpha)
|
353 |
+
else:
|
354 |
+
self.gate_up_proj_alpha=1
|
355 |
+
self.down_proj_alpha=1
|
356 |
+
if config.muP:
|
357 |
+
self.down_proj.__do_scale_tager__ = True
|
358 |
+
self.gate_proj.__do_scale_tager_mu_dim_model__ = True
|
359 |
+
self.up_proj.__do_scale_tager_mu_dim_model__ = True
|
360 |
+
self.down_proj.__do_scale_tager_mu_ffn__ = True
|
361 |
+
|
362 |
+
|
363 |
def forward(self, hidden_state):
|
364 |
+
hidden_state = hidden_state*self.gate_up_proj_alpha
|
365 |
+
#hidden_state = self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))*
|
366 |
+
return self.down_proj_alpha*self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
|
367 |
+
|
368 |
+
|
369 |
+
class MorehMoeFusedMLP(nn.Module):
|
370 |
+
def __init__(self,
|
371 |
+
ffn_dim,
|
372 |
+
hidden_dim,
|
373 |
+
hidden_act_moe,
|
374 |
+
num_experts,
|
375 |
+
num_groups=1,
|
376 |
+
device=None,
|
377 |
+
continual_training=False):
|
378 |
+
super().__init__()
|
379 |
+
self.ffn_dim = ffn_dim
|
380 |
+
self.hidden_dim = hidden_dim
|
381 |
+
self.hidden_act_moe = hidden_act_moe
|
382 |
+
|
383 |
+
self.num_experts = num_experts
|
384 |
+
self.num_groups = num_groups
|
385 |
+
|
386 |
+
assert self.num_experts % self.num_groups == 0
|
387 |
+
self.num_experts_per_group = self.num_experts // self.num_groups
|
388 |
+
|
389 |
+
## bsz, seq, group size, 2*ffn_size
|
390 |
+
|
391 |
+
moreh_ops = torch.ops.moreh
|
392 |
+
self.w13 = nn.ModuleList([
|
393 |
+
moreh_ops.MoeFanInLinear(self.hidden_dim,
|
394 |
+
self.ffn_dim * 2,
|
395 |
+
bias=False,
|
396 |
+
num_experts=self.num_experts_per_group,
|
397 |
+
device=device)
|
398 |
+
for _ in range(self.num_groups)
|
399 |
+
])
|
400 |
+
|
401 |
+
self.w2 = nn.ModuleList([
|
402 |
+
moreh_ops.MoeFanOutLinear(self.ffn_dim,
|
403 |
+
self.hidden_dim,
|
404 |
+
bias=False,
|
405 |
+
num_experts=self.num_experts_per_group,
|
406 |
+
device=device)
|
407 |
+
for _ in range(self.num_groups)
|
408 |
+
])
|
409 |
+
|
410 |
+
## use silu?
|
411 |
+
self.act_fn = ACT2FN[self.hidden_act_moe]
|
412 |
+
|
413 |
+
if continual_training:
|
414 |
+
logger.info('two optipons 1. zero init all weights, 2. add scaling param to moe output.')
|
415 |
+
self._zero_init()
|
416 |
+
|
417 |
+
def _zero_init(self):
|
418 |
+
for module in self.w2:
|
419 |
+
for n,param in module.named_parameters():
|
420 |
+
logger.info(f'{n} {param.shape}')
|
421 |
+
param.data.zero_()
|
422 |
+
|
423 |
+
|
424 |
+
def forward(self, hidden_states, selected_experts, routing_weights):
|
425 |
+
w13_final_output = None
|
426 |
+
for group_idx in range(self.num_groups):
|
427 |
+
w13_output_in_group = self._get_w13_output(hidden_states,
|
428 |
+
selected_experts,
|
429 |
+
group_idx)
|
430 |
+
if w13_final_output is None:
|
431 |
+
w13_final_output = w13_output_in_group
|
432 |
+
else:
|
433 |
+
w13_final_output += w13_output_in_group
|
434 |
+
|
435 |
+
current_hidden_states = self.act_fn(
|
436 |
+
w13_final_output[:, :, :, :self.ffn_dim]
|
437 |
+
) * w13_final_output[:, :, :, self.ffn_dim:]
|
438 |
+
|
439 |
+
final_hidden_states = None
|
440 |
+
for group_idx in range(self.num_groups):
|
441 |
+
w2_output_in_group = self._get_w2_output(current_hidden_states,
|
442 |
+
selected_experts,
|
443 |
+
routing_weights, group_idx)
|
444 |
+
if final_hidden_states is None:
|
445 |
+
final_hidden_states = w2_output_in_group
|
446 |
+
else:
|
447 |
+
final_hidden_states += w2_output_in_group
|
448 |
+
return final_hidden_states
|
449 |
+
|
450 |
+
def _get_w13_output(self, hidden_states, selected_experts, group_idx):
|
451 |
+
selected_experts_in_group = selected_experts - (
|
452 |
+
group_idx * self.num_experts_per_group)
|
453 |
+
|
454 |
+
w13_output = self.w13[group_idx](hidden_states,
|
455 |
+
selected_experts_in_group)
|
456 |
+
return w13_output
|
457 |
+
|
458 |
+
def _get_w2_output(self, hidden_states, selected_experts, routing_weights,
|
459 |
+
group_idx):
|
460 |
+
selected_experts_in_group = selected_experts - (
|
461 |
+
group_idx * self.num_experts_per_group)
|
462 |
+
output = self.w2[group_idx](hidden_states, selected_experts_in_group,
|
463 |
+
routing_weights)
|
464 |
+
return output
|
465 |
+
|
466 |
+
|
467 |
+
class MoEGate(nn.Module):
|
468 |
+
|
469 |
+
def __init__(self, config):
|
470 |
+
super().__init__()
|
471 |
+
self.config = config
|
472 |
+
self.top_k = config.num_experts_per_tok
|
473 |
+
self.n_routed_experts = config.n_routed_experts
|
474 |
+
self.routed_scaling_factor = config.routed_scaling_factor
|
475 |
+
self.scoring_func = config.scoring_func
|
476 |
+
self.seq_aux = config.seq_aux
|
477 |
+
self.topk_method = config.topk_method
|
478 |
+
self.n_group = config.n_group
|
479 |
+
self.topk_group = config.topk_group
|
480 |
+
|
481 |
+
# topk selection algorithm
|
482 |
+
self.norm_topk_prob = config.norm_topk_prob
|
483 |
+
self.gating_dim = config.hidden_size
|
484 |
+
self.weight = nn.Parameter(
|
485 |
+
torch.empty((self.n_routed_experts, self.gating_dim)))
|
486 |
+
if self.topk_method == "noaux_tc":
|
487 |
+
self.e_score_correction_bias = nn.Parameter(
|
488 |
+
torch.empty((self.n_routed_experts)))
|
489 |
+
self.reset_parameters()
|
490 |
+
|
491 |
+
def reset_parameters(self) -> None:
|
492 |
+
import torch.nn.init as init
|
493 |
+
|
494 |
+
init.kaiming_uniform_(self.weight, a=math.sqrt(5))
|
495 |
+
|
496 |
+
def forward(self, hidden_states):
|
497 |
+
bsz, seq_len, h = hidden_states.shape
|
498 |
+
### compute gating score
|
499 |
+
hidden_states = hidden_states.view(-1, h)
|
500 |
+
logits = F.linear(hidden_states.type(torch.float32),
|
501 |
+
self.weight.type(torch.float32), None)
|
502 |
+
if self.scoring_func == "sigmoid":
|
503 |
+
scores = logits.sigmoid()
|
504 |
+
else:
|
505 |
+
raise NotImplementedError(
|
506 |
+
f"insupportable scoring function for MoE gating: {self.scoring_func}"
|
507 |
+
)
|
508 |
+
|
509 |
+
### select top-k experts
|
510 |
+
if self.topk_method == "greedy":
|
511 |
+
topk_weight, topk_idx = torch.topk(scores,
|
512 |
+
k=self.top_k,
|
513 |
+
dim=-1,
|
514 |
+
sorted=False)
|
515 |
+
elif self.topk_method == "group_limited_greedy":
|
516 |
+
group_scores = (scores.view(bsz * seq_len, self.n_group,
|
517 |
+
-1).max(dim=-1).values) # [n, n_group]
|
518 |
+
group_idx = torch.topk(group_scores,
|
519 |
+
k=self.topk_group,
|
520 |
+
dim=-1,
|
521 |
+
sorted=False)[1] # [n, top_k_group]
|
522 |
+
group_mask = torch.zeros_like(group_scores) # [n, n_group]
|
523 |
+
group_mask.scatter_(1, group_idx, 1) # [n, n_group]
|
524 |
+
score_mask = (group_mask.unsqueeze(-1).expand(
|
525 |
+
bsz * seq_len, self.n_group,
|
526 |
+
self.n_routed_experts // self.n_group).reshape(
|
527 |
+
bsz * seq_len, -1)) # [n, e]
|
528 |
+
tmp_scores = scores.masked_fill(~score_mask.bool(), 0.0) # [n, e]
|
529 |
+
topk_weight, topk_idx = torch.topk(tmp_scores,
|
530 |
+
k=self.top_k,
|
531 |
+
dim=-1,
|
532 |
+
sorted=False)
|
533 |
+
elif self.topk_method == "noaux_tc":
|
534 |
+
### will be used. ###
|
535 |
+
scores_for_choice = scores.view(
|
536 |
+
bsz * seq_len, -1) + self.e_score_correction_bias.unsqueeze(0)
|
537 |
+
group_scores = (scores_for_choice.view(
|
538 |
+
bsz * seq_len, self.n_group,
|
539 |
+
-1).topk(2, dim=-1)[0].sum(dim=-1)) # [n, n_group]
|
540 |
+
group_idx = torch.topk(group_scores,
|
541 |
+
k=self.topk_group,
|
542 |
+
dim=-1,
|
543 |
+
sorted=False)[1] # [n, top_k_group]
|
544 |
+
group_mask = torch.zeros_like(group_scores) # [n, n_group]
|
545 |
+
group_mask.scatter_(1, group_idx, 1) # [n, n_group]
|
546 |
+
score_mask = (group_mask.unsqueeze(-1).expand(
|
547 |
+
bsz * seq_len, self.n_group,
|
548 |
+
self.n_routed_experts // self.n_group).reshape(
|
549 |
+
bsz * seq_len, -1)) # [n, e]
|
550 |
+
tmp_scores = scores_for_choice.masked_fill(~score_mask.bool(),
|
551 |
+
0.0) # [n, e]
|
552 |
+
_, topk_idx = torch.topk(tmp_scores,
|
553 |
+
k=self.top_k,
|
554 |
+
dim=-1,
|
555 |
+
sorted=False)
|
556 |
+
topk_weight = scores.gather(1, topk_idx)
|
557 |
+
else:
|
558 |
+
raise NotImplementedError(
|
559 |
+
f"insupportable TopK function for MoE gating: {self.topk_method}"
|
560 |
+
)
|
561 |
+
|
562 |
+
### norm gate to sum 1
|
563 |
+
if self.top_k > 1 and self.norm_topk_prob:
|
564 |
+
denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
|
565 |
+
topk_weight = topk_weight / denominator
|
566 |
+
topk_weight = topk_weight * self.routed_scaling_factor # must multiply the scaling factor
|
567 |
+
|
568 |
+
return topk_idx, topk_weight
|
569 |
+
|
570 |
+
|
571 |
+
class MotifMoE(nn.Module):
|
572 |
+
"""
|
573 |
+
A mixed expert module containing shared experts.
|
574 |
+
"""
|
575 |
+
def __init__(self, config):
|
576 |
+
super().__init__()
|
577 |
+
self.config = config
|
578 |
+
self.num_experts_per_tok = config.num_experts_per_tok
|
579 |
+
self.use_moreh_moe = config.use_moreh_moe
|
580 |
+
self.use_fused_mlp = config.use_fused_mlp
|
581 |
+
|
582 |
+
if hasattr(config, "ep_size") and config.ep_size > 1:
|
583 |
+
assert config.ep_size == dist.get_world_size()
|
584 |
+
assert not config.use_moreh_moe
|
585 |
+
self.ep_size = config.ep_size
|
586 |
+
self.experts_per_rank = config.n_routed_experts // config.ep_size
|
587 |
+
self.ep_rank = dist.get_rank()
|
588 |
+
self.experts = nn.ModuleList([
|
589 |
+
(DeepseekV3MLP(config,
|
590 |
+
intermediate_size=config.moe_intermediate_size)
|
591 |
+
if i >= self.ep_rank * self.experts_per_rank and i <
|
592 |
+
(self.ep_rank + 1) * self.experts_per_rank else None)
|
593 |
+
for i in range(config.n_routed_experts)
|
594 |
+
])
|
595 |
+
else:
|
596 |
+
self.ep_size = 1
|
597 |
+
self.experts_per_rank = config.n_routed_experts
|
598 |
+
self.ep_rank = 0
|
599 |
+
if self.use_moreh_moe:
|
600 |
+
if not self.use_fused_mlp:
|
601 |
+
self.experts = MorehMoeMLP(
|
602 |
+
ffn_dim=config.moe_intermediate_size,
|
603 |
+
hidden_dim=config.hidden_size,
|
604 |
+
hidden_act_moe=config.hidden_act_moe,
|
605 |
+
num_experts=config.n_routed_experts,
|
606 |
+
device=None)
|
607 |
+
else:
|
608 |
+
## group expert.
|
609 |
+
self.experts = MorehMoeFusedMLP(
|
610 |
+
ffn_dim=config.moe_intermediate_size,
|
611 |
+
hidden_dim=config.hidden_size,
|
612 |
+
hidden_act_moe=config.hidden_act_moe,
|
613 |
+
num_experts=config.n_routed_experts,
|
614 |
+
num_groups=config.n_group,
|
615 |
+
device=None,
|
616 |
+
continual_training=config.continual_training,
|
617 |
+
)
|
618 |
+
else:
|
619 |
+
self.experts = nn.ModuleList([
|
620 |
+
DeepseekV3MLP(
|
621 |
+
config, intermediate_size=config.moe_intermediate_size)
|
622 |
+
for i in range(config.n_routed_experts)
|
623 |
+
])
|
624 |
+
|
625 |
+
self.gate = MoEGate(config)
|
626 |
+
|
627 |
+
def forward(self, hidden_states):
|
628 |
+
identity = hidden_states
|
629 |
+
orig_shape = hidden_states.shape
|
630 |
+
topk_idx, topk_weight = self.gate(hidden_states)
|
631 |
+
if self.use_moreh_moe:
|
632 |
+
y = self.experts(hidden_states, topk_idx.view(*orig_shape[:-1], -1),
|
633 |
+
topk_weight.view(*orig_shape[:-1], -1))
|
634 |
+
y = y.type(hidden_states.dtype)
|
635 |
+
else:
|
636 |
+
hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
|
637 |
+
flat_topk_idx = topk_idx.view(-1)
|
638 |
+
if self.training:
|
639 |
+
hidden_states = hidden_states.repeat_interleave(
|
640 |
+
self.num_experts_per_tok, dim=0)
|
641 |
+
y = torch.empty_like(hidden_states)
|
642 |
+
for i, expert in enumerate(self.experts):
|
643 |
+
y[flat_topk_idx == i] = expert(
|
644 |
+
hidden_states[flat_topk_idx == i])
|
645 |
+
y = (y.view(*topk_weight.shape, -1) *
|
646 |
+
topk_weight.unsqueeze(-1)).sum(dim=1)
|
647 |
+
y = y.type(hidden_states.dtype)
|
648 |
+
y = y.view(*orig_shape)
|
649 |
+
# y = AddAuxiliaryLoss.apply(y, aux_loss)
|
650 |
+
else:
|
651 |
+
y = self.moe_infer(hidden_states, topk_idx,
|
652 |
+
topk_weight).view(*orig_shape)
|
653 |
+
return y, identity
|
654 |
+
|
655 |
+
@torch.no_grad()
|
656 |
+
def moe_infer(self, x, topk_ids, topk_weight):
|
657 |
+
cnts = topk_ids.new_zeros((topk_ids.shape[0], len(self.experts)))
|
658 |
+
cnts.scatter_(1, topk_ids, 1)
|
659 |
+
tokens_per_expert = cnts.sum(dim=0)
|
660 |
+
idxs = topk_ids.view(-1).argsort()
|
661 |
+
sorted_tokens = x[idxs // topk_ids.shape[1]]
|
662 |
+
sorted_tokens_shape = sorted_tokens.shape
|
663 |
+
if self.ep_size > 1:
|
664 |
+
tokens_per_ep_rank = tokens_per_expert.view(self.ep_size,
|
665 |
+
-1).sum(dim=1)
|
666 |
+
tokens_per_expert_group = tokens_per_expert.new_empty(
|
667 |
+
tokens_per_expert.shape[0])
|
668 |
+
dist.all_to_all_single(tokens_per_expert_group, tokens_per_expert)
|
669 |
+
output_splits = (tokens_per_expert_group.view(
|
670 |
+
self.ep_size, -1).sum(1).cpu().numpy().tolist())
|
671 |
+
gathered_tokens = sorted_tokens.new_empty(
|
672 |
+
tokens_per_expert_group.sum(dim=0).cpu().item(),
|
673 |
+
sorted_tokens.shape[1])
|
674 |
+
input_split_sizes = tokens_per_ep_rank.cpu().numpy().tolist()
|
675 |
+
dist.all_to_all(
|
676 |
+
list(gathered_tokens.split(output_splits)),
|
677 |
+
list(sorted_tokens.split(input_split_sizes)),
|
678 |
+
)
|
679 |
+
tokens_per_expert_post_gather = tokens_per_expert_group.view(
|
680 |
+
self.ep_size, self.experts_per_rank).sum(dim=0)
|
681 |
+
gatherd_idxs = np.zeros(shape=(gathered_tokens.shape[0],),
|
682 |
+
dtype=np.int32)
|
683 |
+
s = 0
|
684 |
+
for i, k in enumerate(tokens_per_expert_group.cpu().numpy()):
|
685 |
+
gatherd_idxs[s:s + k] = i % self.experts_per_rank
|
686 |
+
s += k
|
687 |
+
gatherd_idxs = gatherd_idxs.argsort()
|
688 |
+
sorted_tokens = gathered_tokens[gatherd_idxs]
|
689 |
+
tokens_per_expert = tokens_per_expert_post_gather
|
690 |
+
tokens_per_expert = tokens_per_expert.cpu().numpy()
|
691 |
+
|
692 |
+
outputs = []
|
693 |
+
start_idx = 0
|
694 |
+
for i, num_tokens in enumerate(tokens_per_expert):
|
695 |
+
end_idx = start_idx + num_tokens
|
696 |
+
if num_tokens == 0:
|
697 |
+
continue
|
698 |
+
expert = self.experts[i + self.ep_rank * self.experts_per_rank]
|
699 |
+
tokens_for_this_expert = sorted_tokens[start_idx:end_idx]
|
700 |
+
expert_out = expert(tokens_for_this_expert)
|
701 |
+
outputs.append(expert_out)
|
702 |
+
start_idx = end_idx
|
703 |
+
|
704 |
+
outs = torch.cat(outputs,
|
705 |
+
dim=0) if len(outputs) else sorted_tokens.new_empty(0)
|
706 |
+
if self.ep_size > 1:
|
707 |
+
new_x = torch.empty_like(outs)
|
708 |
+
new_x[gatherd_idxs] = outs
|
709 |
+
gathered_tokens = new_x.new_empty(*sorted_tokens_shape)
|
710 |
+
dist.all_to_all(
|
711 |
+
list(gathered_tokens.split(input_split_sizes)),
|
712 |
+
list(new_x.split(output_splits)),
|
713 |
+
)
|
714 |
+
outs = gathered_tokens
|
715 |
+
|
716 |
+
new_x = torch.empty_like(outs)
|
717 |
+
new_x[idxs] = outs
|
718 |
+
final_out = (new_x.view(
|
719 |
+
*topk_ids.shape, -1).type(topk_weight.dtype).mul_(
|
720 |
+
topk_weight.unsqueeze(dim=-1)).sum(dim=1).type(new_x.dtype))
|
721 |
+
return final_out
|
722 |
|
723 |
|
724 |
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
725 |
+
|
726 |
+
|
727 |
+
"""
|
728 |
+
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
729 |
+
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
730 |
+
|
731 |
+
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
732 |
+
if n_rep == 1:
|
733 |
+
return hidden_states
|
734 |
+
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
735 |
+
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
736 |
+
"""
|
737 |
|
738 |
+
return torch.repeat_interleave(hidden_states, dim=1, repeats=n_rep)
|
739 |
+
|
740 |
|
|
|
741 |
class MotifAttention(nn.Module):
|
742 |
"""
|
743 |
Differential Attention (DiffAttention) module.
|
744 |
+
|
745 |
+
Implements the Differential Attention from
|
746 |
"DIFFERENTIAL TRANSFORMER" (https://arxiv.org/pdf/2410.05258).
|
747 |
+
|
748 |
Overview
|
749 |
Standard transformers often over-allocate attention to irrelevant context.
|
750 |
+
DiffAttention addresses this by computing attention as the difference between
|
751 |
+
two separate softmax attention maps, effectively canceling noise and promoting
|
752 |
sparse, structured attention patterns.
|
753 |
+
|
754 |
Reference Implementation
|
755 |
https://github.com/microsoft/unilm/tree/master/Diff-Transformer
|
756 |
+
|
757 |
Args
|
758 |
+
The differential attention mechanism computes attention as the difference of two softmax attention scores, weighted by a learnable scalar λ.
|
759 |
λ is re-parameterized as λ = exp(λ_q1 · λ_k1) − exp(λ_q2 · λ_k2) + λ_init.
|
760 |
- lambda_q1, lambda_q2 (nn.Parameter): Learnable vectors used to compute the first and second components of λ for query transformations.
|
761 |
- lambda_k1, lambda_k2 (nn.Parameter): Learnable vectors used to compute the first and second components of λ for key transformations.
|
762 |
- lambda_init (float): A constant used for initializing λ, typically set as λ_init = 0.8 − 0.6 × exp(−0.3 × (layer_index − 1)).
|
763 |
+
|
764 |
"""
|
765 |
|
766 |
def __init__(self, config: MotifConfig, layer_idx: Optional[int] = None):
|
|
|
840 |
self.subln = MotifRMSNorm(2 * self.head_dim, eps=1e-5)
|
841 |
self.lambda_init = 0.8 - 0.6 * math.exp(-0.3 * (layer_idx - 1))
|
842 |
|
843 |
+
self.rotary_emb = MotifRotaryEmbeddingWithCache(self.head_dim,
|
844 |
max_position_embeddings=self.max_position_embeddings,
|
845 |
base=self.rope_theta)
|
846 |
|
|
|
886 |
cos, sin = (self.rotary_emb(value_states, q_len + past_key_value.get_usable_length(q_len, self.layer_idx))
|
887 |
if use_cache else position_embeddings)
|
888 |
|
889 |
+
query_states, key_states = apply_rotary_pos_emb(query_states,
|
890 |
+
key_states,
|
891 |
+
cos,
|
892 |
+
sin,
|
893 |
+
position_ids=position_ids,
|
894 |
+
fused_rope=self.config.fused_rope)
|
895 |
|
896 |
if past_key_value is not None:
|
897 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
|
|
|
960 |
return attn_output, attn_weights, past_key_value
|
961 |
|
962 |
|
|
|
963 |
class MotifFlashAttention2(MotifAttention):
|
964 |
"""
|
965 |
Motif flash attention module, following Motif attention module. This module inherits from `MotifAttention`
|
|
|
973 |
def __init__(self, *args, **kwargs):
|
974 |
super().__init__(*args, **kwargs)
|
975 |
|
976 |
+
|
977 |
|
978 |
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
|
979 |
# 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.
|
|
|
981 |
|
982 |
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
|
983 |
|
984 |
+
logger.info(f'flash attention is used {not self._flash_attn_uses_top_left_mask}')
|
985 |
+
|
986 |
def _reshape_heads(self, tensor, batch_size, seq_len):
|
987 |
"""2-way head split tensor reshape"""
|
988 |
return tensor.reshape(batch_size, seq_len, self.num_heads, 2, self.head_dim)
|
|
|
992 |
return tensor.reshape(batch_size, seq_len, self.num_heads, self.head_dim)
|
993 |
|
994 |
def _compute_attention(self, query_states, key_states, value_states, attention_mask, q_len, position_ids,
|
995 |
+
dropout_rate, sliding_window, is_moreh_attention, batch_num):
|
996 |
"""Flash Attention 2 implements"""
|
997 |
+
|
998 |
scale_factor = 1.0 / math.sqrt(self.head_dim)
|
999 |
+
# Copied from _flash_attention_forward
|
1000 |
if not self._flash_attn_uses_top_left_mask:
|
1001 |
causal = self.is_causal
|
1002 |
else:
|
1003 |
causal = self.is_causal and q_len != 1
|
1004 |
+
|
1005 |
+
if is_moreh_attention:
|
1006 |
+
bsz = query_states.shape[0]
|
1007 |
|
1008 |
+
if batch_num:
|
1009 |
+
query_states = query_states.reshape(bsz*q_len,self.num_heads,self.head_dim)
|
1010 |
+
key_states = key_states.reshape(bsz*q_len,self.num_heads,self.head_dim)
|
1011 |
+
value_states = value_states.reshape(bsz*q_len,self.num_heads,self.head_dim)
|
1012 |
|
1013 |
+
attn_out = moreh_ops.flash_attention_varlen_dp(query_states,
|
1014 |
+
key_states,
|
1015 |
+
value_states,
|
1016 |
+
attention_mask,
|
1017 |
+
attention_mask,
|
1018 |
+
max_seqlen_q=q_len,
|
1019 |
+
max_seqlen_kv=q_len,
|
1020 |
+
dropout_p=dropout_rate,
|
1021 |
+
softmax_scale=scale_factor,
|
1022 |
+
is_causal=causal,
|
1023 |
+
batch_num=batch_num)
|
1024 |
+
attn_out = attn_out.reshape(bsz, q_len, self.num_heads, -1)
|
1025 |
+
else:
|
1026 |
+
return MorehFlashAttention(query_states,
|
1027 |
+
key_states,
|
1028 |
+
value_states,
|
1029 |
+
padding_mask=attention_mask,
|
1030 |
+
dropout_p=dropout_rate,
|
1031 |
+
softmax_scale=scale_factor,
|
1032 |
+
causal=causal)
|
1033 |
+
return attn_out
|
1034 |
+
else:
|
1035 |
+
attn_out = _flash_attention_forward(query_states,
|
1036 |
+
key_states,
|
1037 |
+
value_states,
|
1038 |
+
attention_mask,
|
1039 |
+
q_len,
|
1040 |
+
position_ids=position_ids,
|
1041 |
+
dropout=dropout_rate,
|
1042 |
+
sliding_window=sliding_window,
|
1043 |
+
is_causal=True,
|
1044 |
+
softmax_scale=scale_factor,
|
1045 |
+
use_top_left_mask=self._flash_attn_uses_top_left_mask)
|
1046 |
+
#logger.info(attn_out)
|
1047 |
+
return attn_out
|
1048 |
|
1049 |
def forward(
|
1050 |
self,
|
|
|
1078 |
cos, sin = (self.rotary_emb(value_states, q_len + past_key_value.get_usable_length(q_len, self.layer_idx))
|
1079 |
if use_cache else position_embeddings)
|
1080 |
|
1081 |
+
query_states, key_states = apply_rotary_pos_emb(query_states,
|
1082 |
+
key_states,
|
1083 |
+
cos,
|
1084 |
+
sin,
|
1085 |
+
position_ids=position_ids,
|
1086 |
+
fused_rope=False)
|
1087 |
|
1088 |
if past_key_value is not None:
|
1089 |
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
|
|
|
1094 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
1095 |
dropout_rate = 0.0 if not self.training else self.attention_dropout
|
1096 |
|
1097 |
+
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
|
1098 |
+
# therefore the input hidden states gets silently casted in float32. Hence, we need
|
1099 |
+
# cast them back in float16 just to be sure everything works as expected.
|
1100 |
+
input_dtype = query_states.dtype
|
1101 |
+
if input_dtype == torch.float32 and MorehFlashAttention is None:
|
1102 |
+
if torch.is_autocast_enabled():
|
1103 |
+
target_dtype = torch.get_autocast_gpu_dtype()
|
1104 |
+
# Handle the case where the model is quantized
|
1105 |
+
elif hasattr(self.config, "_pre_quantization_dtype"):
|
1106 |
+
target_dtype = self.config._pre_quantization_dtype
|
1107 |
+
else:
|
1108 |
+
target_dtype = self.q_proj.weight.dtype
|
1109 |
+
|
1110 |
+
logger.warning_once(
|
1111 |
+
f"The input hidden states seems to be silently casted in float32, this might be related to"
|
1112 |
+
f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
|
1113 |
+
f" {target_dtype}.")
|
1114 |
+
|
1115 |
+
query_states = query_states.to(target_dtype)
|
1116 |
+
key_states = key_states.to(target_dtype)
|
1117 |
+
value_states = value_states.to(target_dtype)
|
1118 |
+
|
1119 |
q_len = query_states.shape[-2]
|
1120 |
kv_seq_len = key_states.shape[-2]
|
1121 |
|
|
|
1125 |
value_states = value_states.transpose(1, 2)
|
1126 |
|
1127 |
if (self.config.use_sliding_window and getattr(self.config, "sliding_window", None) is not None
|
1128 |
+
and self.layer_idx >= self.config.max_window_layers and MorehFlashAttention is None):
|
1129 |
sliding_window = self.config.sliding_window
|
1130 |
else:
|
1131 |
sliding_window = None
|
|
|
1145 |
k1, k2 = k1.contiguous(), k2.contiguous()
|
1146 |
v1, v2 = v1.contiguous(), v2.contiguous()
|
1147 |
|
1148 |
+
is_moreh_attention = MorehFlashAttention is not None
|
|
|
|
|
|
|
1149 |
|
1150 |
+
attn11, attn12 = self._compute_attention(q1, k1, v1, attention_mask, q_len, position_ids, dropout_rate, sliding_window, is_moreh_attention, self.batch_num), \
|
1151 |
+
self._compute_attention(q1, k1, v2, attention_mask, q_len, position_ids, dropout_rate, sliding_window, is_moreh_attention, self.batch_num)
|
1152 |
+
attn21, attn22 = self._compute_attention(q2, k2, v1, attention_mask, q_len, position_ids, dropout_rate, sliding_window, is_moreh_attention, self.batch_num), \
|
1153 |
+
self._compute_attention(q2, k2, v2, attention_mask, q_len, position_ids, dropout_rate, sliding_window, is_moreh_attention, self.batch_num)
|
1154 |
+
|
1155 |
+
attn1, attn2 = torch.cat([attn11, attn12], dim=-1), torch.cat([attn21, attn22], dim=-1)
|
1156 |
|
1157 |
lambda_q1 = self.lambda_q1.unsqueeze(0).expand([bsz, self.lambda_q1.shape[0]]) # bsz, num_head
|
1158 |
lambda_q2 = self.lambda_q2.unsqueeze(0).expand([bsz, self.lambda_q2.shape[0]]) # bsz, num_head
|
|
|
1168 |
attn_output = attn_output * (1 - self.lambda_init)
|
1169 |
|
1170 |
if attn_output.size() != (bsz, q_len, self.num_heads, self.head_dim * 2):
|
1171 |
+
raise ValueError(f"`attn_output` should be of size {(bsz, q_len, self.num_heads, 2*self.head_dim)}, but is"
|
1172 |
f" {attn_output.size()}")
|
1173 |
|
1174 |
+
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
|
1175 |
attn_output = self.o_proj(attn_output) * self.o_proj_alpha
|
1176 |
|
1177 |
+
return attn_output, None, past_key_value
|
1178 |
|
1179 |
|
|
|
1180 |
class MotifSdpaAttention(MotifAttention):
|
1181 |
"""
|
1182 |
Motif attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
|
|
|
1270 |
MOTIF_ATTENTION_CLASSES = {
|
1271 |
"eager": MotifAttention,
|
1272 |
"flash_attention_2": MotifFlashAttention2,
|
1273 |
+
"sdpa": MotifAttention,
|
1274 |
}
|
1275 |
|
1276 |
|
1277 |
class MotifDecoderLayer(nn.Module):
|
1278 |
|
1279 |
+
def __init__(self, config: MotifConfig, moe_layer: bool, layer_idx: int):
|
1280 |
super().__init__()
|
1281 |
self.hidden_size = config.hidden_size
|
1282 |
+
if config.use_moreh_attention:
|
1283 |
+
config._attn_implementation = "flash_attention_2"
|
1284 |
if config.sliding_window and config._attn_implementation != "flash_attention_2":
|
1285 |
logger.warning_once(
|
1286 |
f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
|
|
|
1290 |
else:
|
1291 |
self.self_attn = MOTIF_ATTENTION_CLASSES["eager"](config, layer_idx)
|
1292 |
self.mlp = MotifMLP(config)
|
1293 |
+
### moe
|
1294 |
+
self.moe = None
|
1295 |
+
if moe_layer:
|
1296 |
+
self.moe = MotifMoE(config)
|
1297 |
+
|
1298 |
+
RMSNorm = MorehRMSNorm if MorehRMSNorm is not None else MotifRMSNorm
|
1299 |
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
1300 |
self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
1301 |
|
|
|
1364 |
residual = hidden_states
|
1365 |
hidden_states = self.post_attention_layernorm(hidden_states) * self.post_attention_layernorm_alpha
|
1366 |
|
1367 |
+
if self.moe is not None:
|
1368 |
+
hidden_states, identity = self.moe(hidden_states)
|
1369 |
+
## add output of shared expert and output of small moe experts.
|
1370 |
+
## hidden state must be zero tensor (for first forward)
|
1371 |
+
hidden_states += self.mlp(identity)
|
1372 |
+
else:
|
1373 |
+
hidden_states = self.mlp(hidden_states)
|
1374 |
|
1375 |
hidden_states = residual + hidden_states
|
1376 |
|
|
|
1389 |
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
1390 |
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
1391 |
etc.)
|
1392 |
+
|
1393 |
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
1394 |
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
1395 |
and behavior.
|
1396 |
+
|
1397 |
Parameters:
|
1398 |
config ([`MotifConfig`]):
|
1399 |
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
|
|
1443 |
module_std = module_std / math.sqrt(self.config.dim_model_base_lmh) ### lmhead.. 1
|
1444 |
else:
|
1445 |
module_std = module_std
|
1446 |
+
module.weight.data.normal_(mean=0.0, std=module_std)
|
1447 |
+
module.weight.data = torch.where(abs(module.weight.data) > module_std*3, 0, module.weight.data)
|
1448 |
+
#torch.nn.init.trunc_normal_(module.weight.data, mean=0.0, std=module_std, a=-3*module_std, b=3*module_std)
|
1449 |
if module.bias is not None:
|
1450 |
module.bias.data.zero_()
|
1451 |
|
1452 |
elif isinstance(module, nn.Embedding):
|
1453 |
+
module.weight.data.normal_(mean=0.0, std=module_std)
|
1454 |
+
module.weight.data = torch.where(abs(module.weight.data) > module_std*3, 0, module.weight.data)
|
1455 |
+
#torch.nn.init.trunc_normal_(module.weight.data, mean=0.0, std=module_std, a=-3*module_std, b=3*module_std)
|
1456 |
if module.padding_idx is not None:
|
1457 |
module.weight.data[module.padding_idx].zero_()
|
1458 |
|
1459 |
|
1460 |
@dataclass
|
1461 |
class MotifModelOutputWithPast(ModelOutput):
|
1462 |
+
"""
|
1463 |
+
This augments `BaseModelOutputWithPast` in `transformers.modeling_outputs` with new optional keys: `causal_mask`, `position_embeddings`.
|
1464 |
The optional keys are currently used in the following ways:
|
1465 |
+
- pass information to the token-wise last attention layers in multi-token training
|
1466 |
"""
|
1467 |
last_hidden_state: torch.FloatTensor = None
|
1468 |
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
|
|
|
1477 |
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
1478 |
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
1479 |
it.
|
1480 |
+
|
1481 |
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
1482 |
[`PreTrainedTokenizer.__call__`] for details.
|
1483 |
+
|
1484 |
[What are input IDs?](../glossary#input-ids)
|
1485 |
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
1486 |
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
1487 |
+
|
1488 |
- 1 for tokens that are **not masked**,
|
1489 |
- 0 for tokens that are **masked**.
|
1490 |
+
|
1491 |
[What are attention masks?](../glossary#attention-mask)
|
1492 |
+
|
1493 |
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
1494 |
[`PreTrainedTokenizer.__call__`] for details.
|
1495 |
+
|
1496 |
If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
|
1497 |
`past_key_values`).
|
1498 |
+
|
1499 |
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
1500 |
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
1501 |
information on the default strategy.
|
1502 |
+
|
1503 |
- 1 indicates the head is **not masked**,
|
1504 |
- 0 indicates the head is **masked**.
|
1505 |
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
1506 |
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
1507 |
config.n_positions - 1]`.
|
1508 |
+
|
1509 |
[What are position IDs?](../glossary#position-ids)
|
1510 |
past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
|
1511 |
Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
1512 |
blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
|
1513 |
returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
|
1514 |
+
|
1515 |
Two formats are allowed:
|
1516 |
- a [`~cache_utils.Cache`] instance, see our
|
1517 |
[kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
|
1518 |
- Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
|
1519 |
shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
|
1520 |
cache format.
|
1521 |
+
|
1522 |
The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
|
1523 |
legacy cache format will be returned.
|
1524 |
+
|
1525 |
If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
|
1526 |
have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
|
1527 |
of shape `(batch_size, sequence_length)`.
|
|
|
1554 |
class MotifModel(MotifPreTrainedModel):
|
1555 |
"""
|
1556 |
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`MotifDecoderLayer`]
|
1557 |
+
|
1558 |
Args:
|
1559 |
config: MotifConfig
|
1560 |
"""
|
|
|
1566 |
self.multi_token_heads = config.multi_token_heads
|
1567 |
|
1568 |
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
1569 |
+
# NOTE: For multi-token models, the last decoder layers (one for each token index)
|
1570 |
+
# are implemented as a part of `MotifModelForCausalLM` to enable a custom forward-backward procedure.
|
1571 |
|
1572 |
num_hidden_layers = config.num_hidden_layers if self.multi_token_heads is None else config.num_hidden_layers - 1
|
1573 |
+
if config.moe:
|
1574 |
+
moe_layer = [True for i in range(num_hidden_layers)]
|
1575 |
+
else:
|
1576 |
+
moe_layer = [False for i in range(num_hidden_layers)]
|
1577 |
+
logger.info(f'current_moe layer { moe_layer }')
|
1578 |
+
self.layers = nn.ModuleList([MotifDecoderLayer(config = config, moe_layer= moe_layer[layer_idx],
|
1579 |
+
layer_idx=layer_idx) for layer_idx in range(num_hidden_layers)])
|
1580 |
self._attn_implementation = config._attn_implementation
|
1581 |
+
RMSNorm = MorehRMSNorm if MorehRMSNorm is not None else MotifRMSNorm
|
1582 |
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
1583 |
self.hidden_size = config.hidden_size
|
1584 |
self.num_heads = config.num_attention_heads
|
|
|
1592 |
self.gradient_checkpointing = False
|
1593 |
self.post_init()
|
1594 |
|
1595 |
+
self.use_pipeline = config.use_pipeline
|
1596 |
+
if self.use_pipeline:
|
1597 |
+
logger.info('use reinforced pp..')
|
1598 |
+
if config.num_stages==2:
|
1599 |
+
### moe version
|
1600 |
+
if config.decontam_attn:
|
1601 |
+
self.split_layers = [15]
|
1602 |
+
else:
|
1603 |
+
if num_hidden_layers == 32:
|
1604 |
+
self.split_layers = [14] # 14: 15,17 # 13: 14:18
|
1605 |
+
else:
|
1606 |
+
self.split_layers = [6]
|
1607 |
+
elif config.num_stages==3:
|
1608 |
+
self.split_layers = [9,20] ## 10, 11, 11
|
1609 |
+
else:
|
1610 |
+
self.split_layers = [6,15,24] #7(0,7),9(6,15),9(15,24),7(24,31)
|
1611 |
+
logger.info(f' check the split layers (moe): {self.split_layers}')
|
1612 |
+
|
1613 |
+
self.scale_emb = 1
|
1614 |
+
|
1615 |
+
# Reparameterization <|_1_|>
|
1616 |
+
if config.wesar_weights :
|
1617 |
+
logger.info(f'config.wesar_weights {config.wesar_weights}')
|
1618 |
+
self.norm_alpha = nn.Parameter(torch.tensor(1).float())
|
1619 |
+
self.scale_emb = 10
|
1620 |
+
else:
|
1621 |
+
self.norm_alpha = 1
|
1622 |
+
|
1623 |
def get_input_embeddings(self):
|
1624 |
return self.embed_tokens
|
1625 |
|
|
|
1658 |
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...")
|
1659 |
use_cache = False
|
1660 |
|
1661 |
+
# kept for BC (non `Cache` `past_key_values` inputs)
|
1662 |
return_legacy_cache = False
|
1663 |
if use_cache and not isinstance(past_key_values, Cache):
|
1664 |
return_legacy_cache = True
|
|
|
1672 |
"(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)")
|
1673 |
|
1674 |
if inputs_embeds is None:
|
1675 |
+
inputs_embeds = self.embed_tokens(input_ids) * self.scale_emb
|
1676 |
|
1677 |
if cache_position is None:
|
1678 |
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
1679 |
cache_position = torch.arange(past_seen_tokens,
|
1680 |
past_seen_tokens + inputs_embeds.shape[1],
|
1681 |
device=inputs_embeds.device)
|
1682 |
+
#position_ids = None
|
1683 |
if position_ids is None:
|
1684 |
position_ids = cache_position.unsqueeze(0)
|
1685 |
+
|
1686 |
causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position, past_key_values,
|
1687 |
output_attentions)
|
1688 |
|
|
|
1725 |
)
|
1726 |
|
1727 |
hidden_states = layer_outputs[0]
|
1728 |
+
|
1729 |
+
|
1730 |
+
if self.use_pipeline and idx in self.split_layers:
|
1731 |
+
hidden_states = torch.moreh.pipeline_assign(hidden_states)
|
1732 |
|
1733 |
if use_cache:
|
1734 |
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
|
|
|
1736 |
if output_attentions:
|
1737 |
all_self_attns += (layer_outputs[1], )
|
1738 |
|
1739 |
+
# <|_2_|>
|
1740 |
+
hidden_states = self.norm(hidden_states)* self.norm_alpha
|
1741 |
+
|
1742 |
# add hidden states from the last decoder layer
|
1743 |
if output_hidden_states:
|
1744 |
all_hidden_states += (hidden_states, )
|
|
|
1770 |
output_attentions: bool,
|
1771 |
):
|
1772 |
if self.config._attn_implementation == "flash_attention_2":
|
1773 |
+
if MorehFlashAttention is not None:
|
1774 |
+
return attention_mask
|
1775 |
if attention_mask is not None and 0.0 in attention_mask:
|
1776 |
return attention_mask
|
1777 |
return None
|
|
|
1843 |
"""
|
1844 |
Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
|
1845 |
`(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
|
1846 |
+
|
1847 |
Args:
|
1848 |
attention_mask (`torch.Tensor`):
|
1849 |
A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
|
|
|
1901 |
self.vocab_size = config.vocab_size
|
1902 |
self.multi_token_heads = config.multi_token_heads
|
1903 |
|
1904 |
+
if self.multi_token_heads is None:
|
1905 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
1906 |
+
else:
|
1907 |
+
self.tokenwise_last_layers = nn.ModuleList(
|
1908 |
+
[MotifDecoderLayer(config, config.num_hidden_layers - 1) for _ in range(self.multi_token_heads)])
|
1909 |
+
self.tokenwise_lm_heads = nn.ModuleList(
|
1910 |
+
[nn.Linear(config.hidden_size, config.vocab_size, bias=False) for _ in range(self.multi_token_heads)])
|
1911 |
+
self.should_skip_separate_backward_pass = self.multi_token_heads is not None
|
1912 |
|
1913 |
# Initialize weights and apply final processing
|
1914 |
self.post_init()
|
1915 |
+
|
1916 |
+
# <|_3_|>
|
1917 |
+
if config.muP:
|
1918 |
+
self.lm_head.__do_scale_tager_mu_dim_base_model__=True
|
1919 |
+
|
1920 |
+
# <|_4_|>
|
1921 |
+
self.lm_head_alpha = 1
|
1922 |
+
if config.wesar_weights:
|
1923 |
+
self.lm_head_alpha = nn.Parameter(torch.tensor(1).float())
|
1924 |
+
|
1925 |
if getattr(config, "tie_word_embeddings", True):
|
1926 |
logger.info('tie embeddings')
|
1927 |
self.tie_weights()
|
1928 |
+
else:
|
1929 |
+
# <|_5_|>
|
1930 |
+
self.lm_head.__do_scale_tager_mu_dim_base_model__ = False
|
1931 |
|
1932 |
def get_input_embeddings(self):
|
1933 |
return self.model.embed_tokens
|
|
|
1947 |
def get_decoder(self):
|
1948 |
return self.model
|
1949 |
|
1950 |
+
def multi_token_forward_backward(self,
|
1951 |
+
hidden_states: torch.FloatTensor,
|
1952 |
+
outputs: MotifModelOutputWithPast,
|
1953 |
+
labels: torch.LongTensor,
|
1954 |
+
position_ids: Optional[torch.LongTensor],
|
1955 |
+
output_attentions: Optional[bool],
|
1956 |
+
use_cache: Optional[bool],
|
1957 |
+
cache_position: Optional[torch.LongTensor],
|
1958 |
+
return_dict: Optional[bool],
|
1959 |
+
num_logits_to_keep: int = 0) -> CausalLMOutputWithPast:
|
1960 |
+
"""
|
1961 |
+
This implements the main forward-backward procedure for multi-token model training proposed in
|
1962 |
+
the paper https://arxiv.org/abs/2404.19737.
|
1963 |
+
Essentially,
|
1964 |
+
- The multi-token model tries to predict n (instead of 1) tokens at a time.
|
1965 |
+
- Applying this only during training and using first-token prediction during inference is still helpful.
|
1966 |
+
- The change in architecture: when using n-token prediction, each token index (between 1 and n) has its own
|
1967 |
+
(1) last attention layer and (2) lm head.
|
1968 |
+
- The change in loss: sum of cross-entropy losses corresponding to each token index.
|
1969 |
+
- Custom forward-backward procedure for memory efficiency: refer to the implementation of `multi_head_forward_backward`.
|
1970 |
+
"""
|
1971 |
+
if not return_dict:
|
1972 |
+
raise NotImplementedError("return_dict must be True for multi-token training")
|
1973 |
+
|
1974 |
+
past_key_values = outputs.past_key_values
|
1975 |
+
causal_mask = outputs.causal_mask
|
1976 |
+
position_embeddings = outputs.position_embeddings
|
1977 |
+
|
1978 |
+
if labels is not None:
|
1979 |
+
labels = labels.to(hidden_states.device)
|
1980 |
+
|
1981 |
+
def _tokenwise_forward(hidden_states: torch.Tensor, token_idx):
|
1982 |
+
## Model forward
|
1983 |
+
layer = self.tokenwise_last_layers[token_idx]
|
1984 |
+
lm_head = self.tokenwise_lm_heads[token_idx]
|
1985 |
+
|
1986 |
+
layer_outputs = layer(
|
1987 |
+
hidden_states,
|
1988 |
+
attention_mask=causal_mask,
|
1989 |
+
position_ids=position_ids,
|
1990 |
+
past_key_values=past_key_values, # TODO: update past_key_values?
|
1991 |
+
output_attentions=output_attentions,
|
1992 |
+
use_cache=use_cache,
|
1993 |
+
cache_position=cache_position,
|
1994 |
+
position_embeddings=position_embeddings,
|
1995 |
+
)
|
1996 |
+
last_hidden_states = layer_outputs[0]
|
1997 |
+
if num_logits_to_keep > 0:
|
1998 |
+
assert labels is None
|
1999 |
+
last_hidden_states = last_hidden_states[:, -num_logits_to_keep:, :]
|
2000 |
+
tokenwise_logits = lm_head(last_hidden_states)
|
2001 |
+
|
2002 |
+
if labels is None:
|
2003 |
+
return {
|
2004 |
+
"loss": None,
|
2005 |
+
"logits": tokenwise_logits,
|
2006 |
+
}
|
2007 |
+
|
2008 |
+
## Compute loss
|
2009 |
+
shift_n = token_idx + 1
|
2010 |
+
shift_logits = tokenwise_logits[..., :-shift_n, :].contiguous()
|
2011 |
+
shift_labels = labels[..., shift_n:].contiguous()
|
2012 |
+
|
2013 |
+
loss_fct = CrossEntropyLoss()
|
2014 |
+
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
2015 |
+
shift_labels = shift_labels.view(-1)
|
2016 |
+
|
2017 |
+
tokenwise_loss = loss_fct(shift_logits, shift_labels)
|
2018 |
+
|
2019 |
+
return {
|
2020 |
+
"loss": tokenwise_loss,
|
2021 |
+
"logits": tokenwise_logits,
|
2022 |
+
}
|
2023 |
+
|
2024 |
+
head_fns = [
|
2025 |
+
lambda hidden_states, token_idx=token_idx: _tokenwise_forward(hidden_states, token_idx)
|
2026 |
+
for token_idx in range(self.multi_token_heads)
|
2027 |
+
]
|
2028 |
+
loss, logits = multi_head_forward_backward(hidden_states,
|
2029 |
+
head_fns,
|
2030 |
+
return_keys=("loss", "logits"),
|
2031 |
+
return_only_first_head=True)
|
2032 |
+
|
2033 |
+
if not return_dict:
|
2034 |
+
output = (logits, ) + outputs[1:]
|
2035 |
+
return (loss, ) + output
|
2036 |
+
|
2037 |
+
return CausalLMOutputWithPast(
|
2038 |
+
loss=loss,
|
2039 |
+
logits=logits,
|
2040 |
+
past_key_values=outputs.past_key_values,
|
2041 |
+
hidden_states=outputs.hidden_states,
|
2042 |
+
attentions=outputs.attentions,
|
2043 |
+
)
|
2044 |
+
|
2045 |
@add_start_docstrings_to_model_forward(MOTIF_INPUTS_DOCSTRING)
|
2046 |
@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
2047 |
def forward(
|
|
|
2066 |
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
2067 |
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
2068 |
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
2069 |
+
|
2070 |
num_logits_to_keep (`int`, *optional*):
|
2071 |
Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
|
2072 |
`input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
|
2073 |
token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
|
2074 |
+
|
2075 |
Returns:
|
2076 |
+
|
2077 |
Example:
|
2078 |
+
|
2079 |
```python
|
2080 |
>>> from transformers import AutoTokenizer, MotifForCausalLM
|
2081 |
+
|
2082 |
+
>>> model = MotifForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
|
2083 |
+
>>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
|
2084 |
+
|
2085 |
>>> prompt = "Hey, are you conscious? Can you talk to me?"
|
2086 |
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
2087 |
+
|
2088 |
>>> # Generate
|
2089 |
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
2090 |
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
|
|
2097 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
2098 |
|
2099 |
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
2100 |
+
outputs_include_causal_mask = self.multi_token_heads is not None
|
2101 |
+
outputs_include_position_embeddings = self.multi_token_heads is not None
|
2102 |
outputs: MotifModelOutputWithPast = self.model(
|
2103 |
input_ids=input_ids,
|
2104 |
attention_mask=attention_mask,
|
|
|
2110 |
output_hidden_states=output_hidden_states,
|
2111 |
return_dict=return_dict,
|
2112 |
cache_position=cache_position,
|
2113 |
+
outputs_include_causal_mask=outputs_include_causal_mask,
|
2114 |
+
outputs_include_position_embeddings=outputs_include_position_embeddings,
|
2115 |
)
|
2116 |
|
2117 |
hidden_states = outputs[0]
|
2118 |
|
2119 |
+
if self.multi_token_heads is not None:
|
2120 |
+
return self.multi_token_forward_backward(hidden_states,
|
2121 |
+
outputs,
|
2122 |
+
labels,
|
2123 |
+
position_ids,
|
2124 |
+
output_attentions,
|
2125 |
+
use_cache,
|
2126 |
+
cache_position,
|
2127 |
+
return_dict,
|
2128 |
+
num_logits_to_keep=num_logits_to_keep)
|
2129 |
+
|
2130 |
# Only compute necessary logits, and do not upcast them to float if we are not computing the loss
|
2131 |
+
hidden_states = hidden_states * self.lm_head_alpha
|
2132 |
logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
|
2133 |
logits = logits.float()
|
2134 |
|
2135 |
loss = None
|
2136 |
if labels is not None:
|
2137 |
+
logits = logits
|
2138 |
# Shift so that tokens < n predict n
|
2139 |
shift_logits = logits[..., :-1, :].contiguous()
|
2140 |
shift_labels = labels[..., 1:].contiguous()
|
|
|
2156 |
past_key_values=outputs.past_key_values,
|
2157 |
hidden_states=outputs.hidden_states,
|
2158 |
attentions=outputs.attentions,
|
2159 |
+
)
|