upload modeling_keye.py to support non-flash inference
Browse files- modeling_keye.py +226 -765
modeling_keye.py
CHANGED
@@ -31,19 +31,10 @@ import torch.nn.functional as F
|
|
31 |
from torch.nn import CrossEntropyLoss
|
32 |
|
33 |
from transformers.activations import ACT2FN
|
34 |
-
from transformers.cache_utils import
|
35 |
-
Cache,
|
36 |
-
DynamicCache,
|
37 |
-
SlidingWindowCache,
|
38 |
-
StaticCache,
|
39 |
-
)
|
40 |
from transformers.generation import GenerationMixin
|
41 |
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
|
42 |
-
from transformers.modeling_outputs import
|
43 |
-
BaseModelOutputWithPast,
|
44 |
-
BaseModelOutput,
|
45 |
-
BaseModelOutputWithPooling,
|
46 |
-
)
|
47 |
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
|
48 |
from transformers.modeling_utils import PreTrainedModel, sdpa_attention_forward
|
49 |
from transformers.activations import GELUActivation, ACT2FN, PytorchGELUTanh
|
@@ -55,7 +46,7 @@ from transformers.utils import (
|
|
55 |
logging,
|
56 |
replace_return_docstrings,
|
57 |
torch_int,
|
58 |
-
is_flash_attn_greater_or_equal_2_10
|
59 |
)
|
60 |
from .configuration_keye import KeyeConfig, KeyeVisionConfig
|
61 |
|
@@ -64,9 +55,9 @@ import warnings
|
|
64 |
from typing import Any, Callable, Optional, Tuple, Union, List
|
65 |
from torch import nn
|
66 |
from torch.nn.init import _calculate_fan_in_and_fan_out
|
|
|
67 |
|
68 |
|
69 |
-
assert is_flash_attn_2_available()
|
70 |
if is_flash_attn_2_available():
|
71 |
from flash_attn import flash_attn_varlen_func
|
72 |
from flash_attn.layers.rotary import apply_rotary_emb
|
@@ -80,7 +71,6 @@ logger = logging.get_logger(__name__)
|
|
80 |
|
81 |
_CONFIG_FOR_DOC = "KeyeConfig"
|
82 |
|
83 |
-
|
84 |
class KeyeMLP(nn.Module):
|
85 |
def __init__(self, config, bias: bool = False):
|
86 |
super().__init__()
|
@@ -92,9 +82,7 @@ class KeyeMLP(nn.Module):
|
|
92 |
self.act_fn = ACT2FN[config.hidden_act]
|
93 |
|
94 |
def forward(self, hidden_state):
|
95 |
-
return self.down_proj(
|
96 |
-
self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state)
|
97 |
-
)
|
98 |
|
99 |
|
100 |
def _trunc_normal_(tensor, mean, std, a, b):
|
@@ -134,11 +122,7 @@ def _trunc_normal_(tensor, mean, std, a, b):
|
|
134 |
|
135 |
|
136 |
def trunc_normal_tf_(
|
137 |
-
tensor: torch.Tensor,
|
138 |
-
mean: float = 0.0,
|
139 |
-
std: float = 1.0,
|
140 |
-
a: float = -2.0,
|
141 |
-
b: float = 2.0,
|
142 |
) -> torch.Tensor:
|
143 |
"""Fills the input Tensor with values drawn from a truncated
|
144 |
normal distribution. The values are effectively drawn from the
|
@@ -196,39 +180,9 @@ def default_flax_embed_init(tensor):
|
|
196 |
variance_scaling_(tensor, mode="fan_in", distribution="normal")
|
197 |
|
198 |
|
199 |
-
@dataclass
|
200 |
-
# Copied from transformers.models.clip.modeling_clip.CLIPVisionModelOutput with CLIP->Siglip
|
201 |
-
class SiglipVisionModelOutput(ModelOutput):
|
202 |
-
"""
|
203 |
-
Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.
|
204 |
-
|
205 |
-
Args:
|
206 |
-
image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
|
207 |
-
The image embeddings obtained by applying the projection layer to the pooler_output.
|
208 |
-
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
209 |
-
Sequence of hidden-states at the output of the last layer of the model.
|
210 |
-
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
211 |
-
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
212 |
-
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
|
213 |
-
|
214 |
-
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
|
215 |
-
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
216 |
-
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
217 |
-
sequence_length)`.
|
218 |
-
|
219 |
-
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
|
220 |
-
heads.
|
221 |
-
"""
|
222 |
-
|
223 |
-
image_embeds: Optional[torch.FloatTensor] = None
|
224 |
-
last_hidden_state: Optional[torch.FloatTensor] = None
|
225 |
-
hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
|
226 |
-
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
|
227 |
-
|
228 |
-
|
229 |
class Projector(nn.Module):
|
230 |
|
231 |
-
def __init__(self, text_config: KeyeConfig,
|
232 |
super().__init__()
|
233 |
self.text_config = text_config
|
234 |
self.vision_config = vision_config
|
@@ -247,9 +201,7 @@ class Projector(nn.Module):
|
|
247 |
self.hidden_size, self.text_config.hidden_size, bias=True
|
248 |
)
|
249 |
|
250 |
-
def forward(
|
251 |
-
self, image_features: torch.Tensor, image_grid_thw: List[Tuple[int, int, int]]
|
252 |
-
) -> torch.Tensor:
|
253 |
m1, m2 = self.merge_kernel_size
|
254 |
if isinstance(image_features, (list, tuple)):
|
255 |
processed_features = list()
|
@@ -258,15 +210,7 @@ class Projector(nn.Module):
|
|
258 |
t, h, w = image_grid
|
259 |
from einops import rearrange
|
260 |
|
261 |
-
image_feature = rearrange(
|
262 |
-
image_feature,
|
263 |
-
"(t h p1 w p2) d -> (t h w) (p1 p2 d)",
|
264 |
-
t=t,
|
265 |
-
h=h // m1,
|
266 |
-
p1=m1,
|
267 |
-
w=w // m2,
|
268 |
-
p2=m2,
|
269 |
-
)
|
270 |
hidden_states = self.linear_1(image_feature)
|
271 |
hidden_states = self.act(hidden_states)
|
272 |
hidden_states = self.linear_2(hidden_states)
|
@@ -284,7 +228,6 @@ class Projector(nn.Module):
|
|
284 |
|
285 |
return hidden_states.view(*dims, -1)
|
286 |
|
287 |
-
|
288 |
class SiglipVisionEmbeddings(nn.Module):
|
289 |
def __init__(self, config: KeyeVisionConfig):
|
290 |
super().__init__()
|
@@ -308,19 +251,9 @@ class SiglipVisionEmbeddings(nn.Module):
|
|
308 |
self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
|
309 |
self.packing_position_embedding = nn.Embedding(32768, self.embed_dim)
|
310 |
|
311 |
-
self.register_buffer(
|
312 |
-
"position_ids",
|
313 |
-
torch.arange(self.num_positions).expand((1, -1)),
|
314 |
-
persistent=False,
|
315 |
-
)
|
316 |
|
317 |
-
def interpolate_pos_encoding(
|
318 |
-
self,
|
319 |
-
embeddings: torch.Tensor,
|
320 |
-
height: int,
|
321 |
-
width: int,
|
322 |
-
is_after_patchify: bool = False,
|
323 |
-
) -> torch.Tensor:
|
324 |
"""
|
325 |
This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution
|
326 |
images. This method is also adapted to support torch.jit tracing and no class embeddings.
|
@@ -343,9 +276,7 @@ class SiglipVisionEmbeddings(nn.Module):
|
|
343 |
new_width = width // self.patch_size
|
344 |
|
345 |
sqrt_num_positions = torch_int(num_positions**0.5)
|
346 |
-
patch_pos_embed = patch_pos_embed.reshape(
|
347 |
-
1, sqrt_num_positions, sqrt_num_positions, dim
|
348 |
-
)
|
349 |
patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
|
350 |
|
351 |
patch_pos_embed = nn.functional.interpolate(
|
@@ -373,42 +304,33 @@ class SiglipVisionEmbeddings(nn.Module):
|
|
373 |
if grid in self.cache_position_embedding:
|
374 |
self.cache_position_count[grid] += 1
|
375 |
return self.cache_position_embedding[grid]
|
376 |
-
|
377 |
if len(self.cache_position_embedding) >= max_cache:
|
378 |
-
min_hit_grid = min(
|
379 |
-
self.cache_position_count, key=self.cache_position_count.get
|
380 |
-
)
|
381 |
self.cache_position_count.pop(min_hit_grid)
|
382 |
self.cache_position_embedding.pop(min_hit_grid)
|
383 |
-
|
384 |
position_embedding = self.interpolate_pos_encoding(embeddings, h, w, True)
|
385 |
self.cache_position_count[grid] = 1
|
386 |
self.cache_position_embedding[grid] = position_embedding
|
387 |
return position_embedding
|
388 |
|
389 |
def forward(
|
390 |
-
self,
|
391 |
-
pixel_values: torch.FloatTensor,
|
392 |
position_ids: Optional[torch.Tensor] = None,
|
393 |
-
image_grid_thw: Optional[
|
394 |
-
|
395 |
-
] = None,
|
396 |
-
interpolate_pos_encoding=False,
|
397 |
) -> torch.Tensor:
|
398 |
if pixel_values.dim() == 5:
|
399 |
assert position_ids is not None
|
400 |
from einops import rearrange
|
401 |
-
|
402 |
batch_size, squence_len, channel, height, width = pixel_values.shape
|
403 |
target_dtype = self.patch_embedding.weight.dtype
|
404 |
pixel_values = rearrange(pixel_values, "b l c h w -> (b l) c h w")
|
405 |
-
patch_embeds = self.patch_embedding(
|
406 |
-
pixel_values.to(dtype=target_dtype)
|
407 |
-
) # shape = [*, width, grid, grid]
|
408 |
embeddings = patch_embeds.flatten(-2).squeeze(-1)
|
409 |
-
embeddings = rearrange(
|
410 |
-
embeddings, "(b l) d -> b l d", b=batch_size, l=squence_len
|
411 |
-
)
|
412 |
|
413 |
# todo: not dubug
|
414 |
if interpolate_pos_encoding and image_grid_thw is not None:
|
@@ -416,21 +338,15 @@ class SiglipVisionEmbeddings(nn.Module):
|
|
416 |
assert batch_size == 1
|
417 |
start = 0
|
418 |
image_embedding_list = list()
|
419 |
-
assert (
|
420 |
-
sum([np.prod(x) for x in flatten_image_grid_thw])
|
421 |
-
== embeddings.shape[1]
|
422 |
-
), (flatten_image_grid_thw, embeddings.shape)
|
423 |
embeddings = embeddings.squeeze(0)
|
424 |
tmp_embeddings = list()
|
425 |
for image_grid in image_grid_thw:
|
426 |
t, h, w = image_grid
|
427 |
end = start + t * h * w
|
428 |
-
image_embeddings = embeddings[start:end, :]
|
429 |
-
position_embedding = (
|
430 |
-
|
431 |
-
.squeeze(0)
|
432 |
-
.repeat(t, 1)
|
433 |
-
)
|
434 |
image_embeddings = image_embeddings + position_embedding
|
435 |
tmp_embeddings.append(image_embeddings)
|
436 |
start = end
|
@@ -456,12 +372,8 @@ def eager_attention_forward(
|
|
456 |
if attention_mask is not None:
|
457 |
attn_weights = attn_weights + attention_mask
|
458 |
|
459 |
-
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(
|
460 |
-
|
461 |
-
)
|
462 |
-
attn_weights = nn.functional.dropout(
|
463 |
-
attn_weights, p=dropout, training=module.training
|
464 |
-
)
|
465 |
|
466 |
attn_output = torch.matmul(attn_weights, value)
|
467 |
attn_output = attn_output.transpose(1, 2).contiguous()
|
@@ -502,9 +414,7 @@ class SiglipAttention(nn.Module):
|
|
502 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
|
503 |
"""Input shape: Batch x Time x Channel"""
|
504 |
|
505 |
-
use_flash_attn = (
|
506 |
-
cu_seqlens is not None
|
507 |
-
) and self.config._attn_implementation == "flash_attention_2"
|
508 |
|
509 |
batch_size, seq_length, embed_dim = hidden_states.shape
|
510 |
|
@@ -513,28 +423,21 @@ class SiglipAttention(nn.Module):
|
|
513 |
values = self.v_proj(hidden_states)
|
514 |
|
515 |
if rope_emb is None:
|
516 |
-
queries = queries.view(
|
517 |
-
|
518 |
-
).transpose(1, 2)
|
519 |
-
keys = keys.view(
|
520 |
-
batch_size, seq_length, self.num_heads, self.head_dim
|
521 |
-
).transpose(1, 2)
|
522 |
-
values = values.view(
|
523 |
-
batch_size, seq_length, self.num_heads, self.head_dim
|
524 |
-
).transpose(1, 2)
|
525 |
else:
|
526 |
assert cu_seqlens is not None, "Rope support flash attn only."
|
527 |
cos, sin = rope_emb
|
528 |
-
queries = queries.view(
|
529 |
-
batch_size, seq_length, self.num_heads, self.head_dim
|
530 |
-
)
|
531 |
keys = keys.view(batch_size, seq_length, self.num_heads, self.head_dim)
|
532 |
-
|
|
|
|
|
|
|
533 |
queries = queries.transpose(1, 2)
|
534 |
keys = keys.transpose(1, 2)
|
535 |
-
values = values.view(
|
536 |
-
batch_size, seq_length, self.num_heads, self.head_dim
|
537 |
-
).transpose(1, 2)
|
538 |
|
539 |
if not use_flash_attn:
|
540 |
attention_interface: Callable = eager_attention_forward
|
@@ -557,25 +460,16 @@ class SiglipAttention(nn.Module):
|
|
557 |
scaling=self.scale,
|
558 |
dropout=0.0 if not self.training else self.dropout,
|
559 |
)
|
560 |
-
attn_output = attn_output.reshape(
|
561 |
-
batch_size, seq_length, embed_dim
|
562 |
-
).contiguous()
|
563 |
else:
|
564 |
assert batch_size == 1, hidden_states.shape
|
565 |
queries = queries.transpose(1, 2).squeeze(0)
|
566 |
keys = keys.transpose(1, 2).squeeze(0)
|
567 |
values = values.transpose(1, 2).squeeze(0)
|
568 |
|
569 |
-
from flash_attn import flash_attn_func, flash_attn_varlen_func
|
570 |
-
|
571 |
max_seqlen_q = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
|
572 |
max_seqlen_k = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
|
573 |
-
assert (
|
574 |
-
cu_seqlens[-1].item()
|
575 |
-
== queries.shape[0]
|
576 |
-
== keys.shape[0]
|
577 |
-
== values.shape[0]
|
578 |
-
), (cu_seqlens, queries.shape, keys.shape, values.shape)
|
579 |
|
580 |
attn_output = flash_attn_varlen_func(
|
581 |
queries,
|
@@ -841,9 +735,7 @@ class SiglipEncoder(nn.Module):
|
|
841 |
embed_dim = config.hidden_size
|
842 |
num_heads = config.num_attention_heads
|
843 |
head_dim = embed_dim // num_heads
|
844 |
-
self.layers = nn.ModuleList(
|
845 |
-
[SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)]
|
846 |
-
)
|
847 |
self.rotary_pos_emb = SigLIPRotaryEmbedding(head_dim // 2)
|
848 |
self.gradient_checkpointing = False
|
849 |
|
@@ -859,7 +751,6 @@ class SiglipEncoder(nn.Module):
|
|
859 |
|
860 |
def build_window_index(self, image_grid, window_size, device):
|
861 |
from einops import rearrange
|
862 |
-
|
863 |
window_indices = list()
|
864 |
pad_values = -100
|
865 |
start_window_index = 0
|
@@ -871,25 +762,16 @@ class SiglipEncoder(nn.Module):
|
|
871 |
pad_w = (-w) % window_size
|
872 |
assert pad_h >= 0 and pad_w >= 0, (pad_h, pad_w)
|
873 |
window_index = F.pad(window_index, (0, pad_w, 0, pad_h), value=pad_values)
|
874 |
-
window_index = rearrange(
|
875 |
-
window_index,
|
876 |
-
"t (h p1) (w p2) -> t (h w) (p1 p2)",
|
877 |
-
p1=window_size,
|
878 |
-
p2=window_size,
|
879 |
-
)
|
880 |
window_seqlens = (window_index != pad_values).long().sum(-1).reshape(-1)
|
881 |
window_index = window_index.reshape(-1)
|
882 |
window_index = window_index[window_index != pad_values]
|
883 |
window_indices.append(window_index + start_window_index)
|
884 |
-
cu_seqlens_within_windows.append(
|
885 |
-
window_seqlens.cumsum(0) + start_window_index
|
886 |
-
)
|
887 |
start_window_index += t * h * w
|
888 |
window_indices = torch.concat(window_indices, dim=0)
|
889 |
cu_seqlens_within_windows = torch.concat(cu_seqlens_within_windows, dim=0)
|
890 |
-
cu_seqlens_within_windows = F.pad(
|
891 |
-
cu_seqlens_within_windows, (1, 0), value=0
|
892 |
-
).to(torch.int32)
|
893 |
return window_indices, cu_seqlens_within_windows
|
894 |
|
895 |
# Ignore copy
|
@@ -901,9 +783,7 @@ class SiglipEncoder(nn.Module):
|
|
901 |
output_attentions: Optional[bool] = None,
|
902 |
output_hidden_states: Optional[bool] = None,
|
903 |
cu_seqlens: Optional[List[torch.Tensor]] = None,
|
904 |
-
image_grid_thw: Optional[
|
905 |
-
List[Union[Tuple[int, int, int], List[Tuple[int, int, int]]]]
|
906 |
-
] = None,
|
907 |
height_position_ids: Optional[torch.Tensor] = None,
|
908 |
width_position_ids: Optional[torch.Tensor] = None,
|
909 |
use_rope: Optional[bool] = False,
|
@@ -936,17 +816,11 @@ class SiglipEncoder(nn.Module):
|
|
936 |
|
937 |
vision_or_text = "vision"
|
938 |
assert vision_or_text in ["vision", "text"]
|
939 |
-
use_window_attn = window_size > 0 and vision_or_text == "vision"
|
940 |
use_rope = (use_rope is True) and (vision_or_text == "vision")
|
941 |
-
output_attentions =
|
942 |
-
output_attentions
|
943 |
-
if output_attentions is not None
|
944 |
-
else self.config.output_attentions
|
945 |
-
)
|
946 |
output_hidden_states = (
|
947 |
-
output_hidden_states
|
948 |
-
if output_hidden_states is not None
|
949 |
-
else self.config.output_hidden_states
|
950 |
)
|
951 |
|
952 |
encoder_states = () if output_hidden_states else None
|
@@ -954,17 +828,10 @@ class SiglipEncoder(nn.Module):
|
|
954 |
|
955 |
device = inputs_embeds.device
|
956 |
hidden_states = inputs_embeds
|
957 |
-
attention_mask = (
|
958 |
-
attention_mask.to(inputs_embeds.dtype)
|
959 |
-
if attention_mask is not None
|
960 |
-
else None
|
961 |
-
)
|
962 |
if use_rope is True:
|
963 |
flatten_image_grid_thw = self.flatten_list(image_grid_thw)
|
964 |
-
assert (
|
965 |
-
sum([np.prod(x) for x in flatten_image_grid_thw])
|
966 |
-
== hidden_states.shape[1]
|
967 |
-
), (flatten_image_grid_thw, hidden_states.shape)
|
968 |
|
969 |
if width_position_ids is None or height_position_ids is None:
|
970 |
split_hids = list()
|
@@ -977,13 +844,11 @@ class SiglipEncoder(nn.Module):
|
|
977 |
split_wids.append(sample_wids)
|
978 |
width_position_ids = torch.concat(split_wids, dim=0)
|
979 |
height_position_ids = torch.concat(split_hids, dim=0)
|
980 |
-
|
981 |
window_indices, cu_seqlens_within_windows = None, None
|
982 |
|
983 |
if use_window_attn:
|
984 |
-
window_indices, cu_seqlens_within_windows = self.build_window_index(
|
985 |
-
flatten_image_grid_thw, window_size, device
|
986 |
-
)
|
987 |
reversed_window_indices = window_indices.argsort()
|
988 |
height_position_ids = height_position_ids[window_indices]
|
989 |
width_position_ids = width_position_ids[window_indices]
|
@@ -998,17 +863,12 @@ class SiglipEncoder(nn.Module):
|
|
998 |
|
999 |
rope_emb = None
|
1000 |
window_indices, cu_seqlens_within_windows = None, None
|
1001 |
-
|
1002 |
if use_window_attn:
|
1003 |
flatten_image_grid_thw = self.flatten_list(image_grid_thw)
|
1004 |
-
assert (
|
1005 |
-
|
1006 |
-
|
1007 |
-
), (flatten_image_grid_thw, hidden_states.shape)
|
1008 |
-
|
1009 |
-
window_indices, cu_seqlens_within_windows = self.build_window_index(
|
1010 |
-
flatten_image_grid_thw, window_size, device
|
1011 |
-
)
|
1012 |
reversed_window_indices = window_indices.argsort()
|
1013 |
|
1014 |
if use_window_attn:
|
@@ -1020,11 +880,7 @@ class SiglipEncoder(nn.Module):
|
|
1020 |
|
1021 |
for encoder_layer in self.layers:
|
1022 |
if output_hidden_states:
|
1023 |
-
encoder_states = encoder_states + (
|
1024 |
-
(hidden_states[:, reversed_window_indices, :],)
|
1025 |
-
if use_window_attn
|
1026 |
-
else (hidden_states,)
|
1027 |
-
)
|
1028 |
if self.gradient_checkpointing and self.training:
|
1029 |
layer_outputs = self._gradient_checkpointing_func(
|
1030 |
encoder_layer.__call__,
|
@@ -1070,17 +926,13 @@ class SiglipVisionTransformer(nn.Module):
|
|
1070 |
self.embeddings = SiglipVisionEmbeddings(config)
|
1071 |
self.encoder = SiglipEncoder(config)
|
1072 |
self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
|
1073 |
-
self.use_head = (
|
1074 |
-
True if not hasattr(config, "vision_use_head") else config.vision_use_head
|
1075 |
-
)
|
1076 |
if self.use_head:
|
1077 |
self.head = SiglipMultiheadAttentionPoolingHead(config)
|
1078 |
|
1079 |
# @can_return_tuple
|
1080 |
@add_start_docstrings_to_model_forward(SIGLIP_VISION_INPUTS_DOCSTRING)
|
1081 |
-
@replace_return_docstrings(
|
1082 |
-
output_type=BaseModelOutputWithPooling, config_class=KeyeVisionConfig
|
1083 |
-
)
|
1084 |
def forward(
|
1085 |
self,
|
1086 |
pixel_values,
|
@@ -1096,9 +948,7 @@ class SiglipVisionTransformer(nn.Module):
|
|
1096 |
cu_seqlens: Optional[List[torch.Tensor]] = None,
|
1097 |
padding_mask: Optional[torch.Tensor] = None,
|
1098 |
vision_return_embed_list: Optional[bool] = False,
|
1099 |
-
image_grid_thw: Optional[
|
1100 |
-
List[Union[Tuple[int, int, int], List[Tuple[int, int, int]]]]
|
1101 |
-
] = None,
|
1102 |
return_pooler_output: Optional[bool] = True,
|
1103 |
use_rope: Optional[bool] = False,
|
1104 |
window_size: Optional[bool] = -1,
|
@@ -1107,21 +957,15 @@ class SiglipVisionTransformer(nn.Module):
|
|
1107 |
Returns:
|
1108 |
|
1109 |
"""
|
1110 |
-
output_attentions =
|
1111 |
-
output_attentions
|
1112 |
-
if output_attentions is not None
|
1113 |
-
else self.config.output_attentions
|
1114 |
-
)
|
1115 |
output_hidden_states = (
|
1116 |
-
output_hidden_states
|
1117 |
-
if output_hidden_states is not None
|
1118 |
-
else self.config.output_hidden_states
|
1119 |
)
|
1120 |
hidden_states = self.embeddings(
|
1121 |
-
pixel_values,
|
1122 |
-
interpolate_pos_encoding=interpolate_pos_encoding,
|
1123 |
position_ids=position_ids,
|
1124 |
-
image_grid_thw=image_grid_thw
|
1125 |
)
|
1126 |
|
1127 |
encoder_outputs: BaseModelOutput = self.encoder(
|
@@ -1157,32 +1001,22 @@ class SiglipVisionTransformer(nn.Module):
|
|
1157 |
token_indices = (sample_index == sample_idx).nonzero().flatten()
|
1158 |
sample_hidden_state = hidden_state[token_indices]
|
1159 |
sample_hidden_state_list.append(sample_hidden_state)
|
1160 |
-
|
1161 |
if not vision_return_embed_list:
|
1162 |
-
max_length = max(
|
1163 |
-
[_state.shape[0] for _state in sample_hidden_state_list]
|
1164 |
-
)
|
1165 |
tmp_sample_hidden_state_list = list()
|
1166 |
padding_mask = list()
|
1167 |
for idx, _state in enumerate(sample_hidden_state_list):
|
1168 |
padding_length = max_length - _state.shape[0]
|
1169 |
-
mask = _state.new_zeros(size=(max_length,), dtype=torch.int64)
|
1170 |
-
mask[-padding_length:] = 1
|
1171 |
padding_mask.append(mask)
|
1172 |
padding = _state.new_zeros(size=(padding_length, dim))
|
1173 |
new_state = torch.concat([_state, padding], dim=0)
|
1174 |
tmp_sample_hidden_state_list.append(new_state)
|
1175 |
-
sample_hidden_state = torch.stack(
|
1176 |
-
|
1177 |
-
)
|
1178 |
-
padding_mask = (
|
1179 |
-
torch.stack(padding_mask, dim=0)
|
1180 |
-
.float()
|
1181 |
-
.to(last_hidden_state.dtype)
|
1182 |
-
)
|
1183 |
-
pooler_output = self.head(
|
1184 |
-
sample_hidden_state, key_padding_mask=padding_mask
|
1185 |
-
)
|
1186 |
else:
|
1187 |
pooler_output = list()
|
1188 |
for state in sample_hidden_state_list:
|
@@ -1206,15 +1040,15 @@ class SiglipVisionTransformer(nn.Module):
|
|
1206 |
hidden_states=encoder_outputs.hidden_states,
|
1207 |
attentions=encoder_outputs.attentions,
|
1208 |
)
|
1209 |
-
|
1210 |
sample_hidden_state = list()
|
1211 |
assert cu_seqlens is not None
|
1212 |
for i in range(cu_seqlens.shape[0] - 1):
|
1213 |
start = cu_seqlens[i]
|
1214 |
end = cu_seqlens[i + 1]
|
1215 |
-
tensor = last_hidden_state[:, start:end, :].squeeze(0)
|
1216 |
sample_hidden_state.append(tensor)
|
1217 |
-
|
1218 |
return BaseModelOutputWithPooling(
|
1219 |
last_hidden_state=sample_hidden_state,
|
1220 |
pooler_output=None,
|
@@ -1230,9 +1064,7 @@ class SiglipMultiheadAttentionPoolingHead(nn.Module):
|
|
1230 |
super().__init__()
|
1231 |
|
1232 |
self.probe = nn.Parameter(torch.randn(1, 1, config.hidden_size))
|
1233 |
-
self.attention = torch.nn.MultiheadAttention(
|
1234 |
-
config.hidden_size, config.num_attention_heads, batch_first=True
|
1235 |
-
)
|
1236 |
self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
1237 |
self.mlp = SiglipMLP(config)
|
1238 |
|
@@ -1240,9 +1072,7 @@ class SiglipMultiheadAttentionPoolingHead(nn.Module):
|
|
1240 |
batch_size = hidden_state.shape[0]
|
1241 |
probe = self.probe.repeat(batch_size, 1, 1)
|
1242 |
|
1243 |
-
hidden_state = self.attention(
|
1244 |
-
probe, hidden_state, hidden_state, key_padding_mask=key_padding_mask
|
1245 |
-
)[0]
|
1246 |
|
1247 |
residual = hidden_state
|
1248 |
hidden_state = self.layernorm(hidden_state)
|
@@ -1272,9 +1102,7 @@ class SiglipVisionModel(SiglipPreTrainedModel):
|
|
1272 |
|
1273 |
# @can_return_tuple
|
1274 |
@add_start_docstrings_to_model_forward(SIGLIP_VISION_INPUTS_DOCSTRING)
|
1275 |
-
@replace_return_docstrings(
|
1276 |
-
output_type=BaseModelOutputWithPooling, config_class=KeyeVisionConfig
|
1277 |
-
)
|
1278 |
def forward(
|
1279 |
self,
|
1280 |
pixel_values,
|
@@ -1284,9 +1112,7 @@ class SiglipVisionModel(SiglipPreTrainedModel):
|
|
1284 |
interpolate_pos_encoding: bool = False,
|
1285 |
position_ids: Optional[torch.Tensor] = None,
|
1286 |
vision_return_embed_list: Optional[bool] = False,
|
1287 |
-
image_grid_thw: Optional[
|
1288 |
-
List[Union[Tuple[int, int, int], List[Tuple[int, int, int]]]]
|
1289 |
-
] = None,
|
1290 |
cu_seqlens: Optional[List[torch.Tensor]] = None,
|
1291 |
return_pooler_output: Optional[bool] = True,
|
1292 |
use_rope: Optional[bool] = False,
|
@@ -1331,6 +1157,7 @@ class SiglipVisionModel(SiglipPreTrainedModel):
|
|
1331 |
)
|
1332 |
|
1333 |
|
|
|
1334 |
class Qwen3RMSNorm(nn.Module):
|
1335 |
def __init__(self, hidden_size, eps=1e-6):
|
1336 |
"""
|
@@ -1377,6 +1204,7 @@ def apply_rotary_pos_emb_flashatt(
|
|
1377 |
return q_embed, k_embed
|
1378 |
|
1379 |
|
|
|
1380 |
def rotate_half(x):
|
1381 |
"""Rotates half the hidden dims of the input."""
|
1382 |
x1 = x[..., : x.shape[-1] // 2]
|
@@ -1397,156 +1225,6 @@ def apply_rotary_pos_emb_vision(
|
|
1397 |
k_embed = k_embed.to(orig_k_dtype)
|
1398 |
return q_embed, k_embed
|
1399 |
|
1400 |
-
|
1401 |
-
class KeyeVisionAttention(nn.Module):
|
1402 |
-
def __init__(self, dim: int, num_heads: int = 16) -> None:
|
1403 |
-
super().__init__()
|
1404 |
-
self.num_heads = num_heads
|
1405 |
-
self.head_dim = dim // num_heads
|
1406 |
-
self.qkv = nn.Linear(dim, dim * 3, bias=True)
|
1407 |
-
self.proj = nn.Linear(dim, dim)
|
1408 |
-
|
1409 |
-
def forward(
|
1410 |
-
self,
|
1411 |
-
hidden_states: torch.Tensor,
|
1412 |
-
cu_seqlens: torch.Tensor,
|
1413 |
-
rotary_pos_emb: Optional[torch.Tensor] = None,
|
1414 |
-
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
1415 |
-
) -> torch.Tensor:
|
1416 |
-
seq_length = hidden_states.shape[0]
|
1417 |
-
q, k, v = (
|
1418 |
-
self.qkv(hidden_states)
|
1419 |
-
.reshape(seq_length, self.num_heads, 3, -1)
|
1420 |
-
.permute(2, 0, 1, 3)
|
1421 |
-
.unbind(0)
|
1422 |
-
)
|
1423 |
-
if position_embeddings is None:
|
1424 |
-
logger.warning_once(
|
1425 |
-
"The attention layers in this model are transitioning from computing the RoPE embeddings internally "
|
1426 |
-
"through `rotary_pos_emb` (2D tensor of RoPE theta values), to using externally computed "
|
1427 |
-
"`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.54 `rotary_pos_emb` will be "
|
1428 |
-
"removed and `position_embeddings` will be mandatory."
|
1429 |
-
)
|
1430 |
-
emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
|
1431 |
-
cos = emb.cos()
|
1432 |
-
sin = emb.sin()
|
1433 |
-
else:
|
1434 |
-
cos, sin = position_embeddings
|
1435 |
-
q, k = apply_rotary_pos_emb_vision(q, k, cos, sin)
|
1436 |
-
|
1437 |
-
attention_mask = torch.full(
|
1438 |
-
[1, seq_length, seq_length],
|
1439 |
-
torch.finfo(q.dtype).min,
|
1440 |
-
device=q.device,
|
1441 |
-
dtype=q.dtype,
|
1442 |
-
)
|
1443 |
-
for i in range(1, len(cu_seqlens)):
|
1444 |
-
attention_mask[
|
1445 |
-
...,
|
1446 |
-
cu_seqlens[i - 1] : cu_seqlens[i],
|
1447 |
-
cu_seqlens[i - 1] : cu_seqlens[i],
|
1448 |
-
] = 0
|
1449 |
-
|
1450 |
-
q = q.transpose(0, 1)
|
1451 |
-
k = k.transpose(0, 1)
|
1452 |
-
v = v.transpose(0, 1)
|
1453 |
-
attn_weights = torch.matmul(q, k.transpose(1, 2)) / math.sqrt(self.head_dim)
|
1454 |
-
attn_weights = attn_weights + attention_mask
|
1455 |
-
attn_weights = nn.functional.softmax(
|
1456 |
-
attn_weights, dim=-1, dtype=torch.float32
|
1457 |
-
).to(q.dtype)
|
1458 |
-
attn_output = torch.matmul(attn_weights, v)
|
1459 |
-
attn_output = attn_output.transpose(0, 1)
|
1460 |
-
attn_output = attn_output.reshape(seq_length, -1)
|
1461 |
-
attn_output = self.proj(attn_output)
|
1462 |
-
return attn_output
|
1463 |
-
|
1464 |
-
|
1465 |
-
class KeyeVisionSdpaAttention(nn.Module):
|
1466 |
-
def __init__(self, dim: int, num_heads: int = 16) -> None:
|
1467 |
-
super().__init__()
|
1468 |
-
self.num_heads = num_heads
|
1469 |
-
self.qkv = nn.Linear(dim, dim * 3, bias=True)
|
1470 |
-
self.proj = nn.Linear(dim, dim)
|
1471 |
-
|
1472 |
-
def forward(
|
1473 |
-
self,
|
1474 |
-
hidden_states: torch.Tensor,
|
1475 |
-
cu_seqlens: torch.Tensor,
|
1476 |
-
rotary_pos_emb: Optional[torch.Tensor] = None,
|
1477 |
-
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
1478 |
-
) -> torch.Tensor:
|
1479 |
-
seq_length = hidden_states.shape[0]
|
1480 |
-
# q, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
|
1481 |
-
q, k, v = (
|
1482 |
-
self.qkv(hidden_states)
|
1483 |
-
.reshape(seq_length, self.num_heads, 3, -1)
|
1484 |
-
.permute(2, 0, 1, 3)
|
1485 |
-
.unbind(0)
|
1486 |
-
)
|
1487 |
-
if position_embeddings is None:
|
1488 |
-
logger.warning_once(
|
1489 |
-
"The attention layers in this model are transitioning from computing the RoPE embeddings internally "
|
1490 |
-
"through `rotary_pos_emb` (2D tensor of RoPE theta values), to using externally computed "
|
1491 |
-
"`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.54 `rotary_pos_emb` will be "
|
1492 |
-
"removed and `position_embeddings` will be mandatory."
|
1493 |
-
)
|
1494 |
-
emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
|
1495 |
-
cos = emb.cos()
|
1496 |
-
sin = emb.sin()
|
1497 |
-
else:
|
1498 |
-
cos, sin = position_embeddings
|
1499 |
-
q, k = apply_rotary_pos_emb_vision(q, k, cos, sin)
|
1500 |
-
|
1501 |
-
attention_mask = torch.zeros(
|
1502 |
-
[1, seq_length, seq_length], device=q.device, dtype=torch.bool
|
1503 |
-
)
|
1504 |
-
for i in range(1, len(cu_seqlens)):
|
1505 |
-
attention_mask[
|
1506 |
-
...,
|
1507 |
-
cu_seqlens[i - 1] : cu_seqlens[i],
|
1508 |
-
cu_seqlens[i - 1] : cu_seqlens[i],
|
1509 |
-
] = True
|
1510 |
-
q = q.transpose(0, 1)
|
1511 |
-
k = k.transpose(0, 1)
|
1512 |
-
v = v.transpose(0, 1)
|
1513 |
-
attn_output = F.scaled_dot_product_attention(
|
1514 |
-
q, k, v, attention_mask, dropout_p=0.0
|
1515 |
-
)
|
1516 |
-
attn_output = attn_output.transpose(0, 1)
|
1517 |
-
attn_output = attn_output.reshape(seq_length, -1)
|
1518 |
-
attn_output = self.proj(attn_output)
|
1519 |
-
return attn_output
|
1520 |
-
|
1521 |
-
|
1522 |
-
class KeyeVisionBlock(nn.Module):
|
1523 |
-
def __init__(self, config, attn_implementation: str = "sdpa") -> None:
|
1524 |
-
super().__init__()
|
1525 |
-
self.norm1 = Qwen3RMSNorm(config.hidden_size, eps=1e-6)
|
1526 |
-
self.norm2 = Qwen3RMSNorm(config.hidden_size, eps=1e-6)
|
1527 |
-
assert attn_implementation == "flash_attention_2"
|
1528 |
-
self.attn = QWEN3_ATTENTION_CLASSES[attn_implementation](
|
1529 |
-
config.hidden_size, num_heads=config.num_heads
|
1530 |
-
)
|
1531 |
-
self.mlp = KeyeMLP(config, bias=True)
|
1532 |
-
|
1533 |
-
def forward(
|
1534 |
-
self,
|
1535 |
-
hidden_states: torch.Tensor,
|
1536 |
-
cu_seqlens: torch.Tensor,
|
1537 |
-
rotary_pos_emb: Optional[torch.Tensor] = None,
|
1538 |
-
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
1539 |
-
) -> torch.Tensor:
|
1540 |
-
hidden_states = hidden_states + self.attn(
|
1541 |
-
self.norm1(hidden_states),
|
1542 |
-
cu_seqlens=cu_seqlens,
|
1543 |
-
rotary_pos_emb=rotary_pos_emb,
|
1544 |
-
position_embeddings=position_embeddings,
|
1545 |
-
)
|
1546 |
-
hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
|
1547 |
-
return hidden_states
|
1548 |
-
|
1549 |
-
|
1550 |
Keye_START_DOCSTRING = r"""
|
1551 |
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
1552 |
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
@@ -1572,7 +1250,7 @@ class Qwen3PreTrainedModel(PreTrainedModel):
|
|
1572 |
config_class = KeyeConfig
|
1573 |
base_model_prefix = "model"
|
1574 |
supports_gradient_checkpointing = True
|
1575 |
-
_no_split_modules = ["KeyeDecoderLayer"
|
1576 |
_skip_keys_device_placement = "past_key_values"
|
1577 |
_supports_flash_attn_2 = True
|
1578 |
_supports_sdpa = True
|
@@ -1591,6 +1269,7 @@ class Qwen3PreTrainedModel(PreTrainedModel):
|
|
1591 |
module.weight.data[module.padding_idx].zero_()
|
1592 |
|
1593 |
|
|
|
1594 |
class SigLIPRotaryEmbedding(nn.Module):
|
1595 |
def __init__(self, dim: int, theta: float = 10000.0) -> None:
|
1596 |
super().__init__()
|
@@ -1599,15 +1278,11 @@ class SigLIPRotaryEmbedding(nn.Module):
|
|
1599 |
self.rope_init()
|
1600 |
|
1601 |
def rope_init(self):
|
1602 |
-
inv_freq = 1.0 / (
|
1603 |
-
self.theta ** (torch.arange(0, self.dim, 2, dtype=torch.float) / self.dim)
|
1604 |
-
)
|
1605 |
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
1606 |
|
1607 |
def forward(self, seqlen: int) -> torch.Tensor:
|
1608 |
-
seq = torch.arange(
|
1609 |
-
seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype
|
1610 |
-
)
|
1611 |
freqs = torch.outer(seq, self.inv_freq)
|
1612 |
return freqs
|
1613 |
|
@@ -1634,19 +1309,15 @@ class KeyeRotaryEmbedding(nn.Module):
|
|
1634 |
else:
|
1635 |
# BC: "rope_type" was originally "type"
|
1636 |
if config.rope_scaling is not None:
|
1637 |
-
self.rope_type = config.rope_scaling.get(
|
1638 |
-
"rope_type", config.rope_scaling.get("type")
|
1639 |
-
)
|
1640 |
else:
|
1641 |
self.rope_type = "default"
|
1642 |
self.max_seq_len_cached = config.max_position_embeddings
|
1643 |
self.original_max_seq_len = config.max_position_embeddings
|
1644 |
-
|
1645 |
# BC: "rope_type" was originally "type"
|
1646 |
if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
|
1647 |
-
self.rope_type = config.rope_scaling.get(
|
1648 |
-
"rope_type", config.rope_scaling.get("type")
|
1649 |
-
)
|
1650 |
else:
|
1651 |
self.rope_type = "default"
|
1652 |
self.max_seq_len_cached = config.max_position_embeddings
|
@@ -1670,15 +1341,10 @@ class KeyeRotaryEmbedding(nn.Module):
|
|
1670 |
inv_freq, self.attention_scaling = self.rope_init_fn(
|
1671 |
self.config, device, seq_len=seq_len, **self.rope_kwargs
|
1672 |
)
|
1673 |
-
self.register_buffer(
|
1674 |
-
"inv_freq", inv_freq, persistent=False
|
1675 |
-
) # TODO joao: may break with compilation
|
1676 |
self.max_seq_len_cached = seq_len
|
1677 |
|
1678 |
-
if
|
1679 |
-
seq_len < self.original_max_seq_len
|
1680 |
-
and self.max_seq_len_cached > self.original_max_seq_len
|
1681 |
-
): # reset
|
1682 |
self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
|
1683 |
self.max_seq_len_cached = self.original_max_seq_len
|
1684 |
|
@@ -1689,25 +1355,13 @@ class KeyeRotaryEmbedding(nn.Module):
|
|
1689 |
|
1690 |
# Core RoPE block. In contrast to other models, Keye has different position ids for the grids
|
1691 |
# So we expand the inv_freq to shape (3, ...)
|
1692 |
-
inv_freq_expanded = (
|
1693 |
-
|
1694 |
-
.float()
|
1695 |
-
.expand(3, position_ids.shape[1], -1, 1)
|
1696 |
-
)
|
1697 |
-
position_ids_expanded = position_ids[
|
1698 |
-
:, :, None, :
|
1699 |
-
].float() # shape (3, bs, 1, positions)
|
1700 |
# Force float32 (see https://github.com/huggingface/transformers/pull/29285)
|
1701 |
device_type = x.device.type
|
1702 |
-
device_type = (
|
1703 |
-
device_type
|
1704 |
-
if isinstance(device_type, str) and device_type != "mps"
|
1705 |
-
else "cpu"
|
1706 |
-
)
|
1707 |
with torch.autocast(device_type=device_type, enabled=False):
|
1708 |
-
freqs = (
|
1709 |
-
inv_freq_expanded.float() @ position_ids_expanded.float()
|
1710 |
-
).transpose(2, 3)
|
1711 |
emb = torch.cat((freqs, freqs), dim=-1)
|
1712 |
cos = emb.cos()
|
1713 |
sin = emb.sin()
|
@@ -1777,12 +1431,12 @@ def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim
|
|
1777 |
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
1778 |
"""
|
1779 |
mrope_section = mrope_section * 2
|
1780 |
-
cos = torch.cat(
|
1781 |
-
|
1782 |
-
)
|
1783 |
-
sin = torch.cat(
|
1784 |
-
|
1785 |
-
)
|
1786 |
|
1787 |
q_embed = (q * cos) + (rotate_half(q) * sin)
|
1788 |
k_embed = (k * cos) + (rotate_half(k) * sin)
|
@@ -1797,9 +1451,7 @@ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
|
1797 |
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
1798 |
if n_rep == 1:
|
1799 |
return hidden_states
|
1800 |
-
hidden_states = hidden_states[:, :, None, :, :].expand(
|
1801 |
-
batch, num_key_value_heads, n_rep, slen, head_dim
|
1802 |
-
)
|
1803 |
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
1804 |
|
1805 |
|
@@ -1822,43 +1474,27 @@ class KeyeAttention(nn.Module):
|
|
1822 |
|
1823 |
self.hidden_size = config.hidden_size
|
1824 |
self.num_heads = config.num_attention_heads
|
1825 |
-
self.head_dim = getattr(
|
1826 |
-
config, "head_dim", config.hidden_size // config.num_attention_heads
|
1827 |
-
)
|
1828 |
self.num_key_value_heads = config.num_key_value_heads
|
1829 |
-
self.num_key_value_groups =
|
1830 |
-
config.num_attention_heads // config.num_key_value_heads
|
1831 |
-
)
|
1832 |
self.is_causal = True
|
1833 |
self.attention_dropout = config.attention_dropout
|
1834 |
self.rope_scaling = config.rope_scaling
|
1835 |
|
1836 |
self.q_proj = nn.Linear(
|
1837 |
-
config.hidden_size,
|
1838 |
-
config.num_attention_heads * self.head_dim,
|
1839 |
-
bias=config.attention_bias,
|
1840 |
)
|
1841 |
self.k_proj = nn.Linear(
|
1842 |
-
config.hidden_size,
|
1843 |
-
config.num_key_value_heads * self.head_dim,
|
1844 |
-
bias=config.attention_bias,
|
1845 |
)
|
1846 |
self.v_proj = nn.Linear(
|
1847 |
-
config.hidden_size,
|
1848 |
-
config.num_key_value_heads * self.head_dim,
|
1849 |
-
bias=config.attention_bias,
|
1850 |
)
|
1851 |
self.o_proj = nn.Linear(
|
1852 |
-
config.num_attention_heads * self.head_dim,
|
1853 |
-
config.hidden_size,
|
1854 |
-
bias=config.attention_bias,
|
1855 |
)
|
1856 |
-
self.q_norm = Qwen3RMSNorm(
|
1857 |
-
|
1858 |
-
) # unlike olmo, only on the head dim!
|
1859 |
-
self.k_norm = Qwen3RMSNorm(
|
1860 |
-
self.head_dim, eps=config.rms_norm_eps
|
1861 |
-
) # thus post q_norm does not need reshape
|
1862 |
|
1863 |
self.rotary_emb = KeyeRotaryEmbedding(config=config)
|
1864 |
|
@@ -1871,18 +1507,12 @@ class KeyeAttention(nn.Module):
|
|
1871 |
output_attentions: bool = False,
|
1872 |
use_cache: bool = False,
|
1873 |
cache_position: Optional[torch.LongTensor] = None,
|
1874 |
-
position_embeddings: Optional[
|
1875 |
-
Tuple[torch.Tensor, torch.Tensor]
|
1876 |
-
] = None, # necessary, but kept here for BC
|
1877 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
1878 |
bsz, q_len, _ = hidden_states.size()
|
1879 |
|
1880 |
-
query_states = self.q_norm(
|
1881 |
-
|
1882 |
-
)
|
1883 |
-
key_states = self.k_norm(
|
1884 |
-
self.k_proj(hidden_states).view(bsz, q_len, -1, self.head_dim)
|
1885 |
-
)
|
1886 |
value_states = self.v_proj(hidden_states)
|
1887 |
|
1888 |
query_states = query_states.transpose(1, 2)
|
@@ -1895,22 +1525,15 @@ class KeyeAttention(nn.Module):
|
|
1895 |
)
|
1896 |
|
1897 |
if past_key_value is not None:
|
1898 |
-
cache_kwargs = {
|
1899 |
-
|
1900 |
-
"cos": cos,
|
1901 |
-
"cache_position": cache_position,
|
1902 |
-
} # Specific to RoPE models
|
1903 |
-
key_states, value_states = past_key_value.update(
|
1904 |
-
key_states, value_states, self.layer_idx, cache_kwargs
|
1905 |
-
)
|
1906 |
|
1907 |
# repeat k/v heads if n_kv_heads < n_heads
|
1908 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
1909 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
1910 |
|
1911 |
-
attn_weights = torch.matmul(
|
1912 |
-
|
1913 |
-
) / math.sqrt(self.head_dim)
|
1914 |
|
1915 |
if attention_mask is not None: # no matter the length, we just slice it
|
1916 |
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
@@ -1919,17 +1542,11 @@ class KeyeAttention(nn.Module):
|
|
1919 |
# Fix precision issues in float16 inference
|
1920 |
# Replace inf values with zeros in attention weights to prevent NaN propagation
|
1921 |
if query_states.dtype == torch.float16:
|
1922 |
-
attn_weights = torch.where(
|
1923 |
-
torch.isinf(attn_weights), torch.zeros_like(attn_weights), attn_weights
|
1924 |
-
)
|
1925 |
|
1926 |
# upcast attention to fp32
|
1927 |
-
attn_weights = nn.functional.softmax(
|
1928 |
-
|
1929 |
-
).to(query_states.dtype)
|
1930 |
-
attn_weights = nn.functional.dropout(
|
1931 |
-
attn_weights, p=self.attention_dropout, training=self.training
|
1932 |
-
)
|
1933 |
attn_output = torch.matmul(attn_weights, value_states)
|
1934 |
|
1935 |
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
@@ -1975,19 +1592,15 @@ class KeyeFlashAttention2(KeyeAttention):
|
|
1975 |
output_attentions: bool = False,
|
1976 |
use_cache: bool = False,
|
1977 |
cache_position: Optional[torch.LongTensor] = None,
|
1978 |
-
position_embeddings: Optional[
|
1979 |
-
Tuple[torch.Tensor, torch.Tensor]
|
1980 |
-
] = None, # necessary, but kept here for BC
|
1981 |
cu_seqlens: Optional[torch.Tensor] = None,
|
1982 |
-
sliding_window
|
1983 |
**kwargs,
|
1984 |
):
|
1985 |
bsz, q_len, _ = hidden_states.size()
|
1986 |
-
q
|
1987 |
query_states = self.q_norm(q)
|
1988 |
-
key_states = self.k_norm(
|
1989 |
-
self.k_proj(hidden_states).view(bsz, q_len, -1, self.head_dim)
|
1990 |
-
)
|
1991 |
value_states = self.v_proj(hidden_states)
|
1992 |
|
1993 |
query_states = query_states.transpose(1, 2)
|
@@ -2001,20 +1614,14 @@ class KeyeFlashAttention2(KeyeAttention):
|
|
2001 |
)
|
2002 |
|
2003 |
if past_key_value is not None:
|
2004 |
-
cache_kwargs = {
|
2005 |
-
|
2006 |
-
"cos": cos,
|
2007 |
-
"cache_position": cache_position,
|
2008 |
-
} # Specific to RoPE models
|
2009 |
-
key_states, value_states = past_key_value.update(
|
2010 |
-
key_states, value_states, self.layer_idx, cache_kwargs
|
2011 |
-
)
|
2012 |
|
2013 |
# repeat k/v heads if n_kv_heads < n_heads
|
2014 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
2015 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
2016 |
dropout_rate = 0.0 if not self.training else self.attention_dropout
|
2017 |
-
|
2018 |
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
|
2019 |
# therefore the input hidden states gets silently casted in float32. Hence, we need
|
2020 |
# cast them back in float16 just to be sure everything works as expected.
|
@@ -2068,7 +1675,7 @@ class KeyeFlashAttention2(KeyeAttention):
|
|
2068 |
max_seqlen,
|
2069 |
dropout_p=dropout_rate,
|
2070 |
window_size=(sliding_window, sliding_window),
|
2071 |
-
causal=self.is_causal
|
2072 |
)
|
2073 |
else:
|
2074 |
attn_output = _flash_attention_forward(
|
@@ -2108,9 +1715,7 @@ class KeyeSdpaAttention(KeyeAttention):
|
|
2108 |
output_attentions: bool = False,
|
2109 |
use_cache: bool = False,
|
2110 |
cache_position: Optional[torch.LongTensor] = None,
|
2111 |
-
position_embeddings: Optional[
|
2112 |
-
Tuple[torch.Tensor, torch.Tensor]
|
2113 |
-
] = None, # necessary, but kept here for BC
|
2114 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
2115 |
if output_attentions:
|
2116 |
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
|
@@ -2131,12 +1736,8 @@ class KeyeSdpaAttention(KeyeAttention):
|
|
2131 |
|
2132 |
bsz, q_len, _ = hidden_states.size()
|
2133 |
|
2134 |
-
query_states = self.q_norm(
|
2135 |
-
|
2136 |
-
)
|
2137 |
-
key_states = self.k_norm(
|
2138 |
-
self.k_proj(hidden_states).view(bsz, q_len, -1, self.head_dim)
|
2139 |
-
)
|
2140 |
value_states = self.v_proj(hidden_states)
|
2141 |
|
2142 |
query_states = query_states.transpose(1, 2)
|
@@ -2149,14 +1750,8 @@ class KeyeSdpaAttention(KeyeAttention):
|
|
2149 |
)
|
2150 |
|
2151 |
if past_key_value is not None:
|
2152 |
-
cache_kwargs = {
|
2153 |
-
|
2154 |
-
"cos": cos,
|
2155 |
-
"cache_position": cache_position,
|
2156 |
-
} # Specific to RoPE models
|
2157 |
-
key_states, value_states = past_key_value.update(
|
2158 |
-
key_states, value_states, self.layer_idx, cache_kwargs
|
2159 |
-
)
|
2160 |
|
2161 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
2162 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
@@ -2194,6 +1789,7 @@ class KeyeSdpaAttention(KeyeAttention):
|
|
2194 |
return attn_output, None, past_key_value
|
2195 |
|
2196 |
|
|
|
2197 |
QWEN3_ATTENTION_CLASSES = {
|
2198 |
"eager": KeyeAttention,
|
2199 |
"flash_attention_2": KeyeFlashAttention2,
|
@@ -2205,24 +1801,17 @@ class KeyeDecoderLayer(nn.Module):
|
|
2205 |
def __init__(self, config: KeyeConfig, layer_idx: int):
|
2206 |
super().__init__()
|
2207 |
self.hidden_size = config.hidden_size
|
2208 |
-
|
2209 |
-
if
|
2210 |
-
config.use_sliding_window
|
2211 |
-
and config._attn_implementation != "flash_attention_2"
|
2212 |
-
):
|
2213 |
logger.warning_once(
|
2214 |
f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
|
2215 |
"unexpected results may be encountered."
|
2216 |
)
|
2217 |
|
2218 |
-
self.self_attn = QWEN3_ATTENTION_CLASSES[config._attn_implementation](
|
2219 |
-
config, layer_idx
|
2220 |
-
)
|
2221 |
self.mlp = Qwen3MLP(config)
|
2222 |
self.input_layernorm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
2223 |
-
self.post_attention_layernorm = Qwen3RMSNorm(
|
2224 |
-
config.hidden_size, eps=config.rms_norm_eps
|
2225 |
-
)
|
2226 |
|
2227 |
def forward(
|
2228 |
self,
|
@@ -2233,13 +1822,9 @@ class KeyeDecoderLayer(nn.Module):
|
|
2233 |
output_attentions: Optional[bool] = False,
|
2234 |
use_cache: Optional[bool] = False,
|
2235 |
cache_position: Optional[torch.LongTensor] = None,
|
2236 |
-
position_embeddings: Optional[
|
2237 |
-
Tuple[torch.Tensor, torch.Tensor]
|
2238 |
-
] = None, # necessary, but kept here for BC
|
2239 |
**kwargs,
|
2240 |
-
) -> Tuple[
|
2241 |
-
torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]
|
2242 |
-
]:
|
2243 |
"""
|
2244 |
Args:
|
2245 |
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
@@ -2275,7 +1860,7 @@ class KeyeDecoderLayer(nn.Module):
|
|
2275 |
use_cache=use_cache,
|
2276 |
cache_position=cache_position,
|
2277 |
position_embeddings=position_embeddings,
|
2278 |
-
**kwargs
|
2279 |
)
|
2280 |
|
2281 |
hidden_states = residual + hidden_states
|
@@ -2291,6 +1876,7 @@ class KeyeDecoderLayer(nn.Module):
|
|
2291 |
if output_attentions:
|
2292 |
outputs += (self_attn_weights,)
|
2293 |
|
|
|
2294 |
if use_cache:
|
2295 |
outputs += (present_key_value,)
|
2296 |
|
@@ -2307,14 +1893,9 @@ class Qwen3Model(Qwen3PreTrainedModel):
|
|
2307 |
self.padding_idx = config.pad_token_id
|
2308 |
self.vocab_size = config.vocab_size
|
2309 |
|
2310 |
-
self.embed_tokens = nn.Embedding(
|
2311 |
-
config.vocab_size, config.hidden_size, self.padding_idx
|
2312 |
-
)
|
2313 |
self.layers = nn.ModuleList(
|
2314 |
-
[
|
2315 |
-
KeyeDecoderLayer(config, layer_idx)
|
2316 |
-
for layer_idx in range(config.num_hidden_layers)
|
2317 |
-
]
|
2318 |
)
|
2319 |
self._attn_implementation = config._attn_implementation
|
2320 |
self.norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
@@ -2342,28 +1923,18 @@ class Qwen3Model(Qwen3PreTrainedModel):
|
|
2342 |
output_hidden_states: Optional[bool] = None,
|
2343 |
return_dict: Optional[bool] = None,
|
2344 |
cache_position: Optional[torch.LongTensor] = None,
|
2345 |
-
**kwargs
|
2346 |
) -> Union[Tuple, BaseModelOutputWithPast]:
|
2347 |
-
output_attentions =
|
2348 |
-
output_attentions
|
2349 |
-
if output_attentions is not None
|
2350 |
-
else self.config.output_attentions
|
2351 |
-
)
|
2352 |
output_hidden_states = (
|
2353 |
-
output_hidden_states
|
2354 |
-
if output_hidden_states is not None
|
2355 |
-
else self.config.output_hidden_states
|
2356 |
)
|
2357 |
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
2358 |
|
2359 |
-
return_dict =
|
2360 |
-
return_dict if return_dict is not None else self.config.use_return_dict
|
2361 |
-
)
|
2362 |
|
2363 |
if (input_ids is None) ^ (inputs_embeds is not None):
|
2364 |
-
raise ValueError(
|
2365 |
-
"You must specify exactly one of input_ids or inputs_embeds"
|
2366 |
-
)
|
2367 |
|
2368 |
if self.gradient_checkpointing and self.training:
|
2369 |
if use_cache:
|
@@ -2380,29 +1951,19 @@ class Qwen3Model(Qwen3PreTrainedModel):
|
|
2380 |
inputs_embeds = self.embed_tokens(input_ids)
|
2381 |
|
2382 |
if cache_position is None:
|
2383 |
-
past_seen_tokens = (
|
2384 |
-
past_key_values.get_seq_length() if past_key_values is not None else 0
|
2385 |
-
)
|
2386 |
cache_position = torch.arange(
|
2387 |
-
past_seen_tokens,
|
2388 |
-
past_seen_tokens + inputs_embeds.shape[1],
|
2389 |
-
device=inputs_embeds.device,
|
2390 |
)
|
2391 |
|
2392 |
# the hard coded `3` is for temporal, height and width.
|
2393 |
if position_ids is None:
|
2394 |
-
position_ids = cache_position.view(1, 1, -1).expand(
|
2395 |
-
3, inputs_embeds.shape[0], -1
|
2396 |
-
)
|
2397 |
elif position_ids.dim() == 2:
|
2398 |
position_ids = position_ids[None, ...].expand(3, position_ids.shape[0], -1)
|
2399 |
|
2400 |
causal_mask = self._update_causal_mask(
|
2401 |
-
attention_mask,
|
2402 |
-
inputs_embeds,
|
2403 |
-
cache_position,
|
2404 |
-
past_key_values,
|
2405 |
-
output_attentions,
|
2406 |
)
|
2407 |
hidden_states = inputs_embeds
|
2408 |
|
@@ -2462,11 +2023,7 @@ class Qwen3Model(Qwen3PreTrainedModel):
|
|
2462 |
next_cache = next_decoder_cache if use_cache else None
|
2463 |
|
2464 |
if not return_dict:
|
2465 |
-
return tuple(
|
2466 |
-
v
|
2467 |
-
for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]
|
2468 |
-
if v is not None
|
2469 |
-
)
|
2470 |
return BaseModelOutputWithPast(
|
2471 |
last_hidden_state=hidden_states,
|
2472 |
past_key_values=next_cache,
|
@@ -2484,9 +2041,7 @@ class Qwen3Model(Qwen3PreTrainedModel):
|
|
2484 |
):
|
2485 |
if self.config._attn_implementation == "flash_attention_2":
|
2486 |
if attention_mask is not None and past_key_values is not None:
|
2487 |
-
is_padding_right = (
|
2488 |
-
attention_mask[:, -1].sum().item() != input_tensor.size()[0]
|
2489 |
-
)
|
2490 |
if is_padding_right:
|
2491 |
raise ValueError(
|
2492 |
"You are attempting to perform batched generation with padding_side='right'"
|
@@ -2500,9 +2055,7 @@ class Qwen3Model(Qwen3PreTrainedModel):
|
|
2500 |
# For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
|
2501 |
# order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
|
2502 |
# to infer the attention mask.
|
2503 |
-
past_seen_tokens = (
|
2504 |
-
past_key_values.get_seq_length() if past_key_values is not None else 0
|
2505 |
-
)
|
2506 |
using_static_cache = isinstance(past_key_values, StaticCache)
|
2507 |
using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)
|
2508 |
|
@@ -2557,9 +2110,7 @@ class Qwen3Model(Qwen3PreTrainedModel):
|
|
2557 |
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
|
2558 |
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
|
2559 |
# Details: https://github.com/pytorch/pytorch/issues/110213
|
2560 |
-
causal_mask = AttentionMaskConverter._unmask_unattended(
|
2561 |
-
causal_mask, min_dtype
|
2562 |
-
)
|
2563 |
|
2564 |
return causal_mask
|
2565 |
|
@@ -2605,41 +2156,31 @@ class Qwen3Model(Qwen3PreTrainedModel):
|
|
2605 |
else:
|
2606 |
min_dtype = torch.finfo(dtype).min
|
2607 |
causal_mask = torch.full(
|
2608 |
-
(sequence_length, target_length),
|
2609 |
-
fill_value=min_dtype,
|
2610 |
-
dtype=dtype,
|
2611 |
-
device=device,
|
2612 |
)
|
2613 |
-
diagonal_attend_mask = torch.arange(
|
2614 |
-
target_length, device=device
|
2615 |
-
) > cache_position.reshape(-1, 1)
|
2616 |
if config.sliding_window is not None:
|
2617 |
# if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
|
2618 |
# the check is needed to verify is current checkpoint was trained with sliding window or not
|
2619 |
-
if (
|
2620 |
-
|
2621 |
-
|
2622 |
-
|
2623 |
-
sliding_attend_mask = torch.arange(
|
2624 |
-
target_length, device=device
|
2625 |
-
) <= (cache_position.reshape(-1, 1) - config.sliding_window)
|
2626 |
diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
|
2627 |
causal_mask *= diagonal_attend_mask
|
2628 |
causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
|
2629 |
if attention_mask is not None:
|
2630 |
-
causal_mask = (
|
2631 |
-
causal_mask.clone()
|
2632 |
-
) # copy to contiguous memory for in-place edit
|
2633 |
if attention_mask.shape[-1] > target_length:
|
2634 |
attention_mask = attention_mask[:, :target_length]
|
2635 |
mask_length = attention_mask.shape[-1]
|
2636 |
-
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[
|
2637 |
-
|
2638 |
-
|
2639 |
padding_mask = padding_mask == 0
|
2640 |
-
causal_mask[:, :, :, :mask_length] = causal_mask[
|
2641 |
-
|
2642 |
-
|
2643 |
return causal_mask
|
2644 |
|
2645 |
|
@@ -2699,6 +2240,7 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
2699 |
# Initialize weights and apply final processing
|
2700 |
self.post_init()
|
2701 |
|
|
|
2702 |
def get_input_embeddings(self):
|
2703 |
return self.model.embed_tokens
|
2704 |
|
@@ -2783,9 +2325,7 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
2783 |
video_token_id = self.config.video_token_id
|
2784 |
vision_start_token_id = self.config.vision_start_token_id
|
2785 |
mrope_position_deltas = []
|
2786 |
-
if input_ids is not None and (
|
2787 |
-
image_grid_thw is not None or video_grid_thw is not None
|
2788 |
-
):
|
2789 |
total_input_ids = input_ids
|
2790 |
if attention_mask is None:
|
2791 |
attention_mask = torch.ones_like(total_input_ids)
|
@@ -2801,9 +2341,7 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
2801 |
for i, input_ids in enumerate(total_input_ids):
|
2802 |
input_ids = input_ids[attention_mask[i] == 1]
|
2803 |
image_nums, video_nums = 0, 0
|
2804 |
-
vision_start_indices = torch.argwhere(
|
2805 |
-
input_ids == vision_start_token_id
|
2806 |
-
).squeeze(1)
|
2807 |
vision_tokens = input_ids[vision_start_indices + 1]
|
2808 |
image_nums = (vision_tokens == image_token_id).sum()
|
2809 |
video_nums = (vision_tokens == video_token_id).sum()
|
@@ -2851,80 +2389,39 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
2851 |
)
|
2852 |
text_len = ed - st
|
2853 |
|
2854 |
-
st_idx = (
|
2855 |
-
|
2856 |
-
if len(llm_pos_ids_list) > 0
|
2857 |
-
else 0
|
2858 |
-
)
|
2859 |
-
llm_pos_ids_list.append(
|
2860 |
-
torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx
|
2861 |
-
)
|
2862 |
|
2863 |
-
if torch.is_tensor(second_per_grid_t):
|
2864 |
-
second_per_grid_t = second_per_grid_t.detach().item()
|
2865 |
range_tensor = torch.arange(llm_grid_t).view(-1, 1)
|
2866 |
expanded_range = range_tensor.expand(-1, llm_grid_h * llm_grid_w)
|
2867 |
|
2868 |
-
time_tensor =
|
2869 |
-
expanded_range
|
2870 |
-
* second_per_grid_t
|
2871 |
-
* self.config.vision_config.tokens_per_second
|
2872 |
-
)
|
2873 |
|
2874 |
time_tensor_long = time_tensor.long()
|
2875 |
t_index = time_tensor_long.flatten()
|
2876 |
|
2877 |
-
h_index = (
|
2878 |
-
|
2879 |
-
|
2880 |
-
.expand(llm_grid_t, -1, llm_grid_w)
|
2881 |
-
.flatten()
|
2882 |
-
)
|
2883 |
-
w_index = (
|
2884 |
-
torch.arange(llm_grid_w)
|
2885 |
-
.view(1, 1, -1)
|
2886 |
-
.expand(llm_grid_t, llm_grid_h, -1)
|
2887 |
-
.flatten()
|
2888 |
-
)
|
2889 |
-
llm_pos_ids_list.append(
|
2890 |
-
torch.stack([t_index, h_index, w_index]) + text_len + st_idx
|
2891 |
-
)
|
2892 |
st = ed + llm_grid_t * llm_grid_h * llm_grid_w
|
2893 |
|
2894 |
if st < len(input_tokens):
|
2895 |
-
st_idx = (
|
2896 |
-
llm_pos_ids_list[-1].max() + 1
|
2897 |
-
if len(llm_pos_ids_list) > 0
|
2898 |
-
else 0
|
2899 |
-
)
|
2900 |
text_len = len(input_tokens) - st
|
2901 |
-
llm_pos_ids_list.append(
|
2902 |
-
torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx
|
2903 |
-
)
|
2904 |
|
2905 |
llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
|
2906 |
-
position_ids[..., i, attention_mask[i] == 1] = llm_positions.to(
|
2907 |
-
|
2908 |
-
|
2909 |
-
mrope_position_deltas.append(
|
2910 |
-
llm_positions.max() + 1 - len(total_input_ids[i])
|
2911 |
-
)
|
2912 |
-
mrope_position_deltas = torch.tensor(
|
2913 |
-
mrope_position_deltas, device=input_ids.device
|
2914 |
-
).unsqueeze(1)
|
2915 |
return position_ids, mrope_position_deltas
|
2916 |
else:
|
2917 |
if attention_mask is not None:
|
2918 |
position_ids = attention_mask.long().cumsum(-1) - 1
|
2919 |
position_ids.masked_fill_(attention_mask == 0, 1)
|
2920 |
-
position_ids = (
|
2921 |
-
|
2922 |
-
.expand(3, -1, -1)
|
2923 |
-
.to(attention_mask.device)
|
2924 |
-
)
|
2925 |
-
max_position_ids = position_ids.max(0, keepdim=False)[0].max(
|
2926 |
-
-1, keepdim=True
|
2927 |
-
)[0]
|
2928 |
mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
|
2929 |
else:
|
2930 |
position_ids = (
|
@@ -2940,9 +2437,7 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
2940 |
|
2941 |
return position_ids, mrope_position_deltas
|
2942 |
|
2943 |
-
@replace_return_docstrings(
|
2944 |
-
output_type=KeyeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
|
2945 |
-
)
|
2946 |
def forward(
|
2947 |
self,
|
2948 |
input_ids: torch.LongTensor = None,
|
@@ -2962,7 +2457,7 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
2962 |
rope_deltas: Optional[torch.LongTensor] = None,
|
2963 |
cache_position: Optional[torch.LongTensor] = None,
|
2964 |
second_per_grid_ts: Optional[torch.Tensor] = None,
|
2965 |
-
**kwargs
|
2966 |
) -> Union[Tuple, KeyeCausalLMOutputWithPast]:
|
2967 |
r"""
|
2968 |
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
@@ -3003,19 +2498,11 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
3003 |
"The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
|
3004 |
```"""
|
3005 |
|
3006 |
-
output_attentions =
|
3007 |
-
output_attentions
|
3008 |
-
if output_attentions is not None
|
3009 |
-
else self.config.output_attentions
|
3010 |
-
)
|
3011 |
output_hidden_states = (
|
3012 |
-
output_hidden_states
|
3013 |
-
if output_hidden_states is not None
|
3014 |
-
else self.config.output_hidden_states
|
3015 |
-
)
|
3016 |
-
return_dict = (
|
3017 |
-
return_dict if return_dict is not None else self.config.use_return_dict
|
3018 |
)
|
|
|
3019 |
|
3020 |
if inputs_embeds is None:
|
3021 |
inputs_embeds = self.model.embed_tokens(input_ids)
|
@@ -3034,21 +2521,15 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
3034 |
image_grid_hws.append(thw_tuple)
|
3035 |
image_position_ids = torch.arange(numel) % np.prod(thw_tuple[1:])
|
3036 |
siglip_position_ids.append(image_position_ids)
|
3037 |
-
sample_indices.append(torch.full((numel,), idx, dtype=torch.int64))
|
3038 |
cu_seqlens.append(cu_seqlens[-1] + numel)
|
3039 |
-
|
3040 |
-
siglip_position_ids = torch.concat(siglip_position_ids, dim=0).to(
|
3041 |
-
|
3042 |
-
)
|
3043 |
-
cu_seqlens = torch.tensor(cu_seqlens, dtype=torch.int32).to(
|
3044 |
-
pixel_values.device
|
3045 |
-
)
|
3046 |
-
sample_indices = torch.concat(sample_indices, dim=0).to(
|
3047 |
-
pixel_values.device
|
3048 |
-
)
|
3049 |
|
3050 |
vision_outputs = self.visual(
|
3051 |
-
pixel_values=pixel_values,
|
3052 |
image_grid_thw=image_grid_hws,
|
3053 |
position_ids=siglip_position_ids,
|
3054 |
vision_return_embed_list=True,
|
@@ -3057,29 +2538,27 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
3057 |
cu_seqlens=cu_seqlens,
|
3058 |
return_pooler_output=False,
|
3059 |
use_rope=True,
|
3060 |
-
window_size=-1,
|
3061 |
)
|
3062 |
image_embeds = vision_outputs.last_hidden_state
|
3063 |
|
3064 |
image_embeds = self.mlp_AR(image_embeds, image_grid_thw)
|
3065 |
-
|
3066 |
n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
|
3067 |
-
#
|
3068 |
-
image_embeds = torch.cat(image_embeds,
|
3069 |
n_image_features = image_embeds.shape[0]
|
3070 |
if n_image_tokens != n_image_features:
|
3071 |
raise ValueError(
|
3072 |
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
|
3073 |
)
|
3074 |
|
3075 |
-
mask = input_ids == self.config.image_token_id
|
3076 |
mask_unsqueezed = mask.unsqueeze(-1)
|
3077 |
mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
|
3078 |
image_mask = mask_expanded.to(inputs_embeds.device)
|
3079 |
|
3080 |
-
image_embeds = image_embeds.to(
|
3081 |
-
inputs_embeds.device, inputs_embeds.dtype
|
3082 |
-
)
|
3083 |
|
3084 |
inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
|
3085 |
|
@@ -3098,20 +2577,14 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
3098 |
video_grid_hws.append(thw_tuple)
|
3099 |
video_position_ids = torch.arange(numel) % np.prod(thw_tuple[1:])
|
3100 |
siglip_position_ids.append(video_position_ids)
|
3101 |
-
sample_indices.append(torch.full((numel,), idx, dtype=torch.int64))
|
3102 |
cu_seqlens.append(cu_seqlens[-1] + numel)
|
3103 |
-
siglip_position_ids = torch.concat(siglip_position_ids, dim=0).to(
|
3104 |
-
|
3105 |
-
)
|
3106 |
-
cu_seqlens = torch.tensor(cu_seqlens, dtype=torch.int32).to(
|
3107 |
-
pixel_values_videos.device
|
3108 |
-
)
|
3109 |
-
sample_indices = torch.concat(sample_indices, dim=0).to(
|
3110 |
-
pixel_values_videos.device
|
3111 |
-
)
|
3112 |
|
3113 |
vision_outputs = self.visual(
|
3114 |
-
pixel_values=pixel_values_videos,
|
3115 |
image_grid_thw=video_grid_hws,
|
3116 |
position_ids=siglip_position_ids,
|
3117 |
vision_return_embed_list=True,
|
@@ -3120,12 +2593,12 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
3120 |
cu_seqlens=cu_seqlens,
|
3121 |
return_pooler_output=False,
|
3122 |
use_rope=True,
|
3123 |
-
window_size
|
3124 |
)
|
3125 |
video_embeds = vision_outputs.last_hidden_state
|
3126 |
video_embeds = self.mlp_AR(video_embeds, video_grid_thw)
|
3127 |
n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
|
3128 |
-
video_embeds = torch.cat(video_embeds,
|
3129 |
n_video_features = video_embeds.shape[0]
|
3130 |
if n_video_tokens != n_video_features:
|
3131 |
raise ValueError(
|
@@ -3137,18 +2610,14 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
3137 |
mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
|
3138 |
video_mask = mask_expanded.to(inputs_embeds.device)
|
3139 |
|
3140 |
-
video_embeds = video_embeds.to(
|
3141 |
-
inputs_embeds.device, inputs_embeds.dtype
|
3142 |
-
)
|
3143 |
inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
|
3144 |
|
3145 |
if attention_mask is not None:
|
3146 |
attention_mask = attention_mask.to(inputs_embeds.device)
|
3147 |
|
3148 |
# if we get 4D attention mask we cannot calculate rope deltas anymore. TODO @raushan fixme
|
3149 |
-
if position_ids is None and (
|
3150 |
-
attention_mask is None or attention_mask.ndim == 2
|
3151 |
-
):
|
3152 |
# calculate RoPE index once per generation in the pre-fill stage only
|
3153 |
if (
|
3154 |
(cache_position is not None and cache_position[0] == 0)
|
@@ -3189,7 +2658,7 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
3189 |
output_hidden_states=output_hidden_states,
|
3190 |
return_dict=return_dict,
|
3191 |
cache_position=cache_position,
|
3192 |
-
**kwargs
|
3193 |
)
|
3194 |
|
3195 |
hidden_states = outputs[0]
|
@@ -3309,13 +2778,7 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
3309 |
if expand_size == 1:
|
3310 |
return input_ids, model_kwargs
|
3311 |
|
3312 |
-
visual_keys = [
|
3313 |
-
"pixel_values",
|
3314 |
-
"image_grid_thw",
|
3315 |
-
"pixel_values_videos",
|
3316 |
-
"video_grid_thw",
|
3317 |
-
"second_per_grid_ts",
|
3318 |
-
]
|
3319 |
|
3320 |
def _expand_dict_for_generation_visual(dict_to_expand):
|
3321 |
image_grid_thw = model_kwargs.get("image_grid_thw", None)
|
@@ -3325,9 +2788,7 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
3325 |
def _repeat_interleave_samples(x, lengths, repeat_times):
|
3326 |
samples = torch.split(x, lengths)
|
3327 |
repeat_args = [repeat_times] + [1] * (x.dim() - 1)
|
3328 |
-
result = torch.cat(
|
3329 |
-
[sample.repeat(*repeat_args) for sample in samples], dim=0
|
3330 |
-
)
|
3331 |
return result
|
3332 |
|
3333 |
for key in dict_to_expand:
|
@@ -3363,9 +2824,7 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
3363 |
)
|
3364 |
tensor = torch.tensor(dict_to_expand[key])
|
3365 |
lengths = list(video_nums)
|
3366 |
-
tensor = _repeat_interleave_samples(
|
3367 |
-
tensor, lengths=lengths, repeat_times=expand_size
|
3368 |
-
)
|
3369 |
dict_to_expand[key] = tensor.tolist()
|
3370 |
return dict_to_expand
|
3371 |
|
@@ -3377,9 +2836,7 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
3377 |
and isinstance(dict_to_expand[key], torch.Tensor)
|
3378 |
and key not in visual_keys
|
3379 |
):
|
3380 |
-
dict_to_expand[key] = dict_to_expand[key].repeat_interleave(
|
3381 |
-
expand_size, dim=0
|
3382 |
-
)
|
3383 |
return dict_to_expand
|
3384 |
|
3385 |
# input_ids is required for expanding visual inputs
|
@@ -3394,11 +2851,15 @@ class KeyeForConditionalGeneration(Qwen3PreTrainedModel, GenerationMixin):
|
|
3394 |
|
3395 |
if is_encoder_decoder:
|
3396 |
if model_kwargs.get("encoder_outputs") is None:
|
3397 |
-
raise ValueError(
|
3398 |
-
|
3399 |
-
)
|
3400 |
-
model_kwargs["encoder_outputs"] = _expand_dict_for_generation(
|
3401 |
-
model_kwargs["encoder_outputs"]
|
3402 |
-
)
|
3403 |
|
3404 |
return input_ids, model_kwargs
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
from torch.nn import CrossEntropyLoss
|
32 |
|
33 |
from transformers.activations import ACT2FN
|
34 |
+
from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
|
|
|
|
|
|
|
|
|
|
|
35 |
from transformers.generation import GenerationMixin
|
36 |
from transformers.modeling_attn_mask_utils import AttentionMaskConverter
|
37 |
+
from transformers.modeling_outputs import BaseModelOutputWithPast, BaseModelOutput, BaseModelOutputWithPooling
|
|
|
|
|
|
|
|
|
38 |
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
|
39 |
from transformers.modeling_utils import PreTrainedModel, sdpa_attention_forward
|
40 |
from transformers.activations import GELUActivation, ACT2FN, PytorchGELUTanh
|
|
|
46 |
logging,
|
47 |
replace_return_docstrings,
|
48 |
torch_int,
|
49 |
+
is_flash_attn_greater_or_equal_2_10
|
50 |
)
|
51 |
from .configuration_keye import KeyeConfig, KeyeVisionConfig
|
52 |
|
|
|
55 |
from typing import Any, Callable, Optional, Tuple, Union, List
|
56 |
from torch import nn
|
57 |
from torch.nn.init import _calculate_fan_in_and_fan_out
|
58 |
+
from einops import repeat
|
59 |
|
60 |
|
|
|
61 |
if is_flash_attn_2_available():
|
62 |
from flash_attn import flash_attn_varlen_func
|
63 |
from flash_attn.layers.rotary import apply_rotary_emb
|
|
|
71 |
|
72 |
_CONFIG_FOR_DOC = "KeyeConfig"
|
73 |
|
|
|
74 |
class KeyeMLP(nn.Module):
|
75 |
def __init__(self, config, bias: bool = False):
|
76 |
super().__init__()
|
|
|
82 |
self.act_fn = ACT2FN[config.hidden_act]
|
83 |
|
84 |
def forward(self, hidden_state):
|
85 |
+
return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
|
|
|
|
|
86 |
|
87 |
|
88 |
def _trunc_normal_(tensor, mean, std, a, b):
|
|
|
122 |
|
123 |
|
124 |
def trunc_normal_tf_(
|
125 |
+
tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
|
|
|
|
|
|
|
|
|
126 |
) -> torch.Tensor:
|
127 |
"""Fills the input Tensor with values drawn from a truncated
|
128 |
normal distribution. The values are effectively drawn from the
|
|
|
180 |
variance_scaling_(tensor, mode="fan_in", distribution="normal")
|
181 |
|
182 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
183 |
class Projector(nn.Module):
|
184 |
|
185 |
+
def __init__(self, text_config: KeyeConfig,vision_config: KeyeVisionConfig):
|
186 |
super().__init__()
|
187 |
self.text_config = text_config
|
188 |
self.vision_config = vision_config
|
|
|
201 |
self.hidden_size, self.text_config.hidden_size, bias=True
|
202 |
)
|
203 |
|
204 |
+
def forward(self, image_features: torch.Tensor, image_grid_thw: List[Tuple[int, int, int]]) -> torch.Tensor:
|
|
|
|
|
205 |
m1, m2 = self.merge_kernel_size
|
206 |
if isinstance(image_features, (list, tuple)):
|
207 |
processed_features = list()
|
|
|
210 |
t, h, w = image_grid
|
211 |
from einops import rearrange
|
212 |
|
213 |
+
image_feature = rearrange(image_feature, "(t h p1 w p2) d -> (t h w) (p1 p2 d)", t=t, h=h // m1, p1=m1, w=w // m2, p2=m2)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
214 |
hidden_states = self.linear_1(image_feature)
|
215 |
hidden_states = self.act(hidden_states)
|
216 |
hidden_states = self.linear_2(hidden_states)
|
|
|
228 |
|
229 |
return hidden_states.view(*dims, -1)
|
230 |
|
|
|
231 |
class SiglipVisionEmbeddings(nn.Module):
|
232 |
def __init__(self, config: KeyeVisionConfig):
|
233 |
super().__init__()
|
|
|
251 |
self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
|
252 |
self.packing_position_embedding = nn.Embedding(32768, self.embed_dim)
|
253 |
|
254 |
+
self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
|
|
|
|
|
|
|
|
|
255 |
|
256 |
+
def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int, is_after_patchify: bool = False) -> torch.Tensor:
|
|
|
|
|
|
|
|
|
|
|
|
|
257 |
"""
|
258 |
This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher resolution
|
259 |
images. This method is also adapted to support torch.jit tracing and no class embeddings.
|
|
|
276 |
new_width = width // self.patch_size
|
277 |
|
278 |
sqrt_num_positions = torch_int(num_positions**0.5)
|
279 |
+
patch_pos_embed = patch_pos_embed.reshape(1, sqrt_num_positions, sqrt_num_positions, dim)
|
|
|
|
|
280 |
patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
|
281 |
|
282 |
patch_pos_embed = nn.functional.interpolate(
|
|
|
304 |
if grid in self.cache_position_embedding:
|
305 |
self.cache_position_count[grid] += 1
|
306 |
return self.cache_position_embedding[grid]
|
307 |
+
|
308 |
if len(self.cache_position_embedding) >= max_cache:
|
309 |
+
min_hit_grid = min(self.cache_position_count, key=self.cache_position_count.get)
|
|
|
|
|
310 |
self.cache_position_count.pop(min_hit_grid)
|
311 |
self.cache_position_embedding.pop(min_hit_grid)
|
312 |
+
|
313 |
position_embedding = self.interpolate_pos_encoding(embeddings, h, w, True)
|
314 |
self.cache_position_count[grid] = 1
|
315 |
self.cache_position_embedding[grid] = position_embedding
|
316 |
return position_embedding
|
317 |
|
318 |
def forward(
|
319 |
+
self,
|
320 |
+
pixel_values: torch.FloatTensor,
|
321 |
position_ids: Optional[torch.Tensor] = None,
|
322 |
+
image_grid_thw: Optional[List[Union[Tuple[int, int, int], List[Tuple[int, int, int]]]]] = None,
|
323 |
+
interpolate_pos_encoding=False
|
|
|
|
|
324 |
) -> torch.Tensor:
|
325 |
if pixel_values.dim() == 5:
|
326 |
assert position_ids is not None
|
327 |
from einops import rearrange
|
|
|
328 |
batch_size, squence_len, channel, height, width = pixel_values.shape
|
329 |
target_dtype = self.patch_embedding.weight.dtype
|
330 |
pixel_values = rearrange(pixel_values, "b l c h w -> (b l) c h w")
|
331 |
+
patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype)) # shape = [*, width, grid, grid]
|
|
|
|
|
332 |
embeddings = patch_embeds.flatten(-2).squeeze(-1)
|
333 |
+
embeddings = rearrange(embeddings, "(b l) d -> b l d", b=batch_size, l=squence_len)
|
|
|
|
|
334 |
|
335 |
# todo: not dubug
|
336 |
if interpolate_pos_encoding and image_grid_thw is not None:
|
|
|
338 |
assert batch_size == 1
|
339 |
start = 0
|
340 |
image_embedding_list = list()
|
341 |
+
assert sum([np.prod(x) for x in flatten_image_grid_thw]) == embeddings.shape[1], (flatten_image_grid_thw, embeddings.shape)
|
|
|
|
|
|
|
342 |
embeddings = embeddings.squeeze(0)
|
343 |
tmp_embeddings = list()
|
344 |
for image_grid in image_grid_thw:
|
345 |
t, h, w = image_grid
|
346 |
end = start + t * h * w
|
347 |
+
image_embeddings = embeddings[start: end, :]
|
348 |
+
position_embedding = self.interpolate_pos_encoding(image_embeddings, h, w, True).squeeze(0).repeat(
|
349 |
+
t, 1)
|
|
|
|
|
|
|
350 |
image_embeddings = image_embeddings + position_embedding
|
351 |
tmp_embeddings.append(image_embeddings)
|
352 |
start = end
|
|
|
372 |
if attention_mask is not None:
|
373 |
attn_weights = attn_weights + attention_mask
|
374 |
|
375 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
|
376 |
+
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
|
|
|
|
|
|
|
|
377 |
|
378 |
attn_output = torch.matmul(attn_weights, value)
|
379 |
attn_output = attn_output.transpose(1, 2).contiguous()
|
|
|
414 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
|
415 |
"""Input shape: Batch x Time x Channel"""
|
416 |
|
417 |
+
use_flash_attn = (cu_seqlens is not None) and self.config._attn_implementation == "flash_attention_2"
|
|
|
|
|
418 |
|
419 |
batch_size, seq_length, embed_dim = hidden_states.shape
|
420 |
|
|
|
423 |
values = self.v_proj(hidden_states)
|
424 |
|
425 |
if rope_emb is None:
|
426 |
+
queries = queries.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
|
427 |
+
keys = keys.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
|
428 |
+
values = values.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
|
|
|
|
|
|
|
|
|
|
|
|
|
429 |
else:
|
430 |
assert cu_seqlens is not None, "Rope support flash attn only."
|
431 |
cos, sin = rope_emb
|
432 |
+
queries = queries.view(batch_size, seq_length, self.num_heads, self.head_dim)
|
|
|
|
|
433 |
keys = keys.view(batch_size, seq_length, self.num_heads, self.head_dim)
|
434 |
+
if use_flash_attn:
|
435 |
+
queries, keys = apply_rotary_pos_emb_flashatt(queries, keys, cos, sin)
|
436 |
+
else:
|
437 |
+
queries, keys = apply_rotary_pos_emb_vision(queries, keys, cos, sin)
|
438 |
queries = queries.transpose(1, 2)
|
439 |
keys = keys.transpose(1, 2)
|
440 |
+
values = values.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
|
|
|
|
|
441 |
|
442 |
if not use_flash_attn:
|
443 |
attention_interface: Callable = eager_attention_forward
|
|
|
460 |
scaling=self.scale,
|
461 |
dropout=0.0 if not self.training else self.dropout,
|
462 |
)
|
463 |
+
attn_output = attn_output.reshape(batch_size, seq_length, embed_dim).contiguous()
|
|
|
|
|
464 |
else:
|
465 |
assert batch_size == 1, hidden_states.shape
|
466 |
queries = queries.transpose(1, 2).squeeze(0)
|
467 |
keys = keys.transpose(1, 2).squeeze(0)
|
468 |
values = values.transpose(1, 2).squeeze(0)
|
469 |
|
|
|
|
|
470 |
max_seqlen_q = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
|
471 |
max_seqlen_k = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
|
472 |
+
assert cu_seqlens[-1].item() == queries.shape[0] == keys.shape[0] == values.shape[0], (cu_seqlens, queries.shape, keys.shape, values.shape)
|
|
|
|
|
|
|
|
|
|
|
473 |
|
474 |
attn_output = flash_attn_varlen_func(
|
475 |
queries,
|
|
|
735 |
embed_dim = config.hidden_size
|
736 |
num_heads = config.num_attention_heads
|
737 |
head_dim = embed_dim // num_heads
|
738 |
+
self.layers = nn.ModuleList([SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
|
|
|
|
|
739 |
self.rotary_pos_emb = SigLIPRotaryEmbedding(head_dim // 2)
|
740 |
self.gradient_checkpointing = False
|
741 |
|
|
|
751 |
|
752 |
def build_window_index(self, image_grid, window_size, device):
|
753 |
from einops import rearrange
|
|
|
754 |
window_indices = list()
|
755 |
pad_values = -100
|
756 |
start_window_index = 0
|
|
|
762 |
pad_w = (-w) % window_size
|
763 |
assert pad_h >= 0 and pad_w >= 0, (pad_h, pad_w)
|
764 |
window_index = F.pad(window_index, (0, pad_w, 0, pad_h), value=pad_values)
|
765 |
+
window_index = rearrange(window_index, "t (h p1) (w p2) -> t (h w) (p1 p2)", p1=window_size, p2=window_size)
|
|
|
|
|
|
|
|
|
|
|
766 |
window_seqlens = (window_index != pad_values).long().sum(-1).reshape(-1)
|
767 |
window_index = window_index.reshape(-1)
|
768 |
window_index = window_index[window_index != pad_values]
|
769 |
window_indices.append(window_index + start_window_index)
|
770 |
+
cu_seqlens_within_windows.append(window_seqlens.cumsum(0) + start_window_index)
|
|
|
|
|
771 |
start_window_index += t * h * w
|
772 |
window_indices = torch.concat(window_indices, dim=0)
|
773 |
cu_seqlens_within_windows = torch.concat(cu_seqlens_within_windows, dim=0)
|
774 |
+
cu_seqlens_within_windows = F.pad(cu_seqlens_within_windows, (1, 0), value=0).to(torch.int32)
|
|
|
|
|
775 |
return window_indices, cu_seqlens_within_windows
|
776 |
|
777 |
# Ignore copy
|
|
|
783 |
output_attentions: Optional[bool] = None,
|
784 |
output_hidden_states: Optional[bool] = None,
|
785 |
cu_seqlens: Optional[List[torch.Tensor]] = None,
|
786 |
+
image_grid_thw: Optional[List[Union[Tuple[int, int, int], List[Tuple[int, int, int]]]]] = None,
|
|
|
|
|
787 |
height_position_ids: Optional[torch.Tensor] = None,
|
788 |
width_position_ids: Optional[torch.Tensor] = None,
|
789 |
use_rope: Optional[bool] = False,
|
|
|
816 |
|
817 |
vision_or_text = "vision"
|
818 |
assert vision_or_text in ["vision", "text"]
|
819 |
+
use_window_attn = (window_size > 0 and vision_or_text == "vision")
|
820 |
use_rope = (use_rope is True) and (vision_or_text == "vision")
|
821 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
|
|
|
|
|
|
|
|
822 |
output_hidden_states = (
|
823 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
|
|
|
|
824 |
)
|
825 |
|
826 |
encoder_states = () if output_hidden_states else None
|
|
|
828 |
|
829 |
device = inputs_embeds.device
|
830 |
hidden_states = inputs_embeds
|
831 |
+
attention_mask = attention_mask.to(inputs_embeds.dtype) if attention_mask is not None else None
|
|
|
|
|
|
|
|
|
832 |
if use_rope is True:
|
833 |
flatten_image_grid_thw = self.flatten_list(image_grid_thw)
|
834 |
+
assert sum([np.prod(x) for x in flatten_image_grid_thw]) == hidden_states.shape[1], (flatten_image_grid_thw, hidden_states.shape)
|
|
|
|
|
|
|
835 |
|
836 |
if width_position_ids is None or height_position_ids is None:
|
837 |
split_hids = list()
|
|
|
844 |
split_wids.append(sample_wids)
|
845 |
width_position_ids = torch.concat(split_wids, dim=0)
|
846 |
height_position_ids = torch.concat(split_hids, dim=0)
|
847 |
+
|
848 |
window_indices, cu_seqlens_within_windows = None, None
|
849 |
|
850 |
if use_window_attn:
|
851 |
+
window_indices, cu_seqlens_within_windows = self.build_window_index(flatten_image_grid_thw, window_size, device)
|
|
|
|
|
852 |
reversed_window_indices = window_indices.argsort()
|
853 |
height_position_ids = height_position_ids[window_indices]
|
854 |
width_position_ids = width_position_ids[window_indices]
|
|
|
863 |
|
864 |
rope_emb = None
|
865 |
window_indices, cu_seqlens_within_windows = None, None
|
866 |
+
|
867 |
if use_window_attn:
|
868 |
flatten_image_grid_thw = self.flatten_list(image_grid_thw)
|
869 |
+
assert sum([np.prod(x) for x in flatten_image_grid_thw]) == hidden_states.shape[1], (flatten_image_grid_thw, hidden_states.shape)
|
870 |
+
|
871 |
+
window_indices, cu_seqlens_within_windows = self.build_window_index(flatten_image_grid_thw, window_size, device)
|
|
|
|
|
|
|
|
|
|
|
872 |
reversed_window_indices = window_indices.argsort()
|
873 |
|
874 |
if use_window_attn:
|
|
|
880 |
|
881 |
for encoder_layer in self.layers:
|
882 |
if output_hidden_states:
|
883 |
+
encoder_states = encoder_states + ((hidden_states[:, reversed_window_indices, :],) if use_window_attn else (hidden_states, ))
|
|
|
|
|
|
|
|
|
884 |
if self.gradient_checkpointing and self.training:
|
885 |
layer_outputs = self._gradient_checkpointing_func(
|
886 |
encoder_layer.__call__,
|
|
|
926 |
self.embeddings = SiglipVisionEmbeddings(config)
|
927 |
self.encoder = SiglipEncoder(config)
|
928 |
self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
|
929 |
+
self.use_head = True if not hasattr(config, "vision_use_head") else config.vision_use_head
|
|
|
|
|
930 |
if self.use_head:
|
931 |
self.head = SiglipMultiheadAttentionPoolingHead(config)
|
932 |
|
933 |
# @can_return_tuple
|
934 |
@add_start_docstrings_to_model_forward(SIGLIP_VISION_INPUTS_DOCSTRING)
|
935 |
+
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=KeyeVisionConfig)
|
|
|
|
|
936 |
def forward(
|
937 |
self,
|
938 |
pixel_values,
|
|
|
948 |
cu_seqlens: Optional[List[torch.Tensor]] = None,
|
949 |
padding_mask: Optional[torch.Tensor] = None,
|
950 |
vision_return_embed_list: Optional[bool] = False,
|
951 |
+
image_grid_thw: Optional[List[Union[Tuple[int, int, int], List[Tuple[int, int, int]]]]] = None,
|
|
|
|
|
952 |
return_pooler_output: Optional[bool] = True,
|
953 |
use_rope: Optional[bool] = False,
|
954 |
window_size: Optional[bool] = -1,
|
|
|
957 |
Returns:
|
958 |
|
959 |
"""
|
960 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
|
|
|
|
|
|
|
|
961 |
output_hidden_states = (
|
962 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
|
|
|
|
963 |
)
|
964 |
hidden_states = self.embeddings(
|
965 |
+
pixel_values,
|
966 |
+
interpolate_pos_encoding=interpolate_pos_encoding,
|
967 |
position_ids=position_ids,
|
968 |
+
image_grid_thw=image_grid_thw
|
969 |
)
|
970 |
|
971 |
encoder_outputs: BaseModelOutput = self.encoder(
|
|
|
1001 |
token_indices = (sample_index == sample_idx).nonzero().flatten()
|
1002 |
sample_hidden_state = hidden_state[token_indices]
|
1003 |
sample_hidden_state_list.append(sample_hidden_state)
|
1004 |
+
|
1005 |
if not vision_return_embed_list:
|
1006 |
+
max_length = max([_state.shape[0] for _state in sample_hidden_state_list])
|
|
|
|
|
1007 |
tmp_sample_hidden_state_list = list()
|
1008 |
padding_mask = list()
|
1009 |
for idx, _state in enumerate(sample_hidden_state_list):
|
1010 |
padding_length = max_length - _state.shape[0]
|
1011 |
+
mask = _state.new_zeros(size=(max_length, ), dtype=torch.int64)
|
1012 |
+
mask[-padding_length: ] = 1
|
1013 |
padding_mask.append(mask)
|
1014 |
padding = _state.new_zeros(size=(padding_length, dim))
|
1015 |
new_state = torch.concat([_state, padding], dim=0)
|
1016 |
tmp_sample_hidden_state_list.append(new_state)
|
1017 |
+
sample_hidden_state = torch.stack(tmp_sample_hidden_state_list, dim=0)
|
1018 |
+
padding_mask = torch.stack(padding_mask, dim=0).float().to(last_hidden_state.dtype)
|
1019 |
+
pooler_output = self.head(sample_hidden_state, key_padding_mask=padding_mask)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1020 |
else:
|
1021 |
pooler_output = list()
|
1022 |
for state in sample_hidden_state_list:
|
|
|
1040 |
hidden_states=encoder_outputs.hidden_states,
|
1041 |
attentions=encoder_outputs.attentions,
|
1042 |
)
|
1043 |
+
|
1044 |
sample_hidden_state = list()
|
1045 |
assert cu_seqlens is not None
|
1046 |
for i in range(cu_seqlens.shape[0] - 1):
|
1047 |
start = cu_seqlens[i]
|
1048 |
end = cu_seqlens[i + 1]
|
1049 |
+
tensor = last_hidden_state[:, start: end, :].squeeze(0)
|
1050 |
sample_hidden_state.append(tensor)
|
1051 |
+
|
1052 |
return BaseModelOutputWithPooling(
|
1053 |
last_hidden_state=sample_hidden_state,
|
1054 |
pooler_output=None,
|
|
|
1064 |
super().__init__()
|
1065 |
|
1066 |
self.probe = nn.Parameter(torch.randn(1, 1, config.hidden_size))
|
1067 |
+
self.attention = torch.nn.MultiheadAttention(config.hidden_size, config.num_attention_heads, batch_first=True)
|
|
|
|
|
1068 |
self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
1069 |
self.mlp = SiglipMLP(config)
|
1070 |
|
|
|
1072 |
batch_size = hidden_state.shape[0]
|
1073 |
probe = self.probe.repeat(batch_size, 1, 1)
|
1074 |
|
1075 |
+
hidden_state = self.attention(probe, hidden_state, hidden_state, key_padding_mask=key_padding_mask)[0]
|
|
|
|
|
1076 |
|
1077 |
residual = hidden_state
|
1078 |
hidden_state = self.layernorm(hidden_state)
|
|
|
1102 |
|
1103 |
# @can_return_tuple
|
1104 |
@add_start_docstrings_to_model_forward(SIGLIP_VISION_INPUTS_DOCSTRING)
|
1105 |
+
@replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=KeyeVisionConfig)
|
|
|
|
|
1106 |
def forward(
|
1107 |
self,
|
1108 |
pixel_values,
|
|
|
1112 |
interpolate_pos_encoding: bool = False,
|
1113 |
position_ids: Optional[torch.Tensor] = None,
|
1114 |
vision_return_embed_list: Optional[bool] = False,
|
1115 |
+
image_grid_thw: Optional[List[Union[Tuple[int, int, int], List[Tuple[int, int, int]]]]] = None,
|
|
|
|
|
1116 |
cu_seqlens: Optional[List[torch.Tensor]] = None,
|
1117 |
return_pooler_output: Optional[bool] = True,
|
1118 |
use_rope: Optional[bool] = False,
|
|
|
1157 |
)
|
1158 |
|
1159 |
|
1160 |
+
|
1161 |
class Qwen3RMSNorm(nn.Module):
|
1162 |
def __init__(self, hidden_size, eps=1e-6):
|
1163 |
"""
|
|
|
1204 |
return q_embed, k_embed
|
1205 |
|
1206 |
|
1207 |
+
|
1208 |
def rotate_half(x):
|
1209 |
"""Rotates half the hidden dims of the input."""
|
1210 |
x1 = x[..., : x.shape[-1] // 2]
|
|
|
1225 |
k_embed = k_embed.to(orig_k_dtype)
|
1226 |
return q_embed, k_embed
|
1227 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1228 |
Keye_START_DOCSTRING = r"""
|
1229 |
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
1230 |
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
|
|
1250 |
config_class = KeyeConfig
|
1251 |
base_model_prefix = "model"
|
1252 |
supports_gradient_checkpointing = True
|
1253 |
+
_no_split_modules = ["KeyeDecoderLayer"]
|
1254 |
_skip_keys_device_placement = "past_key_values"
|
1255 |
_supports_flash_attn_2 = True
|
1256 |
_supports_sdpa = True
|
|
|
1269 |
module.weight.data[module.padding_idx].zero_()
|
1270 |
|
1271 |
|
1272 |
+
|
1273 |
class SigLIPRotaryEmbedding(nn.Module):
|
1274 |
def __init__(self, dim: int, theta: float = 10000.0) -> None:
|
1275 |
super().__init__()
|
|
|
1278 |
self.rope_init()
|
1279 |
|
1280 |
def rope_init(self):
|
1281 |
+
inv_freq = 1.0 / (self.theta ** (torch.arange(0, self.dim, 2, dtype=torch.float) / self.dim))
|
|
|
|
|
1282 |
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
1283 |
|
1284 |
def forward(self, seqlen: int) -> torch.Tensor:
|
1285 |
+
seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
|
|
|
|
|
1286 |
freqs = torch.outer(seq, self.inv_freq)
|
1287 |
return freqs
|
1288 |
|
|
|
1309 |
else:
|
1310 |
# BC: "rope_type" was originally "type"
|
1311 |
if config.rope_scaling is not None:
|
1312 |
+
self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
|
|
|
|
|
1313 |
else:
|
1314 |
self.rope_type = "default"
|
1315 |
self.max_seq_len_cached = config.max_position_embeddings
|
1316 |
self.original_max_seq_len = config.max_position_embeddings
|
1317 |
+
|
1318 |
# BC: "rope_type" was originally "type"
|
1319 |
if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
|
1320 |
+
self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
|
|
|
|
|
1321 |
else:
|
1322 |
self.rope_type = "default"
|
1323 |
self.max_seq_len_cached = config.max_position_embeddings
|
|
|
1341 |
inv_freq, self.attention_scaling = self.rope_init_fn(
|
1342 |
self.config, device, seq_len=seq_len, **self.rope_kwargs
|
1343 |
)
|
1344 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False) # TODO joao: may break with compilation
|
|
|
|
|
1345 |
self.max_seq_len_cached = seq_len
|
1346 |
|
1347 |
+
if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len: # reset
|
|
|
|
|
|
|
1348 |
self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
|
1349 |
self.max_seq_len_cached = self.original_max_seq_len
|
1350 |
|
|
|
1355 |
|
1356 |
# Core RoPE block. In contrast to other models, Keye has different position ids for the grids
|
1357 |
# So we expand the inv_freq to shape (3, ...)
|
1358 |
+
inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1)
|
1359 |
+
position_ids_expanded = position_ids[:, :, None, :].float() # shape (3, bs, 1, positions)
|
|
|
|
|
|
|
|
|
|
|
|
|
1360 |
# Force float32 (see https://github.com/huggingface/transformers/pull/29285)
|
1361 |
device_type = x.device.type
|
1362 |
+
device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
|
|
|
|
|
|
|
|
|
1363 |
with torch.autocast(device_type=device_type, enabled=False):
|
1364 |
+
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
|
|
|
|
|
1365 |
emb = torch.cat((freqs, freqs), dim=-1)
|
1366 |
cos = emb.cos()
|
1367 |
sin = emb.sin()
|
|
|
1431 |
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
1432 |
"""
|
1433 |
mrope_section = mrope_section * 2
|
1434 |
+
cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
|
1435 |
+
unsqueeze_dim
|
1436 |
+
)
|
1437 |
+
sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
|
1438 |
+
unsqueeze_dim
|
1439 |
+
)
|
1440 |
|
1441 |
q_embed = (q * cos) + (rotate_half(q) * sin)
|
1442 |
k_embed = (k * cos) + (rotate_half(k) * sin)
|
|
|
1451 |
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
1452 |
if n_rep == 1:
|
1453 |
return hidden_states
|
1454 |
+
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
|
|
|
|
1455 |
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
1456 |
|
1457 |
|
|
|
1474 |
|
1475 |
self.hidden_size = config.hidden_size
|
1476 |
self.num_heads = config.num_attention_heads
|
1477 |
+
self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
|
|
|
|
|
1478 |
self.num_key_value_heads = config.num_key_value_heads
|
1479 |
+
self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
|
|
|
|
|
1480 |
self.is_causal = True
|
1481 |
self.attention_dropout = config.attention_dropout
|
1482 |
self.rope_scaling = config.rope_scaling
|
1483 |
|
1484 |
self.q_proj = nn.Linear(
|
1485 |
+
config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
|
|
|
|
|
1486 |
)
|
1487 |
self.k_proj = nn.Linear(
|
1488 |
+
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
|
|
|
|
|
1489 |
)
|
1490 |
self.v_proj = nn.Linear(
|
1491 |
+
config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
|
|
|
|
|
1492 |
)
|
1493 |
self.o_proj = nn.Linear(
|
1494 |
+
config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
|
|
|
|
|
1495 |
)
|
1496 |
+
self.q_norm = Qwen3RMSNorm(self.head_dim, eps=config.rms_norm_eps) # unlike olmo, only on the head dim!
|
1497 |
+
self.k_norm = Qwen3RMSNorm(self.head_dim, eps=config.rms_norm_eps) # thus post q_norm does not need reshape
|
|
|
|
|
|
|
|
|
1498 |
|
1499 |
self.rotary_emb = KeyeRotaryEmbedding(config=config)
|
1500 |
|
|
|
1507 |
output_attentions: bool = False,
|
1508 |
use_cache: bool = False,
|
1509 |
cache_position: Optional[torch.LongTensor] = None,
|
1510 |
+
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC
|
|
|
|
|
1511 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
1512 |
bsz, q_len, _ = hidden_states.size()
|
1513 |
|
1514 |
+
query_states = self.q_norm(self.q_proj(hidden_states).view(bsz, q_len, -1, self.head_dim))
|
1515 |
+
key_states = self.k_norm(self.k_proj(hidden_states).view(bsz, q_len, -1, self.head_dim))
|
|
|
|
|
|
|
|
|
1516 |
value_states = self.v_proj(hidden_states)
|
1517 |
|
1518 |
query_states = query_states.transpose(1, 2)
|
|
|
1525 |
)
|
1526 |
|
1527 |
if past_key_value is not None:
|
1528 |
+
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
|
1529 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
|
|
|
|
|
|
|
|
|
|
|
|
1530 |
|
1531 |
# repeat k/v heads if n_kv_heads < n_heads
|
1532 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
1533 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
1534 |
|
1535 |
+
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
1536 |
+
|
|
|
1537 |
|
1538 |
if attention_mask is not None: # no matter the length, we just slice it
|
1539 |
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
|
|
1542 |
# Fix precision issues in float16 inference
|
1543 |
# Replace inf values with zeros in attention weights to prevent NaN propagation
|
1544 |
if query_states.dtype == torch.float16:
|
1545 |
+
attn_weights = torch.where(torch.isinf(attn_weights), torch.zeros_like(attn_weights), attn_weights)
|
|
|
|
|
1546 |
|
1547 |
# upcast attention to fp32
|
1548 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
1549 |
+
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
|
|
|
|
|
|
|
|
|
1550 |
attn_output = torch.matmul(attn_weights, value_states)
|
1551 |
|
1552 |
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
|
|
1592 |
output_attentions: bool = False,
|
1593 |
use_cache: bool = False,
|
1594 |
cache_position: Optional[torch.LongTensor] = None,
|
1595 |
+
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC
|
|
|
|
|
1596 |
cu_seqlens: Optional[torch.Tensor] = None,
|
1597 |
+
sliding_window = -1,
|
1598 |
**kwargs,
|
1599 |
):
|
1600 |
bsz, q_len, _ = hidden_states.size()
|
1601 |
+
q= self.q_proj(hidden_states).view(bsz, q_len, -1, self.head_dim)
|
1602 |
query_states = self.q_norm(q)
|
1603 |
+
key_states = self.k_norm(self.k_proj(hidden_states).view(bsz, q_len, -1, self.head_dim))
|
|
|
|
|
1604 |
value_states = self.v_proj(hidden_states)
|
1605 |
|
1606 |
query_states = query_states.transpose(1, 2)
|
|
|
1614 |
)
|
1615 |
|
1616 |
if past_key_value is not None:
|
1617 |
+
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
|
1618 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
|
|
|
|
|
|
|
|
|
|
|
|
1619 |
|
1620 |
# repeat k/v heads if n_kv_heads < n_heads
|
1621 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
1622 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
1623 |
dropout_rate = 0.0 if not self.training else self.attention_dropout
|
1624 |
+
|
1625 |
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
|
1626 |
# therefore the input hidden states gets silently casted in float32. Hence, we need
|
1627 |
# cast them back in float16 just to be sure everything works as expected.
|
|
|
1675 |
max_seqlen,
|
1676 |
dropout_p=dropout_rate,
|
1677 |
window_size=(sliding_window, sliding_window),
|
1678 |
+
causal=self.is_causal
|
1679 |
)
|
1680 |
else:
|
1681 |
attn_output = _flash_attention_forward(
|
|
|
1715 |
output_attentions: bool = False,
|
1716 |
use_cache: bool = False,
|
1717 |
cache_position: Optional[torch.LongTensor] = None,
|
1718 |
+
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC
|
|
|
|
|
1719 |
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
1720 |
if output_attentions:
|
1721 |
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
|
|
|
1736 |
|
1737 |
bsz, q_len, _ = hidden_states.size()
|
1738 |
|
1739 |
+
query_states = self.q_norm(self.q_proj(hidden_states).view(bsz, q_len, -1, self.head_dim))
|
1740 |
+
key_states = self.k_norm(self.k_proj(hidden_states).view(bsz, q_len, -1, self.head_dim))
|
|
|
|
|
|
|
|
|
1741 |
value_states = self.v_proj(hidden_states)
|
1742 |
|
1743 |
query_states = query_states.transpose(1, 2)
|
|
|
1750 |
)
|
1751 |
|
1752 |
if past_key_value is not None:
|
1753 |
+
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
|
1754 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
|
|
|
|
|
|
|
|
|
|
|
|
1755 |
|
1756 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
1757 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
|
|
1789 |
return attn_output, None, past_key_value
|
1790 |
|
1791 |
|
1792 |
+
|
1793 |
QWEN3_ATTENTION_CLASSES = {
|
1794 |
"eager": KeyeAttention,
|
1795 |
"flash_attention_2": KeyeFlashAttention2,
|
|
|
1801 |
def __init__(self, config: KeyeConfig, layer_idx: int):
|
1802 |
super().__init__()
|
1803 |
self.hidden_size = config.hidden_size
|
1804 |
+
|
1805 |
+
if config.use_sliding_window and config._attn_implementation != "flash_attention_2":
|
|
|
|
|
|
|
1806 |
logger.warning_once(
|
1807 |
f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
|
1808 |
"unexpected results may be encountered."
|
1809 |
)
|
1810 |
|
1811 |
+
self.self_attn = QWEN3_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
|
|
|
|
|
1812 |
self.mlp = Qwen3MLP(config)
|
1813 |
self.input_layernorm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
1814 |
+
self.post_attention_layernorm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
|
|
|
1815 |
|
1816 |
def forward(
|
1817 |
self,
|
|
|
1822 |
output_attentions: Optional[bool] = False,
|
1823 |
use_cache: Optional[bool] = False,
|
1824 |
cache_position: Optional[torch.LongTensor] = None,
|
1825 |
+
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC
|
|
|
|
|
1826 |
**kwargs,
|
1827 |
+
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
|
|
|
|
1828 |
"""
|
1829 |
Args:
|
1830 |
hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
|
|
|
1860 |
use_cache=use_cache,
|
1861 |
cache_position=cache_position,
|
1862 |
position_embeddings=position_embeddings,
|
1863 |
+
**kwargs
|
1864 |
)
|
1865 |
|
1866 |
hidden_states = residual + hidden_states
|
|
|
1876 |
if output_attentions:
|
1877 |
outputs += (self_attn_weights,)
|
1878 |
|
1879 |
+
|
1880 |
if use_cache:
|
1881 |
outputs += (present_key_value,)
|
1882 |
|
|
|
1893 |
self.padding_idx = config.pad_token_id
|
1894 |
self.vocab_size = config.vocab_size
|
1895 |
|
1896 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
|
|
|
|
1897 |
self.layers = nn.ModuleList(
|
1898 |
+
[KeyeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
|
|
|
|
|
|
1899 |
)
|
1900 |
self._attn_implementation = config._attn_implementation
|
1901 |
self.norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
|
|
1923 |
output_hidden_states: Optional[bool] = None,
|
1924 |
return_dict: Optional[bool] = None,
|
1925 |
cache_position: Optional[torch.LongTensor] = None,
|
1926 |
+
**kwargs
|
1927 |
) -> Union[Tuple, BaseModelOutputWithPast]:
|
1928 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
|
|
|
|
|
|
|
|
1929 |
output_hidden_states = (
|
1930 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
|
|
|
|
1931 |
)
|
1932 |
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
1933 |
|
1934 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
|
|
|
|
1935 |
|
1936 |
if (input_ids is None) ^ (inputs_embeds is not None):
|
1937 |
+
raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
|
|
|
|
|
1938 |
|
1939 |
if self.gradient_checkpointing and self.training:
|
1940 |
if use_cache:
|
|
|
1951 |
inputs_embeds = self.embed_tokens(input_ids)
|
1952 |
|
1953 |
if cache_position is None:
|
1954 |
+
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
|
|
|
|
1955 |
cache_position = torch.arange(
|
1956 |
+
past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
|
|
|
|
|
1957 |
)
|
1958 |
|
1959 |
# the hard coded `3` is for temporal, height and width.
|
1960 |
if position_ids is None:
|
1961 |
+
position_ids = cache_position.view(1, 1, -1).expand(3, inputs_embeds.shape[0], -1)
|
|
|
|
|
1962 |
elif position_ids.dim() == 2:
|
1963 |
position_ids = position_ids[None, ...].expand(3, position_ids.shape[0], -1)
|
1964 |
|
1965 |
causal_mask = self._update_causal_mask(
|
1966 |
+
attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
|
|
|
|
|
|
|
|
|
1967 |
)
|
1968 |
hidden_states = inputs_embeds
|
1969 |
|
|
|
2023 |
next_cache = next_decoder_cache if use_cache else None
|
2024 |
|
2025 |
if not return_dict:
|
2026 |
+
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
|
|
|
|
|
|
|
|
|
2027 |
return BaseModelOutputWithPast(
|
2028 |
last_hidden_state=hidden_states,
|
2029 |
past_key_values=next_cache,
|
|
|
2041 |
):
|
2042 |
if self.config._attn_implementation == "flash_attention_2":
|
2043 |
if attention_mask is not None and past_key_values is not None:
|
2044 |
+
is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0]
|
|
|
|
|
2045 |
if is_padding_right:
|
2046 |
raise ValueError(
|
2047 |
"You are attempting to perform batched generation with padding_side='right'"
|
|
|
2055 |
# For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
|
2056 |
# order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
|
2057 |
# to infer the attention mask.
|
2058 |
+
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
|
|
|
|
2059 |
using_static_cache = isinstance(past_key_values, StaticCache)
|
2060 |
using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)
|
2061 |
|
|
|
2110 |
# Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
|
2111 |
# using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
|
2112 |
# Details: https://github.com/pytorch/pytorch/issues/110213
|
2113 |
+
causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
|
|
|
|
|
2114 |
|
2115 |
return causal_mask
|
2116 |
|
|
|
2156 |
else:
|
2157 |
min_dtype = torch.finfo(dtype).min
|
2158 |
causal_mask = torch.full(
|
2159 |
+
(sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
|
|
|
|
|
|
|
2160 |
)
|
2161 |
+
diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
|
|
|
|
|
2162 |
if config.sliding_window is not None:
|
2163 |
# if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
|
2164 |
# the check is needed to verify is current checkpoint was trained with sliding window or not
|
2165 |
+
if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length:
|
2166 |
+
sliding_attend_mask = torch.arange(target_length, device=device) <= (
|
2167 |
+
cache_position.reshape(-1, 1) - config.sliding_window
|
2168 |
+
)
|
|
|
|
|
|
|
2169 |
diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
|
2170 |
causal_mask *= diagonal_attend_mask
|
2171 |
causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
|
2172 |
if attention_mask is not None:
|
2173 |
+
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
|
|
|
|
|
2174 |
if attention_mask.shape[-1] > target_length:
|
2175 |
attention_mask = attention_mask[:, :target_length]
|
2176 |
mask_length = attention_mask.shape[-1]
|
2177 |
+
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
|
2178 |
+
causal_mask.device
|
2179 |
+
)
|
2180 |
padding_mask = padding_mask == 0
|
2181 |
+
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
|
2182 |
+
padding_mask, min_dtype
|
2183 |
+
)
|
2184 |
return causal_mask
|
2185 |
|
2186 |
|
|
|
2240 |
# Initialize weights and apply final processing
|
2241 |
self.post_init()
|
2242 |
|
2243 |
+
|
2244 |
def get_input_embeddings(self):
|
2245 |
return self.model.embed_tokens
|
2246 |
|
|
|
2325 |
video_token_id = self.config.video_token_id
|
2326 |
vision_start_token_id = self.config.vision_start_token_id
|
2327 |
mrope_position_deltas = []
|
2328 |
+
if input_ids is not None and (image_grid_thw is not None or video_grid_thw is not None):
|
|
|
|
|
2329 |
total_input_ids = input_ids
|
2330 |
if attention_mask is None:
|
2331 |
attention_mask = torch.ones_like(total_input_ids)
|
|
|
2341 |
for i, input_ids in enumerate(total_input_ids):
|
2342 |
input_ids = input_ids[attention_mask[i] == 1]
|
2343 |
image_nums, video_nums = 0, 0
|
2344 |
+
vision_start_indices = torch.argwhere(input_ids == vision_start_token_id).squeeze(1)
|
|
|
|
|
2345 |
vision_tokens = input_ids[vision_start_indices + 1]
|
2346 |
image_nums = (vision_tokens == image_token_id).sum()
|
2347 |
video_nums = (vision_tokens == video_token_id).sum()
|
|
|
2389 |
)
|
2390 |
text_len = ed - st
|
2391 |
|
2392 |
+
st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
|
2393 |
+
llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)
|
|
|
|
|
|
|
|
|
|
|
|
|
2394 |
|
2395 |
+
if torch.is_tensor(second_per_grid_t): second_per_grid_t = second_per_grid_t.detach().item()
|
|
|
2396 |
range_tensor = torch.arange(llm_grid_t).view(-1, 1)
|
2397 |
expanded_range = range_tensor.expand(-1, llm_grid_h * llm_grid_w)
|
2398 |
|
2399 |
+
time_tensor = expanded_range * second_per_grid_t * self.config.vision_config.tokens_per_second
|
|
|
|
|
|
|
|
|
2400 |
|
2401 |
time_tensor_long = time_tensor.long()
|
2402 |
t_index = time_tensor_long.flatten()
|
2403 |
|
2404 |
+
h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten()
|
2405 |
+
w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten()
|
2406 |
+
llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + text_len + st_idx)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2407 |
st = ed + llm_grid_t * llm_grid_h * llm_grid_w
|
2408 |
|
2409 |
if st < len(input_tokens):
|
2410 |
+
st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
|
|
|
|
|
|
|
|
|
2411 |
text_len = len(input_tokens) - st
|
2412 |
+
llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)
|
|
|
|
|
2413 |
|
2414 |
llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
|
2415 |
+
position_ids[..., i, attention_mask[i] == 1] = llm_positions.to(position_ids.device)
|
2416 |
+
mrope_position_deltas.append(llm_positions.max() + 1 - len(total_input_ids[i]))
|
2417 |
+
mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1)
|
|
|
|
|
|
|
|
|
|
|
|
|
2418 |
return position_ids, mrope_position_deltas
|
2419 |
else:
|
2420 |
if attention_mask is not None:
|
2421 |
position_ids = attention_mask.long().cumsum(-1) - 1
|
2422 |
position_ids.masked_fill_(attention_mask == 0, 1)
|
2423 |
+
position_ids = position_ids.unsqueeze(0).expand(3, -1, -1).to(attention_mask.device)
|
2424 |
+
max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0]
|
|
|
|
|
|
|
|
|
|
|
|
|
2425 |
mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
|
2426 |
else:
|
2427 |
position_ids = (
|
|
|
2437 |
|
2438 |
return position_ids, mrope_position_deltas
|
2439 |
|
2440 |
+
@replace_return_docstrings(output_type=KeyeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
|
|
|
|
2441 |
def forward(
|
2442 |
self,
|
2443 |
input_ids: torch.LongTensor = None,
|
|
|
2457 |
rope_deltas: Optional[torch.LongTensor] = None,
|
2458 |
cache_position: Optional[torch.LongTensor] = None,
|
2459 |
second_per_grid_ts: Optional[torch.Tensor] = None,
|
2460 |
+
**kwargs
|
2461 |
) -> Union[Tuple, KeyeCausalLMOutputWithPast]:
|
2462 |
r"""
|
2463 |
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
|
|
2498 |
"The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
|
2499 |
```"""
|
2500 |
|
2501 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
|
|
|
|
|
|
|
|
2502 |
output_hidden_states = (
|
2503 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
|
|
|
|
|
|
|
|
|
|
2504 |
)
|
2505 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
2506 |
|
2507 |
if inputs_embeds is None:
|
2508 |
inputs_embeds = self.model.embed_tokens(input_ids)
|
|
|
2521 |
image_grid_hws.append(thw_tuple)
|
2522 |
image_position_ids = torch.arange(numel) % np.prod(thw_tuple[1:])
|
2523 |
siglip_position_ids.append(image_position_ids)
|
2524 |
+
sample_indices.append(torch.full((numel, ), idx, dtype=torch.int64))
|
2525 |
cu_seqlens.append(cu_seqlens[-1] + numel)
|
2526 |
+
|
2527 |
+
siglip_position_ids = torch.concat(siglip_position_ids, dim=0).to(pixel_values.device)
|
2528 |
+
cu_seqlens = torch.tensor(cu_seqlens, dtype=torch.int32).to(pixel_values.device)
|
2529 |
+
sample_indices = torch.concat(sample_indices, dim=0).to(pixel_values.device)
|
|
|
|
|
|
|
|
|
|
|
|
|
2530 |
|
2531 |
vision_outputs = self.visual(
|
2532 |
+
pixel_values=pixel_values,
|
2533 |
image_grid_thw=image_grid_hws,
|
2534 |
position_ids=siglip_position_ids,
|
2535 |
vision_return_embed_list=True,
|
|
|
2538 |
cu_seqlens=cu_seqlens,
|
2539 |
return_pooler_output=False,
|
2540 |
use_rope=True,
|
2541 |
+
window_size =-1,
|
2542 |
)
|
2543 |
image_embeds = vision_outputs.last_hidden_state
|
2544 |
|
2545 |
image_embeds = self.mlp_AR(image_embeds, image_grid_thw)
|
2546 |
+
|
2547 |
n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
|
2548 |
+
#image_embeds is a list of tensor, each tensor is a image feature,I want to concat them all into a tensor
|
2549 |
+
image_embeds = torch.cat(image_embeds,dim=0)
|
2550 |
n_image_features = image_embeds.shape[0]
|
2551 |
if n_image_tokens != n_image_features:
|
2552 |
raise ValueError(
|
2553 |
f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
|
2554 |
)
|
2555 |
|
2556 |
+
mask = (input_ids == self.config.image_token_id)
|
2557 |
mask_unsqueezed = mask.unsqueeze(-1)
|
2558 |
mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
|
2559 |
image_mask = mask_expanded.to(inputs_embeds.device)
|
2560 |
|
2561 |
+
image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
|
|
|
|
|
2562 |
|
2563 |
inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
|
2564 |
|
|
|
2577 |
video_grid_hws.append(thw_tuple)
|
2578 |
video_position_ids = torch.arange(numel) % np.prod(thw_tuple[1:])
|
2579 |
siglip_position_ids.append(video_position_ids)
|
2580 |
+
sample_indices.append(torch.full((numel, ), idx, dtype=torch.int64))
|
2581 |
cu_seqlens.append(cu_seqlens[-1] + numel)
|
2582 |
+
siglip_position_ids = torch.concat(siglip_position_ids, dim=0).to(pixel_values_videos.device)
|
2583 |
+
cu_seqlens = torch.tensor(cu_seqlens, dtype=torch.int32).to(pixel_values_videos.device)
|
2584 |
+
sample_indices = torch.concat(sample_indices, dim=0).to(pixel_values_videos.device)
|
|
|
|
|
|
|
|
|
|
|
|
|
2585 |
|
2586 |
vision_outputs = self.visual(
|
2587 |
+
pixel_values=pixel_values_videos,
|
2588 |
image_grid_thw=video_grid_hws,
|
2589 |
position_ids=siglip_position_ids,
|
2590 |
vision_return_embed_list=True,
|
|
|
2593 |
cu_seqlens=cu_seqlens,
|
2594 |
return_pooler_output=False,
|
2595 |
use_rope=True,
|
2596 |
+
window_size = -1,
|
2597 |
)
|
2598 |
video_embeds = vision_outputs.last_hidden_state
|
2599 |
video_embeds = self.mlp_AR(video_embeds, video_grid_thw)
|
2600 |
n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
|
2601 |
+
video_embeds = torch.cat(video_embeds,dim=0)
|
2602 |
n_video_features = video_embeds.shape[0]
|
2603 |
if n_video_tokens != n_video_features:
|
2604 |
raise ValueError(
|
|
|
2610 |
mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
|
2611 |
video_mask = mask_expanded.to(inputs_embeds.device)
|
2612 |
|
2613 |
+
video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
|
|
|
|
|
2614 |
inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
|
2615 |
|
2616 |
if attention_mask is not None:
|
2617 |
attention_mask = attention_mask.to(inputs_embeds.device)
|
2618 |
|
2619 |
# if we get 4D attention mask we cannot calculate rope deltas anymore. TODO @raushan fixme
|
2620 |
+
if position_ids is None and (attention_mask is None or attention_mask.ndim == 2):
|
|
|
|
|
2621 |
# calculate RoPE index once per generation in the pre-fill stage only
|
2622 |
if (
|
2623 |
(cache_position is not None and cache_position[0] == 0)
|
|
|
2658 |
output_hidden_states=output_hidden_states,
|
2659 |
return_dict=return_dict,
|
2660 |
cache_position=cache_position,
|
2661 |
+
**kwargs
|
2662 |
)
|
2663 |
|
2664 |
hidden_states = outputs[0]
|
|
|
2778 |
if expand_size == 1:
|
2779 |
return input_ids, model_kwargs
|
2780 |
|
2781 |
+
visual_keys = ["pixel_values", "image_grid_thw", "pixel_values_videos", "video_grid_thw", "second_per_grid_ts"]
|
|
|
|
|
|
|
|
|
|
|
|
|
2782 |
|
2783 |
def _expand_dict_for_generation_visual(dict_to_expand):
|
2784 |
image_grid_thw = model_kwargs.get("image_grid_thw", None)
|
|
|
2788 |
def _repeat_interleave_samples(x, lengths, repeat_times):
|
2789 |
samples = torch.split(x, lengths)
|
2790 |
repeat_args = [repeat_times] + [1] * (x.dim() - 1)
|
2791 |
+
result = torch.cat([sample.repeat(*repeat_args) for sample in samples], dim=0)
|
|
|
|
|
2792 |
return result
|
2793 |
|
2794 |
for key in dict_to_expand:
|
|
|
2824 |
)
|
2825 |
tensor = torch.tensor(dict_to_expand[key])
|
2826 |
lengths = list(video_nums)
|
2827 |
+
tensor = _repeat_interleave_samples(tensor, lengths=lengths, repeat_times=expand_size)
|
|
|
|
|
2828 |
dict_to_expand[key] = tensor.tolist()
|
2829 |
return dict_to_expand
|
2830 |
|
|
|
2836 |
and isinstance(dict_to_expand[key], torch.Tensor)
|
2837 |
and key not in visual_keys
|
2838 |
):
|
2839 |
+
dict_to_expand[key] = dict_to_expand[key].repeat_interleave(expand_size, dim=0)
|
|
|
|
|
2840 |
return dict_to_expand
|
2841 |
|
2842 |
# input_ids is required for expanding visual inputs
|
|
|
2851 |
|
2852 |
if is_encoder_decoder:
|
2853 |
if model_kwargs.get("encoder_outputs") is None:
|
2854 |
+
raise ValueError("If `is_encoder_decoder` is True, make sure that `encoder_outputs` is defined.")
|
2855 |
+
model_kwargs["encoder_outputs"] = _expand_dict_for_generation(model_kwargs["encoder_outputs"])
|
|
|
|
|
|
|
|
|
2856 |
|
2857 |
return input_ids, model_kwargs
|
2858 |
+
|
2859 |
+
|
2860 |
+
|
2861 |
+
|
2862 |
+
|
2863 |
+
|
2864 |
+
|
2865 |
+
|