g-h-chen commited on
Commit
e2a8201
1 Parent(s): 8f83af3

upload llava_arch.py

Browse files
Files changed (1) hide show
  1. llava_arch.py +531 -0
llava_arch.py ADDED
@@ -0,0 +1,531 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 Haotian Liu
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+
16
+ from abc import ABC, abstractmethod
17
+
18
+ import torch
19
+ import torch.nn as nn
20
+
21
+ # from .multimodal_encoder.builder import build_vision_tower
22
+ # from .multimodal_projector.builder import build_vision_projector
23
+
24
+ # from .builders import build_vision_tower, build_vision_projector
25
+ # from .constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
26
+ import pdb
27
+
28
+
29
+ #############################################################################
30
+ # builders
31
+ #############################################################################
32
+
33
+ ###################################################################
34
+
35
+ import torch
36
+ import torch.nn as nn
37
+ import re
38
+
39
+
40
+ class IdentityMap(nn.Module):
41
+ def __init__(self):
42
+ super().__init__()
43
+
44
+ def forward(self, x, *args, **kwargs):
45
+ return x
46
+
47
+ @property
48
+ def config(self):
49
+ return {"mm_projector_type": 'identity'}
50
+
51
+
52
+ class SimpleResBlock(nn.Module):
53
+ def __init__(self, channels):
54
+ super().__init__()
55
+ self.pre_norm = nn.LayerNorm(channels)
56
+
57
+ self.proj = nn.Sequential(
58
+ nn.Linear(channels, channels),
59
+ nn.GELU(),
60
+ nn.Linear(channels, channels)
61
+ )
62
+ def forward(self, x):
63
+ x = self.pre_norm(x)
64
+ return x + self.proj(x)
65
+
66
+
67
+ def build_vision_projector(config, delay_load=False, **kwargs):
68
+ projector_type = getattr(config, 'mm_projector_type', 'linear')
69
+
70
+ if projector_type == 'linear':
71
+ return nn.Linear(config.mm_hidden_size, config.hidden_size)
72
+
73
+ mlp_gelu_match = re.match(r'^mlp(\d+)x_gelu$', projector_type)
74
+ if mlp_gelu_match:
75
+ mlp_depth = int(mlp_gelu_match.group(1))
76
+ modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
77
+ for _ in range(1, mlp_depth):
78
+ modules.append(nn.GELU())
79
+ modules.append(nn.Linear(config.hidden_size, config.hidden_size))
80
+ return nn.Sequential(*modules)
81
+
82
+ if projector_type == 'identity':
83
+ return IdentityMap()
84
+
85
+ raise ValueError(f'Unknown projector type: {projector_type}')
86
+ ###################################################################
87
+
88
+
89
+
90
+ ###################################################################
91
+
92
+ import os
93
+ from transformers import CLIPVisionModel, CLIPImageProcessor, CLIPVisionConfig
94
+ from transformers import AutoModel
95
+
96
+
97
+ class CLIPVisionTower(nn.Module):
98
+ def __init__(self, vision_tower, args, delay_load=False):
99
+ super().__init__()
100
+
101
+ self.is_loaded = False
102
+
103
+ self.vision_tower_name = vision_tower
104
+ self.select_layer = args.mm_vision_select_layer
105
+ self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch')
106
+
107
+ if not delay_load:
108
+ self.load_model()
109
+ else:
110
+ self.cfg_only = CLIPVisionConfig.from_pretrained(self.vision_tower_name)
111
+
112
+ def load_model(self):
113
+ print(f'loading vision model from {self.vision_tower_name}')
114
+ self.image_processor = CLIPImageProcessor.from_pretrained(self.vision_tower_name)
115
+ if 'clip' in self.vision_tower_name.lower():
116
+ self.vision_tower = CLIPVisionModel.from_pretrained(self.vision_tower_name)
117
+
118
+ elif 'internvit' in self.vision_tower_name.lower():
119
+ self.vision_tower = AutoModel.from_pretrained(self.vision_tower_name, trust_remote_code=True)
120
+ else:
121
+ raise ValueError(f'Please implement the loading of vision encoder here')
122
+
123
+ self.vision_tower.requires_grad_(False)
124
+
125
+ self.is_loaded = True
126
+
127
+ def feature_select(self, image_forward_outs):
128
+ image_features = image_forward_outs.hidden_states[self.select_layer]
129
+ if self.select_feature == 'patch':
130
+ image_features = image_features[:, 1:]
131
+ elif self.select_feature == 'cls_patch':
132
+ image_features = image_features
133
+ else:
134
+ raise ValueError(f'Unexpected select feature: {self.select_feature}')
135
+ return image_features
136
+
137
+ @torch.no_grad()
138
+ def forward(self, images):
139
+ if type(images) is list:
140
+ image_features = []
141
+ for image in images:
142
+ image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
143
+ image_feature = self.feature_select(image_forward_out).to(image.dtype)
144
+ image_features.append(image_feature)
145
+ else:
146
+ image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
147
+ image_features = self.feature_select(image_forward_outs).to(images.dtype)
148
+
149
+ return image_features
150
+
151
+ @property
152
+ def dummy_feature(self):
153
+ return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
154
+
155
+ @property
156
+ def dtype(self):
157
+ return self.vision_tower.dtype
158
+
159
+ @property
160
+ def device(self):
161
+ return self.vision_tower.device
162
+
163
+ @property
164
+ def config(self):
165
+ if self.is_loaded:
166
+ return self.vision_tower.config
167
+ else:
168
+ return self.cfg_only
169
+
170
+ @property
171
+ def hidden_size(self):
172
+ return self.config.hidden_size
173
+
174
+ @property
175
+ def num_patches(self):
176
+ return (self.config.image_size // self.config.patch_size) ** 2
177
+
178
+
179
+
180
+ def build_vision_tower(vision_tower_cfg, **kwargs):
181
+ vision_tower = getattr(vision_tower_cfg, 'mm_vision_tower', getattr(vision_tower_cfg, 'vision_tower', None))
182
+ is_absolute_path_exists = os.path.exists(vision_tower)
183
+ if is_absolute_path_exists or vision_tower.startswith("openai") or vision_tower.startswith("laion"):
184
+ return CLIPVisionTower(vision_tower, args=vision_tower_cfg, **kwargs)
185
+
186
+ raise ValueError(f'Unknown vision tower: {vision_tower}')
187
+
188
+
189
+ #############################################################################
190
+ # builders
191
+ #############################################################################
192
+
193
+
194
+
195
+ #############################################################################
196
+ # constants
197
+ #############################################################################
198
+
199
+ CONTROLLER_HEART_BEAT_EXPIRATION = 30
200
+ WORKER_HEART_BEAT_INTERVAL = 15
201
+
202
+ LOGDIR = "."
203
+
204
+ # Model Constants
205
+ IGNORE_INDEX = -100
206
+ IMAGE_TOKEN_INDEX = -200
207
+ DEFAULT_IMAGE_TOKEN = "<image>"
208
+ DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
209
+ DEFAULT_IM_START_TOKEN = "<im_start>"
210
+ DEFAULT_IM_END_TOKEN = "<im_end>"
211
+ IMAGE_PLACEHOLDER = "<image-placeholder>"
212
+
213
+ #############################################################################
214
+ # constants
215
+ #############################################################################
216
+
217
+
218
+ class LlavaMetaModel:
219
+
220
+ def __init__(self, config):
221
+ super(LlavaMetaModel, self).__init__(config)
222
+
223
+ if hasattr(config, "mm_vision_tower"):
224
+ self.vision_tower = build_vision_tower(config, delay_load=True)
225
+ self.mm_projector = build_vision_projector(config)
226
+
227
+ def get_vision_tower(self):
228
+ vision_tower = getattr(self, 'vision_tower', None)
229
+ if type(vision_tower) is list:
230
+ vision_tower = vision_tower[0]
231
+ return vision_tower
232
+
233
+ def initialize_vision_modules(self, model_args, fsdp=None):
234
+ vision_tower = model_args.vision_tower
235
+ mm_vision_select_layer = model_args.mm_vision_select_layer
236
+ mm_vision_select_feature = model_args.mm_vision_select_feature
237
+ pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter
238
+
239
+ self.config.mm_vision_tower = vision_tower
240
+
241
+ if self.get_vision_tower() is None:
242
+ vision_tower = build_vision_tower(model_args)
243
+
244
+ if fsdp is not None and len(fsdp) > 0:
245
+ self.vision_tower = [vision_tower]
246
+ else:
247
+ self.vision_tower = vision_tower
248
+ else:
249
+ if fsdp is not None and len(fsdp) > 0:
250
+ vision_tower = self.vision_tower[0]
251
+ else:
252
+ vision_tower = self.vision_tower
253
+ vision_tower.load_model()
254
+
255
+ self.config.use_mm_proj = True
256
+ self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
257
+ self.config.mm_hidden_size = vision_tower.hidden_size
258
+ self.config.mm_vision_select_layer = mm_vision_select_layer
259
+ self.config.mm_vision_select_feature = mm_vision_select_feature
260
+
261
+ if getattr(self, 'mm_projector', None) is None:
262
+ self.mm_projector = build_vision_projector(self.config)
263
+ else:
264
+ # In case it is frozen by LoRA
265
+ for p in self.mm_projector.parameters():
266
+ p.requires_grad = True
267
+
268
+ if pretrain_mm_mlp_adapter is not None:
269
+ mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
270
+ def get_w(weights, keyword):
271
+ return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}
272
+
273
+ self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'))
274
+
275
+
276
+ class LlavaMetaForCausalLM(ABC):
277
+
278
+ @abstractmethod
279
+ def get_model(self):
280
+ pass
281
+
282
+ @abstractmethod
283
+ def get_tokenizer(self):
284
+ pass
285
+
286
+ def get_vision_tower(self):
287
+ return self.get_model().get_vision_tower()
288
+
289
+ def encode_images(self, images):
290
+ image_features = self.get_model().get_vision_tower()(images)
291
+ image_features = self.get_model().mm_projector(image_features)
292
+ return image_features
293
+
294
+ def prepare_inputs_labels_for_multimodal_new(
295
+ self, input_ids: list[torch.tensor], position_ids, attention_mask: list[torch.tensor], past_key_values, labels, images
296
+ ):
297
+ vision_tower = self.get_vision_tower()
298
+ if not self.training: # TODO: check this out!!
299
+ # pdb.set_trace()
300
+ if vision_tower is None or images is None or input_ids.shape[1] == 1:
301
+ if past_key_values is not None and vision_tower is not None and images is not None and input_ids.shape[1] == 1:
302
+
303
+ if attention_mask is None:
304
+ # only happen for qwen at inference
305
+ # raise ValueError(f'should not be here except for Qwen!')
306
+ return input_ids, None, attention_mask, past_key_values, None, labels
307
+
308
+ target_shape = past_key_values[-1][-1].shape[-2] + 1
309
+ attention_mask = torch.cat((attention_mask, torch.ones(
310
+ (attention_mask.shape[0], target_shape - attention_mask.shape[1]),
311
+ dtype=attention_mask.dtype,
312
+ device=attention_mask.device
313
+ )), dim=1)
314
+ position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
315
+ return input_ids, position_ids, attention_mask, past_key_values, None, labels
316
+
317
+
318
+ # ####################### this block must be optimized! #######################
319
+ # if type(images) is list or images.ndim == 5:
320
+ # concat_images = torch.cat([image for image in images], dim=0)
321
+ # image_features = self.encode_images(concat_images)
322
+ # split_sizes = [image.shape[0] for image in images]
323
+ # image_features = torch.split(image_features, split_sizes, dim=0)
324
+ # image_features = [x.flatten(0, 1).to(self.device) for x in image_features]
325
+ # else:
326
+ # image_features = self.encode_images(images).to(self.device)
327
+ # ####################### this block must be optimized! #######################
328
+
329
+ # ####################### optimized #######################
330
+ if getattr(self, 'cached_image_features', None) is None:
331
+ # this attribute should be cleared in bot.clear_history()
332
+ if type(images) is list or images.ndim == 5:
333
+ concat_images = torch.cat([image for image in images], dim=0)
334
+ image_features = self.encode_images(concat_images)
335
+ split_sizes = [image.shape[0] for image in images]
336
+ image_features = torch.split(image_features, split_sizes, dim=0)
337
+ image_features = [x.flatten(0, 1).to(self.device) for x in image_features]
338
+ else:
339
+ image_features = self.encode_images(images).to(self.device)
340
+ self.cached_image_features = image_features
341
+ image_features = self.cached_image_features
342
+ # ####################### optimized #######################
343
+
344
+
345
+ # TODO: image start / end is not implemented here to support pretraining.
346
+ if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
347
+ raise NotImplementedError
348
+
349
+ # Let's just add dummy tensors if they do not exist,
350
+ # it is a headache to deal with None all the time.
351
+ # But it is not ideal, and if you have a better idea,
352
+ # please open an issue / submit a PR, thanks.
353
+ _labels = labels
354
+ _position_ids = position_ids
355
+ _attention_mask = attention_mask
356
+ if attention_mask is None:
357
+ # attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
358
+ attention_mask = [torch.tensor([1]*l).to(input_ids).bool() for l in map(len, [ip for ip in input_ids])]
359
+ else:
360
+ # attention_mask = attention_mask.bool()
361
+ attention_mask = [att.bool() for att in attention_mask]
362
+
363
+ # if position_ids is None:
364
+ # position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
365
+
366
+ if labels is None:
367
+ labels = [torch.tensor([IGNORE_INDEX]*l).to(input_ids) for l in map(len, [ip for ip in input_ids])]
368
+ # labels = torch.full_like(input_ids, IGNORE_INDEX)
369
+ else:
370
+ labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]
371
+ # remove the padding using attention_mask -- TODO: double check
372
+ # pdb.set_trace()
373
+ input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
374
+
375
+ new_input_embeds = []
376
+ new_labels = []
377
+ cur_image_idx = 0
378
+ for batch_idx, cur_input_ids in enumerate(input_ids):
379
+ num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
380
+ if num_images == 0:
381
+
382
+ ############### FIXME ###############
383
+ if cur_image_idx > len(image_features)-1:
384
+ cur_image_idx = len(image_features)-1
385
+ print(f'warning: {input_ids}')
386
+ ############### FIXME ###############
387
+
388
+ cur_image_features = image_features[cur_image_idx]
389
+ cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
390
+ cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
391
+ new_input_embeds.append(cur_input_embeds)
392
+ new_labels.append(labels[batch_idx])
393
+ cur_image_idx += 1
394
+ continue
395
+
396
+ image_token_indices = [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
397
+ cur_input_ids_noim = []
398
+ cur_labels = labels[batch_idx]
399
+ cur_labels_noim = []
400
+ for i in range(len(image_token_indices) - 1):
401
+ cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
402
+ cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
403
+ split_sizes = [x.shape[0] for x in cur_labels_noim]
404
+ cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
405
+ cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
406
+ cur_new_input_embeds = []
407
+ cur_new_labels = []
408
+
409
+ # you have 10 images, but you have 11 placeholders
410
+ for i in range(num_images + 1):
411
+ cur_new_input_embeds.append(cur_input_embeds_no_im[i])
412
+ cur_new_labels.append(cur_labels_noim[i])
413
+ if i < num_images:
414
+ ############### FIXME ###############
415
+ if cur_image_idx > len(image_features)-1:
416
+ cur_image_idx = len(image_features)-1
417
+ print(f'warning: {input_ids}')
418
+ ############### FIXME ###############
419
+
420
+ cur_image_features = image_features[cur_image_idx]
421
+ cur_image_idx += 1
422
+ cur_new_input_embeds.append(cur_image_features)
423
+ cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
424
+
425
+ cur_new_input_embeds = torch.cat(cur_new_input_embeds)
426
+ cur_new_labels = torch.cat(cur_new_labels)
427
+
428
+ new_input_embeds.append(cur_new_input_embeds)
429
+ new_labels.append(cur_new_labels)
430
+
431
+ # Truncate sequences to max length as image embeddings can make the sequence longer
432
+ tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
433
+ if tokenizer_model_max_length is not None:
434
+ new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
435
+ new_labels = [x[:tokenizer_model_max_length] for x in new_labels]
436
+
437
+ # Combine them
438
+ max_len = max(x.shape[0] for x in new_input_embeds)
439
+ batch_size = len(new_input_embeds)
440
+
441
+ new_input_embeds_padded = []
442
+ new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
443
+ attention_mask = torch.zeros((batch_size, max_len), dtype=torch.bool, device=attention_mask[0].device)
444
+ position_ids = torch.zeros((batch_size, max_len), dtype=torch.long, device=attention_mask[0].device)
445
+
446
+ for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
447
+ cur_len = cur_new_embed.shape[0]
448
+ # print(f'cur_len[{i}]before padding: {cur_len}')
449
+ # if i==0:
450
+ # print(f"{getattr(self.config, 'tokenizer_padding_side', 'right')} {self.get_tokenizer().padding_side}")
451
+ if getattr(self.config, 'tokenizer_padding_side', 'right') == "left": # checked, this is correct
452
+ # if self.get_tokenizer().padding_side == 'left':
453
+ new_input_embeds_padded.append(torch.cat((
454
+ torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
455
+ cur_new_embed
456
+ ), dim=0))
457
+ if cur_len > 0:
458
+ new_labels_padded[i, -cur_len:] = cur_new_labels
459
+ attention_mask[i, -cur_len:] = True
460
+ position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
461
+ else:
462
+ new_input_embeds_padded.append(torch.cat((
463
+ cur_new_embed,
464
+ torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
465
+ ), dim=0))
466
+ if cur_len > 0:
467
+ new_labels_padded[i, :cur_len] = cur_new_labels
468
+ attention_mask[i, :cur_len] = True
469
+ position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
470
+
471
+ new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)
472
+
473
+ if _labels is None:
474
+ new_labels = None
475
+ else:
476
+ new_labels = new_labels_padded
477
+
478
+ if _attention_mask is None:
479
+ attention_mask = None
480
+ else:
481
+ # attention_mask = attention_mask.to(dtype=_attention_mask.dtype)
482
+ attention_mask = attention_mask.to(dtype=torch.bool)
483
+
484
+ if _position_ids is None:
485
+ position_ids = None
486
+
487
+ return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels
488
+
489
+ def initialize_vision_tokenizer(self, model_args, tokenizer):
490
+ if model_args.mm_use_im_patch_token:
491
+ tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
492
+ self.resize_token_embeddings(len(tokenizer))
493
+
494
+ if model_args.mm_use_im_start_end:
495
+ num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
496
+ self.resize_token_embeddings(len(tokenizer))
497
+
498
+ if num_new_tokens > 0:
499
+ input_embeddings = self.get_input_embeddings().weight.data
500
+ output_embeddings = self.get_output_embeddings().weight.data
501
+
502
+ input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
503
+ dim=0, keepdim=True)
504
+ output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
505
+ dim=0, keepdim=True)
506
+
507
+ input_embeddings[-num_new_tokens:] = input_embeddings_avg
508
+ output_embeddings[-num_new_tokens:] = output_embeddings_avg
509
+
510
+ if model_args.tune_mm_mlp_adapter:
511
+ for p in self.get_input_embeddings().parameters():
512
+ p.requires_grad = True
513
+ for p in self.get_output_embeddings().parameters():
514
+ p.requires_grad = False
515
+
516
+ if model_args.pretrain_mm_mlp_adapter:
517
+ mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
518
+ embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight']
519
+ assert num_new_tokens == 2
520
+ if input_embeddings.shape == embed_tokens_weight.shape:
521
+ input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
522
+ elif embed_tokens_weight.shape[0] == num_new_tokens:
523
+ input_embeddings[-num_new_tokens:] = embed_tokens_weight
524
+ else:
525
+ raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
526
+ elif model_args.mm_use_im_patch_token:
527
+ if model_args.tune_mm_mlp_adapter:
528
+ for p in self.get_input_embeddings().parameters():
529
+ p.requires_grad = False
530
+ for p in self.get_output_embeddings().parameters():
531
+ p.requires_grad = False