File size: 29,024 Bytes
61f3f56
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
#    Copyright 2023 Haotian Liu
#
#    Licensed under the Apache License, Version 2.0 (the "License");
#    you may not use this file except in compliance with the License.
#    You may obtain a copy of the License at
#
#        http://www.apache.org/licenses/LICENSE-2.0
#
#    Unless required by applicable law or agreed to in writing, software
#    distributed under the License is distributed on an "AS IS" BASIS,
#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#    See the License for the specific language governing permissions and
#    limitations under the License.


from abc import ABC, abstractmethod

import torch
import torch.nn as nn

from .multimodal_encoder.builder import build_image_tower, build_video_tower
from .multimodal_projector.builder import build_vision_projector

from llava.constants import IGNORE_INDEX, X_TOKEN_INDEX, DEFAULT_X_PATCH_TOKEN, DEFAULT_X_START_TOKEN, DEFAULT_X_END_TOKEN


class LlavaMetaModel:

    def __init__(self, config):
        super(LlavaMetaModel, self).__init__(config)

        if hasattr(config, "mm_image_tower"):
            self.image_tower = build_image_tower(config, delay_load=True)
            self.mm_projector = build_vision_projector(config)
        if hasattr(config, "mm_video_tower"):
            self.video_tower = build_video_tower(config, delay_load=True)
            self.mm_projector = build_vision_projector(config)

    def get_image_tower(self):
        image_tower = getattr(self, 'image_tower', None)
        if type(image_tower) is list:
            image_tower = image_tower[0]
        return image_tower

    def get_video_tower(self):
        video_tower = getattr(self, 'video_tower', None)
        if type(video_tower) is list:
            video_tower = video_tower[0]
        return video_tower

    def initialize_image_modules(self, model_args, fsdp=None):
        image_tower = model_args.image_tower
        mm_vision_select_layer = model_args.mm_vision_select_layer
        mm_vision_select_feature = model_args.mm_vision_select_feature
        pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter

        self.config.mm_image_tower = image_tower

        image_tower = build_image_tower(model_args)

        if fsdp is not None and len(fsdp) > 0:
            self.image_tower = [image_tower]
        else:
            self.image_tower = image_tower

        self.config.use_mm_proj = True
        self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
        self.config.mm_hidden_size = image_tower.hidden_size
        self.config.mm_vision_select_layer = mm_vision_select_layer
        self.config.mm_vision_select_feature = mm_vision_select_feature

        self.mm_projector = build_vision_projector(self.config)

        if pretrain_mm_mlp_adapter is not None:
            mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
            def get_w(weights, keyword):
                return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}

            self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'))

    def initialize_video_modules(self, model_args, fsdp=None):
        video_tower = model_args.video_tower
        mm_vision_select_layer = model_args.mm_vision_select_layer
        mm_vision_select_feature = model_args.mm_vision_select_feature
        pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter

        self.config.mm_video_tower = video_tower

        video_tower = build_video_tower(model_args)

        if fsdp is not None and len(fsdp) > 0:
            self.video_tower = [video_tower]
        else:
            self.video_tower = video_tower

        self.config.use_mm_proj = True
        self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
        self.config.mm_hidden_size = video_tower.hidden_size
        self.config.mm_vision_select_layer = mm_vision_select_layer
        self.config.mm_vision_select_feature = mm_vision_select_feature

        self.mm_projector = build_vision_projector(self.config)

        if pretrain_mm_mlp_adapter is not None:
            mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
            def get_w(weights, keyword):
                return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}

            self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'))

class LlavaMetaForCausalLM(ABC):

    @abstractmethod
    def get_model(self):
        pass

    def get_image_tower(self):
        return self.get_model().get_image_tower()

    def get_video_tower(self):
        return self.get_model().get_video_tower()

    def get_all_tower(self, keys):
        tower = {key: getattr(self, f'get_{key}_tower') for key in keys}
        return tower

    def encode_images(self, images):
        image_features = self.get_model().get_image_tower()(images)
        image_features = self.get_model().mm_projector(image_features)
        return image_features

    def encode_videos(self, videos):
        video_features = self.get_model().get_video_tower()(videos)
        video_features = self.get_model().mm_projector(video_features)
        return video_features
    #
    # def prepare_inputs_labels_for_multimodal(
    #     self, input_ids, attention_mask, past_key_values, labels, images
    # ):
    #     vision_tower = self.get_vision_tower()
    #     if vision_tower is None or images is None or input_ids.shape[1] == 1:
    #         if past_key_values is not None and vision_tower is not None and images is not None and input_ids.shape[1] == 1:
    #             attention_mask = torch.ones((attention_mask.shape[0], past_key_values[-1][-1].shape[-2] + 1), dtype=attention_mask.dtype, device=attention_mask.device)
    #         return input_ids, attention_mask, past_key_values, None, labels
    #
    #     if type(images) is list or images.ndim == 5:
    #         concat_images = torch.cat([image for image in images], dim=0)
    #         image_features = self.encode_images(concat_images)
    #         split_sizes = [image.shape[0] for image in images]
    #         image_features = torch.split(image_features, split_sizes, dim=0)
    #         image_features = [x.flatten(0, 1) for x in image_features]
    #     else:
    #         image_features = self.encode_images(images)
    #
    #     new_input_embeds = []
    #     new_labels = [] if labels is not None else None
    #     cur_image_idx = 0
    #     for batch_idx, cur_input_ids in enumerate(input_ids):
    #         if (cur_input_ids == IMAGE_TOKEN_INDEX).sum() == 0:
    #             # multimodal LLM, but the current sample is not multimodal
    #             # FIXME: this is a hacky fix, for deepspeed zero3 to work
    #             half_len = cur_input_ids.shape[0] // 2
    #             cur_image_features = image_features[cur_image_idx]
    #             cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids[:half_len])
    #             cur_input_embeds_2 = self.get_model().embed_tokens(cur_input_ids[half_len:])
    #             cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0], cur_input_embeds_2], dim=0)
    #             new_input_embeds.append(cur_input_embeds)
    #             if labels is not None:
    #                 new_labels.append(labels[batch_idx])
    #             cur_image_idx += 1
    #             continue
    #         image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]  # 把中间的imgtoken的位置找到
    #         cur_new_input_embeds = []
    #         if labels is not None:
    #             cur_labels = labels[batch_idx]
    #             cur_new_labels = []
    #             assert cur_labels.shape == cur_input_ids.shape
    #         while image_token_indices.numel() > 0:
    #             cur_image_features = image_features[cur_image_idx]
    #             image_token_start = image_token_indices[0]
    #             if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
    #                 cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:image_token_start-1]).detach())
    #                 cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[image_token_start-1:image_token_start]))
    #                 cur_new_input_embeds.append(cur_image_features)
    #                 cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[image_token_start+1:image_token_start+2]))
    #                 if labels is not None:
    #                     cur_new_labels.append(cur_labels[:image_token_start])
    #                     cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype))
    #                     cur_new_labels.append(cur_labels[image_token_start:image_token_start+1])
    #                     cur_labels = cur_labels[image_token_start+2:]
    #             else:
    #                 cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:image_token_start]))  # imgtoken之前的text拿出来,好像都是模板套话
    #                 cur_new_input_embeds.append(cur_image_features)
    #                 if labels is not None:
    #                     cur_new_labels.append(cur_labels[:image_token_start])
    #                     cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype))
    #                     cur_labels = cur_labels[image_token_start+1:]
    #             cur_image_idx += 1
    #             if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
    #                 cur_input_ids = cur_input_ids[image_token_start+2:]
    #             else:
    #                 cur_input_ids = cur_input_ids[image_token_start+1:]   # imgtoken之后的text拿出来,是真的question
    #             image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
    #         if cur_input_ids.numel() > 0:
    #             if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
    #                 cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids).detach())
    #             else:
    #                 cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids))
    #             if labels is not None:
    #                 cur_new_labels.append(cur_labels)
    #         cur_new_input_embeds = [x.to(device=self.device) for x in cur_new_input_embeds]  # 前面text+图片+后面question
    #         cur_new_input_embeds = torch.cat(cur_new_input_embeds, dim=0)
    #         new_input_embeds.append(cur_new_input_embeds)
    #         if labels is not None:
    #             cur_new_labels = torch.cat(cur_new_labels, dim=0)
    #             new_labels.append(cur_new_labels)
    #
    #     if any(x.shape != new_input_embeds[0].shape for x in new_input_embeds):
    #         max_len = max(x.shape[0] for x in new_input_embeds)
    #
    #         new_input_embeds_align = []
    #         for cur_new_embed in new_input_embeds:
    #             cur_new_embed = torch.cat((cur_new_embed, torch.zeros((max_len - cur_new_embed.shape[0], cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)), dim=0)
    #             new_input_embeds_align.append(cur_new_embed)
    #         new_input_embeds = torch.stack(new_input_embeds_align, dim=0)
    #
    #         if labels is not None:
    #             new_labels_align = []
    #             _new_labels = new_labels
    #             for cur_new_label in new_labels:
    #                 cur_new_label = torch.cat((cur_new_label, torch.full((max_len - cur_new_label.shape[0],), IGNORE_INDEX, dtype=cur_new_label.dtype, device=cur_new_label.device)), dim=0)
    #                 new_labels_align.append(cur_new_label)
    #             new_labels = torch.stack(new_labels_align, dim=0)
    #
    #         if attention_mask is not None:
    #             new_attention_mask = []
    #             for cur_attention_mask, cur_new_labels, cur_new_labels_align in zip(attention_mask, _new_labels, new_labels):
    #                 new_attn_mask_pad_left = torch.full((cur_new_labels.shape[0] - labels.shape[1],), True, dtype=attention_mask.dtype, device=attention_mask.device)
    #                 new_attn_mask_pad_right = torch.full((cur_new_labels_align.shape[0] - cur_new_labels.shape[0],), False, dtype=attention_mask.dtype, device=attention_mask.device)
    #                 cur_new_attention_mask = torch.cat((new_attn_mask_pad_left, cur_attention_mask, new_attn_mask_pad_right), dim=0)
    #                 new_attention_mask.append(cur_new_attention_mask)
    #             attention_mask = torch.stack(new_attention_mask, dim=0)
    #             assert attention_mask.shape == new_labels.shape
    #     else:
    #         new_input_embeds = torch.stack(new_input_embeds, dim=0)
    #         if labels is not None:
    #             new_labels  = torch.stack(new_labels, dim=0)
    #
    #         if attention_mask is not None:
    #             new_attn_mask_pad_left = torch.full((attention_mask.shape[0], new_input_embeds.shape[1] - input_ids.shape[1]), True, dtype=attention_mask.dtype, device=attention_mask.device)
    #             attention_mask = torch.cat((new_attn_mask_pad_left, attention_mask), dim=1)
    #             assert attention_mask.shape == new_input_embeds.shape[:2]
    #
    #     return None, attention_mask, past_key_values, new_input_embeds, new_labels
    #
    # def initialize_vision_tokenizer(self, model_args, tokenizer):
    #     if model_args.mm_use_im_patch_token:
    #         tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
    #         self.resize_token_embeddings(len(tokenizer))
    #
    #     if model_args.mm_use_im_start_end:
    #         num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
    #         self.resize_token_embeddings(len(tokenizer))
    #
    #         if num_new_tokens > 0:
    #             input_embeddings = self.get_input_embeddings().weight.data
    #             output_embeddings = self.get_output_embeddings().weight.data
    #
    #             input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
    #                 dim=0, keepdim=True)
    #             output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
    #                 dim=0, keepdim=True)
    #
    #             input_embeddings[-num_new_tokens:] = input_embeddings_avg
    #             output_embeddings[-num_new_tokens:] = output_embeddings_avg
    #
    #         if model_args.tune_mm_mlp_adapter:
    #             for p in self.get_input_embeddings().parameters():
    #                 p.requires_grad = True
    #             for p in self.get_output_embeddings().parameters():
    #                 p.requires_grad = False
    #
    #         if model_args.pretrain_mm_mlp_adapter:
    #             mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
    #             embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight']
    #             assert num_new_tokens == 2
    #             if input_embeddings.shape == embed_tokens_weight.shape:
    #                 input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
    #             elif embed_tokens_weight.shape[0] == num_new_tokens:
    #                 input_embeddings[-num_new_tokens:] = embed_tokens_weight
    #             else:
    #                 raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
    #     elif model_args.mm_use_im_patch_token:
    #         if model_args.tune_mm_mlp_adapter:
    #             for p in self.get_input_embeddings().parameters():
    #                 p.requires_grad = False
    #             for p in self.get_output_embeddings().parameters():
    #                 p.requires_grad = False

    def prepare_inputs_labels_for_multimodal(
        self, input_ids, attention_mask, past_key_values, labels, X_modalities
    ):
        '''
        X_modalities [
        [img_feature, img_feature, video_feature, audio_feature],
        ['image', 'image', 'video', 'audio']
        ]
        '''
        Xs, keys = X_modalities
        all_tower = self.get_all_tower(set(keys)) if len(keys) > 0 else None
        
        # print(2.5)
        if all_tower is None or X_modalities[0][0] is None or input_ids.shape[1] == 1:
            if past_key_values is not None and all_tower is not None and Xs is not None and input_ids.shape[1] == 1:
                attention_mask = torch.ones((attention_mask.shape[0], past_key_values[-1][-1].shape[-2] + 1), dtype=attention_mask.dtype, device=attention_mask.device)
            return input_ids, attention_mask, past_key_values, None, labels

        # if type(images) is list or images.ndim == 5:
        #     concat_images = torch.cat([image for image in images], dim=0)
        #     image_features = self.encode_images(concat_images)
        #     split_sizes = [image.shape[0] for image in images]
        #     image_features = torch.split(image_features, split_sizes, dim=0)
        #     image_features = [x.flatten(0, 1) for x in image_features]
        # else:
        print(keys)
        X_features = [getattr(self, f'encode_{key}s')(X.unsqueeze(0)) for X, key in zip(Xs, keys)]  # expand to get batchsize
        # X_features = []
        # # print(2.5, *[i.shape for i in Xs], keys)  
        # for X, key in zip(Xs, keys):
        #     temp_X = X.unsqueeze(0)
        #     # print(2.6)
        #     # fn = getattr(self, f'encode_{key}s') 
        #     if key == 'image':
        #         out = self.encode_images(temp_X)
        #         # print(2.65, 'image', out.shape)
        #     elif key == 'video':
        #         out = self.encode_videos(temp_X)
        #         # print(2.65, 'video', out.shape)
        #     else:
        #         raise NameError(f'{key}')
        #     # print(2.8, out.shape)
        #     X_features.append(out)
        X_features = [x.flatten(0, 1) for x in X_features]
        # print([[j, i.shape] for i, j in zip(X_features, keys)])


        new_input_embeds = []
        new_labels = [] if labels is not None else None
        cur_X_idx = 0
        # print(2.9, input_ids.shape)
        for batch_idx, cur_input_ids in enumerate(input_ids):
            # print(333333)
            if (torch.any(torch.stack([cur_input_ids == X_TOKEN_INDEX[key.upper()] for key in keys]), dim=0)).sum() == 0:
                # multimodal LLM, but the current sample is not multimodal
                # FIXME: this is a hacky fix, for deepspeed zero3 to work
                half_len = cur_input_ids.shape[0] // 2
                cur_X_features = X_features[cur_X_idx]
                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids[:half_len])
                cur_input_embeds_2 = self.get_model().embed_tokens(cur_input_ids[half_len:])
                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_X_features[0:0], cur_input_embeds_2], dim=0)
                new_input_embeds.append(cur_input_embeds)
                if labels is not None:
                    new_labels.append(labels[batch_idx])
                cur_X_idx += 1 ############## 注意这里跳过了,如果一个sample是一个modal,那么就跳过1个全zero的modal,如果一个sample对应多个modal,这里的训练逻辑不对!!!
                ###### 但似乎不影响1个sample的inference
                ###### 一个text对应视频和图片,直接走下边了。只有1个text,传入none或者1个/2个全zero都无所谓,反正没有下一个数据了。
                continue
            X_token_indices = torch.where(torch.any(torch.stack([cur_input_ids == X_TOKEN_INDEX[key.upper()] for key in keys]), dim=0))[0]  # 把中间的imgtoken的位置找到
            cur_new_input_embeds = []
            if labels is not None:
                cur_labels = labels[batch_idx]
                cur_new_labels = []
                assert cur_labels.shape == cur_input_ids.shape
            # print(4444444444)
            while X_token_indices.numel() > 0:
                cur_X_features = X_features[cur_X_idx]
                X_token_start = X_token_indices[0]
                if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_x_start_end', False):
                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:X_token_start-1]).detach())
                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[X_token_start-1:X_token_start]))
                    cur_new_input_embeds.append(cur_X_features)
                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[X_token_start+1:X_token_start+2]))
                    if labels is not None:
                        cur_new_labels.append(cur_labels[:X_token_start])
                        cur_new_labels.append(torch.full((cur_X_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype))
                        cur_new_labels.append(cur_labels[X_token_start:X_token_start+1])
                        cur_labels = cur_labels[X_token_start+2:]
                else:
                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:X_token_start]))  # imgtoken之前的text拿出来,好像都是模板套话
                    cur_new_input_embeds.append(cur_X_features)
                    if labels is not None:
                        cur_new_labels.append(cur_labels[:X_token_start])
                        cur_new_labels.append(torch.full((cur_X_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype))
                        cur_labels = cur_labels[X_token_start+1:]
                cur_X_idx += 1
                if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_x_start_end', False):
                    cur_input_ids = cur_input_ids[X_token_start+2:]
                else:
                    cur_input_ids = cur_input_ids[X_token_start+1:]   # imgtoken之后的text拿出来,是真的question
                X_token_indices = torch.where(torch.any(torch.stack([cur_input_ids == X_TOKEN_INDEX[key.upper()] for key in keys]), dim=0))[0]
            
            # print(55555555555555555)
            if cur_input_ids.numel() > 0:
                if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_x_start_end', False):
                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids).detach())
                else:
                    cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids))
                if labels is not None:
                    cur_new_labels.append(cur_labels)
            cur_new_input_embeds = [x.to(device=self.device) for x in cur_new_input_embeds]  # 前面text+图片+后面question
            cur_new_input_embeds = torch.cat(cur_new_input_embeds, dim=0)
            new_input_embeds.append(cur_new_input_embeds)
            if labels is not None:
                cur_new_labels = torch.cat(cur_new_labels, dim=0)
                new_labels.append(cur_new_labels)

        if any(x.shape != new_input_embeds[0].shape for x in new_input_embeds):
            max_len = max(x.shape[0] for x in new_input_embeds)

            new_input_embeds_align = []
            for cur_new_embed in new_input_embeds:
                cur_new_embed = torch.cat((cur_new_embed, torch.zeros((max_len - cur_new_embed.shape[0], cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)), dim=0)
                new_input_embeds_align.append(cur_new_embed)
            new_input_embeds = torch.stack(new_input_embeds_align, dim=0)

            if labels is not None:
                new_labels_align = []
                _new_labels = new_labels
                for cur_new_label in new_labels:
                    cur_new_label = torch.cat((cur_new_label, torch.full((max_len - cur_new_label.shape[0],), IGNORE_INDEX, dtype=cur_new_label.dtype, device=cur_new_label.device)), dim=0)
                    new_labels_align.append(cur_new_label)
                new_labels = torch.stack(new_labels_align, dim=0)

            if attention_mask is not None:
                new_attention_mask = []
                for cur_attention_mask, cur_new_labels, cur_new_labels_align in zip(attention_mask, _new_labels, new_labels):
                    new_attn_mask_pad_left = torch.full((cur_new_labels.shape[0] - labels.shape[1],), True, dtype=attention_mask.dtype, device=attention_mask.device)
                    new_attn_mask_pad_right = torch.full((cur_new_labels_align.shape[0] - cur_new_labels.shape[0],), False, dtype=attention_mask.dtype, device=attention_mask.device)
                    cur_new_attention_mask = torch.cat((new_attn_mask_pad_left, cur_attention_mask, new_attn_mask_pad_right), dim=0)
                    new_attention_mask.append(cur_new_attention_mask)
                attention_mask = torch.stack(new_attention_mask, dim=0)
                assert attention_mask.shape == new_labels.shape
        else:
            new_input_embeds = torch.stack(new_input_embeds, dim=0)
            if labels is not None:
                new_labels  = torch.stack(new_labels, dim=0)

            if attention_mask is not None:
                new_attn_mask_pad_left = torch.full((attention_mask.shape[0], new_input_embeds.shape[1] - input_ids.shape[1]), True, dtype=attention_mask.dtype, device=attention_mask.device)
                attention_mask = torch.cat((new_attn_mask_pad_left, attention_mask), dim=1)
                assert attention_mask.shape == new_input_embeds.shape[:2]

        return None, attention_mask, past_key_values, new_input_embeds, new_labels

    def initialize_X_tokenizer(self, model_args, tokenizer):
        if model_args.mm_use_x_patch_token:
            for x in model_args.X:
                tokenizer.add_tokens([DEFAULT_X_PATCH_TOKEN[x.upper()]], special_tokens=True)
            # tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
            self.resize_token_embeddings(len(tokenizer))

        if model_args.mm_use_x_start_end:
            num_new_tokens = 0
            for x in model_args.X:
                num_new_tokens += tokenizer.add_tokens([DEFAULT_X_START_TOKEN[x.upper()], DEFAULT_X_END_TOKEN[x.upper()]], special_tokens=True)
            self.resize_token_embeddings(len(tokenizer))

            if num_new_tokens > 0:
                input_embeddings = self.get_input_embeddings().weight.data
                output_embeddings = self.get_output_embeddings().weight.data

                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
                    dim=0, keepdim=True)
                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
                    dim=0, keepdim=True)

                input_embeddings[-num_new_tokens:] = input_embeddings_avg
                output_embeddings[-num_new_tokens:] = output_embeddings_avg

            if model_args.tune_mm_mlp_adapter:
                for p in self.get_input_embeddings().parameters():
                    p.requires_grad = True
                for p in self.get_output_embeddings().parameters():
                    p.requires_grad = False

            if model_args.pretrain_mm_mlp_adapter:
                mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
                embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight']
                assert num_new_tokens == 2
                if input_embeddings.shape == embed_tokens_weight.shape:
                    input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
                elif embed_tokens_weight.shape[0] == num_new_tokens:
                    input_embeddings[-num_new_tokens:] = embed_tokens_weight
                else:
                    raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
        elif model_args.mm_use_x_patch_token:
            if model_args.tune_mm_mlp_adapter:
                for p in self.get_input_embeddings().parameters():
                    p.requires_grad = False
                for p in self.get_output_embeddings().parameters():
                    p.requires_grad = False