File size: 14,582 Bytes
8155cef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import math
import copy

import torch
import torch.nn.functional as F
from torch.nn import CrossEntropyLoss

from PIL import Image
from functools import partial
from typing import List, Optional, Tuple, Union, Dict
from dataclasses import dataclass

import transformers
from transformers.modeling_outputs import ModelOutput
from transformers.modeling_utils import PreTrainedModel
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, PretrainedConfig, Qwen2Config, SiglipVisionModel

from .adapters import AdapterSigLIP
from .mm_constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX
from .processing_FlashVL import tokenizer_image_token_qwen
from .configuration_FlashVLStatic import FlashVLStaticConfig

@dataclass
class FlashVLStaticOutputWithPast(ModelOutput):
    loss: Optional[torch.FloatTensor] = None
    logits: torch.FloatTensor = None


class FlashVLStatic(PreTrainedModel):
    config_class = FlashVLStaticConfig
    
    def __init__(self, config):
        super().__init__(config)
        self.llm = AutoModelForCausalLM.from_config(config.llm_config, trust_remote_code=True)
        self.vit = SiglipVisionModel(config.vision_config).vision_model
        self.adp = AdapterSigLIP(config)
        self.image_token_num = config.image_token_num
        self.image_size = config.vision_config.image_size

    def merge_text_image_tokens(self, inputs):
        input_ids, image_features, targets, attn_mask, loss_mask = inputs
        micro_batch_size, tokens_len = input_ids.shape
        device = input_ids.device
        
        img_rows, img_cols = torch.where(input_ids == IMAGE_TOKEN_INDEX)
        image_idxs = {i: [] for i in range(micro_batch_size)}
        for row, col in zip(img_rows.tolist(), img_cols.tolist()):
            image_idxs[row].append(col)
        for row in range(micro_batch_size):
            image_idxs[row] = sorted(image_idxs[row])

        split_sizes = []
        for row in range(micro_batch_size):
            image_num = len(image_idxs[row])
            if image_num == 0:
                split_sizes.append(tokens_len)
                continue

            if image_idxs[row][0] != 0:
                split_sizes.append(image_idxs[row][0])

            for idx in range(image_num - 1):
                split_sizes.append(self.image_token_num)
                if image_idxs[row][idx + 1] > image_idxs[row][idx] + self.image_token_num:
                    split_sizes.append(image_idxs[row][idx + 1] - (image_idxs[row][idx] + self.image_token_num))

            if image_idxs[row][image_num - 1] + self.image_token_num >= tokens_len:
                split_sizes.append(tokens_len - image_idxs[row][image_num - 1])
            else:
                split_sizes.append(self.image_token_num)
                split_sizes.append(tokens_len - (image_idxs[row][image_num - 1] + self.image_token_num))

        input_ids_noim = torch.where(input_ids < 0, 151643, input_ids)
        input_ids_noim = input_ids_noim.view(-1)
        input_embeds = self.llm.model.embed_tokens(input_ids_noim)
        input_embeds_split = torch.split(input_embeds, split_sizes, dim=0)

        vl_embeds_list = []
        cur_language_idx = 0
        cur_image_idx = 0
        for row in range(micro_batch_size):
            image_num = len(image_idxs[row])
            if image_num == 0:
                vl_embeds_list.append(input_embeds_split[cur_language_idx])
                cur_language_idx += 1
                vl_embeds_list.append(image_features[cur_image_idx][0:0])
                cur_image_idx += 1
                continue

            if image_idxs[row][0] != 0:
                vl_embeds_list.append(input_embeds_split[cur_language_idx])
                cur_language_idx += 1

            for idx in range(image_num - 1):
                vl_embeds_list.append(image_features[cur_image_idx])
                cur_language_idx += 1
                cur_image_idx += 1

                if image_idxs[row][idx + 1] > image_idxs[row][idx] + self.image_token_num:
                    vl_embeds_list.append(input_embeds_split[cur_language_idx])
                    cur_language_idx += 1

            if image_idxs[row][image_num - 1] + self.image_token_num >= tokens_len:
                vl_embeds_list.append(image_features[cur_image_idx][0 : tokens_len - image_idxs[row][image_num - 1]])
                cur_language_idx += 1
                cur_image_idx += 1
            else:
                vl_embeds_list.append(image_features[cur_image_idx])
                cur_language_idx += 1
                cur_image_idx += 1
                vl_embeds_list.append(input_embeds_split[cur_language_idx])
                cur_language_idx += 1

        vl_embeds = torch.cat(vl_embeds_list)
        vl_embeds = vl_embeds.view(micro_batch_size, tokens_len, vl_embeds.shape[-1])
        return (input_ids, vl_embeds, targets, attn_mask, loss_mask)
    
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        pixel_values: torch.FloatTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        local_pos_batch: Optional[torch.LongTensor] = None,
        image_idx_batch: Optional[torch.Tensor] = None,
        loss_mask: Optional[torch.Tensor] = None,
        use_cache: Optional[bool] = None,
    ):
        inputs = [input_ids, pixel_values, labels, attention_mask, loss_mask]
        
        if isinstance(inputs[1], list):
            pixel_values = [p.bfloat16() for p in inputs[1]]
        else:
            pixel_values = inputs[1].bfloat16()
        img_token = self.vit.forward(pixel_values)
        
        if hasattr(img_token, 'last_hidden_state'):
            img_token = img_token.last_hidden_state
        
        inputs = self.adp(inputs[:1]+[img_token]+inputs[2:])

        inputs = self.merge_text_image_tokens(inputs)
        tokens, hidden_states, targets, attn_mask, loss_mask = inputs

        outputs = self.llm.forward(
            inputs_embeds = hidden_states,
            attention_mask = attn_mask,
            use_cache = use_cache)
        
        lm_logits = outputs.logits
        
        loss = None
        if targets is not None:
            labels = targets.to(lm_logits.device)
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()

            loss_fct = CrossEntropyLoss(reduction='none')
            loss = loss_fct(
                shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
            )

            batch_size = labels.size(0)
            loss_mask = loss_mask[:, 1:].to(loss.dtype)
            loss = (loss.view(batch_size, -1) * loss_mask).sum() / loss_mask.sum()

        return FlashVLStaticOutputWithPast(
            loss=loss,
            logits=lm_logits
        )
    
    def get_input_embeddings(self):
        return self.llm.get_input_embeddings()
    
    def generate(
        self,
        input_ids=None,
        pixel_values=None,
        attention_mask=None,
        **kwargs
        ):  
        image = pixel_values
        img_token = self.vit.forward(image.bfloat16())
        if hasattr(img_token, 'last_hidden_state'):
            img_token = img_token.last_hidden_state
        inputs = self.adp((
                    input_ids.to(self.device),
                    img_token,
                    None, None, None))
        inputs = self.merge_text_image_tokens(inputs)
        tokens, hidden_states, targets, attn_mask, loss_mask = inputs

        keys_to_pop = ['loss_mask', 'labels','attention_mask']
        kwargs = {k: v for k, v in kwargs.items() if k not in keys_to_pop}
            
        outputs = self.llm.generate(
            inputs_embeds=hidden_states.bfloat16(),
            max_new_tokens=2048,
            do_sample=False,
            **kwargs
        )
        
        return outputs
    
    def chat(self, pil_image, messages, answer_prompt=None, do_sample=True, max_new_tokens=256):
        data={}
        data['img'] = pil_image
        data['text_only'] = (pil_image is None)
        data['messages'] = messages

        sources = self.to_llava_format(data)
        sources = [sources]
        has_image = not sources[0]['text_only']

        if has_image:
            img_list = sources[0]['image']
            if not isinstance(img_list, list):
                img_list = [img_list]
            image = torch.stack([torch.from_numpy(self.im_trans(i)['pixel_values'][0]) for i in img_list], dim=0)
            
        sources = copy.deepcopy([e["conversations"] for e in sources])
        
        data_dict = self.preprocess_qwen(
                sources,
                self.tokenizer,
                has_image=has_image,
                )
        
        input_ids_data = data_dict["input_ids"][0]
        data_dict["input_ids"] = [ input_ids_data, ]
        
        if not has_image:
            image = torch.zeros(1, 3, self.image_size, self.image_size)
        data_dict = dict(tokens=data_dict["input_ids"][0],)

        img_token = self.vit.forward(image.cuda().bfloat16())
        
        if hasattr(img_token, 'last_hidden_state'):
            img_token = img_token.last_hidden_state

        inputs = self.adp((
                    data_dict['tokens'].unsqueeze(0).to(self.device),
                    img_token, 
                    None, None, None))
        
        inputs = self.merge_text_image_tokens(inputs)
        tokens, hidden_states, targets, attn_mask, loss_mask = inputs

        outputs = self.llm.generate(
            inputs_embeds=hidden_states.bfloat16(),
            return_dict_in_generate=False,
            max_new_tokens=max_new_tokens,
            do_sample=do_sample,
            pad_token_id=False,
        )
        decoded = self.tokenizer.decode(outputs[0])

        stop_words_ids = [self.llm.generation_config.bos_token_id,
                          self.llm.generation_config.eos_token_id,
                          self.tokenizer.convert_tokens_to_ids('<|im_start|>')]
        stop_words = [self.tokenizer.decode(w) for w in stop_words_ids]

        for stop_word in stop_words:
            decoded = decoded.replace(stop_word, "").strip()

        return decoded
    
    def preprocess_qwen(
        self,
        sources, 
        tokenizer: transformers.PreTrainedTokenizer, 
        has_image: bool = False, 
        max_len=2048, 
        system_message: str = "You are a helpful assistant.",) -> Dict:

        roles = {"human": "user", "gpt": "assistant"}
        tokenizer = copy.deepcopy(tokenizer)
        
        tokenizer.add_tokens(["<image>"], special_tokens=True)
        image_token_index = tokenizer.convert_tokens_to_ids("<image>")
        im_start, im_end = tokenizer.additional_special_tokens_ids[:2]
        
        unmask_tokens_idx =  [198, im_start, im_end]
        nl_tokens = tokenizer("\n").input_ids

        chat_template = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
        tokenizer.chat_template = chat_template
        
        input_ids, targets = [], []
        for i, source in enumerate(sources):
            if roles[source[0]["from"]] != roles["human"]:
                source = source[1:]
            input_id, target = [], []
        
            input_id += tokenizer.apply_chat_template([{"role" : "system", "content" : system_message}])
            target += [IGNORE_INDEX] * len(input_id)
            i=0
            for conv in source:
                try:
                    role = conv["role"]
                    content = conv["content"]
                except:
                    role = conv["from"]
                    content = conv["value"]
                role =  roles.get(role, role)
                if i==len(source)-1:
                    conv = [{"role" : role, "content" : content}]
                    encode_id = tokenizer.apply_chat_template(conv,add_generation_prompt=True)
                else:     
                    conv = [{"role" : role, "content" : content}]
                    encode_id = tokenizer.apply_chat_template(conv)
                i=i+1

                if image_token_index in encode_id:
                    encode_id = tokenizer_image_token_qwen(encode_id, tokenizer, image_token_index, image_token_num=self.image_token_num)
                input_id += encode_id 
                if role in ["user", "system"]:
                    target += [IGNORE_INDEX] * len(encode_id)
                else:
                    target += encode_id

            assert len(input_id) == len(target), f"{len(input_id)} != {len(target)}"
            for idx, encode_id in enumerate(input_id):
                if encode_id in unmask_tokens_idx:
                    target[idx] = encode_id
                if encode_id == image_token_index:
                    input_id[idx] = IMAGE_TOKEN_INDEX
            input_ids.append(input_id)
            targets.append(target)
            
        input_ids = torch.tensor(input_ids, dtype=torch.long)
        targets = torch.tensor(targets, dtype=torch.long)

        return dict(
            input_ids=input_ids,
            labels=targets,
        )
    
    def to_llava_format(self, data):
        img_pil = data['img']
        messages = data['messages']
        text_only = data['text_only']
        is_video=False
        if 'is_video' in data:
            is_video=data['is_video']
            
        messages.append({'role': 'assistant', 'content': ''})
        conversations = []
        for i,m in enumerate(messages):
            if m['role'] == 'user':
                value = str(m['content']).replace('<image>', '')
                if i == 0 and not text_only:
                    value = '<image>\n' + value
                
                conversations.append({'from': 'human', 'value': value})
            elif m['role'] == 'assistant':
                conversations.append({'from': 'gpt', 'value': str(m['content']).replace('<image>', '')})
            else:
                raise ValueError(f"Wrong role in conversation. {m['role']}")

        return {'image': img_pil,
                'text_only': text_only,
                'is_video':is_video,
                'conversations': conversations}