File size: 22,348 Bytes
24df943
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
# Copyright (c) The InternLM team and The HuggingFace Inc. team. All rights reserved.
#
# This code is based on transformers/src/transformers/models/llama/modeling_llama.py
#
# 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.

"""PyTorch InternLMXComposer2 model."""
import copy
import queue
import threading
from typing import List, Optional, Tuple, Union

import torch
import torch.utils.checkpoint
from PIL import Image
from torch import nn
from torch.nn import CrossEntropyLoss
from torchvision import transforms
from torchvision.transforms.functional import InterpolationMode
from transformers.modeling_outputs import CausalLMOutputWithPast
from transformers.utils import (add_start_docstrings_to_model_forward,
                                replace_return_docstrings)

try:
    from transformers.generation.streamers import BaseStreamer
except:  # noqa # pylint: disable=bare-except
    BaseStreamer = None

from .build_mlp import build_vision_projector, build_vision_tower
from .ixc_utils import HD_transform
from .configuration_internlm_xcomposer2 import InternLMXcomposer2Config
from .modeling_internlm2 import (InternLM2_INPUTS_DOCSTRING, InternLM2Model,
                                 InternLM2PreTrainedModel)

_CONFIG_FOR_DOC = 'InternLMXcomposer2Config'


class InternLMXComposer2ForCausalLM(InternLM2PreTrainedModel):
    _auto_class = 'AutoModelForCausalLM'

    _tied_weights_keys = ['output.weight']

    def __init__(self, config):
        super().__init__(config)
        self.model = InternLM2Model(config)
        self.vocab_size = config.vocab_size
        self.output = nn.Linear(
            config.hidden_size, config.vocab_size, bias=False)
        self.tokenizer = None

        self.max_length = config.max_length
        print(f'Set max length to {self.max_length}')
        # Initialize weights and apply final processing
        self.post_init()
        self.plora_glb_GN = nn.Parameter(torch.zeros([1, 1, 4096]))
        self.plora_sub_GN = nn.Parameter(torch.zeros([1, 1, 1, 4096]))

        self.vit = build_vision_tower()
        self.vision_proj = build_vision_projector()

        self.vis_processor = transforms.Compose([
            transforms.ToTensor(),
            transforms.Normalize((0.48145466, 0.4578275, 0.40821073),
                                 (0.26862954, 0.26130258, 0.27577711)),
        ])

    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, InternLM2Model):
            module.gradient_checkpointing = value
        if value:
            self.vit.vision_tower.vision_model.encoder.gradient_checkpointing = value

    def get_input_embeddings(self):
        return self.model.tok_embeddings

    def set_input_embeddings(self, value):
        self.model.tok_embeddings = value

    def get_output_embeddings(self):
        return self.output

    def set_output_embeddings(self, new_embeddings):
        self.output = new_embeddings

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model

    def encode_text(self, text, add_special_tokens=False):
        token = self.tokenizer(
            text, return_tensors='pt',
            add_special_tokens=add_special_tokens).input_ids.to(self.device)
        embs = self.model.tok_embeddings(token)
        return embs

    def encode_img(self, image, hd_num=25):
        if image is None:
            return None
        if isinstance(image, str):
            image = Image.open(image).convert('RGB')
            image = HD_transform(image, hd_num = hd_num)
            image = self.vis_processor(image).unsqueeze(0).to(self.device)

        img_embeds, atts_img, img_target = self.img2emb(image)
        return img_embeds

    def img2emb(self, image):
        img_embeds, img_split = self.vit([image], 
            self.plora_glb_GN, self.plora_sub_GN)
        if len(img_split) > 1:
            print ('Batch Size >1 is not supported.')
            assert 0
        print (img_embeds.shape)
        img_embeds = self.vision_proj(img_embeds)
        atts_img = torch.ones(
            img_embeds.size()[:-1], dtype=torch.long).to(img_embeds.device)

        img_target = torch.ones(
            img_embeds.size()[:2], dtype=torch.long).to(
                img_embeds.device) * -100

        return img_embeds, atts_img, img_target

    def prompt_wrap(self, img_embeds, prompt):
        batch_size = img_embeds.shape[0]
        p_before, p_after = prompt.split('<ImageHere>')
        p_before_tokens = self.tokenizer(
            p_before, return_tensors='pt',
            add_special_tokens=True).to(img_embeds.device)

        p_before_embeds = self.model.tok_embeddings(
            p_before_tokens.input_ids).expand(batch_size, -1, -1)
        wrapped_img_embeds = torch.cat([p_before_embeds, img_embeds], dim=1)

        wrapped_atts_img = torch.ones(
            wrapped_img_embeds.size()[:-1],
            dtype=torch.long).to(img_embeds.device)

        wrapped_target = torch.ones(
            batch_size, wrapped_img_embeds.shape[1], dtype=torch.long).to(
                img_embeds.device) * -100

        return wrapped_img_embeds, wrapped_atts_img, wrapped_target

    def text2emb(self, text, add_special=False):
        to_regress_tokens = self.tokenizer(
            text,
            return_tensors='pt',
            padding='longest',
            truncation=True,
            max_length=self.max_length,
            add_special_tokens=add_special).to(self.device)

        targets = self.mask_human_targets(to_regress_tokens.input_ids)
        targets = targets.to(self.device)
        return to_regress_tokens, targets

    def interleav_wrap_chat(self, tokenizer, query, image, history, meta_instruction):
        prompt = ''
        if meta_instruction:
            prompt += f"""[UNUSED_TOKEN_146]system\n{meta_instruction}[UNUSED_TOKEN_145]\n"""
        for record in history:
            prompt += f"""[UNUSED_TOKEN_146]user\n{record[0]}[UNUSED_TOKEN_145]\n[UNUSED_TOKEN_146]assistant\n{record[1]}[UNUSED_TOKEN_145]\n"""
        prompt += f"""[UNUSED_TOKEN_146]user\n{query}[UNUSED_TOKEN_145]\n[UNUSED_TOKEN_146]assistant\n"""

        im_len = image.shape[1]
        image_nums = len(image)
        parts = prompt.split('<ImageHere>')
        wrap_embeds, wrap_im_mask = [], []
        temp_len = 0

        if len(parts) != image_nums + 1:
            raise ValueError('Invalid <ImageHere> prompt format.')
    
        for idx, part in enumerate(parts):
            if len(part) > 0:
                part_tokens = tokenizer(part, return_tensors='pt').to(self.device)
                part_embeds = self.model.tok_embeddings(
                    part_tokens.input_ids)
                wrap_embeds.append(part_embeds)
                wrap_im_mask.append(torch.zeros(part_embeds.shape[:2]))
                temp_len += part_embeds.shape[1]
            if idx < image_nums:
                wrap_embeds.append(image[idx].unsqueeze(0))
                wrap_im_mask.append(torch.ones(1, image[idx].shape[0]))
                temp_len += im_len
    
            if temp_len > self.max_length:
                break
    
        wrap_embeds = torch.cat(wrap_embeds, dim=1)
        wrap_im_mask = torch.cat(wrap_im_mask, dim=1)
        wrap_embeds = wrap_embeds[:, :self.max_length].to(self.device)
        wrap_im_mask = wrap_im_mask[:, :self.max_length].to(self.device).bool()
        inputs = {
            'inputs_embeds': wrap_embeds
        }
        return inputs, wrap_im_mask

    def interleav_wrap(self, img_list, text_list):
        wrap_embeds_list, wrap_atts_list = [], []
        wrap_target_list, wrap_im_mask_list = [], []

        for image, text in zip(img_list, text_list):
            img_embeds, atts_img, img_target = self.img2emb(image)
            text = text[0]
            parts = text.split('<ImageHere>')
            wrap_tokens, wrap_embeds, wrap_atts, wrap_im_mask = [], [], [], []
            temp_len = 0
            image_nums, im_len = img_embeds.shape[:2]
            need_bos = True
            for idx, part in enumerate(parts):
                if len(part) > 0:
                    part_tokens = self.tokenizer(
                        part,
                        return_tensors='pt',
                        padding='longest',
                        add_special_tokens=need_bos).to(self.device)
                    if need_bos:
                        need_bos = False
                    wrap_tokens.append(part_tokens.input_ids)
                    part_embeds = self.model.tok_embeddings(
                        part_tokens.input_ids)
                    wrap_embeds.append(part_embeds)
                    wrap_atts.append(part_tokens.attention_mask)
                    wrap_im_mask.append(
                        torch.zeros(part_embeds.shape[:2]).to(self.device))

                    temp_len += part_embeds.shape[1]
                if idx < image_nums:
                    wrap_tokens.append(img_target[idx].unsqueeze(0))
                    wrap_embeds.append(img_embeds[idx].unsqueeze(0))
                    wrap_atts.append(atts_img[idx].unsqueeze(0))
                    wrap_im_mask.append(
                        torch.ones_like(atts_img[idx].unsqueeze(0)))

                    temp_len += im_len
                if temp_len > self.max_length:
                    break

            wrap_tokens = torch.cat(wrap_tokens, dim=1)
            wrap_embeds = torch.cat(wrap_embeds, dim=1)
            wrap_atts = torch.cat(wrap_atts, dim=1)
            wrap_im_mask = torch.cat(wrap_im_mask, dim=1)

            wrap_target = self.mask_human_targets(wrap_tokens).to(self.device)

            wrap_embeds = wrap_embeds[:, :self.max_length].to(self.device)
            wrap_atts = wrap_atts[:, :self.max_length].to(self.device)
            wrap_target = wrap_target[:, :self.max_length].to(self.device)
            wrap_im_mask = wrap_im_mask[:, :self.max_length].to(self.device)

            wrap_embeds_list.append(wrap_embeds)
            wrap_atts_list.append(wrap_atts)
            wrap_target_list.append(wrap_target)
            wrap_im_mask_list.append(wrap_im_mask)

        wrap_embeds = torch.cat(wrap_embeds_list)
        wrap_atts = torch.cat(wrap_atts_list)
        wrap_target = torch.cat(wrap_target_list)
        wrap_im_mask = torch.cat(wrap_im_mask_list)
        return wrap_embeds, wrap_atts, wrap_target, wrap_im_mask

    def mask_human_targets(self, input_ids, pure=False):
        target_batch = []
        for bs in range(input_ids.shape[0]):
            ids = input_ids[bs]
            targets = copy.deepcopy(ids)
            end_count = 0
            last_eoa = 0
            for i, temp_id in enumerate(ids):
                if temp_id == 92542:
                    if end_count % 2 == 0:
                        targets[last_eoa:i + 6] = -100
                    else:
                        last_eoa = i + 1
                    end_count += 1
                # # eos and following pad
                elif temp_id == 2:
                    # loss on eos, but not on pad
                    targets[i + 1:] = -100
                    break
            # trunction, end at last question
            if temp_id != 2 and end_count % 2 == 0:
                # mask all after the last answer
                targets[last_eoa + 1:] = -100
            target_batch.append(targets.unsqueeze(0))
        target_batch = torch.cat(target_batch, dim=0)
        return target_batch

    @add_start_docstrings_to_model_forward(InternLM2_INPUTS_DOCSTRING)
    @replace_return_docstrings(
        output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
    def forward(self,
                input_ids: torch.LongTensor = None,
                attention_mask: Optional[torch.Tensor] = None,
                position_ids: Optional[torch.LongTensor] = None,
                past_key_values: Optional[List[torch.FloatTensor]] = None,
                inputs_embeds: Optional[torch.FloatTensor] = None,
                labels: Optional[torch.LongTensor] = None,
                use_cache: Optional[bool] = None,
                output_attentions: Optional[bool] = None,
                output_hidden_states: Optional[bool] = None,
                return_dict: Optional[bool] = None,
                **kwargs) -> Union[Tuple, CausalLMOutputWithPast]:
        r"""
        Args:
            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
        Returns:
        """

        samples = kwargs.get('samples', None)
        if samples:
            if samples['data_type'][0] == 'text':
                has_img = False
            elif samples['data_type'][0] == 'multi':
                has_img = True
            else:
                raise NotImplementedError

            # encode text
            text = samples['text_input']
            # encode image
            if has_img:
                image = samples['image']
                to_regress_embeds, attention_mask, targets, im_mask = self.interleav_wrap(
                    image, text)
            else:
                to_regress_tokens, targets = self.text2emb(
                    text, add_special=True)
                to_regress_embeds = self.model.tok_embeddings(
                    to_regress_tokens.input_ids)
                attention_mask = to_regress_tokens.attention_mask
                im_mask = torch.zeros(to_regress_embeds.shape[:2]).cuda()

            inputs_embeds = to_regress_embeds[:, :self.max_length]
            attention_mask = attention_mask[:, :self.max_length]
            targets = targets[:, :self.max_length]
            im_mask = im_mask[:, :self.max_length].bool()
            labels = targets
        else:
            im_mask = kwargs.get('im_mask', None)
            if im_mask is None and inputs_embeds is not None:
                im_mask = torch.zeros(inputs_embeds.shape[:2]).to(
                    inputs_embeds.device)
                im_mask = im_mask.bool()

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else
            self.config.output_hidden_states)
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            im_mask=im_mask,
        )

        hidden_states = outputs[0]
        logits = self.output(hidden_states)
        logits = logits.float()

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Enable model parallelism
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

        if not return_dict:
            output = (logits, ) + outputs[1:]
            return (loss, ) + output if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    def prepare_inputs_for_generation(self,
                                      input_ids,
                                      past_key_values=None,
                                      attention_mask=None,
                                      inputs_embeds=None,
                                      im_mask=None,
                                      **kwargs):
        if past_key_values is not None:
            past_length = past_key_values[0][0].shape[2]

            # Some generation methods already pass only the last input ID
            if input_ids.shape[1] > past_length:
                remove_prefix_length = past_length
            else:
                # Default to old behavior: keep only final ID
                remove_prefix_length = input_ids.shape[1] - 1

            input_ids = input_ids[:, remove_prefix_length:]

        position_ids = kwargs.get('position_ids', None)
        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past_key_values:
                position_ids = position_ids[:, -input_ids.shape[1]:]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {'inputs_embeds': inputs_embeds}
        else:
            model_inputs = {'input_ids': input_ids}

        im_mask = im_mask

        model_inputs.update({
            'position_ids': position_ids,
            'past_key_values': past_key_values,
            'use_cache': kwargs.get('use_cache'),
            'attention_mask': attention_mask,
            'im_mask': im_mask,
        })
        return model_inputs

    @staticmethod
    def _reorder_cache(past_key_values, beam_idx):
        reordered_past = ()
        for layer_past in past_key_values:
            reordered_past += (tuple(
                past_state.index_select(0, beam_idx.to(past_state.device))
                for past_state in layer_past), )
        return reordered_past

    def build_inputs(self,
                     tokenizer,
                     query: str,
                     history: List[Tuple[str, str]] = [],
                     meta_instruction=''):
        prompt = ''
        if meta_instruction:
            prompt += f"""<s>[UNUSED_TOKEN_146]system\n{meta_instruction}[UNUSED_TOKEN_145]\n"""
        else:
            prompt += '<s>'
        for record in history:
            prompt += f"""[UNUSED_TOKEN_146]user\n{record[0]}[UNUSED_TOKEN_145]\n[UNUSED_TOKEN_146]assistant\n{record[1]}[UNUSED_TOKEN_145]\n"""
        prompt += f"""[UNUSED_TOKEN_146]user\n{query}[UNUSED_TOKEN_145]\n[UNUSED_TOKEN_146]assistant\n"""
        return tokenizer([prompt], return_tensors='pt')

    @torch.no_grad()
    def chat(
        self,
        tokenizer,
        query: str,
        image: torch.Tensor = None,
        hd_num: int = 25,
        history: List[Tuple[str, str]] = [],
        streamer: Optional[BaseStreamer] = None,
        max_new_tokens: int = 1024,
        do_sample: bool = True,
        num_beams: int = 1,
        temperature: float = 1.0,
        top_p: float = 0.8,
        repetition_penalty: float=1.005,
        meta_instruction:
        str = 'You are an AI assistant whose name is InternLM-XComposer (浦语·灵笔).\n'
        '- InternLM-XComposer (浦语·灵笔) is a multi-modality conversational language model that is developed by Shanghai AI Laboratory (上海人工智能实验室). It is designed to be helpful, honest, and harmless.\n'
        '- InternLM-XComposer (浦语·灵笔) can understand and communicate fluently in the language chosen by the user such as English and 中文.\n'
        '- InternLM-XComposer (浦语·灵笔) is capable of comprehending and articulating responses effectively based on the provided image.',
        **kwargs,
    ):
        if image is None:
            inputs = self.build_inputs(tokenizer, query, history, meta_instruction)
            im_mask = torch.zeros(inputs['input_ids'].shape[:2]).cuda().bool()
        else:
            image = self.encode_img(image, hd_num=hd_num)
            inputs, im_mask = self.interleav_wrap_chat(tokenizer, query, image, history, meta_instruction)
        inputs = {
            k: v.to(self.device)
            for k, v in inputs.items() if torch.is_tensor(v)
        }
        # also add end-of-assistant token in eos token id to avoid unnecessary generation
        eos_token_id = [
            tokenizer.eos_token_id,
            tokenizer.convert_tokens_to_ids(['[UNUSED_TOKEN_145]'])[0]
        ]
        outputs = self.generate(
            **inputs,
            streamer=streamer,
            max_new_tokens=max_new_tokens,
            num_beams=num_beams,
            do_sample=do_sample,
            temperature=temperature,
            top_p=top_p,
            eos_token_id=eos_token_id,
            repetition_penalty=repetition_penalty,
            im_mask=im_mask,
            **kwargs,
        )
        if image is None:
            outputs = outputs[0].cpu().tolist()[len(inputs['input_ids'][0]):]
        else:
            outputs = outputs[0].cpu().tolist()
        response = tokenizer.decode(outputs, skip_special_tokens=True)
        response = response.split('[UNUSED_TOKEN_145]')[0]
        history = history + [(query, response)]
        return response, history