zwgao commited on
Commit
d2a4995
1 Parent(s): 449f3b5

Upload model

Browse files
config.json ADDED
@@ -0,0 +1,193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_commit_hash": null,
3
+ "_name_or_path": "/mnt/petrelfs/wangweiyun/workspace_gzw/internvl_small/mini-internvl-da/Mini-InternVL2-1B-DA-RS",
4
+ "architectures": [
5
+ "InternVLChatModel"
6
+ ],
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_internvl_chat.InternVLChatConfig",
9
+ "AutoModel": "modeling_internvl_chat.InternVLChatModel",
10
+ "AutoModelForCausalLM": "modeling_internvl_chat.InternVLChatModel"
11
+ },
12
+ "downsample_ratio": 0.5,
13
+ "dynamic_image_size": true,
14
+ "force_image_size": 448,
15
+ "llm_config": {
16
+ "_name_or_path": "Qwen/Qwen2-0.5B-Instruct",
17
+ "add_cross_attention": false,
18
+ "architectures": [
19
+ "Qwen2ForCausalLM"
20
+ ],
21
+ "attention_dropout": 0.0,
22
+ "bad_words_ids": null,
23
+ "begin_suppress_tokens": null,
24
+ "bos_token_id": 151643,
25
+ "chunk_size_feed_forward": 0,
26
+ "cross_attention_hidden_size": null,
27
+ "decoder_start_token_id": null,
28
+ "diversity_penalty": 0.0,
29
+ "do_sample": false,
30
+ "early_stopping": false,
31
+ "encoder_no_repeat_ngram_size": 0,
32
+ "eos_token_id": 151645,
33
+ "exponential_decay_length_penalty": null,
34
+ "finetuning_task": null,
35
+ "forced_bos_token_id": null,
36
+ "forced_eos_token_id": null,
37
+ "hidden_act": "silu",
38
+ "hidden_size": 896,
39
+ "id2label": {
40
+ "0": "LABEL_0",
41
+ "1": "LABEL_1"
42
+ },
43
+ "initializer_range": 0.02,
44
+ "intermediate_size": 4864,
45
+ "is_decoder": false,
46
+ "is_encoder_decoder": false,
47
+ "label2id": {
48
+ "LABEL_0": 0,
49
+ "LABEL_1": 1
50
+ },
51
+ "length_penalty": 1.0,
52
+ "max_length": 20,
53
+ "max_position_embeddings": 32768,
54
+ "max_window_layers": 24,
55
+ "min_length": 0,
56
+ "model_type": "qwen2",
57
+ "no_repeat_ngram_size": 0,
58
+ "num_attention_heads": 14,
59
+ "num_beam_groups": 1,
60
+ "num_beams": 1,
61
+ "num_hidden_layers": 24,
62
+ "num_key_value_heads": 2,
63
+ "num_return_sequences": 1,
64
+ "output_attentions": false,
65
+ "output_hidden_states": false,
66
+ "output_scores": false,
67
+ "pad_token_id": null,
68
+ "prefix": null,
69
+ "problem_type": null,
70
+ "pruned_heads": {},
71
+ "remove_invalid_values": false,
72
+ "repetition_penalty": 1.0,
73
+ "return_dict": true,
74
+ "return_dict_in_generate": false,
75
+ "rms_norm_eps": 1e-06,
76
+ "rope_theta": 1000000.0,
77
+ "sep_token_id": null,
78
+ "sliding_window": 32768,
79
+ "suppress_tokens": null,
80
+ "task_specific_params": null,
81
+ "temperature": 1.0,
82
+ "tf_legacy_loss": false,
83
+ "tie_encoder_decoder": false,
84
+ "tie_word_embeddings": true,
85
+ "tokenizer_class": null,
86
+ "top_k": 50,
87
+ "top_p": 1.0,
88
+ "torch_dtype": "bfloat16",
89
+ "torchscript": false,
90
+ "transformers_version": "4.37.2",
91
+ "typical_p": 1.0,
92
+ "use_bfloat16": true,
93
+ "use_cache": false,
94
+ "use_sliding_window": false,
95
+ "vocab_size": 151655
96
+ },
97
+ "max_dynamic_patch": 12,
98
+ "min_dynamic_patch": 1,
99
+ "model_type": "internvl_chat",
100
+ "pad2square": false,
101
+ "ps_version": "v2",
102
+ "select_layer": -1,
103
+ "template": "Hermes-2",
104
+ "torch_dtype": "bfloat16",
105
+ "transformers_version": null,
106
+ "use_backbone_lora": 0,
107
+ "use_llm_lora": 0,
108
+ "use_thumbnail": true,
109
+ "vision_config": {
110
+ "_name_or_path": "",
111
+ "add_cross_attention": false,
112
+ "architectures": [
113
+ "InternVisionModel"
114
+ ],
115
+ "attention_dropout": 0.0,
116
+ "bad_words_ids": null,
117
+ "begin_suppress_tokens": null,
118
+ "bos_token_id": null,
119
+ "chunk_size_feed_forward": 0,
120
+ "cross_attention_hidden_size": null,
121
+ "decoder_start_token_id": null,
122
+ "diversity_penalty": 0.0,
123
+ "do_sample": false,
124
+ "drop_path_rate": 0.1,
125
+ "dropout": 0.0,
126
+ "early_stopping": false,
127
+ "encoder_no_repeat_ngram_size": 0,
128
+ "eos_token_id": null,
129
+ "exponential_decay_length_penalty": null,
130
+ "finetuning_task": null,
131
+ "forced_bos_token_id": null,
132
+ "forced_eos_token_id": null,
133
+ "hidden_act": "gelu",
134
+ "hidden_size": 1024,
135
+ "id2label": {
136
+ "0": "LABEL_0",
137
+ "1": "LABEL_1"
138
+ },
139
+ "image_size": 448,
140
+ "initializer_factor": 1.0,
141
+ "initializer_range": 0.02,
142
+ "intermediate_size": 4096,
143
+ "is_decoder": false,
144
+ "is_encoder_decoder": false,
145
+ "label2id": {
146
+ "LABEL_0": 0,
147
+ "LABEL_1": 1
148
+ },
149
+ "layer_norm_eps": 1e-06,
150
+ "length_penalty": 1.0,
151
+ "max_length": 20,
152
+ "min_length": 0,
153
+ "model_type": "intern_vit_6b",
154
+ "no_repeat_ngram_size": 0,
155
+ "norm_type": "layer_norm",
156
+ "num_attention_heads": 16,
157
+ "num_beam_groups": 1,
158
+ "num_beams": 1,
159
+ "num_channels": 3,
160
+ "num_hidden_layers": 24,
161
+ "num_return_sequences": 1,
162
+ "output_attentions": false,
163
+ "output_hidden_states": false,
164
+ "output_scores": false,
165
+ "pad_token_id": null,
166
+ "patch_size": 14,
167
+ "prefix": null,
168
+ "problem_type": null,
169
+ "pruned_heads": {},
170
+ "qk_normalization": false,
171
+ "qkv_bias": true,
172
+ "remove_invalid_values": false,
173
+ "repetition_penalty": 1.0,
174
+ "return_dict": true,
175
+ "return_dict_in_generate": false,
176
+ "sep_token_id": null,
177
+ "suppress_tokens": null,
178
+ "task_specific_params": null,
179
+ "temperature": 1.0,
180
+ "tf_legacy_loss": false,
181
+ "tie_encoder_decoder": false,
182
+ "tie_word_embeddings": true,
183
+ "tokenizer_class": null,
184
+ "top_k": 50,
185
+ "top_p": 1.0,
186
+ "torch_dtype": "bfloat16",
187
+ "torchscript": false,
188
+ "transformers_version": "4.37.2",
189
+ "typical_p": 1.0,
190
+ "use_bfloat16": true,
191
+ "use_flash_attn": true
192
+ }
193
+ }
configuration_intern_vit.py ADDED
@@ -0,0 +1,120 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # --------------------------------------------------------
2
+ # InternVL
3
+ # Copyright (c) 2024 OpenGVLab
4
+ # Licensed under The MIT License [see LICENSE for details]
5
+ # --------------------------------------------------------
6
+
7
+ import os
8
+ from typing import Union
9
+
10
+ from transformers.configuration_utils import PretrainedConfig
11
+ from transformers.utils import logging
12
+
13
+ logger = logging.get_logger(__name__)
14
+
15
+
16
+ class InternVisionConfig(PretrainedConfig):
17
+ r"""
18
+ This is the configuration class to store the configuration of a [`InternVisionModel`]. It is used to
19
+ instantiate a vision encoder according to the specified arguments, defining the model architecture.
20
+
21
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
22
+ documentation from [`PretrainedConfig`] for more information.
23
+
24
+ Args:
25
+ num_channels (`int`, *optional*, defaults to 3):
26
+ Number of color channels in the input images (e.g., 3 for RGB).
27
+ patch_size (`int`, *optional*, defaults to 14):
28
+ The size (resolution) of each patch.
29
+ image_size (`int`, *optional*, defaults to 224):
30
+ The size (resolution) of each image.
31
+ qkv_bias (`bool`, *optional*, defaults to `False`):
32
+ Whether to add a bias to the queries and values in the self-attention layers.
33
+ hidden_size (`int`, *optional*, defaults to 3200):
34
+ Dimensionality of the encoder layers and the pooler layer.
35
+ num_attention_heads (`int`, *optional*, defaults to 25):
36
+ Number of attention heads for each attention layer in the Transformer encoder.
37
+ intermediate_size (`int`, *optional*, defaults to 12800):
38
+ Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
39
+ qk_normalization (`bool`, *optional*, defaults to `True`):
40
+ Whether to normalize the queries and keys in the self-attention layers.
41
+ num_hidden_layers (`int`, *optional*, defaults to 48):
42
+ Number of hidden layers in the Transformer encoder.
43
+ use_flash_attn (`bool`, *optional*, defaults to `True`):
44
+ Whether to use flash attention mechanism.
45
+ hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
46
+ The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
47
+ `"relu"`, `"selu"` and `"gelu_new"` ``"gelu"` are supported.
48
+ layer_norm_eps (`float`, *optional*, defaults to 1e-6):
49
+ The epsilon used by the layer normalization layers.
50
+ dropout (`float`, *optional*, defaults to 0.0):
51
+ The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
52
+ drop_path_rate (`float`, *optional*, defaults to 0.0):
53
+ Dropout rate for stochastic depth.
54
+ attention_dropout (`float`, *optional*, defaults to 0.0):
55
+ The dropout ratio for the attention probabilities.
56
+ initializer_range (`float`, *optional*, defaults to 0.02):
57
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
58
+ initializer_factor (`float`, *optional*, defaults to 0.1):
59
+ A factor for layer scale.
60
+ """
61
+
62
+ model_type = 'intern_vit_6b'
63
+
64
+ def __init__(
65
+ self,
66
+ num_channels=3,
67
+ patch_size=14,
68
+ image_size=224,
69
+ qkv_bias=False,
70
+ hidden_size=3200,
71
+ num_attention_heads=25,
72
+ intermediate_size=12800,
73
+ qk_normalization=True,
74
+ num_hidden_layers=48,
75
+ use_flash_attn=True,
76
+ hidden_act='gelu',
77
+ norm_type='rms_norm',
78
+ layer_norm_eps=1e-6,
79
+ dropout=0.0,
80
+ drop_path_rate=0.0,
81
+ attention_dropout=0.0,
82
+ initializer_range=0.02,
83
+ initializer_factor=0.1,
84
+ **kwargs,
85
+ ):
86
+ super().__init__(**kwargs)
87
+
88
+ self.hidden_size = hidden_size
89
+ self.intermediate_size = intermediate_size
90
+ self.dropout = dropout
91
+ self.drop_path_rate = drop_path_rate
92
+ self.num_hidden_layers = num_hidden_layers
93
+ self.num_attention_heads = num_attention_heads
94
+ self.num_channels = num_channels
95
+ self.patch_size = patch_size
96
+ self.image_size = image_size
97
+ self.initializer_range = initializer_range
98
+ self.initializer_factor = initializer_factor
99
+ self.attention_dropout = attention_dropout
100
+ self.layer_norm_eps = layer_norm_eps
101
+ self.hidden_act = hidden_act
102
+ self.norm_type = norm_type
103
+ self.qkv_bias = qkv_bias
104
+ self.qk_normalization = qk_normalization
105
+ self.use_flash_attn = use_flash_attn
106
+
107
+ @classmethod
108
+ def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> 'PretrainedConfig':
109
+ config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
110
+
111
+ if 'vision_config' in config_dict:
112
+ config_dict = config_dict['vision_config']
113
+
114
+ if 'model_type' in config_dict and hasattr(cls, 'model_type') and config_dict['model_type'] != cls.model_type:
115
+ logger.warning(
116
+ f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
117
+ f'{cls.model_type}. This is not supported for all configurations of models and can yield errors.'
118
+ )
119
+
120
+ return cls.from_dict(config_dict, **kwargs)
configuration_internvl_chat.py ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # --------------------------------------------------------
2
+ # InternVL
3
+ # Copyright (c) 2024 OpenGVLab
4
+ # Licensed under The MIT License [see LICENSE for details]
5
+ # --------------------------------------------------------
6
+
7
+ import copy
8
+
9
+ from transformers import AutoConfig, LlamaConfig, Qwen2Config
10
+ from transformers.configuration_utils import PretrainedConfig
11
+ from transformers.utils import logging
12
+
13
+ from .configuration_intern_vit import InternVisionConfig
14
+
15
+ logger = logging.get_logger(__name__)
16
+
17
+
18
+ class InternVLChatConfig(PretrainedConfig):
19
+ model_type = 'internvl_chat'
20
+ is_composition = True
21
+
22
+ def __init__(
23
+ self,
24
+ vision_config=None,
25
+ llm_config=None,
26
+ use_backbone_lora=0,
27
+ use_llm_lora=0,
28
+ select_layer=-1,
29
+ force_image_size=None,
30
+ downsample_ratio=0.5,
31
+ template=None,
32
+ dynamic_image_size=False,
33
+ use_thumbnail=False,
34
+ ps_version='v1',
35
+ min_dynamic_patch=1,
36
+ max_dynamic_patch=6,
37
+ **kwargs):
38
+ super().__init__(**kwargs)
39
+
40
+ if vision_config is None:
41
+ vision_config = {'architectures': ['InternVisionModel']}
42
+ logger.info('vision_config is None. Initializing the InternVisionConfig with default values.')
43
+
44
+ if llm_config is None:
45
+ llm_config = {'architectures': ['Qwen2ForCausalLM']}
46
+ logger.info('llm_config is None. Initializing the LlamaConfig config with default values (`LlamaConfig`).')
47
+
48
+ self.vision_config = InternVisionConfig(**vision_config)
49
+ if llm_config.get('architectures')[0] == 'LlamaForCausalLM':
50
+ self.llm_config = LlamaConfig(**llm_config)
51
+ elif llm_config.get('architectures')[0] == 'Qwen2ForCausalLM':
52
+ self.llm_config = Qwen2Config(**llm_config)
53
+ else:
54
+ raise ValueError('Unsupported architecture: {}'.format(llm_config.get('architectures')[0]))
55
+ self.use_backbone_lora = use_backbone_lora
56
+ self.use_llm_lora = use_llm_lora
57
+ self.select_layer = select_layer
58
+ self.force_image_size = force_image_size
59
+ self.downsample_ratio = downsample_ratio
60
+ self.template = template
61
+ self.dynamic_image_size = dynamic_image_size
62
+ self.use_thumbnail = use_thumbnail
63
+ self.ps_version = ps_version # pixel shuffle version
64
+ self.min_dynamic_patch = min_dynamic_patch
65
+ self.max_dynamic_patch = max_dynamic_patch
66
+
67
+ logger.info(f'vision_select_layer: {self.select_layer}')
68
+ logger.info(f'ps_version: {self.ps_version}')
69
+ logger.info(f'min_dynamic_patch: {self.min_dynamic_patch}')
70
+ logger.info(f'max_dynamic_patch: {self.max_dynamic_patch}')
71
+
72
+ def to_dict(self):
73
+ """
74
+ Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
75
+
76
+ Returns:
77
+ `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
78
+ """
79
+ output = copy.deepcopy(self.__dict__)
80
+ output['vision_config'] = self.vision_config.to_dict()
81
+ output['llm_config'] = self.llm_config.to_dict()
82
+ output['model_type'] = self.__class__.model_type
83
+ output['use_backbone_lora'] = self.use_backbone_lora
84
+ output['use_llm_lora'] = self.use_llm_lora
85
+ output['select_layer'] = self.select_layer
86
+ output['force_image_size'] = self.force_image_size
87
+ output['downsample_ratio'] = self.downsample_ratio
88
+ output['template'] = self.template
89
+ output['dynamic_image_size'] = self.dynamic_image_size
90
+ output['use_thumbnail'] = self.use_thumbnail
91
+ output['ps_version'] = self.ps_version
92
+ output['min_dynamic_patch'] = self.min_dynamic_patch
93
+ output['max_dynamic_patch'] = self.max_dynamic_patch
94
+
95
+ return output
conversation.py ADDED
@@ -0,0 +1,391 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ Conversation prompt templates.
3
+
4
+ We kindly request that you import fastchat instead of copying this file if you wish to use it.
5
+ If you have changes in mind, please contribute back so the community can benefit collectively and continue to maintain these valuable templates.
6
+
7
+ Modified from https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
8
+ """
9
+
10
+ import dataclasses
11
+ from enum import IntEnum, auto
12
+ from typing import Dict, List, Tuple, Union
13
+
14
+
15
+ class SeparatorStyle(IntEnum):
16
+ """Separator styles."""
17
+
18
+ ADD_COLON_SINGLE = auto()
19
+ ADD_COLON_TWO = auto()
20
+ ADD_COLON_SPACE_SINGLE = auto()
21
+ NO_COLON_SINGLE = auto()
22
+ NO_COLON_TWO = auto()
23
+ ADD_NEW_LINE_SINGLE = auto()
24
+ LLAMA2 = auto()
25
+ CHATGLM = auto()
26
+ CHATML = auto()
27
+ CHATINTERN = auto()
28
+ DOLLY = auto()
29
+ RWKV = auto()
30
+ PHOENIX = auto()
31
+ ROBIN = auto()
32
+ FALCON_CHAT = auto()
33
+ CHATGLM3 = auto()
34
+ INTERNVL_ZH = auto()
35
+ MPT = auto()
36
+
37
+
38
+ @dataclasses.dataclass
39
+ class Conversation:
40
+ """A class that manages prompt templates and keeps all conversation history."""
41
+
42
+ # The name of this template
43
+ name: str
44
+ # The template of the system prompt
45
+ system_template: str = '{system_message}'
46
+ # The system message
47
+ system_message: str = ''
48
+ # The names of two roles
49
+ roles: Tuple[str] = ('USER', 'ASSISTANT')
50
+ # All messages. Each item is (role, message).
51
+ messages: List[List[str]] = ()
52
+ # The number of few shot examples
53
+ offset: int = 0
54
+ # The separator style and configurations
55
+ sep_style: SeparatorStyle = SeparatorStyle.ADD_COLON_SINGLE
56
+ sep: str = '\n'
57
+ sep2: str = None
58
+ # Stop criteria (the default one is EOS token)
59
+ stop_str: Union[str, List[str]] = None
60
+ # Stops generation if meeting any token in this list
61
+ stop_token_ids: List[int] = None
62
+
63
+ def get_prompt(self) -> str:
64
+ """Get the prompt for generation."""
65
+ system_prompt = self.system_template.format(system_message=self.system_message)
66
+ if self.sep_style == SeparatorStyle.ADD_COLON_SINGLE:
67
+ ret = system_prompt + self.sep
68
+ for role, message in self.messages:
69
+ if message:
70
+ ret += role + ': ' + message + self.sep
71
+ else:
72
+ ret += role + ':'
73
+ return ret
74
+ elif self.sep_style == SeparatorStyle.ADD_COLON_TWO:
75
+ seps = [self.sep, self.sep2]
76
+ ret = system_prompt + seps[0]
77
+ for i, (role, message) in enumerate(self.messages):
78
+ if message:
79
+ ret += role + ': ' + message + seps[i % 2]
80
+ else:
81
+ ret += role + ':'
82
+ return ret
83
+ elif self.sep_style == SeparatorStyle.ADD_COLON_SPACE_SINGLE:
84
+ ret = system_prompt + self.sep
85
+ for role, message in self.messages:
86
+ if message:
87
+ ret += role + ': ' + message + self.sep
88
+ else:
89
+ ret += role + ': ' # must be end with a space
90
+ return ret
91
+ elif self.sep_style == SeparatorStyle.ADD_NEW_LINE_SINGLE:
92
+ ret = '' if system_prompt == '' else system_prompt + self.sep
93
+ for role, message in self.messages:
94
+ if message:
95
+ ret += role + '\n' + message + self.sep
96
+ else:
97
+ ret += role + '\n'
98
+ return ret
99
+ elif self.sep_style == SeparatorStyle.NO_COLON_SINGLE:
100
+ ret = system_prompt
101
+ for role, message in self.messages:
102
+ if message:
103
+ ret += role + message + self.sep
104
+ else:
105
+ ret += role
106
+ return ret
107
+ elif self.sep_style == SeparatorStyle.NO_COLON_TWO:
108
+ seps = [self.sep, self.sep2]
109
+ ret = system_prompt
110
+ for i, (role, message) in enumerate(self.messages):
111
+ if message:
112
+ ret += role + message + seps[i % 2]
113
+ else:
114
+ ret += role
115
+ return ret
116
+ elif self.sep_style == SeparatorStyle.RWKV:
117
+ ret = system_prompt
118
+ for i, (role, message) in enumerate(self.messages):
119
+ if message:
120
+ ret += (
121
+ role
122
+ + ': '
123
+ + message.replace('\r\n', '\n').replace('\n\n', '\n')
124
+ )
125
+ ret += '\n\n'
126
+ else:
127
+ ret += role + ':'
128
+ return ret
129
+ elif self.sep_style == SeparatorStyle.LLAMA2:
130
+ seps = [self.sep, self.sep2]
131
+ if self.system_message:
132
+ ret = system_prompt
133
+ else:
134
+ ret = '[INST] '
135
+ for i, (role, message) in enumerate(self.messages):
136
+ tag = self.roles[i % 2]
137
+ if message:
138
+ if i == 0:
139
+ ret += message + ' '
140
+ else:
141
+ ret += tag + ' ' + message + seps[i % 2]
142
+ else:
143
+ ret += tag
144
+ return ret
145
+ elif self.sep_style == SeparatorStyle.CHATGLM:
146
+ # source: https://huggingface.co/THUDM/chatglm-6b/blob/1d240ba371910e9282298d4592532d7f0f3e9f3e/modeling_chatglm.py#L1302-L1308
147
+ # source2: https://huggingface.co/THUDM/chatglm2-6b/blob/e186c891cf64310ac66ef10a87e6635fa6c2a579/modeling_chatglm.py#L926
148
+ round_add_n = 1 if self.name == 'chatglm2' else 0
149
+ if system_prompt:
150
+ ret = system_prompt + self.sep
151
+ else:
152
+ ret = ''
153
+
154
+ for i, (role, message) in enumerate(self.messages):
155
+ if i % 2 == 0:
156
+ ret += f'[Round {i//2 + round_add_n}]{self.sep}'
157
+
158
+ if message:
159
+ ret += f'{role}:{message}{self.sep}'
160
+ else:
161
+ ret += f'{role}:'
162
+ return ret
163
+ elif self.sep_style == SeparatorStyle.CHATML:
164
+ ret = '' if system_prompt == '' else system_prompt + self.sep + '\n'
165
+ for role, message in self.messages:
166
+ if message:
167
+ ret += role + '\n' + message + self.sep + '\n'
168
+ else:
169
+ ret += role + '\n'
170
+ return ret
171
+ elif self.sep_style == SeparatorStyle.CHATGLM3:
172
+ ret = ''
173
+ if self.system_message:
174
+ ret += system_prompt
175
+ for role, message in self.messages:
176
+ if message:
177
+ ret += role + '\n' + ' ' + message
178
+ else:
179
+ ret += role
180
+ return ret
181
+ elif self.sep_style == SeparatorStyle.CHATINTERN:
182
+ # source: https://huggingface.co/internlm/internlm-chat-7b-8k/blob/bd546fa984b4b0b86958f56bf37f94aa75ab8831/modeling_internlm.py#L771
183
+ seps = [self.sep, self.sep2]
184
+ ret = system_prompt
185
+ for i, (role, message) in enumerate(self.messages):
186
+ # if i % 2 == 0:
187
+ # ret += "<s>"
188
+ if message:
189
+ ret += role + ':' + message + seps[i % 2] + '\n'
190
+ else:
191
+ ret += role + ':'
192
+ return ret
193
+ elif self.sep_style == SeparatorStyle.DOLLY:
194
+ seps = [self.sep, self.sep2]
195
+ ret = system_prompt
196
+ for i, (role, message) in enumerate(self.messages):
197
+ if message:
198
+ ret += role + ':\n' + message + seps[i % 2]
199
+ if i % 2 == 1:
200
+ ret += '\n\n'
201
+ else:
202
+ ret += role + ':\n'
203
+ return ret
204
+ elif self.sep_style == SeparatorStyle.PHOENIX:
205
+ ret = system_prompt
206
+ for role, message in self.messages:
207
+ if message:
208
+ ret += role + ': ' + '<s>' + message + '</s>'
209
+ else:
210
+ ret += role + ': ' + '<s>'
211
+ return ret
212
+ elif self.sep_style == SeparatorStyle.ROBIN:
213
+ ret = system_prompt + self.sep
214
+ for role, message in self.messages:
215
+ if message:
216
+ ret += role + ':\n' + message + self.sep
217
+ else:
218
+ ret += role + ':\n'
219
+ return ret
220
+ elif self.sep_style == SeparatorStyle.FALCON_CHAT:
221
+ ret = ''
222
+ if self.system_message:
223
+ ret += system_prompt + self.sep
224
+ for role, message in self.messages:
225
+ if message:
226
+ ret += role + ': ' + message + self.sep
227
+ else:
228
+ ret += role + ':'
229
+
230
+ return ret
231
+ elif self.sep_style == SeparatorStyle.INTERNVL_ZH:
232
+ seps = [self.sep, self.sep2]
233
+ ret = self.system_message + seps[0]
234
+ for i, (role, message) in enumerate(self.messages):
235
+ if message:
236
+ ret += role + ': ' + message + seps[i % 2]
237
+ else:
238
+ ret += role + ':'
239
+ return ret
240
+ elif self.sep_style == SeparatorStyle.MPT:
241
+ ret = system_prompt + self.sep
242
+ for role, message in self.messages:
243
+ if message:
244
+ if type(message) is tuple:
245
+ message, _, _ = message
246
+ ret += role + message + self.sep
247
+ else:
248
+ ret += role
249
+ return ret
250
+ else:
251
+ raise ValueError(f'Invalid style: {self.sep_style}')
252
+
253
+ def set_system_message(self, system_message: str):
254
+ """Set the system message."""
255
+ self.system_message = system_message
256
+
257
+ def append_message(self, role: str, message: str):
258
+ """Append a new message."""
259
+ self.messages.append([role, message])
260
+
261
+ def update_last_message(self, message: str):
262
+ """Update the last output.
263
+
264
+ The last message is typically set to be None when constructing the prompt,
265
+ so we need to update it in-place after getting the response from a model.
266
+ """
267
+ self.messages[-1][1] = message
268
+
269
+ def to_gradio_chatbot(self):
270
+ """Convert the conversation to gradio chatbot format."""
271
+ ret = []
272
+ for i, (role, msg) in enumerate(self.messages[self.offset :]):
273
+ if i % 2 == 0:
274
+ ret.append([msg, None])
275
+ else:
276
+ ret[-1][-1] = msg
277
+ return ret
278
+
279
+ def to_openai_api_messages(self):
280
+ """Convert the conversation to OpenAI chat completion format."""
281
+ ret = [{'role': 'system', 'content': self.system_message}]
282
+
283
+ for i, (_, msg) in enumerate(self.messages[self.offset :]):
284
+ if i % 2 == 0:
285
+ ret.append({'role': 'user', 'content': msg})
286
+ else:
287
+ if msg is not None:
288
+ ret.append({'role': 'assistant', 'content': msg})
289
+ return ret
290
+
291
+ def copy(self):
292
+ return Conversation(
293
+ name=self.name,
294
+ system_template=self.system_template,
295
+ system_message=self.system_message,
296
+ roles=self.roles,
297
+ messages=[[x, y] for x, y in self.messages],
298
+ offset=self.offset,
299
+ sep_style=self.sep_style,
300
+ sep=self.sep,
301
+ sep2=self.sep2,
302
+ stop_str=self.stop_str,
303
+ stop_token_ids=self.stop_token_ids,
304
+ )
305
+
306
+ def dict(self):
307
+ return {
308
+ 'template_name': self.name,
309
+ 'system_message': self.system_message,
310
+ 'roles': self.roles,
311
+ 'messages': self.messages,
312
+ 'offset': self.offset,
313
+ }
314
+
315
+
316
+ # A global registry for all conversation templates
317
+ conv_templates: Dict[str, Conversation] = {}
318
+
319
+
320
+ def register_conv_template(template: Conversation, override: bool = False):
321
+ """Register a new conversation template."""
322
+ if not override:
323
+ assert (
324
+ template.name not in conv_templates
325
+ ), f'{template.name} has been registered.'
326
+
327
+ conv_templates[template.name] = template
328
+
329
+
330
+ def get_conv_template(name: str) -> Conversation:
331
+ """Get a conversation template."""
332
+ return conv_templates[name].copy()
333
+
334
+
335
+ # Both Hermes-2 and internlm2-chat are chatml-format conversation templates. The difference
336
+ # is that during training, the preprocessing function for the Hermes-2 template doesn't add
337
+ # <s> at the beginning of the tokenized sequence, while the internlm2-chat template does.
338
+ # Therefore, they are completely equivalent during inference.
339
+ register_conv_template(
340
+ Conversation(
341
+ name='Hermes-2',
342
+ system_template='<|im_start|>system\n{system_message}',
343
+ # note: The new system prompt was not used here to avoid changes in benchmark performance.
344
+ # system_message='我是书生·万象,英文名是InternVL,是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。',
345
+ system_message='你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型,英文名叫InternVL, 是一个有用无害的人工智能助手。',
346
+ roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
347
+ sep_style=SeparatorStyle.MPT,
348
+ sep='<|im_end|>',
349
+ stop_str='<|endoftext|>',
350
+ )
351
+ )
352
+
353
+
354
+ register_conv_template(
355
+ Conversation(
356
+ name='internlm2-chat',
357
+ system_template='<|im_start|>system\n{system_message}',
358
+ # note: The new system prompt was not used here to avoid changes in benchmark performance.
359
+ # system_message='我是书生·万象,英文名是InternVL,是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。',
360
+ system_message='你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型,英文名叫InternVL, 是一个有用无害的人工智能助手。',
361
+ roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
362
+ sep_style=SeparatorStyle.MPT,
363
+ sep='<|im_end|>',
364
+ )
365
+ )
366
+
367
+
368
+ register_conv_template(
369
+ Conversation(
370
+ name='phi3-chat',
371
+ system_template='<|system|>\n{system_message}',
372
+ # note: The new system prompt was not used here to avoid changes in benchmark performance.
373
+ # system_message='我是书生·万象,英文名是InternVL,是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。',
374
+ system_message='你是由上海人工智能实验室联合商汤科技开发的书生多模态大模型,英文名叫InternVL, 是一个有用无害的人工智能助手。',
375
+ roles=('<|user|>\n', '<|assistant|>\n'),
376
+ sep_style=SeparatorStyle.MPT,
377
+ sep='<|end|>',
378
+ )
379
+ )
380
+
381
+
382
+ register_conv_template(
383
+ Conversation(
384
+ name='internvl2_5',
385
+ system_template='<|im_start|>system\n{system_message}',
386
+ system_message='你是书生·万象,英文名是InternVL,是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。',
387
+ roles=('<|im_start|>user\n', '<|im_start|>assistant\n'),
388
+ sep_style=SeparatorStyle.MPT,
389
+ sep='<|im_end|>\n',
390
+ )
391
+ )
generation_config.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "eos_token_id": [
4
+ 151644,
5
+ 151645
6
+ ],
7
+ "transformers_version": "4.37.2"
8
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c194c5dc3c97d141d09b9b0c39a3d344cbb3155e3229aef39b1e00412b0eb7a1
3
+ size 1876395376
modeling_intern_vit.py ADDED
@@ -0,0 +1,430 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # --------------------------------------------------------
2
+ # InternVL
3
+ # Copyright (c) 2024 OpenGVLab
4
+ # Licensed under The MIT License [see LICENSE for details]
5
+ # --------------------------------------------------------
6
+
7
+ from typing import Optional, Tuple, Union
8
+
9
+ import torch
10
+ import torch.nn.functional as F
11
+ import torch.utils.checkpoint
12
+ from einops import rearrange
13
+ from timm.models.layers import DropPath
14
+ from torch import nn
15
+ from transformers.activations import ACT2FN
16
+ from transformers.modeling_outputs import (BaseModelOutput,
17
+ BaseModelOutputWithPooling)
18
+ from transformers.modeling_utils import PreTrainedModel
19
+ from transformers.utils import logging
20
+
21
+ from .configuration_intern_vit import InternVisionConfig
22
+
23
+ try:
24
+ from flash_attn.bert_padding import pad_input, unpad_input
25
+ from flash_attn.flash_attn_interface import \
26
+ flash_attn_varlen_qkvpacked_func
27
+ has_flash_attn = True
28
+ except:
29
+ print('FlashAttention2 is not installed.')
30
+ has_flash_attn = False
31
+
32
+ logger = logging.get_logger(__name__)
33
+
34
+
35
+ class FlashAttention(nn.Module):
36
+ """Implement the scaled dot product attention with softmax.
37
+ Arguments
38
+ ---------
39
+ softmax_scale: The temperature to use for the softmax attention.
40
+ (default: 1/sqrt(d_keys) where d_keys is computed at
41
+ runtime)
42
+ attention_dropout: The dropout rate to apply to the attention
43
+ (default: 0.0)
44
+ """
45
+
46
+ def __init__(self, softmax_scale=None, attention_dropout=0.0, device=None, dtype=None):
47
+ super().__init__()
48
+ self.softmax_scale = softmax_scale
49
+ self.dropout_p = attention_dropout
50
+
51
+ def forward(self, qkv, key_padding_mask=None, causal=False, cu_seqlens=None,
52
+ max_s=None, need_weights=False):
53
+ """Implements the multihead softmax attention.
54
+ Arguments
55
+ ---------
56
+ qkv: The tensor containing the query, key, and value. (B, S, 3, H, D) if key_padding_mask is None
57
+ if unpadded: (nnz, 3, h, d)
58
+ key_padding_mask: a bool tensor of shape (B, S)
59
+ """
60
+ assert not need_weights
61
+ assert qkv.dtype in [torch.float16, torch.bfloat16]
62
+ assert qkv.is_cuda
63
+
64
+ if cu_seqlens is None:
65
+ batch_size = qkv.shape[0]
66
+ seqlen = qkv.shape[1]
67
+ if key_padding_mask is None:
68
+ qkv = rearrange(qkv, 'b s ... -> (b s) ...')
69
+ max_s = seqlen
70
+ cu_seqlens = torch.arange(0, (batch_size + 1) * seqlen, step=seqlen, dtype=torch.int32,
71
+ device=qkv.device)
72
+ output = flash_attn_varlen_qkvpacked_func(
73
+ qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
74
+ softmax_scale=self.softmax_scale, causal=causal
75
+ )
76
+ output = rearrange(output, '(b s) ... -> b s ...', b=batch_size)
77
+ else:
78
+ nheads = qkv.shape[-2]
79
+ x = rearrange(qkv, 'b s three h d -> b s (three h d)')
80
+ x_unpad, indices, cu_seqlens, max_s = unpad_input(x, key_padding_mask)
81
+ x_unpad = rearrange(x_unpad, 'nnz (three h d) -> nnz three h d', three=3, h=nheads)
82
+ output_unpad = flash_attn_varlen_qkvpacked_func(
83
+ x_unpad, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
84
+ softmax_scale=self.softmax_scale, causal=causal
85
+ )
86
+ output = rearrange(pad_input(rearrange(output_unpad, 'nnz h d -> nnz (h d)'),
87
+ indices, batch_size, seqlen),
88
+ 'b s (h d) -> b s h d', h=nheads)
89
+ else:
90
+ assert max_s is not None
91
+ output = flash_attn_varlen_qkvpacked_func(
92
+ qkv, cu_seqlens, max_s, self.dropout_p if self.training else 0.0,
93
+ softmax_scale=self.softmax_scale, causal=causal
94
+ )
95
+
96
+ return output, None
97
+
98
+
99
+ class InternRMSNorm(nn.Module):
100
+ def __init__(self, hidden_size, eps=1e-6):
101
+ super().__init__()
102
+ self.weight = nn.Parameter(torch.ones(hidden_size))
103
+ self.variance_epsilon = eps
104
+
105
+ def forward(self, hidden_states):
106
+ input_dtype = hidden_states.dtype
107
+ hidden_states = hidden_states.to(torch.float32)
108
+ variance = hidden_states.pow(2).mean(-1, keepdim=True)
109
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
110
+ return self.weight * hidden_states.to(input_dtype)
111
+
112
+
113
+ try:
114
+ from apex.normalization import FusedRMSNorm
115
+
116
+ InternRMSNorm = FusedRMSNorm # noqa
117
+
118
+ logger.info('Discovered apex.normalization.FusedRMSNorm - will use it instead of InternRMSNorm')
119
+ except ImportError:
120
+ # using the normal InternRMSNorm
121
+ pass
122
+ except Exception:
123
+ logger.warning('discovered apex but it failed to load, falling back to InternRMSNorm')
124
+ pass
125
+
126
+
127
+ NORM2FN = {
128
+ 'rms_norm': InternRMSNorm,
129
+ 'layer_norm': nn.LayerNorm,
130
+ }
131
+
132
+
133
+ class InternVisionEmbeddings(nn.Module):
134
+ def __init__(self, config: InternVisionConfig):
135
+ super().__init__()
136
+ self.config = config
137
+ self.embed_dim = config.hidden_size
138
+ self.image_size = config.image_size
139
+ self.patch_size = config.patch_size
140
+
141
+ self.class_embedding = nn.Parameter(
142
+ torch.randn(1, 1, self.embed_dim),
143
+ )
144
+
145
+ self.patch_embedding = nn.Conv2d(
146
+ in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size
147
+ )
148
+
149
+ self.num_patches = (self.image_size // self.patch_size) ** 2
150
+ self.num_positions = self.num_patches + 1
151
+
152
+ self.position_embedding = nn.Parameter(torch.randn(1, self.num_positions, self.embed_dim))
153
+
154
+ def _get_pos_embed(self, pos_embed, H, W):
155
+ target_dtype = pos_embed.dtype
156
+ pos_embed = pos_embed.float().reshape(
157
+ 1, self.image_size // self.patch_size, self.image_size // self.patch_size, -1).permute(0, 3, 1, 2)
158
+ pos_embed = F.interpolate(pos_embed, size=(H, W), mode='bicubic', align_corners=False). \
159
+ reshape(1, -1, H * W).permute(0, 2, 1).to(target_dtype)
160
+ return pos_embed
161
+
162
+ def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
163
+ target_dtype = self.patch_embedding.weight.dtype
164
+ patch_embeds = self.patch_embedding(pixel_values) # shape = [*, channel, width, height]
165
+ batch_size, _, height, width = patch_embeds.shape
166
+ patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
167
+ class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)
168
+ embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
169
+ position_embedding = torch.cat([
170
+ self.position_embedding[:, :1, :],
171
+ self._get_pos_embed(self.position_embedding[:, 1:, :], height, width)
172
+ ], dim=1)
173
+ embeddings = embeddings + position_embedding.to(target_dtype)
174
+ return embeddings
175
+
176
+
177
+ class InternAttention(nn.Module):
178
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
179
+
180
+ def __init__(self, config: InternVisionConfig):
181
+ super().__init__()
182
+ self.config = config
183
+ self.embed_dim = config.hidden_size
184
+ self.num_heads = config.num_attention_heads
185
+ self.use_flash_attn = config.use_flash_attn and has_flash_attn
186
+ if config.use_flash_attn and not has_flash_attn:
187
+ print('Warning: Flash Attention is not available, use_flash_attn is set to False.')
188
+ self.head_dim = self.embed_dim // self.num_heads
189
+ if self.head_dim * self.num_heads != self.embed_dim:
190
+ raise ValueError(
191
+ f'embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:'
192
+ f' {self.num_heads}).'
193
+ )
194
+
195
+ self.scale = self.head_dim ** -0.5
196
+ self.qkv = nn.Linear(self.embed_dim, 3 * self.embed_dim, bias=config.qkv_bias)
197
+ self.attn_drop = nn.Dropout(config.attention_dropout)
198
+ self.proj_drop = nn.Dropout(config.dropout)
199
+
200
+ self.qk_normalization = config.qk_normalization
201
+
202
+ if self.qk_normalization:
203
+ self.q_norm = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)
204
+ self.k_norm = InternRMSNorm(self.embed_dim, eps=config.layer_norm_eps)
205
+
206
+ if self.use_flash_attn:
207
+ self.inner_attn = FlashAttention(attention_dropout=config.attention_dropout)
208
+ self.proj = nn.Linear(self.embed_dim, self.embed_dim)
209
+
210
+ def _naive_attn(self, x):
211
+ B, N, C = x.shape
212
+ qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
213
+ q, k, v = qkv.unbind(0) # make torchscript happy (cannot use tensor as tuple)
214
+
215
+ if self.qk_normalization:
216
+ B_, H_, N_, D_ = q.shape
217
+ q = self.q_norm(q.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
218
+ k = self.k_norm(k.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
219
+
220
+ attn = ((q * self.scale) @ k.transpose(-2, -1))
221
+ attn = attn.softmax(dim=-1)
222
+ attn = self.attn_drop(attn)
223
+
224
+ x = (attn @ v).transpose(1, 2).reshape(B, N, C)
225
+ x = self.proj(x)
226
+ x = self.proj_drop(x)
227
+ return x
228
+
229
+ def _flash_attn(self, x, key_padding_mask=None, need_weights=False):
230
+ qkv = self.qkv(x)
231
+ qkv = rearrange(qkv, 'b s (three h d) -> b s three h d', three=3, h=self.num_heads)
232
+
233
+ if self.qk_normalization:
234
+ q, k, v = qkv.unbind(2)
235
+ q = self.q_norm(q.flatten(-2, -1)).view(q.shape)
236
+ k = self.k_norm(k.flatten(-2, -1)).view(k.shape)
237
+ qkv = torch.stack([q, k, v], dim=2)
238
+
239
+ context, _ = self.inner_attn(
240
+ qkv, key_padding_mask=key_padding_mask, need_weights=need_weights, causal=False
241
+ )
242
+ outs = self.proj(rearrange(context, 'b s h d -> b s (h d)'))
243
+ outs = self.proj_drop(outs)
244
+ return outs
245
+
246
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
247
+ x = self._naive_attn(hidden_states) if not self.use_flash_attn else self._flash_attn(hidden_states)
248
+ return x
249
+
250
+
251
+ class InternMLP(nn.Module):
252
+ def __init__(self, config: InternVisionConfig):
253
+ super().__init__()
254
+ self.config = config
255
+ self.act = ACT2FN[config.hidden_act]
256
+ self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
257
+ self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
258
+
259
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
260
+ hidden_states = self.fc1(hidden_states)
261
+ hidden_states = self.act(hidden_states)
262
+ hidden_states = self.fc2(hidden_states)
263
+ return hidden_states
264
+
265
+
266
+ class InternVisionEncoderLayer(nn.Module):
267
+ def __init__(self, config: InternVisionConfig, drop_path_rate: float):
268
+ super().__init__()
269
+ self.embed_dim = config.hidden_size
270
+ self.intermediate_size = config.intermediate_size
271
+ self.norm_type = config.norm_type
272
+
273
+ self.attn = InternAttention(config)
274
+ self.mlp = InternMLP(config)
275
+ self.norm1 = NORM2FN[self.norm_type](self.embed_dim, eps=config.layer_norm_eps)
276
+ self.norm2 = NORM2FN[self.norm_type](self.embed_dim, eps=config.layer_norm_eps)
277
+
278
+ self.ls1 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))
279
+ self.ls2 = nn.Parameter(config.initializer_factor * torch.ones(self.embed_dim))
280
+ self.drop_path1 = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
281
+ self.drop_path2 = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
282
+
283
+ def forward(
284
+ self,
285
+ hidden_states: torch.Tensor,
286
+ ) -> Tuple[torch.FloatTensor, Optional[torch.FloatTensor], Optional[Tuple[torch.FloatTensor]]]:
287
+ """
288
+ Args:
289
+ hidden_states (`Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]`): input to the layer of shape `(batch, seq_len, embed_dim)`
290
+ """
291
+ hidden_states = hidden_states + self.drop_path1(self.attn(self.norm1(hidden_states).to(hidden_states.dtype)) * self.ls1)
292
+
293
+ hidden_states = hidden_states + self.drop_path2(self.mlp(self.norm2(hidden_states).to(hidden_states.dtype)) * self.ls2)
294
+
295
+ return hidden_states
296
+
297
+
298
+ class InternVisionEncoder(nn.Module):
299
+ """
300
+ Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
301
+ [`InternEncoderLayer`].
302
+
303
+ Args:
304
+ config (`InternConfig`):
305
+ The corresponding vision configuration for the `InternEncoder`.
306
+ """
307
+
308
+ def __init__(self, config: InternVisionConfig):
309
+ super().__init__()
310
+ self.config = config
311
+ # stochastic depth decay rule
312
+ dpr = [x.item() for x in torch.linspace(0, config.drop_path_rate, config.num_hidden_layers)]
313
+ self.layers = nn.ModuleList([
314
+ InternVisionEncoderLayer(config, dpr[idx]) for idx in range(config.num_hidden_layers)])
315
+ self.gradient_checkpointing = True
316
+
317
+ def forward(
318
+ self,
319
+ inputs_embeds,
320
+ output_hidden_states: Optional[bool] = None,
321
+ return_dict: Optional[bool] = None,
322
+ ) -> Union[Tuple, BaseModelOutput]:
323
+ r"""
324
+ Args:
325
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
326
+ Embedded representation of the inputs. Should be float, not int tokens.
327
+ output_hidden_states (`bool`, *optional*):
328
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
329
+ for more detail.
330
+ return_dict (`bool`, *optional*):
331
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
332
+ """
333
+ output_hidden_states = (
334
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
335
+ )
336
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
337
+
338
+ encoder_states = () if output_hidden_states else None
339
+ hidden_states = inputs_embeds
340
+
341
+ for idx, encoder_layer in enumerate(self.layers):
342
+ if output_hidden_states:
343
+ encoder_states = encoder_states + (hidden_states,)
344
+ if self.gradient_checkpointing and self.training:
345
+ layer_outputs = torch.utils.checkpoint.checkpoint(
346
+ encoder_layer,
347
+ hidden_states)
348
+ else:
349
+ layer_outputs = encoder_layer(
350
+ hidden_states,
351
+ )
352
+ hidden_states = layer_outputs
353
+
354
+ if output_hidden_states:
355
+ encoder_states = encoder_states + (hidden_states,)
356
+
357
+ if not return_dict:
358
+ return tuple(v for v in [hidden_states, encoder_states] if v is not None)
359
+ return BaseModelOutput(
360
+ last_hidden_state=hidden_states, hidden_states=encoder_states
361
+ )
362
+
363
+
364
+ class InternVisionModel(PreTrainedModel):
365
+ main_input_name = 'pixel_values'
366
+ _supports_flash_attn_2 = True
367
+ config_class = InternVisionConfig
368
+ _no_split_modules = ['InternVisionEncoderLayer']
369
+
370
+ def __init__(self, config: InternVisionConfig):
371
+ super().__init__(config)
372
+ self.config = config
373
+
374
+ self.embeddings = InternVisionEmbeddings(config)
375
+ self.encoder = InternVisionEncoder(config)
376
+
377
+ def resize_pos_embeddings(self, old_size, new_size, patch_size):
378
+ pos_emb = self.embeddings.position_embedding
379
+ _, num_positions, embed_dim = pos_emb.shape
380
+ cls_emb = pos_emb[:, :1, :]
381
+ pos_emb = pos_emb[:, 1:, :].reshape(1, old_size // patch_size, old_size // patch_size, -1).permute(0, 3, 1, 2)
382
+ pos_emb = F.interpolate(pos_emb.float(), size=new_size // patch_size, mode='bicubic', align_corners=False)
383
+ pos_emb = pos_emb.to(cls_emb.dtype).reshape(1, embed_dim, -1).permute(0, 2, 1)
384
+ pos_emb = torch.cat([cls_emb, pos_emb], dim=1)
385
+ self.embeddings.position_embedding = nn.Parameter(pos_emb)
386
+ self.embeddings.image_size = new_size
387
+ logger.info('Resized position embeddings from {} to {}'.format(old_size, new_size))
388
+
389
+ def get_input_embeddings(self):
390
+ return self.embeddings
391
+
392
+ def forward(
393
+ self,
394
+ pixel_values: Optional[torch.FloatTensor] = None,
395
+ output_hidden_states: Optional[bool] = None,
396
+ return_dict: Optional[bool] = None,
397
+ pixel_embeds: Optional[torch.FloatTensor] = None,
398
+ ) -> Union[Tuple, BaseModelOutputWithPooling]:
399
+ output_hidden_states = (
400
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
401
+ )
402
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
403
+
404
+ if pixel_values is None and pixel_embeds is None:
405
+ raise ValueError('You have to specify pixel_values or pixel_embeds')
406
+
407
+ if pixel_embeds is not None:
408
+ hidden_states = pixel_embeds
409
+ else:
410
+ if len(pixel_values.shape) == 4:
411
+ hidden_states = self.embeddings(pixel_values)
412
+ else:
413
+ raise ValueError(f'wrong pixel_values size: {pixel_values.shape}')
414
+ encoder_outputs = self.encoder(
415
+ inputs_embeds=hidden_states,
416
+ output_hidden_states=output_hidden_states,
417
+ return_dict=return_dict,
418
+ )
419
+ last_hidden_state = encoder_outputs.last_hidden_state
420
+ pooled_output = last_hidden_state[:, 0, :]
421
+
422
+ if not return_dict:
423
+ return (last_hidden_state, pooled_output) + encoder_outputs[1:]
424
+
425
+ return BaseModelOutputWithPooling(
426
+ last_hidden_state=last_hidden_state,
427
+ pooler_output=pooled_output,
428
+ hidden_states=encoder_outputs.hidden_states,
429
+ attentions=encoder_outputs.attentions,
430
+ )
modeling_internvl_chat.py ADDED
@@ -0,0 +1,348 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # --------------------------------------------------------
2
+ # InternVL
3
+ # Copyright (c) 2024 OpenGVLab
4
+ # Licensed under The MIT License [see LICENSE for details]
5
+ # --------------------------------------------------------
6
+
7
+ import warnings
8
+ from typing import List, Optional, Tuple, Union
9
+
10
+ import torch.utils.checkpoint
11
+ import transformers
12
+ from torch import nn
13
+ from torch.nn import CrossEntropyLoss
14
+ from transformers import (AutoModel, GenerationConfig, LlamaForCausalLM,
15
+ Qwen2ForCausalLM)
16
+ from transformers.modeling_outputs import CausalLMOutputWithPast
17
+ from transformers.modeling_utils import PreTrainedModel
18
+ from transformers.utils import ModelOutput, logging
19
+
20
+ from .configuration_internvl_chat import InternVLChatConfig
21
+ from .conversation import get_conv_template
22
+ from .modeling_intern_vit import InternVisionModel, has_flash_attn
23
+
24
+ logger = logging.get_logger(__name__)
25
+
26
+
27
+ def version_cmp(v1, v2, op='eq'):
28
+ import operator
29
+
30
+ from packaging import version
31
+ op_func = getattr(operator, op)
32
+ return op_func(version.parse(v1), version.parse(v2))
33
+
34
+
35
+ class InternVLChatModel(PreTrainedModel):
36
+ config_class = InternVLChatConfig
37
+ main_input_name = 'pixel_values'
38
+ base_model_prefix = 'language_model'
39
+ _supports_flash_attn_2 = True
40
+ _no_split_modules = ['InternVisionModel', 'LlamaDecoderLayer', 'Qwen2DecoderLayer']
41
+
42
+ def __init__(self, config: InternVLChatConfig, vision_model=None, language_model=None, use_flash_attn=True):
43
+ super().__init__(config)
44
+
45
+ assert version_cmp(transformers.__version__, '4.37.0', 'ge')
46
+ image_size = config.force_image_size or config.vision_config.image_size
47
+ patch_size = config.vision_config.patch_size
48
+ self.patch_size = patch_size
49
+ self.select_layer = config.select_layer
50
+ self.template = config.template
51
+ self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio ** 2))
52
+ self.downsample_ratio = config.downsample_ratio
53
+ self.ps_version = config.ps_version
54
+ use_flash_attn = use_flash_attn if has_flash_attn else False
55
+ config.vision_config.use_flash_attn = True if use_flash_attn else False
56
+ config.llm_config._attn_implementation = 'flash_attention_2' if use_flash_attn else 'eager'
57
+
58
+ logger.info(f'num_image_token: {self.num_image_token}')
59
+ logger.info(f'ps_version: {self.ps_version}')
60
+ if vision_model is not None:
61
+ self.vision_model = vision_model
62
+ else:
63
+ self.vision_model = InternVisionModel(config.vision_config)
64
+ if language_model is not None:
65
+ self.language_model = language_model
66
+ else:
67
+ if config.llm_config.architectures[0] == 'LlamaForCausalLM':
68
+ self.language_model = LlamaForCausalLM(config.llm_config)
69
+ elif config.llm_config.architectures[0] == 'Qwen2ForCausalLM':
70
+ self.language_model = Qwen2ForCausalLM(config.llm_config)
71
+ else:
72
+ raise NotImplementedError(f'{config.llm_config.architectures[0]} is not implemented.')
73
+
74
+ vit_hidden_size = config.vision_config.hidden_size
75
+ llm_hidden_size = config.llm_config.hidden_size
76
+
77
+ self.mlp1 = nn.Sequential(
78
+ nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2),
79
+ nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size),
80
+ nn.GELU(),
81
+ nn.Linear(llm_hidden_size, llm_hidden_size)
82
+ )
83
+
84
+ self.img_context_token_id = None
85
+ self.conv_template = get_conv_template(self.template)
86
+ self.system_message = self.conv_template.system_message
87
+
88
+ def forward(
89
+ self,
90
+ pixel_values: torch.FloatTensor,
91
+ input_ids: torch.LongTensor = None,
92
+ attention_mask: Optional[torch.Tensor] = None,
93
+ position_ids: Optional[torch.LongTensor] = None,
94
+ image_flags: Optional[torch.LongTensor] = None,
95
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
96
+ labels: Optional[torch.LongTensor] = None,
97
+ use_cache: Optional[bool] = None,
98
+ output_attentions: Optional[bool] = None,
99
+ output_hidden_states: Optional[bool] = None,
100
+ return_dict: Optional[bool] = None,
101
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
102
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
103
+
104
+ image_flags = image_flags.squeeze(-1)
105
+ input_embeds = self.language_model.get_input_embeddings()(input_ids).clone()
106
+
107
+ vit_embeds = self.extract_feature(pixel_values)
108
+ vit_embeds = vit_embeds[image_flags == 1]
109
+ vit_batch_size = pixel_values.shape[0]
110
+
111
+ B, N, C = input_embeds.shape
112
+ input_embeds = input_embeds.reshape(B * N, C)
113
+
114
+ if torch.distributed.get_rank() == 0:
115
+ print(f'dynamic ViT batch size: {vit_batch_size}, images per sample: {vit_batch_size / B}, dynamic token length: {N}')
116
+
117
+ input_ids = input_ids.reshape(B * N)
118
+ selected = (input_ids == self.img_context_token_id)
119
+ try:
120
+ input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(-1, C)
121
+ except Exception as e:
122
+ vit_embeds = vit_embeds.reshape(-1, C)
123
+ print(f'warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, '
124
+ f'vit_embeds.shape={vit_embeds.shape}')
125
+ n_token = selected.sum()
126
+ input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds[:n_token]
127
+
128
+ input_embeds = input_embeds.reshape(B, N, C)
129
+
130
+ outputs = self.language_model(
131
+ inputs_embeds=input_embeds,
132
+ attention_mask=attention_mask,
133
+ position_ids=position_ids,
134
+ past_key_values=past_key_values,
135
+ use_cache=use_cache,
136
+ output_attentions=output_attentions,
137
+ output_hidden_states=output_hidden_states,
138
+ return_dict=return_dict,
139
+ )
140
+ logits = outputs.logits
141
+
142
+ loss = None
143
+ if labels is not None:
144
+ # Shift so that tokens < n predict n
145
+ shift_logits = logits[..., :-1, :].contiguous()
146
+ shift_labels = labels[..., 1:].contiguous()
147
+ # Flatten the tokens
148
+ loss_fct = CrossEntropyLoss()
149
+ shift_logits = shift_logits.view(-1, self.language_model.config.vocab_size)
150
+ shift_labels = shift_labels.view(-1)
151
+ # Enable model parallelism
152
+ shift_labels = shift_labels.to(shift_logits.device)
153
+ loss = loss_fct(shift_logits, shift_labels)
154
+
155
+ if not return_dict:
156
+ output = (logits,) + outputs[1:]
157
+ return (loss,) + output if loss is not None else output
158
+
159
+ return CausalLMOutputWithPast(
160
+ loss=loss,
161
+ logits=logits,
162
+ past_key_values=outputs.past_key_values,
163
+ hidden_states=outputs.hidden_states,
164
+ attentions=outputs.attentions,
165
+ )
166
+
167
+ def pixel_shuffle(self, x, scale_factor=0.5):
168
+ n, w, h, c = x.size()
169
+ # N, W, H, C --> N, W, H * scale, C // scale
170
+ x = x.view(n, w, int(h * scale_factor), int(c / scale_factor))
171
+ # N, W, H * scale, C // scale --> N, H * scale, W, C // scale
172
+ x = x.permute(0, 2, 1, 3).contiguous()
173
+ # N, H * scale, W, C // scale --> N, H * scale, W * scale, C // (scale ** 2)
174
+ x = x.view(n, int(h * scale_factor), int(w * scale_factor),
175
+ int(c / (scale_factor * scale_factor)))
176
+ if self.ps_version == 'v1':
177
+ warnings.warn("In ps_version 'v1', the height and width have not been swapped back, "
178
+ 'which results in a transposed image.')
179
+ else:
180
+ x = x.permute(0, 2, 1, 3).contiguous()
181
+ return x
182
+
183
+ def extract_feature(self, pixel_values):
184
+ if self.select_layer == -1:
185
+ vit_embeds = self.vision_model(
186
+ pixel_values=pixel_values,
187
+ output_hidden_states=False,
188
+ return_dict=True).last_hidden_state
189
+ else:
190
+ vit_embeds = self.vision_model(
191
+ pixel_values=pixel_values,
192
+ output_hidden_states=True,
193
+ return_dict=True).hidden_states[self.select_layer]
194
+ vit_embeds = vit_embeds[:, 1:, :]
195
+
196
+ h = w = int(vit_embeds.shape[1] ** 0.5)
197
+ vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
198
+ vit_embeds = self.pixel_shuffle(vit_embeds, scale_factor=self.downsample_ratio)
199
+ vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1])
200
+ vit_embeds = self.mlp1(vit_embeds)
201
+ return vit_embeds
202
+
203
+ def batch_chat(self, tokenizer, pixel_values, questions, generation_config, num_patches_list=None,
204
+ history=None, return_history=False, IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>',
205
+ IMG_CONTEXT_TOKEN='<IMG_CONTEXT>', verbose=False, image_counts=None):
206
+ if history is not None or return_history:
207
+ print('Now multi-turn chat is not supported in batch_chat.')
208
+ raise NotImplementedError
209
+
210
+ if image_counts is not None:
211
+ num_patches_list = image_counts
212
+ print('Warning: `image_counts` is deprecated. Please use `num_patches_list` instead.')
213
+
214
+ img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
215
+ self.img_context_token_id = img_context_token_id
216
+
217
+ if verbose and pixel_values is not None:
218
+ image_bs = pixel_values.shape[0]
219
+ print(f'dynamic ViT batch size: {image_bs}')
220
+
221
+ queries = []
222
+ for idx, num_patches in enumerate(num_patches_list):
223
+ question = questions[idx]
224
+ if pixel_values is not None and '<image>' not in question:
225
+ question = '<image>\n' + question
226
+ template = get_conv_template(self.template)
227
+ template.system_message = self.system_message
228
+ template.append_message(template.roles[0], question)
229
+ template.append_message(template.roles[1], None)
230
+ query = template.get_prompt()
231
+
232
+ image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN
233
+ query = query.replace('<image>', image_tokens, 1)
234
+ queries.append(query)
235
+
236
+ tokenizer.padding_side = 'left'
237
+ model_inputs = tokenizer(queries, return_tensors='pt', padding=True)
238
+ input_ids = model_inputs['input_ids'].to(self.device)
239
+ attention_mask = model_inputs['attention_mask'].to(self.device)
240
+ eos_token_id = tokenizer.convert_tokens_to_ids(template.sep.strip())
241
+ generation_config['eos_token_id'] = eos_token_id
242
+ generation_output = self.generate(
243
+ pixel_values=pixel_values,
244
+ input_ids=input_ids,
245
+ attention_mask=attention_mask,
246
+ **generation_config
247
+ )
248
+ responses = tokenizer.batch_decode(generation_output, skip_special_tokens=True)
249
+ responses = [response.split(template.sep.strip())[0].strip() for response in responses]
250
+ return responses
251
+
252
+ def chat(self, tokenizer, pixel_values, question, generation_config, history=None, return_history=False,
253
+ num_patches_list=None, IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>', IMG_CONTEXT_TOKEN='<IMG_CONTEXT>',
254
+ verbose=False):
255
+
256
+ if history is None and pixel_values is not None and '<image>' not in question:
257
+ question = '<image>\n' + question
258
+
259
+ if num_patches_list is None:
260
+ num_patches_list = [pixel_values.shape[0]] if pixel_values is not None else []
261
+ assert pixel_values is None or len(pixel_values) == sum(num_patches_list)
262
+
263
+ img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
264
+ self.img_context_token_id = img_context_token_id
265
+
266
+ template = get_conv_template(self.template)
267
+ template.system_message = self.system_message
268
+ eos_token_id = tokenizer.convert_tokens_to_ids(template.sep.strip())
269
+
270
+ history = [] if history is None else history
271
+ for (old_question, old_answer) in history:
272
+ template.append_message(template.roles[0], old_question)
273
+ template.append_message(template.roles[1], old_answer)
274
+ template.append_message(template.roles[0], question)
275
+ template.append_message(template.roles[1], None)
276
+ query = template.get_prompt()
277
+
278
+ if verbose and pixel_values is not None:
279
+ image_bs = pixel_values.shape[0]
280
+ print(f'dynamic ViT batch size: {image_bs}')
281
+
282
+ for num_patches in num_patches_list:
283
+ image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN
284
+ query = query.replace('<image>', image_tokens, 1)
285
+
286
+ model_inputs = tokenizer(query, return_tensors='pt')
287
+ input_ids = model_inputs['input_ids'].to(self.device)
288
+ attention_mask = model_inputs['attention_mask'].to(self.device)
289
+ generation_config['eos_token_id'] = eos_token_id
290
+ generation_output = self.generate(
291
+ pixel_values=pixel_values,
292
+ input_ids=input_ids,
293
+ attention_mask=attention_mask,
294
+ **generation_config
295
+ )
296
+ response = tokenizer.batch_decode(generation_output, skip_special_tokens=True)[0]
297
+ response = response.split(template.sep.strip())[0].strip()
298
+ history.append((question, response))
299
+ if return_history:
300
+ return response, history
301
+ else:
302
+ query_to_print = query.replace(IMG_CONTEXT_TOKEN, '')
303
+ query_to_print = query_to_print.replace(f'{IMG_START_TOKEN}{IMG_END_TOKEN}', '<image>')
304
+ if verbose:
305
+ print(query_to_print, response)
306
+ return response
307
+
308
+ @torch.no_grad()
309
+ def generate(
310
+ self,
311
+ pixel_values: Optional[torch.FloatTensor] = None,
312
+ input_ids: Optional[torch.FloatTensor] = None,
313
+ attention_mask: Optional[torch.LongTensor] = None,
314
+ visual_features: Optional[torch.FloatTensor] = None,
315
+ generation_config: Optional[GenerationConfig] = None,
316
+ output_hidden_states: Optional[bool] = None,
317
+ **generate_kwargs,
318
+ ) -> torch.LongTensor:
319
+
320
+ assert self.img_context_token_id is not None
321
+ if pixel_values is not None:
322
+ if visual_features is not None:
323
+ vit_embeds = visual_features
324
+ else:
325
+ vit_embeds = self.extract_feature(pixel_values)
326
+ input_embeds = self.language_model.get_input_embeddings()(input_ids)
327
+ B, N, C = input_embeds.shape
328
+ input_embeds = input_embeds.reshape(B * N, C)
329
+
330
+ input_ids = input_ids.reshape(B * N)
331
+ selected = (input_ids == self.img_context_token_id)
332
+ assert selected.sum() != 0
333
+ input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device)
334
+
335
+ input_embeds = input_embeds.reshape(B, N, C)
336
+ else:
337
+ input_embeds = self.language_model.get_input_embeddings()(input_ids)
338
+
339
+ outputs = self.language_model.generate(
340
+ inputs_embeds=input_embeds,
341
+ attention_mask=attention_mask,
342
+ generation_config=generation_config,
343
+ output_hidden_states=output_hidden_states,
344
+ use_cache=True,
345
+ **generate_kwargs,
346
+ )
347
+
348
+ return outputs