File size: 5,892 Bytes
d2c4c68
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from .vision_encoder import VisionEncoder
from .configuration_moondream import MoondreamConfig
from transformers import PreTrainedModel

from .modeling_phi import PhiForCausalLM
from .configuration_moondream import PhiConfig


class Moondream(PreTrainedModel):
    config_class = MoondreamConfig
    _supports_flash_attn_2 = True

    def __init__(self, config):
        super().__init__(config)
        self.vision_encoder = VisionEncoder(
            use_flash_attn=config._attn_implementation == "flash_attention_2"
        )

        if isinstance(config.text_config, dict):
            phi_config = PhiConfig(
                **config.text_config, attn_implementation=config._attn_implementation
            )
        else:
            phi_config = config.text_config
        self.text_model = PhiForCausalLM(phi_config)

    @property
    def device(self):
        return self.text_model.device

    def encode_image(self, image):
        with torch.no_grad():
            return self.vision_encoder(image)

    def input_embeds(self, prompt, image_embeds, tokenizer):
        def _tokenize(txt):
            return tokenizer(
                txt, return_tensors="pt", add_special_tokens=False
            ).input_ids.to(self.device)

        text_emb = self.text_model.get_input_embeddings()

        # Add BOS token
        embeds = []
        embeds.append(
            text_emb((torch.tensor([[tokenizer.bos_token_id]], device=self.device)))
        )

        if "<image>" not in prompt:
            embeds.append(text_emb(_tokenize(prompt)))
        else:
            assert prompt.count("<image>") == 1
            before, after = prompt.split("<image>")
            if len(before) > 0:
                embeds.append(text_emb(_tokenize(before)))
            embeds.append(image_embeds.to(self.device))
            if len(after) > 0:
                embeds.append(text_emb(_tokenize(after)))

        return torch.cat(embeds, dim=1)

    def get_input_embeddings(self):
        return self.text_model.get_input_embeddings()

    def generate(
        self,
        image_embeds,
        prompt,
        tokenizer,
        max_new_tokens=128,
        **kwargs,
    ):
        generate_config = {
            "eos_token_id": tokenizer.eos_token_id,
            "bos_token_id": tokenizer.bos_token_id,
            "pad_token_id": tokenizer.bos_token_id,
            "max_new_tokens": max_new_tokens,
            **kwargs,
        }

        with torch.no_grad():
            inputs_embeds = self.input_embeds(prompt, image_embeds, tokenizer)
            output_ids = self.text_model.generate(
                inputs_embeds=inputs_embeds, **generate_config
            )

        return tokenizer.batch_decode(output_ids, skip_special_tokens=True)

    def answer_question(
        self,
        image_embeds,
        question,
        tokenizer,
        chat_history="",
        result_queue=None,
        **kwargs,
    ):
        prompt = f"<image>\n\n{chat_history}Question: {question}\n\nAnswer:"
        answer = self.generate(
            image_embeds,
            prompt,
            tokenizer=tokenizer,
            max_new_tokens=512,
            **kwargs,
        )[0]
        cleaned_answer = answer.strip()

        # Use the result_queue to pass the result if it is provided
        if result_queue:
            result_queue.put(cleaned_answer)
        else:
            return cleaned_answer

    def batch_answer(
        self,
        images,
        prompts,
        tokenizer,
        **kwargs,
    ):
        image_embeds = self.encode_image(images)

        templated_prompts = [
            f"<image>\n\nQuestion: {prompt}\n\nAnswer:" for prompt in prompts
        ]
        prompt_embs = [
            self.input_embeds(prompt, image_embed.unsqueeze(0), tokenizer)[0]
            for prompt, image_embed in zip(templated_prompts, image_embeds)
        ]

        bos_emb = prompt_embs[0][0]
        max_len = max([p.shape[0] for p in prompt_embs])

        inputs_embeds = torch.cat(
            [
                torch.cat([bos_emb.repeat(max_len - p.shape[0], 1), p]).unsqueeze(0)
                for p in prompt_embs
            ],
            dim=0,
        )
        attention_mask = torch.cat(
            [
                torch.cat(
                    [
                        torch.zeros(
                            1,
                            max_len - p.shape[0],
                            device=self.device,
                            dtype=torch.long,
                        ),
                        torch.ones(1, p.shape[0], device=self.device, dtype=torch.long),
                    ],
                    dim=1,
                )
                for p in prompt_embs
            ],
            dim=0,
        )

        generate_config = {
            "eos_token_id": tokenizer.eos_token_id,
            "bos_token_id": tokenizer.bos_token_id,
            "pad_token_id": tokenizer.bos_token_id,
            "max_new_tokens": 512,
            **kwargs,
        }

        with torch.no_grad():
            output_ids = self.text_model.generate(
                inputs_embeds=inputs_embeds,
                attention_mask=attention_mask,
                **generate_config,
            )

        return [
            x.strip()
            for x in tokenizer.batch_decode(output_ids, skip_special_tokens=True)
        ]

    def forward(self, images, tokens, attn_mask, labels=None):
        with torch.no_grad():
            img_embs = self.vision_encoder(images)

        tok_embs = self.text_model.get_input_embeddings()(tokens)
        inputs_embeds = torch.cat(
            (tok_embs[:, 0:1, :], img_embs, tok_embs[:, 1:, :]), dim=1
        )
        outputs = self.text_model(
            inputs_embeds=inputs_embeds,
            labels=labels,
            attention_mask=attn_mask,
        )
        return outputs