File size: 7,150 Bytes
d8bb2be
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
from dataclasses import dataclass
from .gemma import GemmaConfig, Gemma, KVCache
from .siglip import SigLIPConfig, SigLIPVisionTower
from typing import Optional
import os
import json
from pathlib import Path
from safetensors import safe_open

@dataclass
class PaliGemmaConfig:
    bos_token_id: int = 2
    eos_token_id: int = 1
    hidden_size: int = 2048
    ignore_index: int = -100
    image_token_index: int = 257152
    pad_token_id: int = 0
    projection_dim: int = 2048
    text_config: GemmaConfig = None
    vision_config: SigLIPConfig = None
    vocab_size: int = 257216
    @classmethod
    def from_dict(cls, data):
        return cls(
            bos_token_id = data['bos_token_id'],
            eos_token_id = data['eos_token_id'],
            hidden_size = data['hidden_size'],
            ignore_index = data['ignore_index'],
            image_token_index = data['image_token_index'],
            pad_token_id = data['pad_token_id'],
            projection_dim = data['projection_dim'],
            text_config = GemmaConfig.from_dict(data['text_config']),
            vision_config = SigLIPConfig.from_dict(data['vision_config'])
        )

class PaliGemmaMultimodalProjector(nn.Module):
    def __init__(self, cfg: PaliGemmaConfig):
        super().__init__()
        self.linear = nn.Linear(cfg.vision_config.hidden_size, cfg.vision_config.projection_dim)
    
    def forward(self, x: torch.Tensor):
        x = self.linear(x)
        return x
    
class PaliGemma(nn.Module):
    def __init__(self, cfg: PaliGemmaConfig):
        super().__init__()
        self.cfg = cfg
        self.language_model = Gemma(cfg.text_config)
        
        self.vision_tower = SigLIPVisionTower(cfg.vision_config)
        
        self.multi_modal_projector = PaliGemmaMultimodalProjector(cfg)
    
    def tie_weights(self):
        self.language_model.tie_weights()
        
    def _merge_img_embeds_and_input_embeds(self, img_embeds: torch.Tensor,
                                                input_embeds: torch.Tensor,
                                                input_tokens: torch.Tensor):
        batch_size, seq_len, embed_dim = input_embeds.shape
        scaled_img = img_embeds / (self.cfg.hidden_size ** 0.5)
        
        final_embeddings = torch.zeros((batch_size, seq_len, embed_dim), dtype=img_embeds.dtype, device=img_embeds.device)
        
        
        # (n, seq_len)
        text_mask = (input_tokens != self.cfg.pad_token_id) & (input_tokens != self.cfg.image_token_index)
        img_mask = input_tokens == self.cfg.image_token_index
        pad_mask = input_tokens == self.cfg.pad_token_id
        
        text_mask = text_mask.unsqueeze(-1).expand(-1, -1, embed_dim)
        img_mask = img_mask.unsqueeze(-1).expand(-1, -1, embed_dim)
        pad_mask = pad_mask.unsqueeze(-1).expand(-1, -1, embed_dim)
        
        # (n, seq_len, embed_dim)
        final_embeddings = torch.where(text_mask, input_embeds, final_embeddings)
        final_embeddings = final_embeddings.masked_scatter(img_mask, scaled_img)
        final_embeddings = torch.where(pad_mask, torch.zeros_like(final_embeddings), final_embeddings)
        
        return final_embeddings

    def _create_position_ids_and_attention_mask(self,
                                                device: str = '',
                                                dtype: torch.dtype = torch.float32,
                                                batch_size: int = 32,
                                                seq_len: int = 1,
                                                attention_mask: Optional[torch.Tensor] = None, 
                                                kv_cache: Optional[KVCache] = None):
        # Create Attention Mask
        if kv_cache is None or kv_cache.num_items() == 0:
            causal_mask = torch.full((batch_size, seq_len, seq_len), 0, dtype=dtype, device=device)
            position_ids = attention_mask.cumsum(dim=-1).masked_fill_((attention_mask == 0), 1).to(device)
        
        else:
            assert seq_len == 1
            kv_len = kv_cache.num_items() + 1
            causal_mask = torch.full((batch_size, 1, kv_len), 0, dtype=dtype, device=device)
            position_ids = attention_mask.cumsum(dim=-1)[:, -1].to(device)
        
        # (n, seq_len, kv_len) -> (n, 1, seq_len, kv_len)
        causal_mask = causal_mask.unsqueeze(1)
        
        return position_ids, causal_mask

    @staticmethod
    def from_pretrained(model_dir):
        with open(os.path.join(model_dir, 'config.json'), "r") as f:
            model_config = json.loads(f.read())
        config = PaliGemmaConfig.from_dict(model_config)
    
        safetensor_files = Path(model_dir).glob("*.safetensors")
        
        weights = {}
        for file in safetensor_files:
            with safe_open(file, framework='pt', device="cpu") as f:
                for key in f.keys():
                    weights[key] = f.get_tensor(key)
        
        model = PaliGemma(config)
        model.load_state_dict(weights, strict=False)
        model.tie_weights()
        return model
        
        
    def forward(self, *args, **kwargs):
        
        # input_tokens: (n, seq_len)
        
        # -> (n, seq_len, embed_dim)
        kv_cache = kwargs['kv_cache'] if 'kv_cache' in kwargs else None
        input_tokens = kwargs['input_ids']
        pixel_values = kwargs['pixel_values'] if 'pixel_values' in kwargs else None
        attention_mask = kwargs['attention_mask']
        input_embeds = self.language_model.model.embed_tokens(input_tokens)
        if pixel_values is not None:
            img_embeds = self.vision_tower(pixel_values.to(input_embeds.dtype))
            img_embeds = self.multi_modal_projector(img_embeds)
            final_embeddings = self._merge_img_embeds_and_input_embeds(img_embeds=img_embeds,
                                                                        input_embeds=input_embeds,
                                                                        input_tokens=input_tokens)
        else:
            final_embeddings = input_embeds

        position_ids, causal_mask = self._create_position_ids_and_attention_mask(device=final_embeddings.device.type,
                                                                                    dtype=final_embeddings.dtype,
                                                                                    batch_size=final_embeddings.shape[0],
                                                                                    seq_len=final_embeddings.shape[1],
                                                                                    attention_mask=attention_mask,
                                                                                    kv_cache=kv_cache)
        
        outputs, kv_cache = self.language_model(
            input_embeds=final_embeddings,
            position_ids=position_ids,
            attention_mask=causal_mask,
            kv_cache=kv_cache
        )
        return outputs, kv_cache