File size: 9,052 Bytes
b1bd80d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Taken from https://github.com/comfyanonymous/ComfyUI
# This file is only for reference, and not used in the backend or runtime.


import torch
from ldm_patched.ldm.modules.attention import optimized_attention_for_device

class CLIPAttention(torch.nn.Module):
    def __init__(self, embed_dim, heads, dtype, device, operations):
        super().__init__()

        self.heads = heads
        self.q_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
        self.k_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
        self.v_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)

        self.out_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)

    def forward(self, x, mask=None, optimized_attention=None):
        q = self.q_proj(x)
        k = self.k_proj(x)
        v = self.v_proj(x)

        out = optimized_attention(q, k, v, self.heads, mask)
        return self.out_proj(out)

ACTIVATIONS = {"quick_gelu": lambda a: a * torch.sigmoid(1.702 * a),
               "gelu": torch.nn.functional.gelu,
}

class CLIPMLP(torch.nn.Module):
    def __init__(self, embed_dim, intermediate_size, activation, dtype, device, operations):
        super().__init__()
        self.fc1 = operations.Linear(embed_dim, intermediate_size, bias=True, dtype=dtype, device=device)
        self.activation = ACTIVATIONS[activation]
        self.fc2 = operations.Linear(intermediate_size, embed_dim, bias=True, dtype=dtype, device=device)

    def forward(self, x):
        x = self.fc1(x)
        x = self.activation(x)
        x = self.fc2(x)
        return x

class CLIPLayer(torch.nn.Module):
    def __init__(self, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations):
        super().__init__()
        self.layer_norm1 = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
        self.self_attn = CLIPAttention(embed_dim, heads, dtype, device, operations)
        self.layer_norm2 = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
        self.mlp = CLIPMLP(embed_dim, intermediate_size, intermediate_activation, dtype, device, operations)

    def forward(self, x, mask=None, optimized_attention=None):
        x += self.self_attn(self.layer_norm1(x), mask, optimized_attention)
        x += self.mlp(self.layer_norm2(x))
        return x


class CLIPEncoder(torch.nn.Module):
    def __init__(self, num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations):
        super().__init__()
        self.layers = torch.nn.ModuleList([CLIPLayer(embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations) for i in range(num_layers)])

    def forward(self, x, mask=None, intermediate_output=None):
        optimized_attention = optimized_attention_for_device(x.device, mask=mask is not None, small_input=True)

        if intermediate_output is not None:
            if intermediate_output < 0:
                intermediate_output = len(self.layers) + intermediate_output

        intermediate = None
        for i, l in enumerate(self.layers):
            x = l(x, mask, optimized_attention)
            if i == intermediate_output:
                intermediate = x.clone()
        return x, intermediate

class CLIPEmbeddings(torch.nn.Module):
    def __init__(self, embed_dim, vocab_size=49408, num_positions=77, dtype=None, device=None):
        super().__init__()
        self.token_embedding = torch.nn.Embedding(vocab_size, embed_dim, dtype=dtype, device=device)
        self.position_embedding = torch.nn.Embedding(num_positions, embed_dim, dtype=dtype, device=device)

    def forward(self, input_tokens):
        return self.token_embedding(input_tokens) + self.position_embedding.weight


class CLIPTextModel_(torch.nn.Module):
    def __init__(self, config_dict, dtype, device, operations):
        num_layers = config_dict["num_hidden_layers"]
        embed_dim = config_dict["hidden_size"]
        heads = config_dict["num_attention_heads"]
        intermediate_size = config_dict["intermediate_size"]
        intermediate_activation = config_dict["hidden_act"]

        super().__init__()
        self.embeddings = CLIPEmbeddings(embed_dim, dtype=torch.float32, device=device)
        self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
        self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device)

    def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True):
        x = self.embeddings(input_tokens)
        mask = None
        if attention_mask is not None:
            mask = 1.0 - attention_mask.to(x.dtype).unsqueeze(1).unsqueeze(1).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
            mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))

        causal_mask = torch.empty(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
        if mask is not None:
            mask += causal_mask
        else:
            mask = causal_mask

        x, i = self.encoder(x, mask=mask, intermediate_output=intermediate_output)
        x = self.final_layer_norm(x)
        if i is not None and final_layer_norm_intermediate:
            i = self.final_layer_norm(i)

        pooled_output = x[torch.arange(x.shape[0], device=x.device), input_tokens.to(dtype=torch.int, device=x.device).argmax(dim=-1),]
        return x, i, pooled_output

class CLIPTextModel(torch.nn.Module):
    def __init__(self, config_dict, dtype, device, operations):
        super().__init__()
        self.num_layers = config_dict["num_hidden_layers"]
        self.text_model = CLIPTextModel_(config_dict, dtype, device, operations)
        self.dtype = dtype

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

    def set_input_embeddings(self, embeddings):
        self.text_model.embeddings.token_embedding = embeddings

    def forward(self, *args, **kwargs):
        return self.text_model(*args, **kwargs)

class CLIPVisionEmbeddings(torch.nn.Module):
    def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, dtype=None, device=None, operations=None):
        super().__init__()
        self.class_embedding = torch.nn.Parameter(torch.empty(embed_dim, dtype=dtype, device=device))

        self.patch_embedding = operations.Conv2d(
            in_channels=num_channels,
            out_channels=embed_dim,
            kernel_size=patch_size,
            stride=patch_size,
            bias=False,
            dtype=dtype,
            device=device
        )

        num_patches = (image_size // patch_size) ** 2
        num_positions = num_patches + 1
        self.position_embedding = torch.nn.Embedding(num_positions, embed_dim, dtype=dtype, device=device)

    def forward(self, pixel_values):
        embeds = self.patch_embedding(pixel_values).flatten(2).transpose(1, 2)
        return torch.cat([self.class_embedding.to(embeds.device).expand(pixel_values.shape[0], 1, -1), embeds], dim=1) + self.position_embedding.weight.to(embeds.device)


class CLIPVision(torch.nn.Module):
    def __init__(self, config_dict, dtype, device, operations):
        super().__init__()
        num_layers = config_dict["num_hidden_layers"]
        embed_dim = config_dict["hidden_size"]
        heads = config_dict["num_attention_heads"]
        intermediate_size = config_dict["intermediate_size"]
        intermediate_activation = config_dict["hidden_act"]

        self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], dtype=torch.float32, device=device, operations=operations)
        self.pre_layrnorm = operations.LayerNorm(embed_dim)
        self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
        self.post_layernorm = operations.LayerNorm(embed_dim)

    def forward(self, pixel_values, attention_mask=None, intermediate_output=None):
        x = self.embeddings(pixel_values)
        x = self.pre_layrnorm(x)
        #TODO: attention_mask?
        x, i = self.encoder(x, mask=None, intermediate_output=intermediate_output)
        pooled_output = self.post_layernorm(x[:, 0, :])
        return x, i, pooled_output

class CLIPVisionModelProjection(torch.nn.Module):
    def __init__(self, config_dict, dtype, device, operations):
        super().__init__()
        self.vision_model = CLIPVision(config_dict, dtype, device, operations)
        self.visual_projection = operations.Linear(config_dict["hidden_size"], config_dict["projection_dim"], bias=False)

    def forward(self, *args, **kwargs):
        x = self.vision_model(*args, **kwargs)
        out = self.visual_projection(x[2])
        return (x[0], x[1], out)