File size: 9,088 Bytes
f06aba5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import Any, Dict, List, Optional, Union

import torch
import torch.nn as nn
import torch.nn.functional as F

from diffusers.models.transformers.transformer_flux import \
    FluxTransformerBlock, FluxSingleTransformerBlock, \
    AdaLayerNormContinuous, Transformer2DModelOutput
from diffusers.models.embeddings import Timesteps, TimestepEmbedding,FluxPosEmbed
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.modeling_utils import ModelMixin
from accelerate.logging import get_logger
from diffusers.loaders import PeftAdapterMixin

logger = get_logger(__name__, log_level="INFO")


class TimestepEmbeddings(nn.Module):
    def __init__(self, embedding_dim):
        super().__init__()

        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)

    def forward(self, timestep, hidden_dtype):
        timesteps_proj = self.time_proj(timestep)
        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, D)

        return timesteps_emb
    

class LongCatImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin ):
    """
    The Transformer model introduced in Flux.
    """

    _supports_gradient_checkpointing = True

    @register_to_config
    def __init__(
        self,
        patch_size: int = 1,
        in_channels: int = 64,
        num_layers: int = 19,
        num_single_layers: int = 38,
        attention_head_dim: int = 128,
        num_attention_heads: int = 24,
        joint_attention_dim: int = 3584,
        pooled_projection_dim: int = 3584,
        axes_dims_rope: List[int] = [16, 56, 56],
    ):
        super().__init__()
        self.out_channels = in_channels
        self.inner_dim = num_attention_heads * attention_head_dim
        self.pooled_projection_dim = pooled_projection_dim

        self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)

        self.time_embed = TimestepEmbeddings(embedding_dim=self.inner_dim)

        self.context_embedder = nn.Linear(joint_attention_dim, self.inner_dim)
        self.x_embedder = torch.nn.Linear(in_channels, self.inner_dim)

        self.transformer_blocks = nn.ModuleList(
            [
                FluxTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=num_attention_heads,
                    attention_head_dim=attention_head_dim,
                )
                for i in range(num_layers)
            ]
        )

        self.single_transformer_blocks = nn.ModuleList(
            [
                FluxSingleTransformerBlock(
                    dim=self.inner_dim,
                    num_attention_heads=num_attention_heads,
                    attention_head_dim=attention_head_dim,
                )
                for i in range(num_single_layers)
            ]
        )

        self.norm_out = AdaLayerNormContinuous(
            self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6)
        self.proj_out = nn.Linear(
            self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)

        self.gradient_checkpointing = False

        self.initialize_weights()

        self.use_checkpoint = [True] * num_layers
        self.use_single_checkpoint = [True] * num_single_layers

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: torch.Tensor = None,
        timestep: torch.LongTensor = None,
        img_ids: torch.Tensor = None,
        txt_ids: torch.Tensor = None,
        guidance: torch.Tensor = None,
        return_dict: bool = True,
    ) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
        """
        The  forward method.

        Args:
            hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
                Input `hidden_states`.
            encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
            timestep ( `torch.LongTensor`):
                Used to indicate denoising step.
            block_controlnet_hidden_states: (`list` of `torch.Tensor`):
                A list of tensors that if specified are added to the residuals of transformer blocks.
            return_dict (`bool`, *optional*, defaults to `True`):
                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
                tuple.

        Returns:
            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
            `tuple` where the first element is the sample tensor.
        """
        hidden_states = self.x_embedder(hidden_states)

        timestep = timestep.to(hidden_states.dtype) * 1000
        if guidance is not None:
            guidance = guidance.to(hidden_states.dtype) * 1000
        else:
            guidance = None

        temb = self.time_embed( timestep, hidden_states.dtype )
        encoder_hidden_states = self.context_embedder(encoder_hidden_states)

        if txt_ids.ndim == 3:
            logger.warning(
                "Passing `txt_ids` 3d torch.Tensor is deprecated."
                "Please remove the batch dimension and pass it as a 2d torch Tensor"
            )
            txt_ids = txt_ids[0]
        if img_ids.ndim == 3:
            logger.warning(
                "Passing `img_ids` 3d torch.Tensor is deprecated."
                "Please remove the batch dimension and pass it as a 2d torch Tensor"
            )
            img_ids = img_ids[0]

        ids = torch.cat((txt_ids, img_ids), dim=0)
        image_rotary_emb = self.pos_embed(ids)

        for index_block, block in enumerate(self.transformer_blocks):
            if torch.is_grad_enabled() and self.gradient_checkpointing and self.use_checkpoint[index_block]:
                encoder_hidden_states, hidden_states = self._gradient_checkpointing_func(
                    block,
                    hidden_states,
                    encoder_hidden_states,
                    temb,
                    image_rotary_emb,
                )
            else:
                encoder_hidden_states, hidden_states = block(
                    hidden_states=hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    temb=temb,
                    image_rotary_emb=image_rotary_emb,
                )

        for index_block, block in enumerate(self.single_transformer_blocks):
            if torch.is_grad_enabled() and self.gradient_checkpointing and self.use_single_checkpoint[index_block]:
                encoder_hidden_states,hidden_states = self._gradient_checkpointing_func(
                    block,
                    hidden_states,
                    encoder_hidden_states,
                    temb,
                    image_rotary_emb,
                )
            else:
                encoder_hidden_states, hidden_states = block(
                    hidden_states=hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    temb=temb,
                    image_rotary_emb=image_rotary_emb,
                )

        hidden_states = self.norm_out(hidden_states, temb)
        output = self.proj_out(hidden_states)

        if not return_dict:
            return (output,)

        return Transformer2DModelOutput(sample=output)

    def initialize_weights(self):
        # Initialize transformer layers:
        def _basic_init(module):
            if isinstance(module, nn.Linear):
                torch.nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0)
        self.apply(_basic_init)

        # Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
        w = self.x_embedder.weight.data
        nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
        nn.init.constant_(self.x_embedder.bias, 0)

        # Initialize caption embedding MLP:
        nn.init.normal_(self.context_embedder.weight, std=0.02)

        # Zero-out adaLN modulation layers in blocks:
        for block in self.transformer_blocks:
            nn.init.constant_(block.norm1.linear.weight, 0)
            nn.init.constant_(block.norm1.linear.bias, 0)
            nn.init.constant_(block.norm1_context.linear.weight, 0)
            nn.init.constant_(block.norm1_context.linear.bias, 0)

        for block in self.single_transformer_blocks:
            nn.init.constant_(block.norm.linear.weight, 0)
            nn.init.constant_(block.norm.linear.bias, 0)

        # Zero-out output layers:
        nn.init.constant_(self.norm_out.linear.weight, 0)
        nn.init.constant_(self.norm_out.linear.bias, 0)
        nn.init.constant_(self.proj_out.weight, 0)
        nn.init.constant_(self.proj_out.bias, 0)