File size: 11,235 Bytes
3324de2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
# Code from https://github.com/ali-vilab/TeaCache/blob/main/TeaCache4TangoFlux/teacache_tango_flux.py

from typing import Any, Dict, Optional, Union
from diffusers.models.modeling_outputs import Transformer2DModelOutput
from diffusers.utils import (
    USE_PEFT_BACKEND,
    is_torch_version,
    logging,
    scale_lora_layers,
    unscale_lora_layers,
)
import torch
import numpy as np


logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


def teacache_forward(
    self,
    hidden_states: torch.Tensor,
    encoder_hidden_states: torch.Tensor = None,
    pooled_projections: torch.Tensor = None,
    timestep: torch.LongTensor = None,
    img_ids: torch.Tensor = None,
    txt_ids: torch.Tensor = None,
    guidance: torch.Tensor = None,
    joint_attention_kwargs: Optional[Dict[str, Any]] = None,
    return_dict: bool = True,
) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
    """
    The [`FluxTransformer2DModel`] 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.
        pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected
            from the embeddings of input conditions.
        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.
        joint_attention_kwargs (`dict`, *optional*):
            A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
            `self.processor` in
            [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
        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.
    """
    if joint_attention_kwargs is not None:
        joint_attention_kwargs = joint_attention_kwargs.copy()
        lora_scale = joint_attention_kwargs.pop("scale", 1.0)
    else:
        lora_scale = 1.0

    if USE_PEFT_BACKEND:
        # weight the lora layers by setting `lora_scale` for each PEFT layer
        scale_lora_layers(self, lora_scale)
    else:
        if (
            joint_attention_kwargs is not None
            and joint_attention_kwargs.get("scale", None) is not None
        ):
            logger.warning(
                "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
            )
    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_text_embed(timestep, pooled_projections)
        if guidance is None
        else self.time_text_embed(timestep, guidance, pooled_projections)
    )
    encoder_hidden_states = self.context_embedder(encoder_hidden_states)

    ids = torch.cat((txt_ids, img_ids), dim=1)
    image_rotary_emb = self.pos_embed(ids)

    if self.enable_teacache:
        inp = hidden_states.clone()
        temb_ = temb.clone()
        modulated_inp, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
            self.transformer_blocks[0].norm1(inp, emb=temb_)
        )
        if self.cnt == 0 or self.cnt == self.num_steps - 1:
            should_calc = True
            self.accumulated_rel_l1_distance = 0
        else:
            coefficients = [
                4.98651651e02,
                -2.83781631e02,
                5.58554382e01,
                -3.82021401e00,
                2.64230861e-01,
            ]
            rescale_func = np.poly1d(coefficients)
            self.accumulated_rel_l1_distance += rescale_func(
                (
                    (modulated_inp - self.previous_modulated_input).abs().mean()
                    / self.previous_modulated_input.abs().mean()
                )
                .cpu()
                .item()
            )
            if self.accumulated_rel_l1_distance < self.rel_l1_thresh:
                should_calc = False
            else:
                should_calc = True
                self.accumulated_rel_l1_distance = 0
        self.previous_modulated_input = modulated_inp
        self.cnt += 1
        if self.cnt == self.num_steps:
            self.cnt = 0

    if self.enable_teacache:
        if not should_calc:
            hidden_states += self.previous_residual
        else:
            ori_hidden_states = hidden_states.clone()
            for index_block, block in enumerate(self.transformer_blocks):
                if self.training and self.gradient_checkpointing:

                    def create_custom_forward(module, return_dict=None):
                        def custom_forward(*inputs):
                            if return_dict is not None:
                                return module(*inputs, return_dict=return_dict)
                            else:
                                return module(*inputs)

                        return custom_forward

                    ckpt_kwargs: Dict[str, Any] = (
                        {"use_reentrant": False}
                        if is_torch_version(">=", "1.11.0")
                        else {}
                    )
                    encoder_hidden_states, hidden_states = (
                        torch.utils.checkpoint.checkpoint(
                            create_custom_forward(block),
                            hidden_states,
                            encoder_hidden_states,
                            temb,
                            image_rotary_emb,
                            **ckpt_kwargs,
                        )
                    )

                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 = torch.cat([encoder_hidden_states, hidden_states], dim=1)

            for index_block, block in enumerate(self.single_transformer_blocks):
                if self.training and self.gradient_checkpointing:

                    def create_custom_forward(module, return_dict=None):
                        def custom_forward(*inputs):
                            if return_dict is not None:
                                return module(*inputs, return_dict=return_dict)
                            else:
                                return module(*inputs)

                        return custom_forward

                    ckpt_kwargs: Dict[str, Any] = (
                        {"use_reentrant": False}
                        if is_torch_version(">=", "1.11.0")
                        else {}
                    )
                    hidden_states = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(block),
                        hidden_states,
                        temb,
                        image_rotary_emb,
                        **ckpt_kwargs,
                    )

                else:
                    hidden_states = block(
                        hidden_states=hidden_states,
                        temb=temb,
                        image_rotary_emb=image_rotary_emb,
                    )

            hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]
            self.previous_residual = hidden_states - ori_hidden_states
    else:
        for index_block, block in enumerate(self.transformer_blocks):
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

                ckpt_kwargs: Dict[str, Any] = (
                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                )
                encoder_hidden_states, hidden_states = (
                    torch.utils.checkpoint.checkpoint(
                        create_custom_forward(block),
                        hidden_states,
                        encoder_hidden_states,
                        temb,
                        image_rotary_emb,
                        **ckpt_kwargs,
                    )
                )

            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 = torch.cat([encoder_hidden_states, hidden_states], dim=1)

        for index_block, block in enumerate(self.single_transformer_blocks):
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

                ckpt_kwargs: Dict[str, Any] = (
                    {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                )
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(block),
                    hidden_states,
                    temb,
                    image_rotary_emb,
                    **ckpt_kwargs,
                )

            else:
                hidden_states = block(
                    hidden_states=hidden_states,
                    temb=temb,
                    image_rotary_emb=image_rotary_emb,
                )

        hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]

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

    if USE_PEFT_BACKEND:
        # remove `lora_scale` from each PEFT layer
        unscale_lora_layers(self, lora_scale)

    if not return_dict:
        return (output,)

    return Transformer2DModelOutput(sample=output)