File size: 17,592 Bytes
3e936b2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from utils.wan_wrapper import WanDiffusionWrapper
from utils.scheduler import SchedulerInterface
from typing import List, Optional, Tuple
import torch
import torch.distributed as dist

class SelfForcingTrainingPipeline:

    def __init__(self, denoising_step_list: List[int], scheduler: SchedulerInterface, generator: WanDiffusionWrapper, num_frame_per_block=3, independent_first_frame: bool=False, same_step_across_blocks: bool=False, last_step_only: bool=False, num_max_frames: int=21, context_noise: int=0, **kwargs):
        super().__init__()
        self.scheduler = scheduler
        self.generator = generator
        self.denoising_step_list = denoising_step_list
        if self.denoising_step_list[-1] == 0:
            self.denoising_step_list = self.denoising_step_list[:-1]
        self.num_transformer_blocks = 30
        self.frame_seq_length = 1560
        self.num_frame_per_block = num_frame_per_block
        self.context_noise = context_noise
        self.i2v = False
        self.kv_cache1 = None
        self.kv_cache2 = None
        self.crossattn_cache = None
        self.independent_first_frame = independent_first_frame
        self.same_step_across_blocks = same_step_across_blocks
        self.last_step_only = last_step_only
        self.local_attn_size = kwargs.get('local_attn_size', -1)
        if not isinstance(self.local_attn_size, int) and hasattr(self.local_attn_size, '__iter__'):
            self.local_attn_size = list(self.local_attn_size)
        if isinstance(self.local_attn_size, (list, tuple)):
            assert len(self.local_attn_size) == len(self.denoising_step_list), f'local_attn_size length ({len(self.local_attn_size)}) must match denoising_step_list length ({len(self.denoising_step_list)}).'
        num_training_frames: Optional[int] = kwargs.get('num_training_frames', 21)
        slice_last_frames: int = int(kwargs.get('slice_last_frames', 21))

        def _resolve_kv_frames(local_cfg):
            if isinstance(local_cfg, (list, tuple)):
                base = int(max(local_cfg)) if len(local_cfg) > 0 else -1
                return min(base + slice_last_frames, num_training_frames)
            else:
                base = int(local_cfg)
                return min(base + slice_last_frames, num_training_frames)
        kv_frames = _resolve_kv_frames(self.local_attn_size)
        self.kv_cache_size = int(kv_frames) * self.frame_seq_length

    def generate_and_sync_list(self, num_blocks, num_denoising_steps, device):
        rank = dist.get_rank() if dist.is_initialized() else 0
        if rank == 0:
            indices = torch.randint(low=0, high=num_denoising_steps, size=(num_blocks,), device=device)
            if self.last_step_only:
                indices = torch.ones_like(indices) * (num_denoising_steps - 1)
        else:
            indices = torch.empty(num_blocks, dtype=torch.long, device=device)
        if dist.is_initialized():
            dist.broadcast(indices, src=0)
        return indices.tolist()

    def generate_chunk_with_cache(self, noise: torch.Tensor, conditional_dict: dict, *, current_start_frame: int=0, requires_grad: bool=True, return_sim_step: bool=False) -> Tuple[torch.Tensor, Optional[int], Optional[int]]:
        batch_size, chunk_frames, num_channels, height, width = noise.shape
        if not self.independent_first_frame or chunk_frames % self.num_frame_per_block == 0:
            assert chunk_frames % self.num_frame_per_block == 0
            num_blocks = chunk_frames // self.num_frame_per_block
            all_num_frames = [self.num_frame_per_block] * num_blocks
        else:
            assert (chunk_frames - 1) % self.num_frame_per_block == 0
            num_blocks = (chunk_frames - 1) // self.num_frame_per_block
            all_num_frames = [1] + [self.num_frame_per_block] * num_blocks
        output = torch.zeros_like(noise)
        num_denoising_steps = len(self.denoising_step_list)
        exit_flags = self.generate_and_sync_list(len(all_num_frames), num_denoising_steps, device=noise.device)
        if not requires_grad:
            start_gradient_frame_index = chunk_frames
        else:
            start_gradient_frame_index = 0
        local_start_frame = 0
        if not (isinstance(self.local_attn_size, (list, tuple)) or (hasattr(self.local_attn_size, '__iter__') and (not isinstance(self.local_attn_size, (str, bytes))))):
            self.generator.model.local_attn_size = int(self.local_attn_size)
            self._set_all_modules_max_attention_size(int(self.local_attn_size))
        for block_index, current_num_frames in enumerate(all_num_frames):
            noisy_input = noise[:, local_start_frame:local_start_frame + current_num_frames]
            for step_idx, current_timestep in enumerate(self.denoising_step_list):
                if isinstance(self.local_attn_size, (list, tuple)) or (hasattr(self.local_attn_size, '__iter__') and (not isinstance(self.local_attn_size, (str, bytes)))):
                    self.generator.model.local_attn_size = int(self.local_attn_size[step_idx])
                    self._set_all_modules_max_attention_size(int(self.local_attn_size[step_idx]))
                exit_flag = step_idx == exit_flags[0] if self.same_step_across_blocks else step_idx == exit_flags[block_index]
                timestep = torch.ones([batch_size, current_num_frames], device=noise.device, dtype=torch.int64) * current_timestep
                if not exit_flag:
                    with torch.no_grad():
                        _, denoised_pred = self.generator(noisy_image_or_video=noisy_input, conditional_dict=conditional_dict, timestep=timestep, kv_cache=self.kv_cache1, crossattn_cache=self.crossattn_cache, current_start=(current_start_frame + local_start_frame) * self.frame_seq_length)
                        if step_idx < len(self.denoising_step_list) - 1:
                            next_timestep = self.denoising_step_list[step_idx + 1]
                            noisy_input = self.scheduler.add_noise(denoised_pred.flatten(0, 1), torch.randn_like(denoised_pred.flatten(0, 1)), next_timestep * torch.ones([batch_size * current_num_frames], device=noise.device, dtype=torch.long)).unflatten(0, denoised_pred.shape[:2])
                else:
                    enable_grad = local_start_frame >= start_gradient_frame_index
                    context_manager = torch.enable_grad() if enable_grad else torch.no_grad()
                    with context_manager:
                        _, denoised_pred = self.generator(noisy_image_or_video=noisy_input, conditional_dict=conditional_dict, timestep=timestep, kv_cache=self.kv_cache1, crossattn_cache=self.crossattn_cache, current_start=(current_start_frame + local_start_frame) * self.frame_seq_length)
                    break
            output[:, local_start_frame:local_start_frame + current_num_frames] = denoised_pred
            context_timestep = torch.ones_like(timestep) * self.context_noise
            context_noisy = self.scheduler.add_noise(denoised_pred.flatten(0, 1), torch.randn_like(denoised_pred.flatten(0, 1)), context_timestep.flatten(0, 1)).unflatten(0, denoised_pred.shape[:2])
            with torch.no_grad():
                self.generator(noisy_image_or_video=context_noisy, conditional_dict=conditional_dict, timestep=context_timestep, kv_cache=self.kv_cache1, crossattn_cache=self.crossattn_cache, current_start=(current_start_frame + local_start_frame) * self.frame_seq_length)
            local_start_frame += current_num_frames
        if not self.same_step_across_blocks:
            denoised_timestep_from, denoised_timestep_to = (None, None)
        elif exit_flags[0] == len(self.denoising_step_list) - 1:
            denoised_timestep_to = 0
            denoised_timestep_from = 1000 - torch.argmin((self.scheduler.timesteps.cuda() - self.denoising_step_list[exit_flags[0]].cuda()).abs(), dim=0).item()
        else:
            denoised_timestep_to = 1000 - torch.argmin((self.scheduler.timesteps.cuda() - self.denoising_step_list[exit_flags[0] + 1].cuda()).abs(), dim=0).item()
            denoised_timestep_from = 1000 - torch.argmin((self.scheduler.timesteps.cuda() - self.denoising_step_list[exit_flags[0]].cuda()).abs(), dim=0).item()
        if return_sim_step:
            return (output, denoised_timestep_from, denoised_timestep_to, exit_flags[0] + 1)
        return (output, denoised_timestep_from, denoised_timestep_to)

    def inference_with_trajectory(self, noise: torch.Tensor, initial_latent: Optional[torch.Tensor]=None, return_sim_step: bool=False, slice_last_frames: int=21, **conditional_dict) -> torch.Tensor:
        batch_size, num_frames, num_channels, height, width = noise.shape
        if not self.independent_first_frame or (self.independent_first_frame and initial_latent is not None):
            assert num_frames % self.num_frame_per_block == 0
            num_blocks = num_frames // self.num_frame_per_block
        else:
            assert (num_frames - 1) % self.num_frame_per_block == 0
            num_blocks = (num_frames - 1) // self.num_frame_per_block
        num_input_frames = initial_latent.shape[1] if initial_latent is not None else 0
        num_output_frames = num_frames + num_input_frames
        output = torch.zeros([batch_size, num_output_frames, num_channels, height, width], device=noise.device, dtype=noise.dtype)
        self._initialize_kv_cache(batch_size=batch_size, dtype=noise.dtype, device=noise.device)
        self._initialize_crossattn_cache(batch_size=batch_size, dtype=noise.dtype, device=noise.device)
        current_start_frame = 0
        if initial_latent is not None:
            timestep = torch.ones([batch_size, 1], device=noise.device, dtype=torch.int64) * 0
            output[:, :1] = initial_latent
            with torch.no_grad():
                self.generator(noisy_image_or_video=initial_latent, conditional_dict=conditional_dict, timestep=timestep * 0, kv_cache=self.kv_cache1, crossattn_cache=self.crossattn_cache, current_start=current_start_frame * self.frame_seq_length)
            current_start_frame += 1
        all_num_frames = [self.num_frame_per_block] * num_blocks
        if self.independent_first_frame and initial_latent is None:
            all_num_frames = [1] + all_num_frames
        num_denoising_steps = len(self.denoising_step_list)
        exit_flags = self.generate_and_sync_list(len(all_num_frames), num_denoising_steps, device=noise.device)
        start_gradient_frame_index = num_output_frames - slice_last_frames
        grad_enable_mask = torch.zeros((batch_size, sum(all_num_frames)), dtype=torch.bool)
        if not isinstance(self.local_attn_size, (list, tuple)):
            self.generator.model.local_attn_size = int(self.local_attn_size)
            self._set_all_modules_max_attention_size(int(self.local_attn_size))
        for block_index, current_num_frames in enumerate(all_num_frames):
            noisy_input = noise[:, current_start_frame - num_input_frames:current_start_frame + current_num_frames - num_input_frames]
            for index, current_timestep in enumerate(self.denoising_step_list):
                if isinstance(self.local_attn_size, (list, tuple)):
                    self.generator.model.local_attn_size = int(self.local_attn_size[index])
                    self._set_all_modules_max_attention_size(int(self.local_attn_size[index]))
                if self.same_step_across_blocks:
                    exit_flag = index == exit_flags[0]
                else:
                    exit_flag = index == exit_flags[block_index]
                timestep = torch.ones([batch_size, current_num_frames], device=noise.device, dtype=torch.int64) * current_timestep
                if not exit_flag:
                    with torch.no_grad():
                        _, denoised_pred = self.generator(noisy_image_or_video=noisy_input, conditional_dict=conditional_dict, timestep=timestep, kv_cache=self.kv_cache1, crossattn_cache=self.crossattn_cache, current_start=current_start_frame * self.frame_seq_length)
                        next_timestep = self.denoising_step_list[index + 1]
                        noisy_input = self.scheduler.add_noise(denoised_pred.flatten(0, 1), torch.randn_like(denoised_pred.flatten(0, 1)), next_timestep * torch.ones([batch_size * current_num_frames], device=noise.device, dtype=torch.long)).unflatten(0, denoised_pred.shape[:2])
                else:
                    if current_start_frame < start_gradient_frame_index:
                        grad_enable_mask[:, current_start_frame:current_start_frame + current_num_frames] = False
                        with torch.no_grad():
                            _, denoised_pred = self.generator(noisy_image_or_video=noisy_input, conditional_dict=conditional_dict, timestep=timestep, kv_cache=self.kv_cache1, crossattn_cache=self.crossattn_cache, current_start=current_start_frame * self.frame_seq_length)
                    else:
                        grad_enable_mask[:, current_start_frame:current_start_frame + current_num_frames] = True
                        _, denoised_pred = self.generator(noisy_image_or_video=noisy_input, conditional_dict=conditional_dict, timestep=timestep, kv_cache=self.kv_cache1, crossattn_cache=self.crossattn_cache, current_start=current_start_frame * self.frame_seq_length)
                    break
            output[:, current_start_frame:current_start_frame + current_num_frames] = denoised_pred
            context_timestep = torch.ones_like(timestep) * self.context_noise
            denoised_pred = self.scheduler.add_noise(denoised_pred.flatten(0, 1), torch.randn_like(denoised_pred.flatten(0, 1)), context_timestep * torch.ones([batch_size * current_num_frames], device=noise.device, dtype=torch.long)).unflatten(0, denoised_pred.shape[:2])
            with torch.no_grad():
                self.generator(noisy_image_or_video=denoised_pred, conditional_dict=conditional_dict, timestep=context_timestep, kv_cache=self.kv_cache1, crossattn_cache=self.crossattn_cache, current_start=current_start_frame * self.frame_seq_length)
            current_start_frame += current_num_frames
        if not self.same_step_across_blocks:
            denoised_timestep_from, denoised_timestep_to = (None, None)
        elif exit_flags[0] == len(self.denoising_step_list) - 1:
            denoised_timestep_to = 0
            denoised_timestep_from = 1000 - torch.argmin((self.scheduler.timesteps.cuda() - self.denoising_step_list[exit_flags[0]].cuda()).abs(), dim=0).item()
        else:
            denoised_timestep_to = 1000 - torch.argmin((self.scheduler.timesteps.cuda() - self.denoising_step_list[exit_flags[0] + 1].cuda()).abs(), dim=0).item()
            denoised_timestep_from = 1000 - torch.argmin((self.scheduler.timesteps.cuda() - self.denoising_step_list[exit_flags[0]].cuda()).abs(), dim=0).item()
        if return_sim_step:
            return (output, denoised_timestep_from, denoised_timestep_to, exit_flags[0] + 1)
        return (output, denoised_timestep_from, denoised_timestep_to)

    def _initialize_kv_cache(self, batch_size, dtype, device):
        kv_cache1 = []
        for _ in range(self.num_transformer_blocks):
            kv_cache1.append({'k': torch.zeros([batch_size, self.kv_cache_size, 12, 128], dtype=dtype, device=device), 'v': torch.zeros([batch_size, self.kv_cache_size, 12, 128], dtype=dtype, device=device), 'global_end_index': torch.tensor([0], dtype=torch.long, device=device), 'local_end_index': torch.tensor([0], dtype=torch.long, device=device)})
        self.kv_cache1 = kv_cache1

    def _initialize_crossattn_cache(self, batch_size, dtype, device):
        crossattn_cache = []
        for _ in range(self.num_transformer_blocks):
            crossattn_cache.append({'k': torch.zeros([batch_size, 512, 12, 128], dtype=dtype, device=device), 'v': torch.zeros([batch_size, 512, 12, 128], dtype=dtype, device=device), 'is_init': False})
        self.crossattn_cache = crossattn_cache

    def clear_kv_cache(self):
        if getattr(self, 'kv_cache1', None) is not None:
            for blk in self.kv_cache1:
                blk['k'].zero_()
                blk['v'].zero_()
                if 'global_end_index' in blk:
                    blk['global_end_index'].zero_()
                if 'local_end_index' in blk:
                    blk['local_end_index'].zero_()
        if getattr(self, 'crossattn_cache', None) is not None:
            for blk in self.crossattn_cache:
                blk['k'].zero_()
                blk['v'].zero_()
                blk['is_init'] = False

    def _set_all_modules_max_attention_size(self, local_attn_size_value: int):
        if isinstance(local_attn_size_value, (list, tuple)):
            raise ValueError('_set_all_modules_max_attention_size expects an int, got list/tuple.')
        if int(local_attn_size_value) == -1:
            target_size = 32760
            policy = 'global'
        else:
            target_size = int(local_attn_size_value) * self.frame_seq_length
            policy = 'local'
        if hasattr(self.generator.model, 'max_attention_size'):
            try:
                _ = getattr(self.generator.model, 'max_attention_size')
            except Exception:
                pass
            setattr(self.generator.model, 'max_attention_size', target_size)
        for name, module in self.generator.model.named_modules():
            if hasattr(module, 'max_attention_size'):
                try:
                    setattr(module, 'max_attention_size', target_size)
                except Exception:
                    pass