File size: 17,365 Bytes
5c93746
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
from models import (
    get_diffusion_wrapper,
    get_text_encoder_wrapper,
    get_vae_wrapper
)
from models.wan.taehv_wrapper import TAEHVWanVAEWrapper
from typing import List
import torch
import torch.distributed as dist
import logging

LOGGER = logging.getLogger(__name__)

class CausalStreamInferencePipeline(torch.nn.Module):
    def __init__(self, args, device):
        super().__init__()
        model_type = args.model_type
        self.device = device
        # Step 1: Initialize all models
        self.generator_model_name = getattr(
            args, "generator_name", args.model_name)
        self.generator = get_diffusion_wrapper(
            model_name=self.generator_model_name)(model_type=model_type)
        self.text_encoder = get_text_encoder_wrapper(
            model_name=args.model_name)(model_type=model_type)
        if getattr(args, "use_taehv", False):
            LOGGER.info("Using TAEHV VAE wrapper for Wan inference")
            self.vae = TAEHVWanVAEWrapper(
                model_type=model_type,
                checkpoint_path=getattr(args, "taehv_checkpoint_path", None),
                use_tensorrt=getattr(args, "use_tensorrt", False),
            )
        else:
            self.vae = get_vae_wrapper(model_name=args.model_name)(model_type=model_type)

        # Step 2: Initialize all causal hyperparmeters
        self._init_denoising_step_list(args, device)

        if model_type == "T2V-1.3B":
            self.num_transformer_blocks = 30
            self.num_heads = 12
        elif model_type == "T2V-14B":
            self.num_transformer_blocks = 40
            self.num_heads = 40
        else:
            raise ValueError(f"Model type {model_type} not supported")
        scale_size = 16
        self.height = args.height//scale_size*2
        self.width = args.width//scale_size*2
        self.frame_seq_length = (args.height//scale_size) * (args.width//scale_size)
        self.num_kv_cache = args.num_kv_cache
        self.kv_cache_length = self.frame_seq_length*self.num_kv_cache
        self.num_sink_tokens = args.num_sink_tokens
        self.adapt_sink_threshold = args.adapt_sink_threshold

        self.conditional_dict = None
        self.kv_cache1 = None
        self.kv_cache2 = None
        self.hidden_states = None
        self.block_x = None
        self.args = args
        self.num_frame_per_block = getattr(
            args, "num_frame_per_block", 1)

        LOGGER.info("KV inference with %s frames per block", self.num_frame_per_block)

        if self.num_frame_per_block > 1:
            self.generator.model.num_frame_per_block = self.num_frame_per_block

        self.generator.model.to(self.device)

    def _init_denoising_step_list(self, args, device):
        self.denoising_step_list = torch.tensor(
            args.denoising_step_list, dtype=torch.long, device=device)
        assert self.denoising_step_list[-1] == 0
        if not args.t2v:
            # remove the last timestep (which equals zero)
            self.denoising_step_list = self.denoising_step_list[:-1]

        self.scheduler = self.generator.get_scheduler()
        if args.warp_denoising_step:  # Warp the denoising step according to the scheduler time shift
            timesteps = torch.cat(
                (self.scheduler.timesteps.cpu(), torch.tensor([0], dtype=torch.float32))
            ).to(device)
            self.denoising_step_list = timesteps[1000 - self.denoising_step_list]

    def _initialize_kv_cache(self, batch_size, dtype, device):
        """
        Initialize a Per-GPU KV cache for the Wan model.
        """
        kv_cache1 = []
        
        for i in range(self.num_transformer_blocks):
            cache_length = self.kv_cache_length
            self.generator.model.blocks[i].self_attn.sink_size = self.num_sink_tokens
            self.generator.model.blocks[i].self_attn.adapt_sink_thr = self.adapt_sink_threshold

            kv_cache1.append({
                "k": torch.zeros([batch_size, cache_length, self.num_heads, 128], dtype=dtype, device=device),
                "v": torch.zeros([batch_size, cache_length, self.num_heads, 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),
                "total_steps": len(self.denoising_step_list),
                "current_step": len(self.denoising_step_list),
            })

        self.kv_cache1 = kv_cache1  # always store the clean cache

    def _initialize_crossattn_cache(self, batch_size, dtype, device):
        """
        Initialize a Per-GPU cross-attention cache for the Wan model.
        """
        crossattn_cache = []

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

        self.crossattn_cache = crossattn_cache  # always store the clean cache
    
    def prepare(
        self,
        text_prompts: List[str],
        device: torch.device,
        dtype: torch.dtype,
        block_mode: str='input',
        noise: torch.Tensor = None,
        current_start: int = 0,
        current_end: int = None,
        block_num: torch.Tensor = None,
        batch_denoise: bool=True,
    ):
        self.device = device
        batch_size = noise.shape[0]

        self.conditional_dict = self.text_encoder(
            text_prompts=text_prompts
        )

        # Step 1: Initialize KV cache
        if self.kv_cache1 is None:
            self._initialize_kv_cache(
                batch_size=batch_size,
                dtype=dtype,
                device=device
            )

            self._initialize_crossattn_cache(
                batch_size=batch_size,
                dtype=dtype,
                device=device
            )
        else:
            # reset cross attn cache
            for block_index in range(self.num_transformer_blocks):
                self.crossattn_cache[block_index]["is_init"] = False
        
        current_start = torch.tensor([current_start], dtype=torch.long, device=device)
        current_end = torch.tensor([current_end], dtype=torch.long, device=device)

        for index, current_timestep in enumerate(self.denoising_step_list):
            # set current timestep
            timestep = torch.ones(
                [batch_size, noise.shape[1]], device=noise.device, dtype=torch.int64) * current_timestep

            if index < len(self.denoising_step_list) - 1:
                denoised_pred = self.generator(
                    noisy_image_or_video=noise,
                    conditional_dict=self.conditional_dict,
                    timestep=timestep,
                    kv_cache=self.kv_cache1,
                    crossattn_cache=self.crossattn_cache,
                    current_start=current_start,
                    current_end=current_end
                )
                next_timestep = self.denoising_step_list[index + 1]
                noise = self.scheduler.add_noise(
                    denoised_pred.flatten(0, 1),
                    torch.randn_like(denoised_pred.flatten(0, 1)),
                    next_timestep *
                    torch.ones([batch_size], device=noise.device,
                                dtype=torch.long)
                ).unflatten(0, denoised_pred.shape[:2])
            else:
                # for getting real output
                denoised_pred = self.generator(
                    noisy_image_or_video=noise,
                    conditional_dict=self.conditional_dict,
                    timestep=timestep,
                    kv_cache=self.kv_cache1,
                    crossattn_cache=self.crossattn_cache,
                    current_start=current_start,
                    current_end=current_end
                )

        if not batch_denoise:
            return denoised_pred

        # Pre-allocate hidden_states tensor to avoid memory allocation during inference
        self.batch_size = len(self.denoising_step_list)

        # Determine which blocks to keep based on block_num range
        blocks_to_keep = []
        if block_num is not None:
            start_block, end_block = block_num[0].item(), block_num[1].item()
            blocks_to_keep = list(range(start_block, end_block))
        else:
            blocks_to_keep = list(range(self.num_transformer_blocks))

        # Process only the blocks in the specified range
        for i in range(self.num_transformer_blocks):
            if dist.is_initialized():
                dist.broadcast(self.crossattn_cache[i]['k'], src=0)
                dist.broadcast(self.crossattn_cache[i]['v'], src=0)
                dist.broadcast(self.kv_cache1[i]['k'], src=0)
                dist.broadcast(self.kv_cache1[i]['v'], src=0)

            self.kv_cache1[i]['k'] = self.kv_cache1[i]['k'].repeat(self.batch_size, 1, 1, 1)
            self.kv_cache1[i]['v'] = self.kv_cache1[i]['v'].repeat(self.batch_size, 1, 1, 1)

            self.kv_cache1[i]['global_end_index'] = self.kv_cache1[i]['global_end_index'].repeat(self.batch_size)
            self.kv_cache1[i]['local_end_index'] = self.kv_cache1[i]['local_end_index'].repeat(self.batch_size)

            self.crossattn_cache[i]['k'] = self.crossattn_cache[i]['k'].expand(self.batch_size, -1, -1, -1)
            self.crossattn_cache[i]['v'] = self.crossattn_cache[i]['v'].expand(self.batch_size, -1, -1, -1)
        
        # Remove blocks outside the range
        if block_num is not None:
            for i in range(self.num_transformer_blocks):
                if i not in blocks_to_keep:
                    self.kv_cache1[i]['k'] = self.kv_cache1[i]['k'].cpu()
                    self.kv_cache1[i]['v'] = self.kv_cache1[i]['v'].cpu()

        self.hidden_states = torch.zeros(
            (self.batch_size, self.num_frame_per_block, *noise.shape[2:]), dtype=noise.dtype, device=device
        )

        if block_mode in ['output', 'middle']:
            self.block_x = torch.zeros(
                (self.batch_size, self.frame_seq_length, self.num_heads*128), dtype=noise.dtype, device=device
            )
        else:
            self.block_x = None

        self.kv_cache_starts = torch.ones(self.batch_size, dtype=torch.long, device=device) * current_end
        self.kv_cache_ends = torch.ones(self.batch_size, dtype=torch.long, device=device) * current_end + self.frame_seq_length

        self.timestep = self.denoising_step_list

        self.conditional_dict['prompt_embeds'] = self.conditional_dict['prompt_embeds'].repeat(self.batch_size, 1, 1)
    
        return denoised_pred
    
    def inference_stream(
        self, 
        noise: torch.Tensor, 
        current_start: int, 
        current_end: int, 
        current_step: int,
        ) -> torch.Tensor:

        self.hidden_states[1:] = self.hidden_states[:-1].clone()
        self.hidden_states[0] = noise[0]

        self.kv_cache_starts[1:] = self.kv_cache_starts[:-1].clone()
        self.kv_cache_starts[0] = current_start
        
        self.kv_cache_ends[1:] = self.kv_cache_ends[:-1].clone()
        self.kv_cache_ends[0] = current_end

        if current_step is not None:
            self.timestep[0] = current_step
        
        self.hidden_states = self.generator(
            noisy_image_or_video=self.hidden_states,
            conditional_dict=self.conditional_dict,
            timestep=self.timestep.unsqueeze(1).expand(-1, self.hidden_states.shape[1]),
            kv_cache=self.kv_cache1,
            crossattn_cache=self.crossattn_cache,
            current_start=self.kv_cache_starts,
            current_end=self.kv_cache_ends,
        )

        for i in range(len(self.denoising_step_list) - 1):
            self.hidden_states[[i]] = self.scheduler.add_noise(
                self.hidden_states[[i]],
                torch.randn_like(self.hidden_states[[i]]),
                self.denoising_step_list[i + 1] *
                torch.ones([1], device=self.hidden_states.device,
                            dtype=torch.long)
            )

        return self.hidden_states
    
    def inference_wo_batch(
        self, 
        noise: torch.Tensor, 
        current_start: int, 
        current_end: int, 
        current_step: int,
        ) -> torch.Tensor:

        batch_size = noise.shape[0]

        current_start = torch.ones(batch_size, dtype=torch.long, device=self.device) * current_start
        current_end = torch.ones(batch_size, dtype=torch.long, device=self.device) * current_end

        # Step 2.1: Spatial denoising loop
        self.denoising_step_list[0] = current_step
        for index, current_timestep in enumerate(self.denoising_step_list):
            # set current timestep
            timestep = torch.ones(
                [batch_size, noise.shape[1]], device=noise.device, dtype=torch.int64) * current_timestep

            if index < len(self.denoising_step_list) - 1:
                denoised_pred = self.generator(
                    noisy_image_or_video=noise,
                    conditional_dict=self.conditional_dict,
                    timestep=timestep,
                    kv_cache=self.kv_cache1,
                    crossattn_cache=self.crossattn_cache,
                    current_start=current_start,
                    current_end=current_end
                )
                next_timestep = self.denoising_step_list[index + 1]
                noise = self.scheduler.add_noise(
                    denoised_pred.flatten(0, 1),
                    torch.randn_like(denoised_pred.flatten(0, 1)),
                    next_timestep *
                    torch.ones([batch_size], device=noise.device,
                                dtype=torch.long)
                ).unflatten(0, denoised_pred.shape[:2])
            else:
                # for getting real output
                denoised_pred = self.generator(
                    noisy_image_or_video=noise,
                    conditional_dict=self.conditional_dict,
                    timestep=timestep,
                    kv_cache=self.kv_cache1,
                    crossattn_cache=self.crossattn_cache,
                    current_start=current_start,
                    current_end=current_end
                )

        return denoised_pred

    def inference(
        self, 
        noise: torch.Tensor, 
        current_start: int, 
        current_end: int, 
        current_step: int, 
        block_mode: str='input', 
        block_num=None,
        patched_x_shape: torch.Tensor=None, 
        block_x: torch.Tensor=None,
        ) -> torch.Tensor:

        if block_mode == 'input':
            self.hidden_states[1:] = self.hidden_states[:-1].clone()
            self.hidden_states[0] = noise[0]

            self.kv_cache_starts[1:] = self.kv_cache_starts[:-1].clone()
            self.kv_cache_starts[0] = current_start
            
            self.kv_cache_ends[1:] = self.kv_cache_ends[:-1].clone()
            self.kv_cache_ends[0] = current_end
        else:
            self.block_x.copy_(block_x)
            self.hidden_states.copy_(noise)
            self.kv_cache_starts.copy_(current_start)
            self.kv_cache_ends.copy_(current_end)

        if current_step is not None:
            self.timestep[0] = current_step
        
        if block_mode == 'output':
            denoised_pred = self.generator.forward_output(
                noisy_image_or_video=self.hidden_states,
                conditional_dict=self.conditional_dict,
                timestep=self.timestep.unsqueeze(1).expand(-1, self.hidden_states.shape[1]),
                kv_cache=self.kv_cache1,
                crossattn_cache=self.crossattn_cache,
                current_start=self.kv_cache_starts,
                current_end=self.kv_cache_ends,
                block_mode=block_mode,
                block_num=block_num,
                patched_x_shape=patched_x_shape,
                block_x=self.block_x
            )

            for i in range(len(self.denoising_step_list) - 1):
                denoised_pred[[i]] = self.scheduler.add_noise(
                    denoised_pred[[i]],
                    torch.randn_like(denoised_pred[[i]]),
                    self.denoising_step_list[i + 1] *
                    torch.ones([1], device=denoised_pred.device,
                                dtype=torch.long)
                )
            patched_x_shape = None

        else:
            denoised_pred, patched_x_shape = self.generator.forward_input(
                noisy_image_or_video=self.hidden_states,
                conditional_dict=self.conditional_dict,
                timestep=self.timestep.unsqueeze(1).expand(-1, self.hidden_states.shape[1]),
                kv_cache=self.kv_cache1,
                crossattn_cache=self.crossattn_cache,
                current_start=self.kv_cache_starts,
                current_end=self.kv_cache_ends,
                block_mode=block_mode,
                block_num=block_num,
                patched_x_shape=patched_x_shape,
                block_x=self.block_x,
            ) 

        return denoised_pred, patched_x_shape