wan_runner.py 10.5 KB
Newer Older
helloyongyang's avatar
helloyongyang committed
1
import os
2
import gc
helloyongyang's avatar
helloyongyang committed
3
4
5
6
7
8
9
import numpy as np
import torch
import torchvision.transforms.functional as TF
from PIL import Image
from lightx2v.utils.registry_factory import RUNNER_REGISTER
from lightx2v.models.runners.default_runner import DefaultRunner
from lightx2v.models.schedulers.wan.scheduler import WanScheduler
10
from lightx2v.models.schedulers.wan.changing_resolution.scheduler import WanScheduler4ChangingResolution
11
12
from lightx2v.models.schedulers.wan.feature_caching.scheduler import (
    WanSchedulerTeaCaching,
13
14
15
    WanSchedulerTaylorCaching,
    WanSchedulerAdaCaching,
    WanSchedulerCustomCaching,
16
)
helloyongyang's avatar
helloyongyang committed
17
18
19
20
21
22
from lightx2v.utils.profiler import ProfilingContext
from lightx2v.models.input_encoders.hf.t5.model import T5EncoderModel
from lightx2v.models.input_encoders.hf.xlm_roberta.model import CLIPModel
from lightx2v.models.networks.wan.model import WanModel
from lightx2v.models.networks.wan.lora_adapter import WanLoraWrapper
from lightx2v.models.video_encoders.hf.wan.vae import WanVAE
23
from lightx2v.models.video_encoders.hf.wan.vae_tiny import WanVAE_tiny
24
from lightx2v.utils.utils import cache_video
root's avatar
root committed
25
from loguru import logger
helloyongyang's avatar
helloyongyang committed
26
27
28
29
30
31
32


@RUNNER_REGISTER("wan2.1")
class WanRunner(DefaultRunner):
    def __init__(self, config):
        super().__init__(config)

33
34
35
36
37
38
    def load_transformer(self):
        model = WanModel(
            self.config.model_path,
            self.config,
            self.init_device,
        )
39
        if self.config.lora_path:
40
            assert not self.config.get("dit_quantized", False) or self.config.mm_config.get("weight_auto_quant", False)
41
            lora_wrapper = WanLoraWrapper(model)
GoatWu's avatar
GoatWu committed
42
43
44
45
            for lora_path in self.config.lora_path:
                lora_name = lora_wrapper.load_lora(lora_path)
                lora_wrapper.apply_lora(lora_name, self.config.strength_model)
                logger.info(f"Loaded LoRA: {lora_name}")
46
47
        return model

48
    def load_image_encoder(self):
helloyongyang's avatar
helloyongyang committed
49
        image_encoder = None
50
51
52
        if self.config.task == "i2v":
            image_encoder = CLIPModel(
                dtype=torch.float16,
53
                device=self.init_device,
54
55
56
57
                checkpoint_path=os.path.join(
                    self.config.model_path,
                    "models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth",
                ),
58
59
60
                clip_quantized=self.config.get("clip_quantized", False),
                clip_quantized_ckpt=self.config.get("clip_quantized_ckpt", None),
                quant_scheme=self.config.get("clip_quant_scheme", None),
61
62
            )
        return image_encoder
helloyongyang's avatar
helloyongyang committed
63

64
    def load_text_encoder(self):
gushiqiao's avatar
gushiqiao committed
65
66
67
68
69
        t5_offload = self.config.get("t5_cpu_offload", False)
        if t5_offload:
            t5_device = torch.device("cpu")
        else:
            t5_device = torch.device("cuda")
helloyongyang's avatar
helloyongyang committed
70
71
72
        text_encoder = T5EncoderModel(
            text_len=self.config["text_len"],
            dtype=torch.bfloat16,
gushiqiao's avatar
gushiqiao committed
73
            device=t5_device,
helloyongyang's avatar
helloyongyang committed
74
75
76
            checkpoint_path=os.path.join(self.config.model_path, "models_t5_umt5-xxl-enc-bf16.pth"),
            tokenizer_path=os.path.join(self.config.model_path, "google/umt5-xxl"),
            shard_fn=None,
gushiqiao's avatar
gushiqiao committed
77
            cpu_offload=t5_offload,
78
79
80
81
            offload_granularity=self.config.get("t5_offload_granularity", "model"),
            t5_quantized=self.config.get("t5_quantized", False),
            t5_quantized_ckpt=self.config.get("t5_quantized_ckpt", None),
            quant_scheme=self.config.get("t5_quant_scheme", None),
helloyongyang's avatar
helloyongyang committed
82
83
        )
        text_encoders = [text_encoder]
84
        return text_encoders
helloyongyang's avatar
helloyongyang committed
85

86
    def load_vae_encoder(self):
87
88
        vae_config = {
            "vae_pth": os.path.join(self.config.model_path, "Wan2.1_VAE.pth"),
89
            "device": self.init_device,
90
91
92
            "parallel": self.config.parallel_vae,
            "use_tiling": self.config.get("use_tiling_vae", False),
        }
93
94
95
96
97
98
99
100
101
102
103
104
105
        if self.config.task != "i2v":
            return None
        else:
            return WanVAE(**vae_config)

    def load_vae_decoder(self):
        vae_config = {
            "vae_pth": os.path.join(self.config.model_path, "Wan2.1_VAE.pth"),
            "device": self.init_device,
            "parallel": self.config.parallel_vae,
            "use_tiling": self.config.get("use_tiling_vae", False),
        }
        if self.config.get("tiny_vae", False):
106
            vae_decoder = WanVAE_tiny(
107
                vae_pth=self.config.tiny_vae_path,
108
                device=self.init_device,
109
            ).to("cuda")
110
        else:
111
            vae_decoder = WanVAE(**vae_config)
112
        return vae_decoder
helloyongyang's avatar
helloyongyang committed
113

114
    def load_vae(self):
gushiqiao's avatar
gushiqiao committed
115
116
117
118
119
120
        vae_encoder = self.load_vae_encoder()
        if vae_encoder is None or self.config.get("tiny_vae", False):
            vae_decoder = self.load_vae_decoder()
        else:
            vae_decoder = vae_encoder
        return vae_encoder, vae_decoder
helloyongyang's avatar
helloyongyang committed
121
122

    def init_scheduler(self):
123
124
        if self.config.get("changing_resolution", False):
            scheduler = WanScheduler4ChangingResolution(self.config)
helloyongyang's avatar
helloyongyang committed
125
        else:
126
127
128
129
130
131
132
133
134
135
136
137
            if self.config.feature_caching == "NoCaching":
                scheduler = WanScheduler(self.config)
            elif self.config.feature_caching == "Tea":
                scheduler = WanSchedulerTeaCaching(self.config)
            elif self.config.feature_caching == "TaylorSeer":
                scheduler = WanSchedulerTaylorCaching(self.config)
            elif self.config.feature_caching == "Ada":
                scheduler = WanSchedulerAdaCaching(self.config)
            elif self.config.feature_caching == "Custom":
                scheduler = WanSchedulerCustomCaching(self.config)
            else:
                raise NotImplementedError(f"Unsupported feature_caching type: {self.config.feature_caching}")
helloyongyang's avatar
helloyongyang committed
138
139
        self.model.set_scheduler(scheduler)

140
    def run_text_encoder(self, text, img):
gushiqiao's avatar
gushiqiao committed
141
        if self.config.get("lazy_load", False) or self.config.get("unload_modules", False):
142
            self.text_encoders = self.load_text_encoder()
helloyongyang's avatar
helloyongyang committed
143
        text_encoder_output = {}
144
145
146
        n_prompt = self.config.get("negative_prompt", "")
        context = self.text_encoders[0].infer([text])
        context_null = self.text_encoders[0].infer([n_prompt if n_prompt else ""])
gushiqiao's avatar
gushiqiao committed
147
        if self.config.get("lazy_load", False) or self.config.get("unload_modules", False):
148
149
150
            del self.text_encoders[0]
            torch.cuda.empty_cache()
            gc.collect()
helloyongyang's avatar
helloyongyang committed
151
152
153
154
        text_encoder_output["context"] = context
        text_encoder_output["context_null"] = context_null
        return text_encoder_output

155
    def run_image_encoder(self, img):
gushiqiao's avatar
gushiqiao committed
156
        if self.config.get("lazy_load", False) or self.config.get("unload_modules", False):
157
            self.image_encoder = self.load_image_encoder()
158
159
        img = TF.to_tensor(img).sub_(0.5).div_(0.5).cuda()
        clip_encoder_out = self.image_encoder.visual([img[:, None, :, :]], self.config).squeeze(0).to(torch.bfloat16)
gushiqiao's avatar
gushiqiao committed
160
        if self.config.get("lazy_load", False) or self.config.get("unload_modules", False):
161
162
163
            del self.image_encoder
            torch.cuda.empty_cache()
            gc.collect()
164
165
166
167
        return clip_encoder_out

    def run_vae_encoder(self, img):
        kwargs = {}
helloyongyang's avatar
helloyongyang committed
168
169
170
        img = TF.to_tensor(img).sub_(0.5).div_(0.5).cuda()
        h, w = img.shape[1:]
        aspect_ratio = h / w
171
172
173
174
175
        max_area = self.config.target_height * self.config.target_width
        lat_h = round(np.sqrt(max_area * aspect_ratio) // self.config.vae_stride[1] // self.config.patch_size[1] * self.config.patch_size[1])
        lat_w = round(np.sqrt(max_area / aspect_ratio) // self.config.vae_stride[2] // self.config.patch_size[2] * self.config.patch_size[2])
        h = lat_h * self.config.vae_stride[1]
        w = lat_w * self.config.vae_stride[2]
helloyongyang's avatar
helloyongyang committed
176

177
178
        self.config.lat_h, kwargs["lat_h"] = lat_h, lat_h
        self.config.lat_w, kwargs["lat_w"] = lat_w, lat_w
helloyongyang's avatar
helloyongyang committed
179

180
181
182
183
184
185
186
        msk = torch.ones(
            1,
            self.config.target_video_length,
            lat_h,
            lat_w,
            device=torch.device("cuda"),
        )
helloyongyang's avatar
helloyongyang committed
187
188
189
190
        msk[:, 1:] = 0
        msk = torch.concat([torch.repeat_interleave(msk[:, 0:1], repeats=4, dim=1), msk[:, 1:]], dim=1)
        msk = msk.view(1, msk.shape[1] // 4, 4, lat_h, lat_w)
        msk = msk.transpose(1, 2)[0]
gushiqiao's avatar
gushiqiao committed
191
        if self.config.get("lazy_load", False) or self.config.get("unload_modules", False):
192
            self.vae_encoder = self.load_vae_encoder()
193
        vae_encode_out = self.vae_encoder.encode(
194
195
196
197
            [
                torch.concat(
                    [
                        torch.nn.functional.interpolate(img[None].cpu(), size=(h, w), mode="bicubic").transpose(0, 1),
198
                        torch.zeros(3, self.config.target_video_length - 1, h, w),
199
200
201
202
                    ],
                    dim=1,
                ).cuda()
            ],
203
            self.config,
helloyongyang's avatar
helloyongyang committed
204
        )[0]
gushiqiao's avatar
gushiqiao committed
205
        if self.config.get("lazy_load", False) or self.config.get("unload_modules", False):
206
207
208
            del self.vae_encoder
            torch.cuda.empty_cache()
            gc.collect()
helloyongyang's avatar
helloyongyang committed
209
        vae_encode_out = torch.concat([msk, vae_encode_out]).to(torch.bfloat16)
210
211
212
        return vae_encode_out, kwargs

    def get_encoder_output_i2v(self, clip_encoder_out, vae_encode_out, text_encoder_output, img):
213
214
215
216
217
218
219
220
        image_encoder_output = {
            "clip_encoder_out": clip_encoder_out,
            "vae_encode_out": vae_encode_out,
        }
        return {
            "text_encoder_output": text_encoder_output,
            "image_encoder_output": image_encoder_output,
        }
helloyongyang's avatar
helloyongyang committed
221
222

    def set_target_shape(self):
223
        ret = {}
224
        num_channels_latents = self.config.get("num_channels_latents", 16)
helloyongyang's avatar
helloyongyang committed
225
        if self.config.task == "i2v":
226
227
            self.config.target_shape = (
                num_channels_latents,
228
                (self.config.target_video_length - 1) // self.config.vae_stride[0] + 1,
229
230
231
                self.config.lat_h,
                self.config.lat_w,
            )
232
233
            ret["lat_h"] = self.config.lat_h
            ret["lat_w"] = self.config.lat_w
helloyongyang's avatar
helloyongyang committed
234
235
        elif self.config.task == "t2v":
            self.config.target_shape = (
236
                num_channels_latents,
237
                (self.config.target_video_length - 1) // self.config.vae_stride[0] + 1,
helloyongyang's avatar
helloyongyang committed
238
239
240
                int(self.config.target_height) // self.config.vae_stride[1],
                int(self.config.target_width) // self.config.vae_stride[2],
            )
241
242
243
244
        ret["target_shape"] = self.config.target_shape
        return ret

    def save_video_func(self, images):
245
246
247
248
249
250
251
252
        cache_video(
            tensor=images,
            save_file=self.config.save_video_path,
            fps=self.config.get("fps", 16),
            nrow=1,
            normalize=True,
            value_range=(-1, 1),
        )