model.py 20.6 KB
Newer Older
gushiqiao's avatar
gushiqiao committed
1
import gc
Yang Yong (雍洋)'s avatar
Yang Yong (雍洋) committed
2
import glob
3
4
import os

helloyongyang's avatar
helloyongyang committed
5
import torch
6
import torch.distributed as dist
helloyongyang's avatar
helloyongyang committed
7
import torch.nn.functional as F
PengGao's avatar
PengGao committed
8
9
10
from loguru import logger
from safetensors import safe_open

11
from lightx2v.models.networks.wan.infer.feature_caching.transformer_infer import (
12
13
    WanTransformerInferAdaCaching,
    WanTransformerInferCustomCaching,
Rongjin Yang's avatar
Rongjin Yang committed
14
15
    WanTransformerInferDualBlock,
    WanTransformerInferDynamicBlock,
PengGao's avatar
PengGao committed
16
    WanTransformerInferFirstBlock,
Musisoul's avatar
Musisoul committed
17
    WanTransformerInferMagCaching,
PengGao's avatar
PengGao committed
18
19
20
    WanTransformerInferTaylorCaching,
    WanTransformerInferTeaCaching,
)
21
22
23
from lightx2v.models.networks.wan.infer.offload.transformer_infer import (
    WanOffloadTransformerInfer,
)
PengGao's avatar
PengGao committed
24
25
26
27
28
29
30
31
from lightx2v.models.networks.wan.infer.post_infer import WanPostInfer
from lightx2v.models.networks.wan.infer.pre_infer import WanPreInfer
from lightx2v.models.networks.wan.infer.transformer_infer import (
    WanTransformerInfer,
)
from lightx2v.models.networks.wan.weights.pre_weights import WanPreWeights
from lightx2v.models.networks.wan.weights.transformer_weights import (
    WanTransformerWeights,
32
)
Yang Yong(雍洋)'s avatar
Yang Yong(雍洋) committed
33
from lightx2v.utils.custom_compiler import CompiledMethodsMixin, compiled_method
34
from lightx2v.utils.envs import *
yihuiwen's avatar
yihuiwen committed
35
from lightx2v.utils.ggml_tensor import load_gguf_sd_ckpt
36
from lightx2v.utils.utils import *
helloyongyang's avatar
helloyongyang committed
37
38


Yang Yong(雍洋)'s avatar
Yang Yong(雍洋) committed
39
class WanModel(CompiledMethodsMixin):
helloyongyang's avatar
helloyongyang committed
40
41
42
    pre_weight_class = WanPreWeights
    transformer_weight_class = WanTransformerWeights

43
    def __init__(self, model_path, config, device, model_type="wan2.1"):
Yang Yong(雍洋)'s avatar
Yang Yong(雍洋) committed
44
        super().__init__()
helloyongyang's avatar
helloyongyang committed
45
46
        self.model_path = model_path
        self.config = config
47
48
        self.cpu_offload = self.config.get("cpu_offload", False)
        self.offload_granularity = self.config.get("offload_granularity", "block")
49
        self.model_type = model_type
50
51
52
53
        self.remove_keys = []
        self.lazy_load = self.config.get("lazy_load", False)
        if self.lazy_load:
            self.remove_keys.extend(["blocks."])
helloyongyang's avatar
helloyongyang committed
54
55
56
57
        if self.config["seq_parallel"]:
            self.seq_p_group = self.config.get("device_mesh").get_group(mesh_dim="seq_p")
        else:
            self.seq_p_group = None
58

gushiqiao's avatar
gushiqiao committed
59
        self.clean_cuda_cache = self.config.get("clean_cuda_cache", False)
60
        self.dit_quantized = self.config.get("dit_quantized", False)
61
        if self.dit_quantized:
62
63
64
65
66
67
68
69
70
71
72
73
74
75
            assert self.config.get("dit_quant_scheme", "Default") in [
                "Default-Force-FP32",
                "fp8-vllm",
                "int8-vllm",
                "fp8-q8f",
                "int8-q8f",
                "fp8-b128-deepgemm",
                "fp8-sgl",
                "int8-sgl",
                "int8-torchao",
                "nvfp4",
                "mxfp4",
                "mxfp6-mxfp8",
                "mxfp8",
Kane's avatar
Kane committed
76
                "int8-tmo",
yihuiwen's avatar
yihuiwen committed
77
78
79
80
81
82
83
84
85
86
87
88
                "gguf-Q8_0",
                "gguf-Q6_K",
                "gguf-Q5_K_S",
                "gguf-Q5_K_M",
                "gguf-Q5_0",
                "gguf-Q5_1",
                "gguf-Q4_K_S",
                "gguf-Q4_K_M",
                "gguf-Q4_0",
                "gguf-Q4_1",
                "gguf-Q3_K_S",
                "gguf-Q3_K_M",
89
            ]
gushiqiao's avatar
gushiqiao committed
90
        self.device = device
helloyongyang's avatar
helloyongyang committed
91
92
93
94
95
96
97
        self._init_infer_class()
        self._init_weights()
        self._init_infer()

    def _init_infer_class(self):
        self.pre_infer_class = WanPreInfer
        self.post_infer_class = WanPostInfer
helloyongyang's avatar
helloyongyang committed
98
99

        if self.config["feature_caching"] == "NoCaching":
100
            self.transformer_infer_class = WanTransformerInfer if not self.cpu_offload else WanOffloadTransformerInfer
helloyongyang's avatar
helloyongyang committed
101
102
103
104
105
106
107
108
109
110
111
112
113
114
        elif self.config["feature_caching"] == "Tea":
            self.transformer_infer_class = WanTransformerInferTeaCaching
        elif self.config["feature_caching"] == "TaylorSeer":
            self.transformer_infer_class = WanTransformerInferTaylorCaching
        elif self.config["feature_caching"] == "Ada":
            self.transformer_infer_class = WanTransformerInferAdaCaching
        elif self.config["feature_caching"] == "Custom":
            self.transformer_infer_class = WanTransformerInferCustomCaching
        elif self.config["feature_caching"] == "FirstBlock":
            self.transformer_infer_class = WanTransformerInferFirstBlock
        elif self.config["feature_caching"] == "DualBlock":
            self.transformer_infer_class = WanTransformerInferDualBlock
        elif self.config["feature_caching"] == "DynamicBlock":
            self.transformer_infer_class = WanTransformerInferDynamicBlock
Musisoul's avatar
Musisoul committed
115
116
        elif self.config["feature_caching"] == "Mag":
            self.transformer_infer_class = WanTransformerInferMagCaching
helloyongyang's avatar
helloyongyang committed
117
        else:
helloyongyang's avatar
helloyongyang committed
118
            raise NotImplementedError(f"Unsupported feature_caching type: {self.config['feature_caching']}")
helloyongyang's avatar
helloyongyang committed
119

gushiqiao's avatar
gushiqiao committed
120
121
122
123
124
125
    def _should_load_weights(self):
        """Determine if current rank should load weights from disk."""
        if self.config.get("device_mesh") is None:
            # Single GPU mode
            return True
        elif dist.is_initialized():
126
127
128
129
130
131
            if self.config.get("load_from_rank0", False):
                # Multi-GPU mode, only rank 0 loads
                if dist.get_rank() == 0:
                    logger.info(f"Loading weights from {self.model_path}")
                    return True
            else:
gushiqiao's avatar
gushiqiao committed
132
133
134
                return True
        return False

135
    def _should_init_empty_model(self):
136
        if self.config.get("lora_configs") and self.config["lora_configs"]:
137
138
139
140
141
142
143
144
145
146
147
148
            if self.model_type in ["wan2.1"]:
                return True
            if self.model_type in ["wan2.2_moe_high_noise"]:
                for lora_config in self.config["lora_configs"]:
                    if lora_config["name"] == "high_noise_model":
                        return True
            if self.model_type in ["wan2.2_moe_low_noise"]:
                for lora_config in self.config["lora_configs"]:
                    if lora_config["name"] == "low_noise_model":
                        return True
        return False

149
    def _load_safetensor_to_dict(self, file_path, unified_dtype, sensitive_layer):
150
151
        remove_keys = self.remove_keys if hasattr(self, "remove_keys") else []

152
        if self.device.type != "cpu" and dist.is_initialized():
153
            device = dist.get_rank()
154
        else:
155
            device = str(self.device)
156

157
        with safe_open(file_path, framework="pt", device=device) as f:
158
159
160
161
162
            return {
                key: (f.get_tensor(key).to(GET_DTYPE()) if unified_dtype or all(s not in key for s in sensitive_layer) else f.get_tensor(key).to(GET_SENSITIVE_DTYPE()))
                for key in f.keys()
                if not any(remove_key in key for remove_key in remove_keys)
            }
helloyongyang's avatar
helloyongyang committed
163

164
    def _load_ckpt(self, unified_dtype, sensitive_layer):
165
166
167
168
169
170
171
172
173
        if self.config.get("dit_original_ckpt", None):
            safetensors_path = self.config["dit_original_ckpt"]
        else:
            safetensors_path = self.model_path

        if os.path.isdir(safetensors_path):
            safetensors_files = glob.glob(os.path.join(safetensors_path, "*.safetensors"))
        else:
            safetensors_files = [safetensors_path]
174

175
176
177
178
        if self.lazy_load:
            assert len(safetensors_files) == 1, "Only support single safetensors file in lazy load mode"
            self.lazy_load_path = safetensors_files[0]

helloyongyang's avatar
helloyongyang committed
179
180
        weight_dict = {}
        for file_path in safetensors_files:
181
            if self.config.get("adapter_model_path", None) is not None:
182
                if self.config["adapter_model_path"] == file_path:
183
                    continue
184
            logger.info(f"Loading weights from {file_path}")
185
            file_weights = self._load_safetensor_to_dict(file_path, unified_dtype, sensitive_layer)
helloyongyang's avatar
helloyongyang committed
186
            weight_dict.update(file_weights)
187

helloyongyang's avatar
helloyongyang committed
188
189
        return weight_dict

190
    def _load_quant_ckpt(self, unified_dtype, sensitive_layer):
191
        remove_keys = self.remove_keys if hasattr(self, "remove_keys") else []
gushiqiao's avatar
Fix  
gushiqiao committed
192

193
194
195
196
        if self.config.get("dit_quantized_ckpt", None):
            safetensors_path = self.config["dit_quantized_ckpt"]
        else:
            safetensors_path = self.model_path
gushiqiao's avatar
Fix  
gushiqiao committed
197

yihuiwen's avatar
yihuiwen committed
198
199
200
201
202
203
204
205
206
207
208
        if "gguf" in self.config.get("dit_quant_scheme", ""):
            gguf_path = ""
            if os.path.isdir(safetensors_path):
                gguf_type = self.config.get("dit_quant_scheme").replace("gguf-", "")
                gguf_files = list(filter(lambda x: gguf_type in x, glob.glob(os.path.join(safetensors_path, "*.gguf"))))
                gguf_path = gguf_files[0]
            else:
                gguf_path = safetensors_path
            weight_dict = self._load_gguf_ckpt(gguf_path)
            return weight_dict

209
210
211
212
        if os.path.isdir(safetensors_path):
            safetensors_files = glob.glob(os.path.join(safetensors_path, "*.safetensors"))
        else:
            safetensors_files = [safetensors_path]
213
            safetensors_path = os.path.dirname(safetensors_path)
gushiqiao's avatar
Fix  
gushiqiao committed
214

215
216
217
218
        if self.lazy_load:
            assert len(safetensors_files) == 1, "Only support single safetensors file in lazy load mode"
            self.lazy_load_path = safetensors_files[0]

gushiqiao's avatar
Fix  
gushiqiao committed
219
        weight_dict = {}
220
221
222
223
        for safetensor_path in safetensors_files:
            if self.config.get("adapter_model_path", None) is not None:
                if self.config["adapter_model_path"] == safetensor_path:
                    continue
yihuiwen's avatar
yihuiwen committed
224

gushiqiao's avatar
Fix  
gushiqiao committed
225
226
227
            with safe_open(safetensor_path, framework="pt") as f:
                logger.info(f"Loading weights from {safetensor_path}")
                for k in f.keys():
228
229
                    if any(remove_key in k for remove_key in remove_keys):
                        continue
230
231
232
233
234
                    if f.get_tensor(k).dtype in [
                        torch.float16,
                        torch.bfloat16,
                        torch.float,
                    ]:
235
                        if unified_dtype or all(s not in k for s in sensitive_layer):
gushiqiao's avatar
gushiqiao committed
236
                            weight_dict[k] = f.get_tensor(k).to(GET_DTYPE()).to(self.device)
gushiqiao's avatar
Fix  
gushiqiao committed
237
                        else:
gushiqiao's avatar
gushiqiao committed
238
                            weight_dict[k] = f.get_tensor(k).to(GET_SENSITIVE_DTYPE()).to(self.device)
gushiqiao's avatar
Fix  
gushiqiao committed
239
                    else:
gushiqiao's avatar
gushiqiao committed
240
                        weight_dict[k] = f.get_tensor(k).to(self.device)
241

242
243
244
245
246
247
248
        if self.config.get("dit_quant_scheme", "Default") == "nvfp4":
            calib_path = os.path.join(safetensors_path, "calib.pt")
            logger.info(f"[CALIB] Loaded calibration data from: {calib_path}")
            calib_data = torch.load(calib_path, map_location="cpu")
            for k, v in calib_data["absmax"].items():
                weight_dict[k.replace(".weight", ".input_absmax")] = v.to(self.device)

249
250
        return weight_dict

yihuiwen's avatar
yihuiwen committed
251
252
253
    def _load_gguf_ckpt(self, gguf_path):
        state_dict = load_gguf_sd_ckpt(gguf_path, to_device=self.device)
        return state_dict
254

lijiaqi2's avatar
lijiaqi2 committed
255
    def _init_weights(self, weight_dict=None):
256
        unified_dtype = GET_DTYPE() == GET_SENSITIVE_DTYPE()
gushiqiao's avatar
Fix  
gushiqiao committed
257
        # Some layers run with float32 to achieve high accuracy
258
        sensitive_layer = {
gushiqiao's avatar
gushiqiao committed
259
260
261
262
263
264
            "norm",
            "embedding",
            "modulation",
            "time",
            "img_emb.proj.0",
            "img_emb.proj.4",
gushiqiao's avatar
gushiqiao committed
265
266
            "before_proj",  # vace
            "after_proj",  # vace
gushiqiao's avatar
gushiqiao committed
267
        }
268

lijiaqi2's avatar
lijiaqi2 committed
269
        if weight_dict is None:
gushiqiao's avatar
gushiqiao committed
270
            is_weight_loader = self._should_load_weights()
271
            if is_weight_loader:
272
                if not self.dit_quantized:
gushiqiao's avatar
gushiqiao committed
273
274
                    # Load original weights
                    weight_dict = self._load_ckpt(unified_dtype, sensitive_layer)
275
                else:
gushiqiao's avatar
gushiqiao committed
276
                    # Load quantized weights
277
                    weight_dict = self._load_quant_ckpt(unified_dtype, sensitive_layer)
278

279
280
            if self.config.get("device_mesh") is not None and self.config.get("load_from_rank0", False):
                weight_dict = self._load_weights_from_rank0(weight_dict, is_weight_loader)
281

282
283
284
            if hasattr(self, "adapter_weights_dict"):
                weight_dict.update(self.adapter_weights_dict)

gushiqiao's avatar
gushiqiao committed
285
            self.original_weight_dict = weight_dict
lijiaqi2's avatar
lijiaqi2 committed
286
287
        else:
            self.original_weight_dict = weight_dict
288

gushiqiao's avatar
gushiqiao committed
289
        # Initialize weight containers
helloyongyang's avatar
helloyongyang committed
290
        self.pre_weight = self.pre_weight_class(self.config)
291
292
293
294
        if self.lazy_load:
            self.transformer_weights = self.transformer_weight_class(self.config, self.lazy_load_path)
        else:
            self.transformer_weights = self.transformer_weight_class(self.config)
295
        if not self._should_init_empty_model():
296
            self._apply_weights()
gushiqiao's avatar
gushiqiao committed
297

298
299
300
301
302
    def _apply_weights(self, weight_dict=None):
        if weight_dict is not None:
            self.original_weight_dict = weight_dict
            del weight_dict
            gc.collect()
gushiqiao's avatar
gushiqiao committed
303
        # Load weights into containers
304
        self.pre_weight.load(self.original_weight_dict)
gushiqiao's avatar
gushiqiao committed
305
        self.transformer_weights.load(self.original_weight_dict)
helloyongyang's avatar
helloyongyang committed
306

gushiqiao's avatar
gushiqiao committed
307
308
309
310
        del self.original_weight_dict
        torch.cuda.empty_cache()
        gc.collect()

311
312
    def _load_weights_from_rank0(self, weight_dict, is_weight_loader):
        logger.info("Loading distributed weights")
gushiqiao's avatar
gushiqiao committed
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
        global_src_rank = 0
        target_device = "cpu" if self.cpu_offload else "cuda"

        if is_weight_loader:
            meta_dict = {}
            for key, tensor in weight_dict.items():
                meta_dict[key] = {"shape": tensor.shape, "dtype": tensor.dtype}

            obj_list = [meta_dict]
            dist.broadcast_object_list(obj_list, src=global_src_rank)
            synced_meta_dict = obj_list[0]
        else:
            obj_list = [None]
            dist.broadcast_object_list(obj_list, src=global_src_rank)
            synced_meta_dict = obj_list[0]

        distributed_weight_dict = {}
        for key, meta in synced_meta_dict.items():
            distributed_weight_dict[key] = torch.empty(meta["shape"], dtype=meta["dtype"], device=target_device)

        if target_device == "cuda":
            dist.barrier(device_ids=[torch.cuda.current_device()])

        for key in sorted(synced_meta_dict.keys()):
            if is_weight_loader:
                distributed_weight_dict[key].copy_(weight_dict[key], non_blocking=True)

gushiqiao's avatar
gushiqiao committed
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
            if target_device == "cpu":
                if is_weight_loader:
                    gpu_tensor = distributed_weight_dict[key].cuda()
                    dist.broadcast(gpu_tensor, src=global_src_rank)
                    distributed_weight_dict[key].copy_(gpu_tensor.cpu(), non_blocking=True)
                    del gpu_tensor
                    torch.cuda.empty_cache()
                else:
                    gpu_tensor = torch.empty_like(distributed_weight_dict[key], device="cuda")
                    dist.broadcast(gpu_tensor, src=global_src_rank)
                    distributed_weight_dict[key].copy_(gpu_tensor.cpu(), non_blocking=True)
                    del gpu_tensor
                    torch.cuda.empty_cache()

                if distributed_weight_dict[key].is_pinned():
                    distributed_weight_dict[key].copy_(distributed_weight_dict[key], non_blocking=True)
            else:
                dist.broadcast(distributed_weight_dict[key], src=global_src_rank)

        if target_device == "cuda":
            torch.cuda.synchronize()
        else:
            for tensor in distributed_weight_dict.values():
                if tensor.is_pinned():
                    tensor.copy_(tensor, non_blocking=False)
gushiqiao's avatar
gushiqiao committed
365
366

        logger.info(f"Weights distributed across {dist.get_world_size()} devices on {target_device}")
367

gushiqiao's avatar
gushiqiao committed
368
369
        return distributed_weight_dict

helloyongyang's avatar
helloyongyang committed
370
371
372
    def _init_infer(self):
        self.pre_infer = self.pre_infer_class(self.config)
        self.post_infer = self.post_infer_class(self.config)
helloyongyang's avatar
helloyongyang committed
373
        self.transformer_infer = self.transformer_infer_class(self.config)
374
        if hasattr(self.transformer_infer, "offload_manager"):
375
376
377
            self.transformer_infer.offload_manager.init_cuda_buffer(self.transformer_weights.offload_block_cuda_buffers, self.transformer_weights.offload_phase_cuda_buffers)
            if self.lazy_load:
                self.transformer_infer.offload_manager.init_cpu_buffer(self.transformer_weights.offload_block_cpu_buffers, self.transformer_weights.offload_phase_cpu_buffers)
helloyongyang's avatar
helloyongyang committed
378
379
380

    def set_scheduler(self, scheduler):
        self.scheduler = scheduler
381
382
        self.pre_infer.set_scheduler(scheduler)
        self.post_infer.set_scheduler(scheduler)
helloyongyang's avatar
helloyongyang committed
383
384
        self.transformer_infer.set_scheduler(scheduler)

TorynCurtis's avatar
TorynCurtis committed
385
386
387
388
389
390
391
392
    def to_cpu(self):
        self.pre_weight.to_cpu()
        self.transformer_weights.to_cpu()

    def to_cuda(self):
        self.pre_weight.to_cuda()
        self.transformer_weights.to_cuda()

helloyongyang's avatar
helloyongyang committed
393
394
    @torch.no_grad()
    def infer(self, inputs):
395
        if self.cpu_offload:
396
            if self.offload_granularity == "model" and self.scheduler.step_index == 0 and "wan2.2_moe" not in self.config["model_cls"]:
397
398
399
                self.to_cuda()
            elif self.offload_granularity != "model":
                self.pre_weight.to_cuda()
gushiqiao's avatar
gushiqiao committed
400
                self.transformer_weights.non_block_weights_to_cuda()
401

402
        if self.config["enable_cfg"]:
helloyongyang's avatar
helloyongyang committed
403
404
405
406
407
408
409
            if self.config["cfg_parallel"]:
                # ==================== CFG Parallel Processing ====================
                cfg_p_group = self.config["device_mesh"].get_group(mesh_dim="cfg_p")
                assert dist.get_world_size(cfg_p_group) == 2, "cfg_p_world_size must be equal to 2"
                cfg_p_rank = dist.get_rank(cfg_p_group)

                if cfg_p_rank == 0:
helloyongyang's avatar
helloyongyang committed
410
                    noise_pred = self._infer_cond_uncond(inputs, infer_condition=True)
helloyongyang's avatar
helloyongyang committed
411
                else:
helloyongyang's avatar
helloyongyang committed
412
                    noise_pred = self._infer_cond_uncond(inputs, infer_condition=False)
helloyongyang's avatar
helloyongyang committed
413

helloyongyang's avatar
helloyongyang committed
414
415
416
417
418
419
                noise_pred_list = [torch.zeros_like(noise_pred) for _ in range(2)]
                dist.all_gather(noise_pred_list, noise_pred, group=cfg_p_group)
                noise_pred_cond = noise_pred_list[0]  # cfg_p_rank == 0
                noise_pred_uncond = noise_pred_list[1]  # cfg_p_rank == 1
            else:
                # ==================== CFG Processing ====================
helloyongyang's avatar
helloyongyang committed
420
421
                noise_pred_cond = self._infer_cond_uncond(inputs, infer_condition=True)
                noise_pred_uncond = self._infer_cond_uncond(inputs, infer_condition=False)
gushiqiao's avatar
gushiqiao committed
422

helloyongyang's avatar
helloyongyang committed
423
424
425
            self.scheduler.noise_pred = noise_pred_uncond + self.scheduler.sample_guide_scale * (noise_pred_cond - noise_pred_uncond)
        else:
            # ==================== No CFG ====================
helloyongyang's avatar
helloyongyang committed
426
            self.scheduler.noise_pred = self._infer_cond_uncond(inputs, infer_condition=True)
427
428

        if self.cpu_offload:
429
            if self.offload_granularity == "model" and self.scheduler.step_index == self.scheduler.infer_steps - 1 and "wan2.2_moe" not in self.config["model_cls"]:
430
431
                self.to_cpu()
            elif self.offload_granularity != "model":
root's avatar
root committed
432
                self.pre_weight.to_cpu()
gushiqiao's avatar
gushiqiao committed
433
                self.transformer_weights.non_block_weights_to_cpu()
gushiqiao's avatar
gushiqiao committed
434

Yang Yong(雍洋)'s avatar
Yang Yong(雍洋) committed
435
    @compiled_method()
436
    @torch.no_grad()
helloyongyang's avatar
helloyongyang committed
437
438
439
440
    def _infer_cond_uncond(self, inputs, infer_condition=True):
        self.scheduler.infer_condition = infer_condition

        pre_infer_out = self.pre_infer.infer(self.pre_weight, inputs)
helloyongyang's avatar
helloyongyang committed
441
442
443
444
445
446
447
448
449

        if self.config["seq_parallel"]:
            pre_infer_out = self._seq_parallel_pre_process(pre_infer_out)

        x = self.transformer_infer.infer(self.transformer_weights, pre_infer_out)

        if self.config["seq_parallel"]:
            x = self._seq_parallel_post_process(x)

gushiqiao's avatar
gushiqiao committed
450
        noise_pred = self.post_infer.infer(x, pre_infer_out)[0]
helloyongyang's avatar
helloyongyang committed
451
452
453
454
455
456
457
458
459

        if self.clean_cuda_cache:
            del x, pre_infer_out
            torch.cuda.empty_cache()

        return noise_pred

    @torch.no_grad()
    def _seq_parallel_pre_process(self, pre_infer_out):
helloyongyang's avatar
helloyongyang committed
460
        x = pre_infer_out.x
helloyongyang's avatar
helloyongyang committed
461
462
463
464
465
        world_size = dist.get_world_size(self.seq_p_group)
        cur_rank = dist.get_rank(self.seq_p_group)

        padding_size = (world_size - (x.shape[0] % world_size)) % world_size
        if padding_size > 0:
helloyongyang's avatar
helloyongyang committed
466
            x = F.pad(x, (0, 0, 0, padding_size))
helloyongyang's avatar
helloyongyang committed
467

helloyongyang's avatar
helloyongyang committed
468
        pre_infer_out.x = torch.chunk(x, world_size, dim=0)[cur_rank]
helloyongyang's avatar
helloyongyang committed
469

470
        if self.config["model_cls"] in ["wan2.2", "wan2.2_audio"] and self.config["task"] in ["i2v", "s2v"]:
helloyongyang's avatar
helloyongyang committed
471
472
473
474
475
476
477
            embed, embed0 = pre_infer_out.embed, pre_infer_out.embed0

            padding_size = (world_size - (embed.shape[0] % world_size)) % world_size
            if padding_size > 0:
                embed = F.pad(embed, (0, 0, 0, padding_size))
                embed0 = F.pad(embed0, (0, 0, 0, 0, 0, padding_size))

helloyongyang's avatar
helloyongyang committed
478
479
            pre_infer_out.embed = torch.chunk(embed, world_size, dim=0)[cur_rank]
            pre_infer_out.embed0 = torch.chunk(embed0, world_size, dim=0)[cur_rank]
helloyongyang's avatar
helloyongyang committed
480
481
482
483
484
485
486
487
488

        return pre_infer_out

    @torch.no_grad()
    def _seq_parallel_post_process(self, x):
        world_size = dist.get_world_size(self.seq_p_group)
        gathered_x = [torch.empty_like(x) for _ in range(world_size)]
        dist.all_gather(gathered_x, x, group=self.seq_p_group)
        combined_output = torch.cat(gathered_x, dim=0)
helloyongyang's avatar
helloyongyang committed
489
        return combined_output