model.py 18.4 KB
Newer Older
1
import json
2
3
import os

helloyongyang's avatar
helloyongyang committed
4
import torch
5
import torch.distributed as dist
helloyongyang's avatar
helloyongyang committed
6
import torch.nn.functional as F
PengGao's avatar
PengGao committed
7
8
9
from loguru import logger
from safetensors import safe_open

10
from lightx2v.models.networks.wan.infer.feature_caching.transformer_infer import (
11
12
    WanTransformerInferAdaCaching,
    WanTransformerInferCustomCaching,
Rongjin Yang's avatar
Rongjin Yang committed
13
14
    WanTransformerInferDualBlock,
    WanTransformerInferDynamicBlock,
PengGao's avatar
PengGao committed
15
    WanTransformerInferFirstBlock,
Musisoul's avatar
Musisoul committed
16
    WanTransformerInferMagCaching,
PengGao's avatar
PengGao committed
17
18
19
    WanTransformerInferTaylorCaching,
    WanTransformerInferTeaCaching,
)
20
21
22
from lightx2v.models.networks.wan.infer.offload.transformer_infer import (
    WanOffloadTransformerInfer,
)
PengGao's avatar
PengGao committed
23
24
25
26
27
28
29
30
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,
31
)
Yang Yong(雍洋)'s avatar
Yang Yong(雍洋) committed
32
from lightx2v.utils.custom_compiler import CompiledMethodsMixin, compiled_method
33
from lightx2v.utils.envs import *
34
from lightx2v.utils.utils import *
helloyongyang's avatar
helloyongyang committed
35

36
37
38
39
40
try:
    import gguf
except ImportError:
    gguf = None

helloyongyang's avatar
helloyongyang committed
41

Yang Yong(雍洋)'s avatar
Yang Yong(雍洋) committed
42
class WanModel(CompiledMethodsMixin):
helloyongyang's avatar
helloyongyang committed
43
44
45
    pre_weight_class = WanPreWeights
    transformer_weight_class = WanTransformerWeights

helloyongyang's avatar
helloyongyang committed
46
    def __init__(self, model_path, config, device):
Yang Yong(雍洋)'s avatar
Yang Yong(雍洋) committed
47
        super().__init__()
helloyongyang's avatar
helloyongyang committed
48
49
        self.model_path = model_path
        self.config = config
50
51
        self.cpu_offload = self.config.get("cpu_offload", False)
        self.offload_granularity = self.config.get("offload_granularity", "block")
helloyongyang's avatar
helloyongyang committed
52
53
54
55
56

        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
57

gushiqiao's avatar
gushiqiao committed
58
        self.clean_cuda_cache = self.config.get("clean_cuda_cache", False)
59
        self.dit_quantized = self.config.mm_config.get("mm_type", "Default") != "Default"
60

gushiqiao's avatar
gushiqiao committed
61
62
        if self.dit_quantized:
            dit_quant_scheme = self.config.mm_config.get("mm_type").split("-")[1]
gushiqiao's avatar
gushiqiao committed
63
64
            if self.config.model_cls == "wan2.1_distill":
                dit_quant_scheme = "distill_" + dit_quant_scheme
65
66
67
68
            if dit_quant_scheme == "gguf":
                self.dit_quantized_ckpt = find_gguf_model_path(config, "dit_quantized_ckpt", subdir=dit_quant_scheme)
                self.config.use_gguf = True
            else:
69
70
71
72
73
74
                self.dit_quantized_ckpt = find_hf_model_path(
                    config,
                    self.model_path,
                    "dit_quantized_ckpt",
                    subdir=dit_quant_scheme,
                )
gushiqiao's avatar
Fix bug  
gushiqiao committed
75
76
77
78
79
            quant_config_path = os.path.join(self.dit_quantized_ckpt, "config.json")
            if os.path.exists(quant_config_path):
                with open(quant_config_path, "r") as f:
                    quant_model_config = json.load(f)
                self.config.update(quant_model_config)
gushiqiao's avatar
gushiqiao committed
80
81
        else:
            self.dit_quantized_ckpt = None
82
83
            assert not self.config.get("lazy_load", False)

84
85
86
87
        self.weight_auto_quant = self.config.mm_config.get("weight_auto_quant", False)
        if self.dit_quantized:
            assert self.weight_auto_quant or self.dit_quantized_ckpt is not None

gushiqiao's avatar
gushiqiao committed
88
        self.device = device
helloyongyang's avatar
helloyongyang committed
89
90
91
92
93
94
95
        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
96
97

        if self.config["feature_caching"] == "NoCaching":
98
            self.transformer_infer_class = WanTransformerInfer if not self.cpu_offload else WanOffloadTransformerInfer
helloyongyang's avatar
helloyongyang committed
99
100
101
102
103
104
105
106
107
108
109
110
111
112
        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
113
114
        elif self.config["feature_caching"] == "Mag":
            self.transformer_infer_class = WanTransformerInferMagCaching
helloyongyang's avatar
helloyongyang committed
115
        else:
helloyongyang's avatar
helloyongyang committed
116
            raise NotImplementedError(f"Unsupported feature_caching type: {self.config['feature_caching']}")
helloyongyang's avatar
helloyongyang committed
117

gushiqiao's avatar
gushiqiao committed
118
119
120
121
122
123
124
125
126
127
128
129
    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():
            # Multi-GPU mode, only rank 0 loads
            if dist.get_rank() == 0:
                logger.info(f"Loading weights from {self.model_path}")
                return True
        return False

130
    def _load_safetensor_to_dict(self, file_path, unified_dtype, sensitive_layer):
helloyongyang's avatar
helloyongyang committed
131
        with safe_open(file_path, framework="pt") as f:
132
133
134
135
            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())).pin_memory().to(self.device)
                for key in f.keys()
            }
helloyongyang's avatar
helloyongyang committed
136

137
    def _load_ckpt(self, unified_dtype, sensitive_layer):
helloyongyang's avatar
helloyongyang committed
138
        safetensors_path = find_hf_model_path(self.config, self.model_path, "dit_original_ckpt", subdir="original")
139
        safetensors_files = glob.glob(os.path.join(safetensors_path, "*.safetensors"))
140

helloyongyang's avatar
helloyongyang committed
141
142
        weight_dict = {}
        for file_path in safetensors_files:
143
144
145
            if self.config.get("adapter_model_path", None) is not None:
                if self.config.adapter_model_path == file_path:
                    continue
146
            file_weights = self._load_safetensor_to_dict(file_path, unified_dtype, sensitive_layer)
helloyongyang's avatar
helloyongyang committed
147
148
149
            weight_dict.update(file_weights)
        return weight_dict

150
    def _load_quant_ckpt(self, unified_dtype, sensitive_layer):
gushiqiao's avatar
gushiqiao committed
151
        ckpt_path = self.dit_quantized_ckpt
152
        logger.info(f"Loading quant dit model from {ckpt_path}")
153

gushiqiao's avatar
Fix  
gushiqiao committed
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
        index_files = [f for f in os.listdir(ckpt_path) if f.endswith(".index.json")]
        if not index_files:
            raise FileNotFoundError(f"No *.index.json found in {ckpt_path}")

        index_path = os.path.join(ckpt_path, index_files[0])
        logger.info(f" Using safetensors index: {index_path}")

        with open(index_path, "r") as f:
            index_data = json.load(f)

        weight_dict = {}
        for filename in set(index_data["weight_map"].values()):
            safetensor_path = os.path.join(ckpt_path, filename)
            with safe_open(safetensor_path, framework="pt") as f:
                logger.info(f"Loading weights from {safetensor_path}")
                for k in f.keys():
170
171
172
173
174
                    if f.get_tensor(k).dtype in [
                        torch.float16,
                        torch.bfloat16,
                        torch.float,
                    ]:
175
176
                        if unified_dtype or all(s not in k for s in sensitive_layer):
                            weight_dict[k] = f.get_tensor(k).pin_memory().to(GET_DTYPE()).to(self.device)
gushiqiao's avatar
Fix  
gushiqiao committed
177
                        else:
178
                            weight_dict[k] = f.get_tensor(k).pin_memory().to(GET_SENSITIVE_DTYPE()).to(self.device)
gushiqiao's avatar
Fix  
gushiqiao committed
179
180
                    else:
                        weight_dict[k] = f.get_tensor(k).pin_memory().to(self.device)
181

182
183
        return weight_dict

184
    def _load_quant_split_ckpt(self, unified_dtype, sensitive_layer):
gushiqiao's avatar
gushiqiao committed
185
        lazy_load_model_path = self.dit_quantized_ckpt
186
        logger.info(f"Loading splited quant model from {lazy_load_model_path}")
gushiqiao's avatar
gushiqiao committed
187
        pre_post_weight_dict = {}
188
189

        safetensor_path = os.path.join(lazy_load_model_path, "non_block.safetensors")
gushiqiao's avatar
gushiqiao committed
190
        with safe_open(safetensor_path, framework="pt", device="cpu") as f:
191
            for k in f.keys():
192
193
194
195
196
                if f.get_tensor(k).dtype in [
                    torch.float16,
                    torch.bfloat16,
                    torch.float,
                ]:
197
198
                    if unified_dtype or all(s not in k for s in sensitive_layer):
                        pre_post_weight_dict[k] = f.get_tensor(k).pin_memory().to(GET_DTYPE()).to(self.device)
gushiqiao's avatar
Fix  
gushiqiao committed
199
                    else:
200
                        pre_post_weight_dict[k] = f.get_tensor(k).pin_memory().to(GET_SENSITIVE_DTYPE()).to(self.device)
gushiqiao's avatar
Fix  
gushiqiao committed
201
202
                else:
                    pre_post_weight_dict[k] = f.get_tensor(k).pin_memory().to(self.device)
203

gushiqiao's avatar
gushiqiao committed
204
        return pre_post_weight_dict
205

206
207
208
209
210
211
212
213
    def _load_gguf_ckpt(self):
        gguf_path = self.dit_quantized_ckpt
        logger.info(f"Loading gguf-quant dit model from {gguf_path}")
        reader = gguf.GGUFReader(gguf_path)
        for tensor in reader.tensors:
            # TODO: implement _load_gguf_ckpt
            pass

lijiaqi2's avatar
lijiaqi2 committed
214
    def _init_weights(self, weight_dict=None):
215
        unified_dtype = GET_DTYPE() == GET_SENSITIVE_DTYPE()
gushiqiao's avatar
Fix  
gushiqiao committed
216
        # Some layers run with float32 to achieve high accuracy
217
        sensitive_layer = {
gushiqiao's avatar
gushiqiao committed
218
219
220
221
222
223
            "norm",
            "embedding",
            "modulation",
            "time",
            "img_emb.proj.0",
            "img_emb.proj.4",
gushiqiao's avatar
gushiqiao committed
224
225
            "before_proj",  # vace
            "after_proj",  # vace
gushiqiao's avatar
gushiqiao committed
226
        }
227

lijiaqi2's avatar
lijiaqi2 committed
228
        if weight_dict is None:
gushiqiao's avatar
gushiqiao committed
229
            is_weight_loader = self._should_load_weights()
230
231
            if is_weight_loader:
                if not self.dit_quantized or self.weight_auto_quant:
gushiqiao's avatar
gushiqiao committed
232
233
                    # Load original weights
                    weight_dict = self._load_ckpt(unified_dtype, sensitive_layer)
234
                else:
gushiqiao's avatar
gushiqiao committed
235
                    # Load quantized weights
236
                    if not self.config.get("lazy_load", False):
gushiqiao's avatar
gushiqiao committed
237
                        weight_dict = self._load_quant_ckpt(unified_dtype, sensitive_layer)
238
                    else:
gushiqiao's avatar
gushiqiao committed
239
                        weight_dict = self._load_quant_split_ckpt(unified_dtype, sensitive_layer)
240

gushiqiao's avatar
gushiqiao committed
241
            if self.config.get("device_mesh") is not None:
gushiqiao's avatar
gushiqiao committed
242
                weight_dict = self._load_weights_distribute(weight_dict, is_weight_loader)
243

244
245
246
            if hasattr(self, "adapter_weights_dict"):
                weight_dict.update(self.adapter_weights_dict)

gushiqiao's avatar
gushiqiao committed
247
            self.original_weight_dict = weight_dict
lijiaqi2's avatar
lijiaqi2 committed
248
249
        else:
            self.original_weight_dict = weight_dict
250

gushiqiao's avatar
gushiqiao committed
251
        # Initialize weight containers
helloyongyang's avatar
helloyongyang committed
252
253
        self.pre_weight = self.pre_weight_class(self.config)
        self.transformer_weights = self.transformer_weight_class(self.config)
gushiqiao's avatar
gushiqiao committed
254
255

        # Load weights into containers
256
        self.pre_weight.load(self.original_weight_dict)
gushiqiao's avatar
gushiqiao committed
257
        self.transformer_weights.load(self.original_weight_dict)
helloyongyang's avatar
helloyongyang committed
258

gushiqiao's avatar
gushiqiao committed
259
    def _load_weights_distribute(self, weight_dict, is_weight_loader):
gushiqiao's avatar
gushiqiao committed
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
        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
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
            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
312
313
314
315

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

helloyongyang's avatar
helloyongyang committed
316
317
318
    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
319
        self.transformer_infer = self.transformer_infer_class(self.config)
helloyongyang's avatar
helloyongyang committed
320
321
322

    def set_scheduler(self, scheduler):
        self.scheduler = scheduler
323
324
        self.pre_infer.set_scheduler(scheduler)
        self.post_infer.set_scheduler(scheduler)
helloyongyang's avatar
helloyongyang committed
325
326
        self.transformer_infer.set_scheduler(scheduler)

TorynCurtis's avatar
TorynCurtis committed
327
328
329
330
331
332
333
334
    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
335
336
    @torch.no_grad()
    def infer(self, inputs):
337
338
339
340
341
        if self.cpu_offload:
            if self.offload_granularity == "model" and self.scheduler.step_index == 0:
                self.to_cuda()
            elif self.offload_granularity != "model":
                self.pre_weight.to_cuda()
gushiqiao's avatar
gushiqiao committed
342
                self.transformer_weights.non_block_weights_to_cuda()
343

344
        if self.config["enable_cfg"]:
helloyongyang's avatar
helloyongyang committed
345
346
347
348
349
350
351
            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
352
                    noise_pred = self._infer_cond_uncond(inputs, infer_condition=True)
helloyongyang's avatar
helloyongyang committed
353
                else:
helloyongyang's avatar
helloyongyang committed
354
                    noise_pred = self._infer_cond_uncond(inputs, infer_condition=False)
helloyongyang's avatar
helloyongyang committed
355

helloyongyang's avatar
helloyongyang committed
356
357
358
359
360
361
                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
362
363
                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
364

helloyongyang's avatar
helloyongyang committed
365
366
367
            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
368
            self.scheduler.noise_pred = self._infer_cond_uncond(inputs, infer_condition=True)
369
370
371
372
373

        if self.cpu_offload:
            if self.offload_granularity == "model" and self.scheduler.step_index == self.scheduler.infer_steps - 1:
                self.to_cpu()
            elif self.offload_granularity != "model":
root's avatar
root committed
374
                self.pre_weight.to_cpu()
gushiqiao's avatar
gushiqiao committed
375
                self.transformer_weights.non_block_weights_to_cpu()
gushiqiao's avatar
gushiqiao committed
376

Yang Yong(雍洋)'s avatar
Yang Yong(雍洋) committed
377
    @compiled_method()
378
    @torch.no_grad()
helloyongyang's avatar
helloyongyang committed
379
380
381
382
    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
383
384
385
386
387
388
389
390
391

        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
392
        noise_pred = self.post_infer.infer(x, pre_infer_out)[0]
helloyongyang's avatar
helloyongyang committed
393
394
395
396
397
398
399
400
401

        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
402
        x = pre_infer_out.x
helloyongyang's avatar
helloyongyang committed
403
404
405
406
407
        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
408
            x = F.pad(x, (0, 0, 0, padding_size))
helloyongyang's avatar
helloyongyang committed
409

helloyongyang's avatar
helloyongyang committed
410
        pre_infer_out.x = torch.chunk(x, world_size, dim=0)[cur_rank]
helloyongyang's avatar
helloyongyang committed
411

sandy's avatar
sandy committed
412
        if self.config["model_cls"] in ["wan2.2", "wan2.2_audio"] and self.config["task"] == "i2v":
helloyongyang's avatar
helloyongyang committed
413
414
415
416
417
418
419
            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
420
421
            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
422
423
424
425
426
427
428
429
430

        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
431
        return combined_output