File size: 22,906 Bytes
c4f8e4a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
"""
InternVLA M1 framework:
Vision-Language-Action diffusion model integrating:
  - Qwen2.5 vision-language backbone
  - Layer-wise QFormer aggregation
  - DINO multi-view visual encoder
  - DiT diffusion head for future action sequence prediction
Primary goal: predict continuous future actions conditioned on multi-view images + instruction.
"""

from typing import List
from tqdm import tqdm
from typing import List, Optional, Tuple
import torch
import torch.nn as nn
import numpy as np
from PIL import Image
from qwen_vl_utils import process_vision_info


from InternVLA.training.trainer_utils import initialize_overwatch

logger = initialize_overwatch(__name__)

# HuggingFace Default / LLaMa-2 IGNORE_INDEX (for labels)
IGNORE_INDEX = -100

from InternVLA.model.framework.base_framework import baseframework
from InternVLA.model.modules.vlm.QWen2_5 import get_qwen2_5_interface
from InternVLA.model.modules.projector.QFormer import get_layerwise_qformer
from InternVLA.model.modules.action_model.DiTActionHeader import get_action_model
from InternVLA.model.modules.dino_model.dino import get_dino_model
from InternVLA.training.trainer_utils.metrics import resize_images


class InternVLA_M1(baseframework):
    """
    Multimodal vision-language-action model.

    Components:
      - Qwen2.5 VL interface for fused language/vision token embeddings
      - Layer-wise QFormer for multi-layer feature aggregation
      - DINO encoder for dense multi-view spatial tokens
      - DiT diffusion head for future action sequence modeling

    Focus: Predict future continuous actions conditioned on images + instruction.
    """

    def __init__(
        self,
        config: Optional[dict] = None,
        **kwargs,
    ) -> None:
        """
        Construct all submodules and cache key configuration values.

        Args:
            config: Hierarchical configuration (OmegaConf/dict) containing framework + trainer sections.
            **kwargs: Reserved for future overrides (unused).
        """
        super().__init__()
        self.config = config
        self.qwen_vl_interface = get_qwen2_5_interface(config=self.config)
        self.layer_qformer = get_layerwise_qformer(config=self.config)
        self.action_model = get_action_model(config=self.config)
        self.dino_encoder = get_dino_model(backone_name=getattr(self.config.framework.dino, "dino_backbone", "dinov2_vits14"))
        self.dino_pro = nn.Linear(
            in_features=self.dino_encoder.num_channels,
            out_features=self.qwen_vl_interface.model.config.hidden_size
        )

        self.future_action_window_size = config.framework.action_model.future_action_window_size
        self.past_action_window_size = config.framework.action_model.past_action_window_size

    def forward(
        self,
        examples: List[dict] = None,
        **kwargs,
    ) -> Tuple:
        """
        Forward pass for training (diffusion objective).

        Flow:
          1. Build QwenVL inputs (images + instruction tokens)
          2. Extract hidden states from configured layer range
          3. Encode images with DINO, flatten multi-view tokens and project
          4. Concatenate per-layer language tokens with visual tokens
          5. Fuse via layer-wise QFormer -> action condition embeddings
          6. Prepare repeated future action windows (for diffusion efficiency)
          7. Predict noise and compute diffusion loss

        Args:
            examples: List[dict], each dict requires:
                - image: List[PIL.Image] (multi-view)
                - lang: str instruction
                - action: np.ndarray or list shaped [T, action_dim]
            **kwargs: Reserved.

        Returns:
            dict:
                action_loss (torch.Tensor): Scalar diffusion noise prediction loss.
        """
        batch_images = [example["image"] for example in examples]  #  [B,[PLT]]
        instructions = [example["lang"] for example in examples]  # [B, str]
        actions = [example["action"] for example in examples]  # label [B, len, 7]

        # Step 1: QWenVL input format
        qwen_inputs = self.qwen_vl_interface.build_qwenvl_inputs(images=batch_images, instructions=instructions)
        with torch.autocast("cuda", dtype=torch.bfloat16):
            qwenvl_outputs = self.qwen_vl_interface(
                **qwen_inputs,
                output_attentions=False,
                output_hidden_states=True,
                return_dict=True,
            )
            pass

        # Step 2: DINO Forward
        image_tensors = self.dino_encoder.prepare_dino_input(batch_images)  #
        B = len(batch_images)
        dino_features = self.dino_encoder(image_tensors)  # DINO output is [B*num_view, token, dim]
        dino_encoded_features = dino_features.reshape(B, -1, dino_features.shape[-1])  # [B, num_view * token, dim]
        dino_encoded_features = self.dino_pro(dino_encoded_features)  # [B, num_view * token, hidden_size]

        # Step 3: aggregation condition for Action expert
        start_layer = self.config.framework.layer_qformer.qformer_start_layer
        end_layer = self.config.framework.layer_qformer.qformer_end_layer
        condition_features = qwenvl_outputs.hidden_states[start_layer:end_layer]

        cat_conditions = []
        for layer_index in range(len(condition_features)):
            layer_features = condition_features[layer_index]  # [B, n_qformer_token, D]
            layer_features = torch.cat(
                [layer_features, dino_encoded_features], dim=1
            )  # [B, n_qformer_token + num_view * token, D]
            cat_conditions.append(layer_features)

        action_condition = self.layer_qformer(cat_conditions)  # [B, 64, D_action]

        # Step 4: Action Expert Forward and Loss
        with torch.autocast("cuda", dtype=torch.float32):

            # here is a tips to accelerate training speed, by repeating each sample for several times @ref to CogACT
            actions = torch.tensor(np.array(actions), device=action_condition.device)  # [B, chunk, 7]
            actions_future = actions[:, -(self.future_action_window_size + 1) :, :]

            # tips: Repeat 'actions' 'repeated_diffusion_steps' times, resulting in [repeated_diffusion_steps*B, T, D]
            repeated_diffusion_steps = (
                self.config.trainer.get("repeated_diffusion_steps", 4) if self.config and self.config.trainer else 4
            )
            actions_repeated = actions_future.repeat(repeated_diffusion_steps, 1, 1)
            action_condition = action_condition.repeat(
                repeated_diffusion_steps, 1, 1
            )  # [repeated_diffusion_steps*B, T, D_action]

            # DiT noise add and predict
            noise_pred, noise, timestep = self.action_model(actions_repeated, action_condition)

            # perdition loss
            action_loss = self.action_model.loss(noise_pred, noise)

        return {"action_loss": action_loss}

    @torch.inference_mode()
    def predict_action(
        self,
        batch_images: List[List[Image.Image]],  # B * List of PIL Image as [view1, view2]
        instructions: List[str],
        cfg_scale: float = 1.5,
        use_ddim: bool = True,
        num_ddim_steps: int = 5,
        resize_image = [224, 224],
        **kwargs: str,
    ) -> np.ndarray:
        """
        Inference: generate future normalized action sequence via diffusion sampling.

        Steps:
          1. Resize images to training resolution (if specified)
          2. Encode with QwenVL (hidden states retained)
          3. Extract DINO tokens and project to vlm hidden size
          4. Build multi-layer fused QwenVL and DINO features via QFormer
          5. Run diffusion sampling (DDIM optional, CFG optional)
          6. Return normalized action trajectory

        Args:
            batch_images: List of samples; each sample is List[PIL.Image] (multi-view).
            instructions: List[str] natural language task instructions.
            cfg_scale: >1 enables classifier-free guidance (scales conditional vs unconditional).
            use_ddim: Whether to use DDIM deterministic sampling.
            num_ddim_steps: Number of DDIM steps if enabled.
            **kwargs: Reserved.

        Returns:
            dict:
                normalized_actions (np.ndarray): Shape [B, T, action_dim], diffusion-sampled normalized actions.
        """
        # align obs and lang # is policy's duty to make sure the image size?
        train_obs_image_size = getattr(self.config.datasets.vla_data, "image_size", None)
        print("resize image from:", batch_images[0][0].size)
        if train_obs_image_size:
            batch_images = resize_images(batch_images, target_size=train_obs_image_size)
            print("resize image to:", batch_images[0][0].size)
        instructions = [instruction.lower() for instruction in instructions]

        inferface_inputs = self.qwen_vl_interface.build_qwenvl_inputs(images=batch_images, instructions=instructions)
        qwen_inputs = inferface_inputs

        with torch.autocast("cuda", dtype=torch.bfloat16):
            qwenvl_outputs = self.qwen_vl_interface(
                **qwen_inputs,
                output_hidden_states=True,
                return_dict=True,
            )

            B = len(batch_images) # dino don't have smart resize in processing
            image_tensors = self.dino_encoder.prepare_dino_input(batch_images)
            dino_features = self.dino_encoder(image_tensors)

            B = len(batch_images)
            dino_encoded_features = dino_features.reshape(B, -1, dino_features.shape[-1])  # [B, num_view * token, dim]
            dino_encoded_features = self.dino_pro(dino_encoded_features)  # [B, 256, D]

        with torch.autocast("cuda", dtype=torch.bfloat16):

            start_layer = self.config.framework.layer_qformer.qformer_start_layer
            end_layer = self.config.framework.layer_qformer.qformer_end_layer
            condition_features = qwenvl_outputs.hidden_states[start_layer:end_layer] # final transformer block
            cat_conditions = []
            for layer_index in range(len(condition_features)):
                layer_features = condition_features[layer_index]  # [B, n_qformer_token, D]
                layer_features = torch.cat(
                    [layer_features, dino_encoded_features], dim=1
                )  # [B, n_qformer_token + num_view * token, D]
                cat_conditions.append(layer_features)

            action_condition_feature = self.layer_qformer(cat_conditions)  # [B, 64, D_action]

            using_cfg = cfg_scale > 1.0

            model_dtype = next(self.action_model.net.parameters()).dtype
            B = action_condition_feature.shape[0]

            # Sample random noise
            noise = torch.randn(
                B,
                self.future_action_window_size + 1,
                self.action_model.in_channels,
                device=action_condition_feature.device,
            ).to(
                model_dtype
            )  # [B, T, D]

            # Setup classifier-free guidance:
            if using_cfg:
                noise = torch.cat([noise, noise], 0)  # [2,16,7]
                uncondition = self.action_model.net.z_embedder.uncondition  # [64, 768]
                uncondition_shape = uncondition.shape
                uncondition = uncondition.unsqueeze(0)  # [1, 64, D]
                uncondition = uncondition.expand(
                    B, uncondition_shape[0], uncondition_shape[1]
                )  # [B, n_qformer_token, D]
                z = torch.cat([action_condition_feature, uncondition], 0)  # [2, 64, 768]
                cfg_scale = cfg_scale
                model_kwargs = dict(z=z, cfg_scale=cfg_scale)
                sample_fn = self.action_model.net.forward_with_cfg
            else:
                model_kwargs = dict(z=action_condition_feature)
                sample_fn = self.action_model.net.forward

            # DDIM Sampling
            if use_ddim and num_ddim_steps is not None:
                if self.action_model.ddim_diffusion is None:
                    self.action_model.create_ddim(ddim_step=num_ddim_steps)
                samples = self.action_model.ddim_diffusion.ddim_sample_loop(
                    sample_fn,
                    noise.shape,
                    noise,
                    clip_denoised=False,
                    model_kwargs=model_kwargs,
                    progress=False,
                    device=action_condition_feature.device,
                    eta=0.0,
                )

            if using_cfg:
                samples, _ = samples.chunk(2, dim=0)  # Remove null class samples
            normalized_actions = samples.cpu().numpy()
            
            raw_actions = None
     
        return {"normalized_actions": normalized_actions}  # [B, T, action_dim]

    # @torch.inference_mode()
    def predict_action_rtc(
        self,
        batch_images: List[List[Image.Image]],  # B * List of PIL Image as [view1, view2]
        instructions: List[str],
        cfg_scale: float = 1.5,
        use_ddim: bool = True,
        num_ddim_steps: int = 5,
        resize_image = [224, 224],
        prev_actions: np.ndarray = None,
        inference_delay: int = 0,
        execution_horizon: int = 0,
        mask_schedule: str = "exponential",
        guidance_weight: float = 5.0,
        **kwargs: str,
    ) -> np.ndarray:
        """
        Inference: generate future normalized action sequence via diffusion sampling.

        Steps:
          1. Resize images to training resolution (if specified)
          2. Encode with QwenVL (hidden states retained)
          3. Extract DINO tokens and project to vlm hidden size
          4. Build multi-layer fused QwenVL and DINO features via QFormer
          5. Run diffusion sampling (DDIM optional, CFG optional)
          6. Return normalized action trajectory

        Args:
            batch_images: List of samples; each sample is List[PIL.Image] (multi-view).
            instructions: List[str] natural language task instructions.
            cfg_scale: >1 enables classifier-free guidance (scales conditional vs unconditional).
            use_ddim: Whether to use DDIM deterministic sampling.
            num_ddim_steps: Number of DDIM steps if enabled.
            prev_actions: Previous action chunk [B, H, D] (from last inference)
            inference_delay: Number of frozen steps d (already executed during inference)
            execution_horizon: Execution horizon s (non-overlapping steps at end).
            mask_schedule: Soft mask schedule - 'exponential' (default), 'linear', or 'hard'
            guidance_weight: Max guidance weight β (default=5.0, as in paper)
            **kwargs: Reserved.

        Returns:
            dict:
                normalized_actions (np.ndarray): Shape [B, T, action_dim], diffusion-sampled normalized actions.
        """
        # align obs and lang # is policy's duty to make sure the image size?
        train_obs_image_size = getattr(self.config.datasets.vla_data, "image_size", None)
        print("resize image from:", batch_images[0][0].size)
        if train_obs_image_size:
            batch_images = resize_images(batch_images, target_size=train_obs_image_size)
            print("resize image to:", batch_images[0][0].size)
        instructions = [instruction.lower() for instruction in instructions]

        inferface_inputs = self.qwen_vl_interface.build_qwenvl_inputs(images=batch_images, instructions=instructions)
        qwen_inputs = inferface_inputs

        with torch.autocast("cuda", dtype=torch.bfloat16):
            qwenvl_outputs = self.qwen_vl_interface(
                **qwen_inputs,
                output_hidden_states=True,
                return_dict=True,
            )

            B = len(batch_images) # dino don't have smart resize in processing
            image_tensors = self.dino_encoder.prepare_dino_input(batch_images)
            dino_features = self.dino_encoder(image_tensors)

            B = len(batch_images)
            dino_encoded_features = dino_features.reshape(B, -1, dino_features.shape[-1])  # [B, num_view * token, dim]
            dino_encoded_features = self.dino_pro(dino_encoded_features)  # [B, 256, D]

        with torch.autocast("cuda", dtype=torch.bfloat16):

            start_layer = self.config.framework.layer_qformer.qformer_start_layer
            end_layer = self.config.framework.layer_qformer.qformer_end_layer
            condition_features = qwenvl_outputs.hidden_states[start_layer:end_layer]
            cat_conditions = []
            for layer_index in range(len(condition_features)):
                layer_features = condition_features[layer_index]  # [B, n_qformer_token, D]
                layer_features = torch.cat(
                    [layer_features, dino_encoded_features], dim=1
                )  # [B, n_qformer_token + num_view * token, D]
                cat_conditions.append(layer_features)

            action_condition_feature = self.layer_qformer(cat_conditions)  # [B, 64, D_action]

            using_cfg = cfg_scale > 1.0

            model_dtype = next(self.action_model.net.parameters()).dtype
            B = action_condition_feature.shape[0]

            # Sample random noise
            noise = torch.randn(
                B,
                self.future_action_window_size + 1,
                self.action_model.in_channels,
                device=action_condition_feature.device,
            ).to(
                model_dtype
            )  # [B, T, D]

            # Setup classifier-free guidance:
            if using_cfg:
                noise = torch.cat([noise, noise], 0)  # [2,16,7]
                uncondition = self.action_model.net.z_embedder.uncondition  # [64, 768]
                uncondition_shape = uncondition.shape
                uncondition = uncondition.unsqueeze(0)  # [1, 64, D]
                uncondition = uncondition.expand(
                    B, uncondition_shape[0], uncondition_shape[1]
                )  # [B, n_qformer_token, D]
                z = torch.cat([action_condition_feature, uncondition], 0)  # [2, 64, 768]
                cfg_scale = cfg_scale
                model_kwargs = dict(z=z, cfg_scale=cfg_scale)
                sample_fn = self.action_model.net.forward_with_cfg
            else:
                model_kwargs = dict(z=action_condition_feature)
                sample_fn = self.action_model.net.forward

            # DDIM Sampling
            if use_ddim and num_ddim_steps is not None:
                if self.action_model.ddim_diffusion is None:
                    self.action_model.create_ddim(ddim_step=num_ddim_steps)
                samples = self.action_model.ddim_diffusion.ddim_sample_loop_rtc(
                    sample_fn,
                    noise.shape,
                    noise,
                    prev_actions=torch.tensor(prev_actions, device=noise.device),
                    inference_delay=inference_delay,
                    execution_horizon=execution_horizon,
                    mask_schedule=mask_schedule,
                    guidance_weight=guidance_weight,
                    clip_denoised=False,
                    model_kwargs=model_kwargs,
                    progress=False,
                    device=action_condition_feature.device,
                    eta=0.0,
                )

            if using_cfg:
                samples, _ = samples.chunk(2, dim=0)  # Remove null class samples
            normalized_actions = samples.cpu().numpy()
            
            raw_actions = None
     
        return {"normalized_actions": normalized_actions}  # [B, T, action_dim]

    @torch.inference_mode()
    def chat_with_M1(
        self,
        image: Image.Image,
        text: str,
        max_new_tokens: int = 128,
        device: Optional[str] = "cuda",
    ) -> List[str]:
        processor = getattr(self.qwen_vl_interface, "processor", None)
        model = getattr(self.qwen_vl_interface, "model", None)
        # if processor is None or model is None:
        #     raise RuntimeError("qwen_vl_interface 缺少 processor 或 model。")

        messages0 = [
            {
                "role": "user",
                "content": [
                    {
                        "type": "image",
                        "image": image,
                    },
                    {"type": "text", "text": text},
                ],
            }
        ]

        messages = [messages0]
        # text info
        texts = [processor.apply_chat_template(msg, tokenize=False, add_generation_prompt=True) for msg in messages]
        # visual info
        image_inputs, video_inputs = process_vision_info(messages)

        # tokenizer
        inputs = processor(
            text=texts,
            images=image_inputs,
            videos=video_inputs,
            padding=True,
            return_tensors="pt",
        ).to(device)

        model.eval()
        generated_ids = model.generate(**inputs, max_new_tokens=max_new_tokens)
        generated_ids_trimmed = [out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
        outputs = processor.batch_decode(
            generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
        )
        return outputs


def build_model_framework(config: dict = {}) -> InternVLA_M1:
    """
    Factory helper to build InternVLA_M1 with provided config.

    Args:
        config: Dict or OmegaConf containing framework/component settings.

    Returns:
        InternVLA_M1: Initialized model instance.
    """
    model = InternVLA_M1(config=config)
    return model


if __name__ == "__main__":
    from omegaconf import OmegaConf

    # model parameters
    import debugpy

    debugpy.listen(("0.0.0.0", 10092))
    print("🔍 Rank 0 waiting for debugger attach on port 10092...")
    debugpy.wait_for_client()

    config_yaml = "InternVLA/config/training/internvla_cotrain_oxe.yaml"
    cfg = OmegaConf.load(config_yaml)

    # try get model
    model = build_model_framework(cfg)
    print(model)

 
    # try forward model
    # can be fake sample, but here get from dataloader for simpler
    from InternVLA.dataloader.lerobot_datasets import get_vla_dataset, collate_fn

    vla_dataset_cfg = cfg.datasets.vla_data
    dataset = get_vla_dataset(data_cfg=vla_dataset_cfg)

    from torch.utils.data import DataLoader

    train_dataloader = DataLoader(
        dataset,
        batch_size=2,
        num_workers=1,  # For Debug
        collate_fn=collate_fn,
    )

    for batch in tqdm(train_dataloader, desc="Processing Batches"):
        batch
        break

    # try get model
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = model.to(device)
    model(batch)