File size: 28,914 Bytes
22d08f8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
import os
from dataclasses import dataclass, field
from pathlib import Path
from typing import Callable, List, Optional

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS

from rosetta.utils import DataClassMixin, PRECISION_TO_TYPE

MODEL_ZOO: dict[str, "QwenViTConfig"] = {}


def register_model_config(name, base=None, **kwargs):
    if base is not None:
        if base not in MODEL_ZOO:
            raise ValueError(f"Base model {base} not found in MODEL_ZOO. Valid models: {list(MODEL_ZOO.keys())}")
        base_config = MODEL_ZOO[base].to_dict()
        base_config.update({**kwargs, "name": name})
        MODEL_ZOO[name] = QwenViTConfig(**base_config)
    else:
        MODEL_ZOO[name] = QwenViTConfig(name=name, **kwargs)


@dataclass
class QwenViTConfig(DataClassMixin):
    name: str = ""
    depth: int = 27
    hidden_size: int = 1152
    intermediate_size: int = 4304
    num_heads: int = 16
    in_channels: int = 3
    patch_size: int = 16
    spatial_merge_size: int = 2
    temporal_patch_size: int = 2
    out_hidden_size: int = 3584
    num_position_embeddings: int = 2304
    deepstack_visual_indexes: List[int] = field(default_factory=lambda: [8, 16, 24])
    initializer_range: float = 0.02

    @classmethod
    def from_name(cls, model_name: str, **kwargs) -> "QwenViTConfig":
        if model_name not in MODEL_ZOO:
            raise ValueError(f"Model {model_name} not found in MODEL_ZOO. Valid models: {list(MODEL_ZOO.keys())}")
        model_config = MODEL_ZOO[model_name].to_dict()
        model_config.update(kwargs)
        return cls(**model_config)


register_model_config(
    name="qwen3vl-vit-for-30b-a3b",
    depth=27,
    hidden_size=1152,
    intermediate_size=4304,
    num_heads=16,
    in_channels=3,
    patch_size=16,
    spatial_merge_size=2,
    temporal_patch_size=2,
    out_hidden_size=2048,
    num_position_embeddings=2304,
    deepstack_visual_indexes=[8, 16, 24],
)

register_model_config(
    name="qwen3vl-vit-for-0.6b",
    base="qwen3vl-vit-for-30b-a3b",
    out_hidden_size=1024,
)

def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: float,
    dropout: float = 0.0,
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()
    return attn_output, attn_weights


class Qwen3VLVisionPatchMerger(nn.Module):
    def __init__(self, hidden_size, spatial_merge_size, out_hidden_size, use_postshuffle_norm=False) -> None:
        super().__init__()
        self.hidden_size = hidden_size * (spatial_merge_size**2)
        self.use_postshuffle_norm = use_postshuffle_norm
        self.norm = nn.LayerNorm(self.hidden_size if use_postshuffle_norm else hidden_size, eps=1e-6)
        self.linear_fc1 = nn.Linear(self.hidden_size, self.hidden_size)
        self.act_fn = nn.GELU()
        self.linear_fc2 = nn.Linear(self.hidden_size, out_hidden_size)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.norm(x.view(-1, self.hidden_size) if self.use_postshuffle_norm else x).view(-1, self.hidden_size)
        x = self.linear_fc2(self.act_fn(self.linear_fc1(x)))
        return x


class Qwen3VLVisionRotaryEmbedding(nn.Module):
    inv_freq: torch.Tensor  # fix linting for `register_buffer`

    def __init__(self, dim: int, theta: float = 10000.0) -> None:
        super().__init__()
        self.dim = dim
        self.theta = theta
        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=True)


    def reset_parameters(self):
        inv_freq = 1.0 / (self.theta ** (torch.arange(0, self.dim, 2, dtype=torch.float) / self.dim))
        self.register_buffer("inv_freq", inv_freq, persistent=True)

    def forward(self, seqlen: int) -> torch.Tensor:
        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
        freqs = torch.outer(seq, self.inv_freq)
        return freqs


class Qwen3VLVisionPatchEmbed(nn.Module):
    def __init__(self, patch_size, temporal_patch_size, in_channels, hidden_size) -> None:
        super().__init__()
        self.patch_size = patch_size
        self.temporal_patch_size = temporal_patch_size
        self.in_channels = in_channels
        self.embed_dim = hidden_size

        kernel_size = [self.temporal_patch_size, self.patch_size, self.patch_size]
        self.proj = nn.Conv3d(self.in_channels, self.embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=True)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        target_dtype = self.proj.weight.dtype
        hidden_states = hidden_states.view(
            -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
        )
        hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
        return hidden_states


def rotate_half(x):
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb_vision(
    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
    orig_q_dtype = q.dtype
    orig_k_dtype = k.dtype
    q, k = q.float(), k.float()
    cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float()
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    q_embed = q_embed.to(orig_q_dtype)
    k_embed = k_embed.to(orig_k_dtype)
    return q_embed, k_embed


class Qwen3VLVisionAttention(nn.Module):
    def __init__(self, hidden_size, num_heads, attn_implementation="sdpa") -> None:
        super().__init__()
        self.dim = hidden_size
        self.num_heads = num_heads
        self.head_dim = self.dim // self.num_heads
        self.num_key_value_groups = 1  # needed for eager attention
        self.qkv = nn.Linear(self.dim, self.dim * 3, bias=True)
        self.proj = nn.Linear(self.dim, self.dim)
        self.scaling = self.head_dim**-0.5
        self._attn_implementation = attn_implementation
        self.attention_dropout = 0.0
        self.is_causal = False
        # Create a minimal config object for transformers flash attention compatibility
        self.config = type('Config', (), {'_attn_implementation': attn_implementation})()

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
        **kwargs,
    ) -> torch.Tensor:
        seq_length = hidden_states.shape[0]
        query_states, key_states, value_states = (
            self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
        )
        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb_vision(query_states, key_states, cos, sin)

        query_states = query_states.transpose(0, 1).unsqueeze(0)
        key_states = key_states.transpose(0, 1).unsqueeze(0)
        value_states = value_states.transpose(0, 1).unsqueeze(0)

        attention_interface: Callable = eager_attention_forward
        if self._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self._attn_implementation]

        if self._attn_implementation == "flash_attention_2":
            # Flash Attention 2: Use cu_seqlens for variable length attention
            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
            attn_output, _ = attention_interface(
                self,
                query_states,
                key_states,
                value_states,
                attention_mask=None,
                scaling=self.scaling,
                dropout=0.0 if not self.training else self.attention_dropout,
                cu_seq_lens_q=cu_seqlens,
                cu_seq_lens_k=cu_seqlens,
                max_length_q=max_seqlen,
                max_length_k=max_seqlen,
                is_causal=False,
                **kwargs,
            )
        else:
            # Other implementations: Process each chunk separately
            lengths = cu_seqlens[1:] - cu_seqlens[:-1]
            splits = [
                torch.split(tensor, lengths.tolist(), dim=2) for tensor in (query_states, key_states, value_states)
            ]

            attn_outputs = [
                attention_interface(
                    self,
                    q,
                    k,
                    v,
                    attention_mask=None,
                    scaling=self.scaling,
                    dropout=0.0 if not self.training else self.attention_dropout,
                    is_causal=False,
                    **kwargs,
                )[0]
                for q, k, v in zip(*splits)
            ]
            attn_output = torch.cat(attn_outputs, dim=1)

        attn_output = attn_output.reshape(seq_length, -1).contiguous()
        attn_output = self.proj(attn_output)
        return attn_output


class GELUTanh(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.act_fn = nn.GELU(approximate="tanh")

    def forward(self, input: torch.Tensor) -> torch.Tensor:
        return self.act_fn(input)


class Qwen3VLVisionMLP(nn.Module):
    def __init__(self, hidden_size, intermediate_size):
        super().__init__()
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.linear_fc1 = nn.Linear(self.hidden_size, self.intermediate_size, bias=True)
        self.linear_fc2 = nn.Linear(self.intermediate_size, self.hidden_size, bias=True)
        self.act_fn = GELUTanh()

    def forward(self, hidden_state):
        return self.linear_fc2(self.act_fn(self.linear_fc1(hidden_state)))


class Qwen3VLVisionBlock(nn.Module):
    def __init__(self, hidden_size, num_heads, intermediate_size, attn_implementation: str = "sdpa") -> None:
        super().__init__()
        self.norm1 = nn.LayerNorm(hidden_size, eps=1e-6)
        self.norm2 = nn.LayerNorm(hidden_size, eps=1e-6)
        self.attn = Qwen3VLVisionAttention(hidden_size, num_heads, attn_implementation)
        self.mlp = Qwen3VLVisionMLP(hidden_size, intermediate_size)

    def forward(
        self,
        hidden_states: torch.Tensor,
        cu_seqlens: torch.Tensor,
        rotary_pos_emb: Optional[torch.Tensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
        **kwargs,
    ) -> torch.Tensor:
        attn = self.attn(
            self.norm1(hidden_states),
            cu_seqlens=cu_seqlens,
            rotary_pos_emb=rotary_pos_emb,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = hidden_states + attn
        mlp = self.mlp(self.norm2(hidden_states))
        hidden_states = hidden_states + mlp
        return hidden_states


class Qwen3VLVisionModel(nn.Module):
    def __init__(self, config: QwenViTConfig, attn_implementation="sdpa", **kwargs):
        super().__init__()
        self.config = config
        self.spatial_merge_size = config.spatial_merge_size
        self.patch_size = config.patch_size
        self.spatial_merge_unit = self.spatial_merge_size * self.spatial_merge_size

        self.patch_embed = Qwen3VLVisionPatchEmbed(
            config.patch_size,
            config.temporal_patch_size,
            config.in_channels,
            config.hidden_size,
        )

        self.pos_embed = nn.Embedding(config.num_position_embeddings, config.hidden_size)
        self.num_grid_per_side = int(config.num_position_embeddings**0.5)

        head_dim = config.hidden_size // config.num_heads
        self.rotary_pos_emb = Qwen3VLVisionRotaryEmbedding(head_dim // 2)

        self.blocks = nn.ModuleList(
            [
                Qwen3VLVisionBlock(config.hidden_size, config.num_heads, config.intermediate_size, attn_implementation) for _ in range(config.depth)
            ]
        )
        self.merger = Qwen3VLVisionPatchMerger(
            config.hidden_size,
            config.spatial_merge_size,
            config.out_hidden_size,
            use_postshuffle_norm=False,
        )

        self.deepstack_visual_indexes = config.deepstack_visual_indexes
        self.deepstack_merger_list = nn.ModuleList(
            [
                Qwen3VLVisionPatchMerger(
                    config.hidden_size,
                    config.spatial_merge_size,
                    config.out_hidden_size,
                    use_postshuffle_norm=True,
                )
                for _ in range(len(config.deepstack_visual_indexes))
            ]
        )
        self.gradient_checkpointing = False

    def rot_pos_emb(self, grid_thw: torch.Tensor) -> torch.Tensor:
        merge_size = self.spatial_merge_size

        max_hw = int(grid_thw[:, 1:].max().item())
        freq_table = self.rotary_pos_emb(max_hw)  # (max_hw, dim // 2)
        device = freq_table.device

        total_tokens = int(torch.prod(grid_thw, dim=1).sum().item())
        pos_ids = torch.empty((total_tokens, 2), dtype=torch.long, device=device)

        offset = 0
        for num_frames, height, width in grid_thw:
            merged_h, merged_w = height // merge_size, width // merge_size

            block_rows = torch.arange(merged_h, device=device)  # block row indices
            block_cols = torch.arange(merged_w, device=device)  # block col indices
            intra_row = torch.arange(merge_size, device=device)  # intra-block row offsets
            intra_col = torch.arange(merge_size, device=device)  # intra-block col offsets

            # Compute full-resolution positions
            row_idx = block_rows[:, None, None, None] * merge_size + intra_row[None, None, :, None]
            col_idx = block_cols[None, :, None, None] * merge_size + intra_col[None, None, None, :]

            row_idx = row_idx.expand(merged_h, merged_w, merge_size, merge_size).reshape(-1)
            col_idx = col_idx.expand(merged_h, merged_w, merge_size, merge_size).reshape(-1)

            coords = torch.stack((row_idx, col_idx), dim=-1)

            if num_frames > 1:
                coords = coords.repeat(num_frames, 1)

            num_tokens = coords.shape[0]
            pos_ids[offset : offset + num_tokens] = coords
            offset += num_tokens

        embeddings = freq_table[pos_ids]  # lookup rotary embeddings
        embeddings = embeddings.flatten(1)
        return embeddings

    def fast_pos_embed_interpolate(self, grid_thw):
        grid_ts, grid_hs, grid_ws = grid_thw[:, 0], grid_thw[:, 1], grid_thw[:, 2]

        idx_list = [[] for _ in range(4)]
        weight_list = [[] for _ in range(4)]

        for t, h, w in zip(grid_ts, grid_hs, grid_ws):
            # Convert tensor to Python int for torch.linspace
            h_int = h.item() if isinstance(h, torch.Tensor) else int(h)
            w_int = w.item() if isinstance(w, torch.Tensor) else int(w)
            h_idxs = torch.linspace(0, self.num_grid_per_side - 1, h_int)
            w_idxs = torch.linspace(0, self.num_grid_per_side - 1, w_int)

            h_idxs_floor = h_idxs.int()
            w_idxs_floor = w_idxs.int()
            h_idxs_ceil = (h_idxs.int() + 1).clip(max=self.num_grid_per_side - 1)
            w_idxs_ceil = (w_idxs.int() + 1).clip(max=self.num_grid_per_side - 1)

            dh = h_idxs - h_idxs_floor
            dw = w_idxs - w_idxs_floor

            base_h = h_idxs_floor * self.num_grid_per_side
            base_h_ceil = h_idxs_ceil * self.num_grid_per_side

            indices = [
                (base_h[None].T + w_idxs_floor[None]).flatten(),
                (base_h[None].T + w_idxs_ceil[None]).flatten(),
                (base_h_ceil[None].T + w_idxs_floor[None]).flatten(),
                (base_h_ceil[None].T + w_idxs_ceil[None]).flatten(),
            ]

            weights = [
                ((1 - dh)[None].T * (1 - dw)[None]).flatten(),
                ((1 - dh)[None].T * dw[None]).flatten(),
                (dh[None].T * (1 - dw)[None]).flatten(),
                (dh[None].T * dw[None]).flatten(),
            ]

            for i in range(4):
                idx_list[i].extend(indices[i].tolist())
                weight_list[i].extend(weights[i].tolist())

        idx_tensor = torch.tensor(idx_list, dtype=torch.long, device=self.pos_embed.weight.device)
        weight_tensor = torch.tensor(
            weight_list, dtype=self.pos_embed.weight.dtype, device=self.pos_embed.weight.device
        )
        pos_embeds = self.pos_embed(idx_tensor) * weight_tensor[:, :, None]
        patch_pos_embeds = pos_embeds[0] + pos_embeds[1] + pos_embeds[2] + pos_embeds[3]

        patch_pos_embeds = patch_pos_embeds.split([h * w for h, w in zip(grid_hs, grid_ws)])

        patch_pos_embeds_permute = []
        merge_size = self.config.spatial_merge_size
        for pos_embed, t, h, w in zip(patch_pos_embeds, grid_ts, grid_hs, grid_ws):
            pos_embed = pos_embed.repeat(t, 1)
            pos_embed = (
                pos_embed.view(t, h // merge_size, merge_size, w // merge_size, merge_size, -1)
                .permute(0, 1, 3, 2, 4, 5)
                .flatten(0, 4)
            )
            patch_pos_embeds_permute.append(pos_embed)
        patch_pos_embeds = torch.cat(patch_pos_embeds_permute)
        return patch_pos_embeds

    def process_single_hidden_states(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, **kwargs) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
                The final hidden states of the model.
            grid_thw (`torch.Tensor` of shape `(num_images_or_videos, 3)`):
                The temporal, height and width of feature shape of each image in LLM.

        Returns:
            `torch.Tensor`: hidden_states.
        """
        if hidden_states.ndim == 3:
            hidden_states = hidden_states.squeeze(0)
        hidden_states = self.patch_embed(hidden_states)

        pos_embeds = self.fast_pos_embed_interpolate(grid_thw)
        hidden_states = hidden_states + pos_embeds

        rotary_pos_emb = self.rot_pos_emb(grid_thw)

        seq_len, _ = hidden_states.size()
        hidden_states = hidden_states.reshape(seq_len, -1)
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
        position_embeddings = (emb.cos(), emb.sin())

        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
            dim=0,
            # Select dtype based on the following factors:
            #  - FA2 requires that cu_seqlens_q must have dtype int32
            #  - torch.onnx.export requires that cu_seqlens_q must have same dtype as grid_thw
            # See https://github.com/huggingface/transformers/pull/34852 for more information
            dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
        )
        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
        deepstack_feature_lists = []
        for layer_num, blk in enumerate(self.blocks):
            hidden_states = blk(
                hidden_states,
                cu_seqlens=cu_seqlens,
                position_embeddings=position_embeddings,
                **kwargs,
            )
            if layer_num in self.deepstack_visual_indexes:
                deepstack_feature = self.deepstack_merger_list[self.deepstack_visual_indexes.index(layer_num)](
                    hidden_states
                )
                deepstack_feature_lists.append(deepstack_feature)


        hidden_states = self.merger(hidden_states)
        hidden_states = hidden_states.unsqueeze(0)

        return hidden_states, deepstack_feature_lists

    def forward(self, hidden_states: torch.Tensor, grid_thw: Optional[torch.Tensor] = None, **kwargs) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.Tensor` of shape `(seq_len, hidden_size)`):
                The final hidden states of the model.
            grid_thw (`torch.Tensor` of shape `(num_images_or_videos, 3)`):
                The temporal, height and width of feature shape of each image in LLM.
            **kwargs: additional keyword arguments.
        Returns:
            `torch.Tensor`: hidden_states, (bsz, seqlen, n_embd).
            `list[torch.Tensor]`: deepstack_feature_lists, [[], [], ...]
        """

        if hidden_states.ndim == 3 and hidden_states.shape[0] > 1:
            # 如果输入的hidden_states是(batch, seq_len, hidden_size)
            batch_size, seq_len, hidden_size = hidden_states.shape

            # 处理**kwargs将其也拆分成和hidden_states_i = hidden_states[i]
            all_kwargs = []
            for i in range(batch_size):
                kwargs_i = {k: v[i:i+1] for k, v in kwargs.items()}
                all_kwargs.append(kwargs_i)

            all_hidden_states = []
            all_deepstack_feature_lists = []
            for i in range(batch_size):
                hidden_states_i = hidden_states[i]
                grid_thw_i = grid_thw[i:i+1]
                kwargs_i = all_kwargs[i]
                hidden_states_i, deepstack_feature_lists_i = self.process_single_hidden_states(hidden_states_i, grid_thw_i, **kwargs_i)
                all_hidden_states.append(hidden_states_i)
                all_deepstack_feature_lists.append(deepstack_feature_lists_i)

            # 合并所有hidden_states和deepstack_feature_lists
            hidden_states = torch.cat(all_hidden_states, dim=0)
            deepstack_feature_lists = all_deepstack_feature_lists
        else:
            if isinstance(grid_thw, list) and len(grid_thw) == 1:
                grid_thw = grid_thw[0]
            elif isinstance(grid_thw, list) and len(grid_thw) > 1:
                assert False, "process_single_hidden_states should be called with a single grid_thw"
            hidden_states, deepstack_feature_lists = self.process_single_hidden_states(hidden_states, grid_thw, **kwargs)

        if not len(self.deepstack_visual_indexes) > 0:
            deepstack_feature_lists = None

        return hidden_states, deepstack_feature_lists

ASSETS_BASE = os.getenv("ASSETS_BASE", "./public_assets").rstrip("/")
VISION_ENCODER_BASE = os.getenv("VISION_ENCODER_BASE", f"{ASSETS_BASE}/vision_encoder").rstrip("/")
VISION_ENCODER_META_INFO = {
    "qwen3vl-vit-for-0.6b": {
        "path": f"{VISION_ENCODER_BASE}/Qwen3-VL-30B-A3B-Instruct",
        "downsample_factor": [32, 32],
        "spatial_merge_size": 2,
        "patch_dim": 1536,
        "dummy_number": 1,
    },
}


def load_vision_model(
        vision_model_type=None,
        vision_model_precision=None,
        device=None,
        logger=None,
        require_grad=False,
        eval_mode=True,
        no_load_pretrained=False,
        vision_model_params=None,
        config=None,
):
    if logger is None:
        from loguru import logger

    if config is not None:
        vision_model_type = config["vision_model_type"]
        vision_model_precision = config.get("vision_model_precision", vision_model_precision)
        require_grad = not config.get("vision_model_freeze", not require_grad)
        eval_mode = config.get("vision_model_freeze", eval_mode)
        no_load_pretrained = config.get("no_load_pretrained_vision_model", False)
        vision_model_params = config.get("vision_model_params", vision_model_params)

    if vision_model_params is None:
        vision_model_params = {}

    if vision_model_type.startswith("qwen3vl-vit"):
        model_config = QwenViTConfig.from_name(vision_model_type)
        vision_model = Qwen3VLVisionModel(model_config, **vision_model_params)
        if not no_load_pretrained and vision_model_type in VISION_ENCODER_META_INFO:
            meta = VISION_ENCODER_META_INFO[vision_model_type]
            if "path" in meta:
                from loguru import logger as _logger
                from safetensors.torch import load_file as st_load_file
                import json
                hf_path = Path(meta["path"])
                index_file = hf_path / "model.safetensors.index.json"
                if index_file.exists():
                    with open(index_file) as _f:
                        shard_map = json.load(_f)["weight_map"]
                    # Only load keys that belong to visual encoder. HF checkpoints may
                    # store them as either "visual.*" or "model.visual.*".
                    vit_shards = set(
                        v for k, v in shard_map.items()
                        if k.startswith("visual.") or k.startswith("model.visual.")
                    )
                    vit_state = {}
                    for shard in sorted(vit_shards):
                        sd = st_load_file(hf_path / shard)
                        for k, v in sd.items():
                            if k.startswith("visual."):
                                vit_state[k[len("visual."):]] = v
                            elif k.startswith("model.visual."):
                                vit_state[k[len("model.visual."):]] = v
                    model_state = vision_model.state_dict()
                    vit_state = {
                        k: v for k, v in vit_state.items()
                        if k in model_state and model_state[k].shape == v.shape
                    }
                    has_meta_params = any(param.is_meta for param in vision_model.parameters())
                    missing, unexpected = vision_model.load_state_dict(
                        vit_state, strict=False, assign=has_meta_params
                    )
                    meta_params = [name for name, param in vision_model.named_parameters() if param.is_meta]
                    if meta_params:
                        raise RuntimeError(
                            "Vision encoder still has meta parameters after loading weights: "
                            f"{meta_params[:8]}"
                        )
                    _logger.info(
                        f"[qwen3vl-vit] Loaded backbone from {hf_path.name}: "
                        f"{len(vit_state)-len(missing)} keys loaded, "
                        f"{len(missing)} missing (e.g. shape-mismatch merger.linear_fc2 → random init), "
                        f"{len(unexpected)} unexpected."
                    )
    else:
        raise NotImplementedError(f"vision_model_type {vision_model_type} not implemented")

    if vision_model_precision is not None:
        logger.warning(f"You are transforming the Vision Encoder to {vision_model_precision}. Please make sure this is what you want.")
        if isinstance(vision_model_precision, str):
            vision_model_precision = PRECISION_TO_TYPE[vision_model_precision]
        vision_model = vision_model.to(dtype=vision_model_precision)

    if device is not None:
        vision_model = vision_model.to(device=device)

    if not require_grad:
        vision_model.requires_grad_(False)

    if eval_mode:
        vision_model.eval()

    return vision_model


def load_vision_model_processor(vision_model_type, **kwargs):
    if vision_model_type.startswith("qwen3vl-vit"):
        from transformers import AutoImageProcessor
        vision_model_meta_info = VISION_ENCODER_META_INFO[vision_model_type]
        vision_model_path = Path(vision_model_meta_info["path"])
        processor = AutoImageProcessor.from_pretrained(vision_model_path, **kwargs)
    else:
        raise NotImplementedError(f"vision_model_type {vision_model_type} not implemented")

    return processor


load_vit = load_vision_model
load_vit_processor = load_vision_model_processor