File size: 31,561 Bytes
c94c8c9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
import torch
from torch import Tensor
from torch.cuda.amp import autocast
from transformers import AutoModelForCausalLM, AutoModel, AutoTokenizer, AutoImageProcessor
from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
from model.build import MODEL_REGISTRY, BaseModel
from modules.build import build_module
from optim.utils import no_decay_param_group
from peft import LoraConfig, get_peft_model
from model.data_augmentation import *
import torch.nn as nn
from typing import List, Optional, Tuple, Union

def last_token_pool(last_hidden_states: Tensor,
                 attention_mask: Tensor) -> Tensor:
    left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
    if left_padding:
        return last_hidden_states[:, -1]
    else:
        sequence_lengths = attention_mask.sum(dim=1) - 1
        batch_size = last_hidden_states.shape[0]
        return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]

class Qwen3RotaryEmbedding(nn.Module):
    def __init__(
            self,
            dim=None,
            max_position_embeddings=2048,
            base=10000,
            device=None,
            scaling_factor=1.0,
            rope_type="default",
    ):
        super().__init__()
        self.rope_type = "default"
        self.max_seq_len_cached = 32768
        self.original_max_seq_len = 32768

        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
        self.rope_kwargs = {
                "rope_type": rope_type,
                "factor": scaling_factor,
                "dim": 32,
                "base": base,
                "max_position_embeddings": max_position_embeddings,
        }
        inv_freq, self.attention_scaling = self.rope_init_fn(**self.rope_kwargs)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    def _dynamic_frequency_update(self, position_ids, device):
        """
        dynamic RoPE layers should recompute `inv_freq` in the following situations:
        1 - growing beyond the cached sequence length (allow scaling)
        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
        """
        seq_len = torch.max(position_ids) + 1
        if seq_len > self.max_seq_len_cached:  # growth
            inv_freq, self.attention_scaling = self.rope_init_fn(**self.rope_kwargs)
            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
            self.max_seq_len_cached = seq_len

        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
            self.max_seq_len_cached = self.original_max_seq_len

    @torch.no_grad()
    def forward(self, x, position_ids):
        if "dynamic" in self.rope_type:
            self._dynamic_frequency_update(position_ids, device=x.device)

        # Core RoPE block
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
        position_ids_expanded = position_ids[:, None, :].float()
        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
        device_type = x.device.type
        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos()
            sin = emb.sin()

        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
        cos = cos * self.attention_scaling
        sin = sin * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


# Copied from transformers.models.llama.modeling_llama.rotate_half
def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2:]
    return torch.cat((-x2, x1), dim=-1)


# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.
    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed

class Qwen3RMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        Qwen3RMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
    
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    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)

class Qwen3Attention(nn.Module):
    """
    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
    and "Generating Long Sequences with Sparse Transformers".
    """

    def __init__(self, layer_idx = None):
        super().__init__()
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        self.hidden_size = 512
        self.num_heads = 16
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = 8
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        self.max_position_embeddings = 32768
        self.rope_theta = 1000000
        self.is_causal = True
        self.attention_dropout = 0.0
        self.use_qk_norm = True
        self.headwise_attn_output_gate = False
        self.elementwise_attn_output_gate = True
        
        qkv_bias = False
        rms_norm_eps = 1e-06
        if self.headwise_attn_output_gate:
            self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim + self.num_heads, bias=qkv_bias)
        elif self.elementwise_attn_output_gate:
            self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim * 2, bias=qkv_bias)
        else:
            self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=qkv_bias)

        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=qkv_bias)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=qkv_bias)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=qkv_bias)
        if self.use_qk_norm:
            self.q_norm = Qwen3RMSNorm(self.head_dim, eps=rms_norm_eps)
            self.k_norm = Qwen3RMSNorm(self.head_dim, eps=rms_norm_eps)

    def forward(
            self,
            hidden_states: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            past_key_value: Optional[Cache] = None,
            output_attentions: bool = False,
            use_cache: bool = False,
            cache_position: Optional[torch.LongTensor] = None,
            position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        bsz, q_len, _ = hidden_states.size()

        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)

        if self.headwise_attn_output_gate:
            query_states = query_states.view(bsz, q_len, self.num_key_value_heads, -1)
            query_states, gate_score = torch.split(query_states, [self.head_dim * self.num_key_value_groups, self.num_key_value_groups], dim=-1)
            gate_score = gate_score.reshape(bsz, q_len, -1, 1)
            query_states = query_states.reshape(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        elif self.elementwise_attn_output_gate:
            query_states = query_states.view(bsz, q_len, self.num_key_value_heads, -1)
            query_states, gate_score = torch.split(query_states, [self.head_dim * self.num_key_value_groups, self.head_dim * self.num_key_value_groups], dim=-1)
            gate_score = gate_score.reshape(bsz, q_len, -1, self.head_dim)
            query_states = query_states.reshape(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        else:
            query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        
        key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)

        if self.use_qk_norm:
            query_states = self.q_norm(query_states)
            key_states = self.k_norm(key_states)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_value is not None:
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

        # repeat k/v heads if n_kv_heads < n_heads
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
        if attention_mask is not None:  # no matter the length, we just slice it
            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
            attn_weights = attn_weights + causal_mask

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)

        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2).contiguous()

        if self.headwise_attn_output_gate or self.elementwise_attn_output_gate:
            attn_output = attn_output * torch.sigmoid(gate_score)

        attn_output = attn_output.reshape(bsz, q_len, -1)

        attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output
    
class _GlobalViewAttnBlock(nn.Module):
    """One pre-norm Transformer-style block over view tokens (B,V,D)."""
    def __init__(
        self,
        dim: int,
        num_heads: int,
        mlp_ratio: float,
        dropout: float,
        zero_init_residual: bool,
        zero_init_attn_out: bool,
    ):
        super().__init__()
        self.zero_init_residual = zero_init_residual
        self.zero_init_attn_out = zero_init_attn_out

        self.norm1 = nn.LayerNorm(dim)
        self.attn = nn.MultiheadAttention(
            embed_dim=dim,
            num_heads=num_heads,
            dropout=dropout,
            batch_first=True,
            bias=True,
        )

        self.norm2 = nn.LayerNorm(dim)
        hidden_dim = int(dim * mlp_ratio)
        self.mlp = nn.Sequential(
            nn.Linear(dim, hidden_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, dim),
            nn.Dropout(dropout),
        )

        self._init_weights()

    def forward(self, x, key_padding_mask=None):
        h = self.norm1(x)
        attn_out, _ = self.attn(
            h, h, h,
            key_padding_mask=key_padding_mask,
            need_weights=False,
        )
        x = x + attn_out
        x = x + self.mlp(self.norm2(x))
        return x

    @torch.no_grad()
    def _init_weights(self):
        # LayerNorm
        for ln in (self.norm1, self.norm2):
            nn.init.ones_(ln.weight)
            nn.init.zeros_(ln.bias)

        # MultiheadAttention: in_proj for qkv (3D, D)
        if getattr(self.attn, "in_proj_weight", None) is not None:
            nn.init.xavier_uniform_(self.attn.in_proj_weight)
        if getattr(self.attn, "in_proj_bias", None) is not None:
            nn.init.zeros_(self.attn.in_proj_bias)

        # out proj
        nn.init.xavier_uniform_(self.attn.out_proj.weight)
        if self.attn.out_proj.bias is not None:
            nn.init.zeros_(self.attn.out_proj.bias)

        # optional: start attn residual near-zero
        if self.zero_init_attn_out:
            nn.init.zeros_(self.attn.out_proj.weight)
            if self.attn.out_proj.bias is not None:
                nn.init.zeros_(self.attn.out_proj.bias)

        # MLP
        fc1: nn.Linear = self.mlp[0]
        fc2: nn.Linear = self.mlp[3]

        nn.init.xavier_uniform_(fc1.weight)
        if fc1.bias is not None:
            nn.init.zeros_(fc1.bias)

        # zero-init last projection for stable residual start (recommended)
        if self.zero_init_residual:
            nn.init.zeros_(fc2.weight)
            if fc2.bias is not None:
                nn.init.zeros_(fc2.bias)
        else:
            nn.init.xavier_uniform_(fc2.weight)
            if fc2.bias is not None:
                nn.init.zeros_(fc2.bias)
                
class _GlobalViewGatedAttnBlock(nn.Module):
    """Pre-norm Transformer block over view tokens (B,V,D) with gated residuals."""
    def __init__(
        self,
        dim: int,
        num_heads: int,
        mlp_ratio: float,
        dropout: float,
        zero_init_residual: bool,
        zero_init_attn_out: bool,
        gate_bias_init: float = -2.0,   # sigmoid(-2)≈0.12, starts near-identity (small updates)
    ):
        super().__init__()
        self.zero_init_residual = zero_init_residual
        self.zero_init_attn_out = zero_init_attn_out

        self.norm1 = nn.LayerNorm(dim)
        self.attn = nn.MultiheadAttention(
            embed_dim=dim,
            num_heads=num_heads,
            dropout=dropout,
            batch_first=True,
            bias=True,
        )

        # --- Gating for attention residual ---
        # Produces per-token, per-channel gates in (0,1)
        self.attn_gate = nn.Linear(dim, dim, bias=True)

        self.norm2 = nn.LayerNorm(dim)
        hidden_dim = int(dim * mlp_ratio)
        self.mlp = nn.Sequential(
            nn.Linear(dim, hidden_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, dim),
            nn.Dropout(dropout),
        )

        # --- Gating for MLP residual ---
        self.mlp_gate = nn.Linear(dim, dim, bias=True)

        self._init_weights(gate_bias_init=gate_bias_init)

    def forward(self, x: torch.Tensor, key_padding_mask=None) -> torch.Tensor:
        # x: (B, V, D)
        h1 = self.norm1(x)
        attn_out, _ = self.attn(
            h1, h1, h1,
            key_padding_mask=key_padding_mask,
            need_weights=False,
        )
        g_attn = torch.sigmoid(self.attn_gate(h1))          # (B, V, D)
        x = x + g_attn * attn_out

        h2 = self.norm2(x)
        mlp_out = self.mlp(h2)
        g_mlp = torch.sigmoid(self.mlp_gate(h2))            # (B, V, D)
        x = x + g_mlp * mlp_out
        return x

    @torch.no_grad()
    def _init_weights(self, gate_bias_init: float):
        # LayerNorm
        for ln in (self.norm1, self.norm2):
            nn.init.ones_(ln.weight)
            nn.init.zeros_(ln.bias)

        # MultiheadAttention: in_proj for qkv
        if getattr(self.attn, "in_proj_weight", None) is not None:
            nn.init.xavier_uniform_(self.attn.in_proj_weight)
        if getattr(self.attn, "in_proj_bias", None) is not None:
            nn.init.zeros_(self.attn.in_proj_bias)

        # out proj
        nn.init.xavier_uniform_(self.attn.out_proj.weight)
        if self.attn.out_proj.bias is not None:
            nn.init.zeros_(self.attn.out_proj.bias)

        # optional: start attn residual near-zero
        if self.zero_init_attn_out:
            nn.init.zeros_(self.attn.out_proj.weight)
            if self.attn.out_proj.bias is not None:
                nn.init.zeros_(self.attn.out_proj.bias)

        # MLP
        fc1: nn.Linear = self.mlp[0]
        fc2: nn.Linear = self.mlp[3]
        nn.init.xavier_uniform_(fc1.weight)
        if fc1.bias is not None:
            nn.init.zeros_(fc1.bias)

        if self.zero_init_residual:
            nn.init.zeros_(fc2.weight)
            if fc2.bias is not None:
                nn.init.zeros_(fc2.bias)
        else:
            nn.init.xavier_uniform_(fc2.weight)
            if fc2.bias is not None:
                nn.init.zeros_(fc2.bias)

        # Gates: start “mostly closed” so training is stable, then learn to open
        nn.init.zeros_(self.attn_gate.weight)
        nn.init.constant_(self.attn_gate.bias, gate_bias_init)

        nn.init.zeros_(self.mlp_gate.weight)
        nn.init.constant_(self.mlp_gate.bias, gate_bias_init)
                
class GlobalViewAttention(nn.Module):
    """
    Multi-layer global self-attention over multi-view tokens.

    Input:  x ∈ (B, V, D)
    Output: x' ∈ (B, V, D)
    """
    def __init__(
        self,
        dim: int,
        num_layers: int = 1,
        num_heads: int = 8,
        mlp_ratio: float = 4.0,
        dropout: float = 0.0,
        zero_init_residual: bool = True,   # recommended (stable when adding layers)
        zero_init_attn_out: bool = False,  # optional extra safety
    ):
        super().__init__()
        assert num_layers >= 1, "num_layers must be >= 1"

        self.dim = dim
        self.num_layers = num_layers
        self.num_heads = num_heads
        # self.layers = nn.ModuleList([Qwen3Attention(layer_idx) for layer_idx in range(num_layers)])
        # self.rotary_emb = Qwen3RotaryEmbedding()
        self.layers = nn.ModuleList([
            _GlobalViewAttnBlock(
                dim=dim,
                num_heads=num_heads,
                mlp_ratio=mlp_ratio,
                dropout=dropout,
                zero_init_residual=zero_init_residual,
                zero_init_attn_out=zero_init_attn_out,
            )
            for _ in range(num_layers)
        ])

    def forward(self, x, key_padding_mask=None):
        """
        x: (B, V, D)
        key_padding_mask: (B, V), True = ignore (padding)
        """
        for layer in self.layers:
            x = layer(x, key_padding_mask=key_padding_mask)
        return x
                
@MODEL_REGISTRY.register()
class OpenVocab(BaseModel):
    def __init__(self, cfg):
        super().__init__(cfg)
        self.cfg = cfg
        model_root = "fg-clip-base"
        self.pm_encoder = AutoModelForCausalLM.from_pretrained(model_root, trust_remote_code=True)
        
        # self.global_attn = GlobalViewAttention(dim=512, num_heads=8, mlp_ratio=4.0, dropout=0.1)
        
        if cfg.mode in ['warmup', 'pretrain']:
            self.frozen_model = AutoModelForCausalLM.from_pretrained(model_root, trust_remote_code=True)
            self.use_scene_cap = self.cfg.data.args.get("use_scene_cap", False)
            self.set_training_mode()
        else:
            self.text_encoder = AutoModel.from_pretrained('jinaai/jina-clip-v2', trust_remote_code=True)
            self.tokenizer =  AutoTokenizer.from_pretrained('jinaai/jina-clip-v2', trust_remote_code=True)
            self.text_encoder.text_model.output_tokens = True
            self.set_downstream_mode()

            self.head_list = self.cfg.model.heads.head_list
            for head in self.head_list:
                setattr(self, head, build_module("heads", getattr(self.cfg.model.heads, head)))

    def set_training_mode(self):
        for name, param in self.frozen_model.named_parameters():
            param.requires_grad = False
            
        for name, param in self.pm_encoder.named_parameters():
            if "text_model" in name:
                param.requires_grad = False
                
        self.pm_encoder.train()
        self.frozen_model.eval()

    def set_downstream_mode(self):
        """Set the model to downstream mode."""
        for param in self.pm_encoder.parameters():
            param.requires_grad = False
            
        for name, param in self.text_encoder.named_parameters():
            if "vision_model" in name:
                param.requires_grad = False
                
        self.pm_encoder.eval()
        self.text_encoder.train()
        
    def forward(self, data_dict, mode=None):        
        # Ensure step counters exist
        if 'cur_step' not in data_dict:
            data_dict['cur_step'] = 1
            data_dict['total_steps'] = 1
    
        data_dict['logit_scale'] = self.pm_encoder.logit_scale.exp()

        if mode == "warmup":
            data_dict['images'] = data_dict['images'].squeeze(1)
            data_dict['point_map'] = data_dict['point_map'].squeeze(1).permute(0, 3, 1, 2)
            B, C, H, W = data_dict["images"].shape
            data_dict["txt_ids"] = data_dict["txt_ids"].view(B, -1)
            with torch.autocast("cuda", dtype=torch.bfloat16):
                pm = data_dict["point_map"]
                _, data_dict["inter_view_pm_embed"] = self.pm_encoder.get_image_features(pm)
                with torch.no_grad():
                    data_dict["inter_view_txt_embed"] = self.frozen_model.get_text_features(data_dict["txt_ids"])
                    _, data_dict["inter_view_rgb_embed"] = self.frozen_model.get_image_features(data_dict["images"])   
        elif mode == 'pretrain':
            pm_basic_features = []
            B, V, H, W, C = data_dict['point_map'].shape
            # point_cloud = data_dict['point_map'].reshape(B, -1, C).contiguous()
            # point_cloud, _ = scale_point_cloud(point_cloud, min_s=0.8, max_s=1.2)
            # point_cloud, _ = rotate_point_cloud(point_cloud)
            # point_cloud, _ = translate_point_cloud(point_cloud, scale=0.1)
            # point_cloud, _ = rotate_point_cloud_z(point_cloud)
            # point_maps = point_cloud.reshape(B, V, H, W, C)
            # data_dict['point_map'] = point_cloud.reshape(B, V, H, W, C).to(torch.bfloat16, non_blocking=True).permute(0, 1, 4, 2, 3)
            
            data_dict['point_map'] = data_dict['point_map'].to(torch.bfloat16, non_blocking=True).permute(0, 1, 4, 2, 3)
            
            for i in range(data_dict['point_map'].shape[0]):  # batch dimension
                with autocast(dtype=torch.bfloat16):
                    pm = data_dict['point_map'][i]   # [8, C, H, W]
                    _, pm_feat = self.pm_encoder.get_image_features(data_dict['point_map'][i])
                    pm_basic_features.append(pm_feat)
                
            pm_basic_features = torch.stack(pm_basic_features, dim=0) 
            data_dict['inter_view_pm_embed'] = pm_basic_features
            # with autocast(dtype=torch.bfloat16):
            #     pm_basic_features = self.global_attn(pm_basic_features)
            
            # data_dict['inter_view_context_pm_embed'] = pm_basic_features
            # data_dict['scene_pm_embed'] = data_dict['inter_view_context_pm_embed'].mean(dim=1)
            
            data_dict['scene_pm_embed'] = data_dict['inter_view_pm_embed'].mean(dim=1)
            
            B_txt = data_dict['txt_ids'].shape[0]
            lang_basic_features = torch.empty((B_txt, 32, 512), dtype=torch.bfloat16, device=data_dict['txt_ids'].device)
            ground_lang_basic_features = torch.empty((B_txt, 32, 512), dtype=torch.bfloat16, device=data_dict['txt_ids'].device)
            rgb_basic_features  = torch.empty((B_txt, 32, 512), dtype=torch.bfloat16, device=data_dict['txt_ids'].device)
            with torch.no_grad():
                with autocast(dtype=torch.bfloat16):
                    for i in range(B_txt):
                        lang_basic_features[i] = self.frozen_model.get_text_features(data_dict['txt_ids'][i], walk_short_pos=True)
                        ground_lang_basic_features[i] = self.frozen_model.get_text_features(data_dict['ground_txt_ids'][i], walk_short_pos=True)
                        rgb_basic_features[i]  = self.frozen_model.get_image_features(data_dict['images'][i])[1]

                    if getattr(self, "use_scene_cap", False):
                        data_dict['scene_text_embed'] = self.frozen_model.get_text_features(data_dict['scene_txt_ids'], walk_short_pos=False)
                    
            data_dict['inter_view_txt_embed'] = lang_basic_features
            data_dict['inter_view_ground_txt_embed'] = ground_lang_basic_features
            data_dict['inter_view_rgb_embed'] = rgb_basic_features
            data_dict['scene_rgb_embed'] = rgb_basic_features.mean(dim=1)
        elif mode == 'qa':
             # B, V, C, H, W
            B, V, C, H, W = data_dict['vision_inputs'].shape
            vision_inputs = data_dict['vision_inputs'].reshape(B * V, C, H, W).contiguous().float()

            with torch.no_grad():
                with autocast(dtype=torch.bfloat16):
                    _, vision_feat = self.pm_encoder.get_image_features(vision_inputs)
                    data_dict['inter_view_pm_embed'] = vision_feat.reshape(B, V, -1)
                    
            # jinaclip
            tokenized = self.tokenizer.batch_encode_plus(
                data_dict['sentence'],
                padding="max_length",
                return_tensors="pt",
                max_length=256,
            ).to(data_dict['inter_view_pm_embed'].device)
            
            # tokenized = self.tokenizer(
            #     data_dict['sentence'],
            #     padding=True,
            #     max_length=256,
            #     truncation = False,
            #     return_tensors="pt",
            # ).to(data_dict['inter_view_pm_embed'].device)
            
            data_dict['txt_ids'] = tokenized['input_ids']
            with autocast(dtype=torch.bfloat16):
                data_dict['inter_view_txt_tokens'] = self.text_encoder.text_model(data_dict['txt_ids'])[-1]
                data_dict['attention_mask'] = tokenized['attention_mask'].ne(1).bool()
                # text_embeddings = self.text_encoder(**tokenized)
                # data_dict['inter_view_txt_tokens'] = text_embeddings.last_hidden_state

                # --- QA Head (if used) ---
                if hasattr(self, "qa_head") and self.qa_head is not None:
                    answer_scores = self.qa_head(
                        data_dict['inter_view_pm_embed'],
                        data_dict['inter_view_txt_tokens'],
                        data_dict['attention_mask']
                    )
                    data_dict['answer_scores'] = answer_scores  
        return data_dict

    def get_vision_params(self, model):
        return [(n, p) for n, p in model.named_parameters() if p.requires_grad]
    
    def get_text_params(self, model):
        text_params = [
            (n, p) for n, p in model.named_parameters()
            if "text_model" in n
        ]
        return text_params
    
    def get_opt_params(self):
        def get_lr(cfg, default_lr):
            return default_lr if cfg.get("lr") is None else cfg.get("lr")

        optimizer_grouped_parameters = []
        if self.cfg.mode == 'warmup':
            optimizer_grouped_parameters += no_decay_param_group(self.get_vision_params(self.pm_encoder),get_lr(self.cfg.model.vision, self.cfg.solver.lr))
        elif self.cfg.mode == 'pretrain':
            optimizer_grouped_parameters += no_decay_param_group(self.get_vision_params(self.pm_encoder),get_lr(self.cfg.model.vision, self.cfg.solver.lr))
            # optimizer_grouped_parameters += no_decay_param_group(self.get_vision_params(self.global_attn), get_lr(self.cfg.model.vision, self.cfg.solver.lr))
        else:
            optimizer_grouped_parameters += no_decay_param_group(self.get_text_params(self.text_encoder), get_lr(self.cfg.model.vision, self.cfg.solver.lr))
            if "qa_head" in self.head_list:
                optimizer_grouped_parameters += no_decay_param_group(
                    self.qa_head.named_parameters(), get_lr(self.cfg.model.heads.qa_head, self.cfg.solver.lr)
            )
            if "ground_head" in self.head_list:
                optimizer_grouped_parameters += no_decay_param_group(
                    self.ground_head.named_parameters(), get_lr(self.cfg.model.heads.ground_head, self.cfg.solver.lr)
            ) 
            
        return optimizer_grouped_parameters