File size: 29,405 Bytes
d58fd10
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Qwen3-VL text encoder: custom HF model + packing-aware forward patches + TextEncoder wrapper."""
from __future__ import annotations

import os
from collections.abc import Callable
from dataclasses import dataclass

try:
    from typing import Unpack
except ImportError:
    from typing_extensions import Unpack

import torch
from loguru import logger
from torch import nn
from transformers import AutoProcessor, AutoTokenizer, Cache, Qwen3VLForConditionalGeneration
from transformers.cache_utils import DynamicCache
from transformers.masking_utils import create_causal_mask
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_outputs import BaseModelOutputWithPast
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
from transformers.models.qwen3_vl.modeling_qwen3_vl import (
    Qwen3VLCausalLMOutputWithPast,
    apply_rotary_pos_emb,
    eager_attention_forward,
)
from transformers.utils import ModelOutput

from ._attn_backend import flash_attn_varlen_func


# ===========================================================================
# Custom Qwen3-VL model (customizable forward output)
# ===========================================================================

@dataclass
class Qwen3VLModelOutput(ModelOutput):
    """Flexible output class for custom Qwen3-VL model."""

    loss: torch.FloatTensor | None = None
    logits: torch.FloatTensor | None = None
    past_key_values: Cache | None = None
    hidden_states: tuple[torch.FloatTensor, ...] | None = None
    last_hidden_state: torch.FloatTensor | None = None
    attentions: tuple[torch.FloatTensor, ...] | None = None
    rope_deltas: torch.LongTensor | None = None


class CustomQwen3VLForConditionalGeneration(Qwen3VLForConditionalGeneration):
    """
    Custom Qwen3-VL model that allows customizing the forward output.

    This class inherits from Qwen3VLForConditionalGeneration and provides
    hooks to customize what is returned from the forward pass.

    Example usage:
        ```python
        model = CustomQwen3VLForConditionalGeneration.from_pretrained(
            "Qwen/Qwen3-VL-8B-Instruct",
            attn_implementation="flash_attention_2"  # Use flash attention for faster inference
        )

        # Option 1: Use built-in output modes
        model.set_output_mode("embedding")  # Only return last hidden state (default)
        model.set_output_mode("full")       # Return everything
        model.set_output_mode("logits")     # Only return logits

        # Option 2: Set a custom output processor
        def my_custom_output(hidden_states, logits, outputs, **kwargs):
            return {"embeddings": hidden_states, "pooled": hidden_states.mean(dim=1)}
        model.set_output_processor(my_custom_output)
        ```
    """

    # Output mode constants
    OUTPUT_MODE_FULL = "full"
    OUTPUT_MODE_EMBEDDING = "embedding"
    OUTPUT_MODE_LOGITS = "logits"
    OUTPUT_MODE_HIDDEN = "hidden"

    def __init__(self, config):
        super().__init__(config)
        self._output_mode = self.OUTPUT_MODE_EMBEDDING
        self._skip_lm_head = True

    def set_output_mode(self, mode: str):
        """
        Set the output mode for the forward pass.

        Args:
            mode: One of:
                - "full": Return full Qwen3VLCausalLMOutputWithPast
                - "embedding": Only return last hidden state (skip lm_head) (default)
                - "logits": Only return logits
                - "hidden": Return all hidden states
        """
        valid_modes = [
            self.OUTPUT_MODE_FULL,
            self.OUTPUT_MODE_EMBEDDING,
            self.OUTPUT_MODE_LOGITS,
            self.OUTPUT_MODE_HIDDEN,
        ]
        if mode not in valid_modes:
            raise ValueError(f"Invalid output mode: {mode}. Must be one of {valid_modes}")
        self._output_mode = mode
        self._skip_lm_head = mode == self.OUTPUT_MODE_EMBEDDING

    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        pixel_values: torch.Tensor | None = None,
        pixel_values_videos: torch.FloatTensor | None = None,
        image_grid_thw: torch.LongTensor | None = None,
        video_grid_thw: torch.LongTensor | None = None,
        cache_position: torch.LongTensor | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> Qwen3VLCausalLMOutputWithPast | Qwen3VLModelOutput | dict | torch.Tensor:
        """
        Forward pass with customizable output.

        Returns different outputs based on the configured output mode or custom processor.
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )

        # Get outputs from the base model (Qwen3VLModel)
        outputs = self.model(
            input_ids=input_ids,
            pixel_values=pixel_values,
            pixel_values_videos=pixel_values_videos,
            image_grid_thw=image_grid_thw,
            video_grid_thw=video_grid_thw,
            position_ids=position_ids,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            cache_position=cache_position,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            **kwargs,
        )

        # Get the last hidden state
        hidden_states = outputs[0]  # This is the last hidden state

        # Compute logits if not skipping lm_head
        logits = None
        if not self._skip_lm_head:
            slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
            logits = self.lm_head(hidden_states[:, slice_indices, :])

        # Compute loss if labels are provided
        loss = None
        if labels is not None and logits is not None:
            loss = self.loss_function(
                logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs
            )

        # Return based on output mode
        if self._output_mode == self.OUTPUT_MODE_EMBEDDING:
            return Qwen3VLModelOutput(
                last_hidden_state=hidden_states,
                past_key_values=outputs.past_key_values,
                attentions=outputs.attentions,
                rope_deltas=outputs.rope_deltas,
            )
        elif self._output_mode == self.OUTPUT_MODE_LOGITS:
            return logits
        elif self._output_mode == self.OUTPUT_MODE_HIDDEN:
            return Qwen3VLModelOutput(
                last_hidden_state=hidden_states,
                hidden_states=outputs.hidden_states,
                past_key_values=outputs.past_key_values,
                attentions=outputs.attentions,
                rope_deltas=outputs.rope_deltas,
            )
        else:  # OUTPUT_MODE_FULL
            return Qwen3VLCausalLMOutputWithPast(
                loss=loss,
                logits=logits,
                past_key_values=outputs.past_key_values,
                hidden_states=outputs.hidden_states,
                attentions=outputs.attentions,
                rope_deltas=outputs.rope_deltas,
            )


# ===========================================================================
# Packing-aware forward patches (cu_seqlens) for the Qwen3-VL text encoder
# ===========================================================================

def model_forward(
    self,
    input_ids: torch.LongTensor | None = None,
    attention_mask: torch.Tensor | None = None,
    position_ids: torch.LongTensor | None = None,
    past_key_values: Cache | None = None,
    inputs_embeds: torch.FloatTensor | None = None,
    use_cache: bool | None = None,
    cache_position: torch.LongTensor | None = None,
    # args for deepstack
    visual_pos_masks: torch.Tensor | None = None,
    deepstack_visual_embeds: list[torch.Tensor] | None = None,
    **kwargs: Unpack[FlashAttentionKwargs],
) -> tuple | BaseModelOutputWithPast:
    r"""
    visual_pos_masks (`torch.Tensor` of shape `(batch_size, seqlen)`, *optional*):
        The mask of the visual positions.
    deepstack_visual_embeds (`list[torch.Tensor]`, *optional*):
        The deepstack visual embeddings. The shape is (num_layers, visual_seqlen, embed_dim).
        The feature is extracted from the different visual encoder layers, and fed to the decoder
        hidden states. It's from the paper DeepStack(https://arxiv.org/abs/2406.04334).
    """
    if (input_ids is None) ^ (inputs_embeds is not None):
        raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

    # torch.jit.trace() doesn't support cache objects in the output
    if use_cache and past_key_values is None and not torch.jit.is_tracing():
        past_key_values = DynamicCache(config=self.config)

    if inputs_embeds is None:
        inputs_embeds = self.embed_tokens(input_ids)

    if cache_position is None:
        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
        cache_position = torch.arange(
            past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
        )

    # the hard coded `3` is for temporal, height and width.
    if position_ids is None:
        position_ids = cache_position.view(1, 1, -1).expand(3, inputs_embeds.shape[0], -1)
    elif position_ids.ndim == 2:
        position_ids = position_ids[None, ...].expand(3, position_ids.shape[0], -1)

    if position_ids.ndim == 3 and position_ids.shape[0] == 4:
        text_position_ids = position_ids[0]
        position_ids = position_ids[1:]
    else:
        text_position_ids = position_ids[0]

    if kwargs.get("cu_seqlens") is None:
        attention_mask = create_causal_mask(
            config=self.config,
            input_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            past_key_values=past_key_values,
            position_ids=text_position_ids,
        )

    hidden_states = inputs_embeds

    # create position embeddings to be shared across the decoder layers
    position_embeddings = self.rotary_emb(hidden_states, position_ids)

    # decoder layers
    for layer_idx, decoder_layer in enumerate(self.layers):
        layer_outputs = decoder_layer(
            hidden_states,
            attention_mask=attention_mask,
            position_ids=text_position_ids,
            past_key_values=past_key_values,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = layer_outputs

        # add visual features to the hidden states of first several layers
        if deepstack_visual_embeds is not None and layer_idx in range(len(deepstack_visual_embeds)):
            hidden_states = self._deepstack_process(
                hidden_states,
                visual_pos_masks,
                deepstack_visual_embeds[layer_idx],
            )

    hidden_states = self.norm(hidden_states)

    return BaseModelOutputWithPast(
        last_hidden_state=hidden_states,
        past_key_values=past_key_values,
    )


def forward(
    self,
    hidden_states: torch.Tensor,
    position_embeddings: tuple[torch.Tensor, torch.Tensor],
    attention_mask: torch.Tensor | None,
    past_key_values: Cache | None = None,
    cache_position: torch.LongTensor | None = None,
    **kwargs: Unpack[FlashAttentionKwargs],
) -> tuple[torch.Tensor, torch.Tensor | None]:
    input_shape = hidden_states.shape[:-1]
    hidden_shape = (*input_shape, -1, self.head_dim)

    query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
    key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
    value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

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

    if past_key_values is not None:
        # sin and cos are specific to RoPE models; cache_position needed for the static cache
        cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
        key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)

    cu_seqlens = kwargs.get("cu_seqlens", None)

    if cu_seqlens is None:
        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            **kwargs,
        )
    else:
        max_seqlen = torch.diff(cu_seqlens).max().item() if cu_seqlens is not None else None
        query_states = query_states.transpose(1, 2).squeeze(0)
        key_states = key_states.transpose(1, 2).squeeze(0)
        value_states = value_states.transpose(1, 2).squeeze(0)
        attn_output = flash_attn_varlen_func(
            q=query_states,
            k=key_states,
            v=value_states,
            cu_seqlens_q=cu_seqlens,
            cu_seqlens_k=cu_seqlens,
            max_seqlen_q=max_seqlen,
            max_seqlen_k=max_seqlen,
            causal=True,
            window_size=(-1, -1),
            softmax_scale=self.head_dim**-0.5,
            dropout_p=0.0,
        )

    attn_output = attn_output.reshape(*input_shape, -1).contiguous()
    attn_output = self.o_proj(attn_output)
    return attn_output, None


def qwen3_patch_forward():
    """Patch the Qwen3-VL text model + attention forwards to support packed
    varlen (cu_seqlens) inputs used by ``TextEncoder.forward``."""
    from transformers.models.qwen3_vl.modeling_qwen3_vl import Qwen3VLTextAttention, Qwen3VLTextModel

    Qwen3VLTextModel.forward = model_forward
    Qwen3VLTextAttention.forward = forward


# ===========================================================================
# TextEncoder wrapper (packed text -> DiT conditioning embeddings)
# ===========================================================================
_FA2_ALIASES = {"flash2", "fa2", "flash_attention_2", "flash_attn_2"}
_FA4_ALIASES = {"flash4", "fa4", "flash_attention_4", "flash_attn_4"}
_SDPA_ALIASES = {"sdpa", "torch_sdpa", "scaled_dot_product_attention"}


def _resolve_hf_attn_impl(attn_type: str) -> str:
    """Map a project-level attn_type to a HuggingFace ``attn_implementation`` string.

    ``VF_HF_ATTN_IMPL`` env var, if set, takes precedence (useful for forcing
    sdpa on machines without flash-attn). For FA4 we additionally probe that
    the CUTE-DSL kernel is importable and (when available) ask the HF helper
    to confirm; if not, fall back to sdpa rather than crashing at load time.
    """
    override = os.environ.get("VF_HF_ATTN_IMPL")
    if override:
        return override

    name = attn_type.lower().strip()
    if name in _FA2_ALIASES:
        return "flash_attention_2"
    if name in _FA4_ALIASES:
        try:
            import flash_attn.cute  # noqa: F401
            fa4_importable = True
        except Exception:
            fa4_importable = False
        if fa4_importable:
            try:
                from transformers.utils.import_utils import is_flash_attn_4_available
                if is_flash_attn_4_available():
                    return "flash_attention_4"
            except ImportError:
                return "flash_attention_4"
        logger.warning(
            "attn_type=flash4 requested but flash_attn.cute is unavailable; "
            "falling back to sdpa for HF text encoder."
        )
        return "sdpa"
    if name in _SDPA_ALIASES:
        return "sdpa"
    raise ValueError(
        f"Unknown attn_type {attn_type!r}; expected one of "
        f"{sorted(_FA2_ALIASES | _FA4_ALIASES | _SDPA_ALIASES)}"
    )


SEQ_MULTI_OF = 32



class TextEncoder(nn.Module):
    def __init__(
        self,
        model_name: str,
        version: str,
        tokenizer_max_length: int,
        prompt_template: dict | None,
        dit_structure: dict,
        use_packed_text_infer: bool = False,
        attn_type: str = "flash2",
        **hf_kwargs,
    ):
        super().__init__()
        self.model_name = model_name
        self.tokenizer_max_length = tokenizer_max_length
        self.tokenizer: AutoTokenizer = AutoTokenizer.from_pretrained(version)
        self.tokenizer.padding_side = "right"

        hf_attn_impl = _resolve_hf_attn_impl(attn_type)
        logger.info(f"TextEncoder attn_type={attn_type} -> attn_implementation={hf_attn_impl}")

        logger.info("init vl model: qwen3")
        self.hf_module: CustomQwen3VLForConditionalGeneration = CustomQwen3VLForConditionalGeneration.from_pretrained(
            version,
            attn_implementation=hf_attn_impl,
            **hf_kwargs
        )

        # Use local_files_only if version is a local path (absolute path or contains path separators)
        is_local = version.startswith("/") or os.sep in version
        self.processor = AutoProcessor.from_pretrained(version, local_files_only=is_local)

        self.hf_module = self.hf_module.eval().requires_grad_(False)

        prompt_template = prompt_template or {}
        self.prompt_template_encode = prompt_template.get("template", "")
        self.prompt_template_encode_start_idx = prompt_template.get("start_idx", 0)
        self.dit_structure = dit_structure
        self.use_packed_text_infer = use_packed_text_infer

    def forward(
        self,
        input_ids: torch.Tensor,
        cu_seqlens: torch.Tensor,
        inputs: dict | None = None,
        drop_idx_override: int | None = None,
    ):
        """Encode packed text (varlen ``cu_seqlens``) into DiT conditioning embeddings.

        This is the sole text-embedding path — both t2i and edit call it. Uses
        Flash-Attention-2's varlen capability (``cu_seqlens``) via the patched
        Qwen3-VL forward to process several concatenated sequences in a single
        launch, with no padding. Verified numerically identical to a padded-batch
        forward with per-sample cu_seqlens isolation (zero cross-contamination).

        Args:
            input_ids: Packed token ids ``[Total_L]``.
            cu_seqlens: Cumulative sequence lengths ``[B+1]``.
            inputs: Optional dict with additional model inputs (e.g. ``pixel_values``,
                ``image_grid_thw`` for the multimodal edit path). Passed through to
                the text encoder.
            drop_idx_override: If set, override the number of leading (system-prompt)
                tokens to drop per sequence. Use 0 for multi-turn where the system
                prompt is embedded in the conversation and should not be stripped.

        Returns:
            dict with keys:
                - ``txt``: text embeddings ``[Total_L - B*drop_idx, D]`` (system prompt dropped)
                - ``vec``: pooled text embeddings ``[B, D]``
                - ``txt_seq_lens``: per-sequence lengths ``[B]`` (after dropping system prompt)
        """
        # Compute seqlens from cu_seqlens
        seqlens = cu_seqlens[1:] - cu_seqlens[:-1]
        seqlens_list = seqlens.cpu().tolist()

        # Build position_ids for packing: each sequence starts from 0
        position_ids_list = []
        for length in seqlens_list:
            position_ids_list.append(torch.arange(length, device=input_ids.device))
        position_ids = torch.cat(position_ids_list)  # [Total_L]

        # Reshape for model input: [1, Total_L]
        input_ids_packed = input_ids.unsqueeze(0)  # [1, Total_L]
        position_ids_packed = position_ids.unsqueeze(0)  # [1, Total_L]

        # Move to text encoder device
        device = self.hf_module.device
        input_ids_packed = input_ids_packed.to(device)
        position_ids_packed = position_ids_packed.to(device)

        # Get text embeddings (the text encoder is always frozen)
        with torch.no_grad():
            forward_kwargs = {
                "input_ids": input_ids_packed,
                "cu_seqlens": cu_seqlens,
                "position_ids": position_ids_packed,
                "output_hidden_states": False,
                "max_seqlen": None,
            }
            # Pass multimodal inputs for edit mode (reference images)
            if inputs is not None:
                for key in ("pixel_values", "image_grid_thw"):
                    if key in inputs and inputs[key] is not None:
                        val = inputs[key]
                        if hasattr(val, "to"):
                            val = val.to(device)
                        forward_kwargs[key] = val
            outputs = self.hf_module(**forward_kwargs)

        # Extract hidden state
        if hasattr(outputs, "last_hidden_state") and outputs.last_hidden_state is not None:
            hidden = outputs.last_hidden_state  # [1, Total_L, D]
        elif hasattr(outputs, "hidden_states"):
            hidden = outputs.hidden_states[-1]

        # Remove batch dimension: [Total_L, D]
        hidden = hidden.squeeze(0)

        # Get drop_idx (system prompt length to skip).
        # For multi-turn, drop_idx_override=0 is passed since system prompt is in the messages.
        if drop_idx_override is not None:
            drop_idx = drop_idx_override
        else:
            drop_idx = self.prompt_template_encode_start_idx

        # Split hidden states by sequence
        hidden_split = torch.split(hidden, seqlens_list, dim=0)

        # Extract valid embeddings (drop system prompt) and compute vec
        txt_list = []
        vec_list = []
        valid_lengths = []

        for h in hidden_split:
            # Drop system prompt tokens
            h_valid = h[drop_idx:]  # [seq_len - drop_idx, D]
            txt_list.append(h_valid)
            valid_lengths.append(h_valid.shape[0])

            # Compute pooled embedding (mean of valid tokens only, after dropping system prompt)
            vec_list.append(h_valid.mean(dim=0))  # [D]

        txt = torch.cat(txt_list, dim=0)  # [Total_valid, D]
        vec = torch.stack(vec_list, dim=0)  # [B, D]
        txt_seq_lens = torch.tensor(valid_lengths, device=input_ids.device)

        result = {
            "txt": txt,
            "vec": vec,
            "txt_seq_lens": txt_seq_lens,
        }

        return result

    # ------------------------------------------------------------------
    # Mandatory content-policy screening (same Qwen3-VL weights)
    # ------------------------------------------------------------------
    # The policy classifier lives HERE, on the text encoder, so it runs on the
    # exact weights that produce the diffusion conditioning and is not a
    # separable, toggleable pre-pass in the pipeline. The classifier needs
    # autoregressive ``.generate()`` (JSON verdict) whereas conditioning is a
    # single embedding forward — they cannot be one GPU forward without a
    # trained classification head, so "fused" here means: same module, same
    # weights, always run, FAIL-CLOSED (any error blocks).

    def screen_text(self, prompt: str, max_new_tokens: int = 160):
        """Classify a text-to-image ``prompt`` against the content policy.

        Returns a ``FilterVerdict``. FAIL-CLOSED: any error (generation, parse)
        returns ``violates=True`` so a broken classifier cannot be used as a
        bypass. An empty prompt is not a violation.
        """
        from .mage_text import (
            CONTENT_FILTER_SYSTEM, FilterVerdict, _extract_json_object,
            _full_output_mode,
        )

        if not prompt or not prompt.strip():
            return FilterVerdict(False, [], "empty prompt", "")
        try:
            tokenizer = self.tokenizer
            hf = self.hf_module
            device = next(hf.parameters()).device

            messages = [
                {"role": "system", "content": CONTENT_FILTER_SYSTEM},
                {"role": "user", "content": f"Prompt to classify:\n{prompt}"},
            ]
            text = tokenizer.apply_chat_template(
                messages, tokenize=False, add_generation_prompt=True)
            inputs = tokenizer(text, return_tensors="pt").to(device)

            eos_id = tokenizer.eos_token_id
            pad_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else eos_id

            with _full_output_mode(hf), torch.no_grad():
                out = hf.generate(
                    **inputs, max_new_tokens=max_new_tokens, do_sample=False,
                    pad_token_id=pad_id, eos_token_id=eos_id)
            gen = tokenizer.decode(
                out[0, inputs.input_ids.shape[1]:], skip_special_tokens=True).strip()

            parsed = _extract_json_object(gen)
            violates = bool(parsed.get("violates", False))
            cats = [c for c in (parsed.get("categories", []) or []) if isinstance(c, str)]
            reason = str(parsed.get("reason", "")).strip()
            return FilterVerdict(violates, cats, reason, gen)
        except Exception as exc:  # noqa: BLE001
            # FAIL-CLOSED: block on any screening error.
            return FilterVerdict(
                True, ["policy"], f"filter error (blocked): {type(exc).__name__}: {exc}", "")

    def screen_edit(self, prompt: str, ref_images, max_new_tokens: int = 192):
        """Classify an image-EDIT request (source image(s) + instruction).

        Considers BOTH the source image(s) and the instruction via multimodal
        Qwen3-VL. Falls back to :meth:`screen_text` when no image is given.
        FAIL-CLOSED: any error returns ``violates=True``.
        """
        from PIL import Image

        from .mage_text import (
            CONTENT_FILTER_EDIT_SYSTEM, FilterVerdict, _extract_json_object,
            _full_output_mode,
        )

        pils = [ref_images] if isinstance(ref_images, Image.Image) else list(ref_images)
        pils = [p.convert("RGB") for p in pils if p is not None]
        if not pils:
            return self.screen_text(prompt, max_new_tokens=max_new_tokens)

        instruction = (prompt or "").strip() or "(no textual instruction)"
        try:
            processor = self.processor
            tokenizer = self.tokenizer
            hf = self.hf_module
            device = next(hf.parameters()).device

            user_content = [{"type": "image"} for _ in pils]
            user_content.append({
                "type": "text",
                "text": (
                    f"There {'is' if len(pils) == 1 else 'are'} {len(pils)} source "
                    f"image(s) above. Edit instruction: {instruction}\n"
                    "Classify this edit request."
                ),
            })
            messages = [
                {"role": "system", "content": CONTENT_FILTER_EDIT_SYSTEM},
                {"role": "user", "content": user_content},
            ]
            text = processor.apply_chat_template(
                messages, tokenize=False, add_generation_prompt=True)
            inputs = processor(
                text=[text], images=pils, padding=True, return_tensors="pt").to(device)

            eos_id = tokenizer.eos_token_id
            pad_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else eos_id

            # Keep only the kwargs Qwen3-VL .generate() consumes.
            gen_inputs = {
                k: inputs[k]
                for k in ("input_ids", "attention_mask", "pixel_values", "image_grid_thw")
                if k in inputs and inputs[k] is not None
            }
            input_len = gen_inputs["input_ids"].shape[1]

            with _full_output_mode(hf), torch.no_grad():
                out = hf.generate(
                    **gen_inputs, max_new_tokens=max_new_tokens, do_sample=False,
                    pad_token_id=pad_id, eos_token_id=eos_id)
            gen = tokenizer.decode(out[0, input_len:], skip_special_tokens=True).strip()

            parsed = _extract_json_object(gen)
            violates = bool(parsed.get("violates", False))
            cats = [c for c in (parsed.get("categories", []) or []) if isinstance(c, str)]
            reason = str(parsed.get("reason", "")).strip()
            return FilterVerdict(violates, cats, reason, gen)
        except Exception as exc:  # noqa: BLE001
            # FAIL-CLOSED: block on any screening error.
            return FilterVerdict(
                True, ["policy"], f"edit filter error (blocked): {type(exc).__name__}: {exc}", "")