qwen3_vl.py 62.9 KB
Newer Older
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
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

# Copyright 2025 The vLLM team.
# Copyright 2025 The Qwen Team.
# Copyright 2025 The HuggingFace Inc. team.
# All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Inference-only Qwen3VL model compatible with HuggingFace weights."""
26

27
from collections.abc import Callable, Iterable, Iterator, Mapping, Sequence
28
from functools import lru_cache, partial
29
from itertools import islice
30
from typing import Any
31
32
33
34
35

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
36
from transformers import BatchFeature
37
from transformers.models.qwen2_vl import Qwen2VLImageProcessorFast
38
from transformers.models.qwen2_vl.image_processing_qwen2_vl import (
39
40
41
    smart_resize as image_smart_resize,
)
from transformers.models.qwen3_vl import Qwen3VLProcessor, Qwen3VLVideoProcessor
42
from transformers.models.qwen3_vl.configuration_qwen3_vl import (
43
44
45
    Qwen3VLConfig,
    Qwen3VLVisionConfig,
)
46
from transformers.models.qwen3_vl.video_processing_qwen3_vl import (
47
48
    smart_resize as video_smart_resize,
)
49
50
from transformers.video_utils import VideoMetadata

51
from vllm.attention.backends.registry import AttentionBackendEnum
52
53
from vllm.compilation.decorators import support_torch_compile
from vllm.config import VllmConfig
54
from vllm.config.multimodal import BaseDummyOptions, VideoDummyOptions
55
56
57
from vllm.distributed import get_pp_group
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import _ACTIVATION_REGISTRY
58
from vllm.model_executor.layers.conv import Conv3dLayer
59
60
61
62
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    RowParallelLinear,
)
63
64
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
65
from vllm.model_executor.layers.rotary_embedding import get_rope
66
67
68
69
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.models.module_mapping import MultiModelKeys
from vllm.multimodal import MULTIMODAL_REGISTRY
70
71
from vllm.multimodal.inputs import (
    MultiModalDataDict,
72
    MultiModalFeatureSpec,
73
74
75
76
77
78
79
80
81
82
83
84
    MultiModalFieldConfig,
    MultiModalKwargsItem,
    MultiModalKwargsItems,
    VideoItem,
)
from vllm.multimodal.parse import ImageSize, MultiModalDataItems, MultiModalDataParser
from vllm.multimodal.processing import (
    BaseMultiModalProcessor,
    PromptReplacement,
    PromptUpdate,
    PromptUpdateDetails,
)
85
86
from vllm.multimodal.profiling import BaseDummyInputsBuilder
from vllm.sequence import IntermediateTensors
87
from vllm.utils.collection_utils import is_list_of
88

89
90
from .interfaces import (
    MultiModalEmbeddings,
91
    SupportsEagle3,
92
    SupportsLoRA,
93
    SupportsMRoPE,
94
95
96
97
98
99
100
101
102
103
104
105
    SupportsMultiModal,
    SupportsPP,
)
from .qwen2_5_vl import (
    Qwen2_5_VisionAttention,
    Qwen2_5_VLImageEmbeddingInputs,
    Qwen2_5_VLImageInputs,
    Qwen2_5_VLImagePixelInputs,
    Qwen2_5_VLVideoEmbeddingInputs,
    Qwen2_5_VLVideoInputs,
    Qwen2_5_VLVideoPixelInputs,
)
106
from .qwen2_vl import Qwen2VLMultiModalDataParser, Qwen2VLProcessingInfo
107
from .qwen3 import Qwen3ForCausalLM, Qwen3Model
108
109
110
111
112
113
114
from .utils import (
    AutoWeightsLoader,
    PPMissingLayer,
    WeightsMapper,
    _merge_multimodal_embeddings,
    maybe_prefix,
)
115
116
117
118
from .vision import (
    get_vit_attn_backend,
    run_dp_sharded_mrope_vision_model,
)
119
120
121

logger = init_logger(__name__)

122
123
124
# Official recommended max pixels is 24576 * 32 * 32
_MAX_FRAMES_PER_VIDEO = 24576

125
126
127
128
129
130
131
132
133
134
135
136
137
138
139

class Qwen3_VisionPatchEmbed(nn.Module):
    def __init__(
        self,
        patch_size: int = 14,
        temporal_patch_size: int = 2,
        in_channels: int = 3,
        hidden_size: int = 1152,
    ) -> None:
        super().__init__()
        self.patch_size = patch_size
        self.temporal_patch_size = temporal_patch_size
        self.hidden_size = hidden_size

        kernel_size = (temporal_patch_size, patch_size, patch_size)
140
141
        self.proj = Conv3dLayer(
            in_channels,
142
            hidden_size,
143
144
            kernel_size=kernel_size,
            stride=kernel_size,
145
146
            bias=True,
        )
147
148

    def forward(self, x: torch.Tensor) -> torch.Tensor:
149
150
151
        L, C = x.shape
        x = x.view(L, -1, self.temporal_patch_size, self.patch_size, self.patch_size)
        x = self.proj(x).view(L, self.hidden_size)
152
153
154
155
        return x


class Qwen3_VisionMLP(nn.Module):
156
157
158
159
160
161
    def __init__(
        self,
        in_features: int,
        hidden_features: int,
        bias: bool = False,
        act_fn: Callable[[torch.Tensor], torch.Tensor] = F.silu,
162
        quant_config: QuantizationConfig | None = None,
163
164
165
        prefix: str = "",
        use_data_parallel: bool = False,
    ):
166
        super().__init__()
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
        self.linear_fc1 = ColumnParallelLinear(
            in_features,
            hidden_features,
            bias=bias,
            quant_config=quant_config,
            return_bias=False,
            prefix=f"{prefix}.linear_fc1",
            disable_tp=use_data_parallel,
        )
        self.linear_fc2 = RowParallelLinear(
            hidden_features,
            in_features,
            bias=bias,
            quant_config=quant_config,
            return_bias=False,
            prefix=f"{prefix}.linear_fc2",
            disable_tp=use_data_parallel,
        )
185
186
187
188
189
190
191
192
193
194
195
196
197
198
        self.act_fn = act_fn

    def forward(self, x: torch.Tensor):
        mlp_output = self.linear_fc2(self.act_fn(self.linear_fc1(x)))
        return mlp_output


class Qwen3_VisionBlock(nn.Module):
    def __init__(
        self,
        dim: int,
        num_heads: int,
        mlp_hidden_dim: int,
        act_fn: Callable[[torch.Tensor], torch.Tensor] = F.silu,
199
200
        norm_layer: Callable[[int], nn.Module] | None = None,
        quant_config: QuantizationConfig | None = None,
201
        prefix: str = "",
202
        use_data_parallel: bool = False,
203
        attn_backend: AttentionBackendEnum = AttentionBackendEnum.TORCH_SDPA,
204
205
206
207
208
209
    ) -> None:
        super().__init__()
        if norm_layer is None:
            norm_layer = partial(nn.LayerNorm, eps=1e-6)
        self.norm1 = norm_layer(dim)
        self.norm2 = norm_layer(dim)
210
211
212
213
214
215
        self.attn = Qwen2_5_VisionAttention(
            embed_dim=dim,
            num_heads=num_heads,
            projection_size=dim,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
216
217
            use_data_parallel=use_data_parallel,
            attn_backend=attn_backend,
218
219
220
221
222
223
224
225
226
227
        )
        self.mlp = Qwen3_VisionMLP(
            dim,
            mlp_hidden_dim,
            act_fn=act_fn,
            bias=True,
            quant_config=quant_config,
            prefix=f"{prefix}.mlp",
            use_data_parallel=use_data_parallel,
        )
228
229

    def forward(
230
231
232
        self,
        x: torch.Tensor,
        cu_seqlens: torch.Tensor,
233
234
        rotary_pos_emb_cos: torch.Tensor,
        rotary_pos_emb_sin: torch.Tensor,
235
        max_seqlen: torch.Tensor,  # Only used for Flash Attention
236
    ) -> torch.Tensor:
237
238
239
        x = x + self.attn(
            self.norm1(x),
            cu_seqlens=cu_seqlens,
240
241
            rotary_pos_emb_cos=rotary_pos_emb_cos,
            rotary_pos_emb_sin=rotary_pos_emb_sin,
242
243
            max_seqlen=max_seqlen,
        )
244
245
246
247
248
249
250
251
252
253

        x = x + self.mlp(self.norm2(x))
        return x


class Qwen3_VisionPatchMerger(nn.Module):
    def __init__(
        self,
        d_model: int,
        context_dim: int,
254
        norm_layer: Callable[[int], nn.Module] | None = None,
255
256
        spatial_merge_size: int = 2,
        use_postshuffle_norm: bool = False,
257
        quant_config: QuantizationConfig | None = None,
258
        prefix: str = "",
259
        use_data_parallel: bool = False,
260
261
262
263
264
265
266
267
268
269
    ) -> None:
        super().__init__()
        self.hidden_size = context_dim * (spatial_merge_size**2)

        self.use_postshuffle_norm = use_postshuffle_norm
        if self.use_postshuffle_norm:
            context_dim = self.hidden_size

        if norm_layer is None:
            norm_layer = partial(nn.LayerNorm, eps=1e-6)
270
        self.norm = norm_layer(context_dim)
271
272
273
274
275
276
277
278
        self.linear_fc1 = ColumnParallelLinear(
            self.hidden_size,
            self.hidden_size,
            bias=True,
            quant_config=quant_config,
            prefix=f"{prefix}.linear_fc1",
            disable_tp=use_data_parallel,
        )
279
        self.act_fn = nn.GELU()
280
281
282
283
284
285
286
287
        self.linear_fc2 = RowParallelLinear(
            self.hidden_size,
            d_model,
            bias=True,
            quant_config=quant_config,
            prefix=f"{prefix}.linear_fc2",
            disable_tp=use_data_parallel,
        )
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305

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

        x_parallel, _ = self.linear_fc1(x)
        x_parallel = self.act_fn(x_parallel)
        out, _ = self.linear_fc2(x_parallel)
        return out


class Qwen3_VisionTransformer(nn.Module):
    def __init__(
        self,
        vision_config: Qwen3VLVisionConfig,
        norm_eps: float = 1e-6,
306
        quant_config: QuantizationConfig | None = None,
307
        prefix: str = "",
308
        use_data_parallel: bool = False,
309
        attn_backend_override: AttentionBackendEnum | None = None,
310
311
312
313
314
315
316
317
318
319
    ) -> None:
        super().__init__()
        self.hidden_size = vision_config.hidden_size
        self.num_heads = vision_config.num_heads
        self.num_position_embeddings = vision_config.num_position_embeddings
        self.patch_size = vision_config.patch_size
        self.spatial_merge_size = vision_config.spatial_merge_size
        self.spatial_merge_unit = self.spatial_merge_size**2
        self.temporal_patch_size = vision_config.temporal_patch_size
        self.deepstack_visual_indexes = vision_config.deepstack_visual_indexes
320
        self.use_data_parallel = use_data_parallel
321
        self.num_grid_per_side = int(self.num_position_embeddings**0.5)
322
323
324

        # NOTE: This is used for creating empty tensor for all_gather for
        # DP ViT. Here out_hidden_size is enlarged due to deepstack
325
326
327
        self.out_hidden_size = vision_config.out_hidden_size * (
            1 + len(self.deepstack_visual_indexes)
        )
328
329
330
331
332
333
334
335

        self.patch_embed = Qwen3_VisionPatchEmbed(
            patch_size=self.patch_size,
            temporal_patch_size=self.temporal_patch_size,
            in_channels=vision_config.in_channels,
            hidden_size=self.hidden_size,
        )

336
        self.pos_embed = nn.Embedding(self.num_position_embeddings, self.hidden_size)
337
338
339

        norm_layer = partial(nn.LayerNorm, eps=norm_eps)
        head_dim = self.hidden_size // self.num_heads
340
341
342
343
344
345
        self.rotary_pos_emb = get_rope(
            head_size=head_dim,
            rotary_dim=head_dim // 2,
            max_position=8192,
            is_neox_style=True,
        )
346
347
348
349
350
351
352
353

        self.merger = Qwen3_VisionPatchMerger(
            d_model=vision_config.out_hidden_size,
            context_dim=self.hidden_size,
            norm_layer=norm_layer,
            spatial_merge_size=self.spatial_merge_size,
            quant_config=quant_config,
            prefix=f"{prefix}.merger",
354
            use_data_parallel=use_data_parallel,
355
356
        )

357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
        self.deepstack_merger_list = nn.ModuleList(
            [
                Qwen3_VisionPatchMerger(
                    d_model=vision_config.out_hidden_size,
                    context_dim=self.hidden_size,
                    spatial_merge_size=self.spatial_merge_size,
                    use_postshuffle_norm=True,
                    norm_layer=norm_layer,
                    quant_config=quant_config,
                    prefix=f"{prefix}.deepstack_merger_list.{layer_idx}",
                    use_data_parallel=use_data_parallel,
                )
                for layer_idx in range(len(self.deepstack_visual_indexes))
            ]
        )
372
373

        self.attn_backend = get_vit_attn_backend(
374
375
376
            head_size=head_dim,
            dtype=torch.get_default_dtype(),
            attn_backend_override=attn_backend_override,
377
        )
378
379

        if self.attn_backend not in {
380
381
382
            AttentionBackendEnum.FLASH_ATTN,
            AttentionBackendEnum.TORCH_SDPA,
            AttentionBackendEnum.ROCM_AITER_FA,
383
384
        }:
            raise RuntimeError(
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
                f"Qwen3-VL does not support {self.attn_backend} backend now."
            )
        self.blocks = nn.ModuleList(
            [
                Qwen3_VisionBlock(
                    dim=self.hidden_size,
                    num_heads=self.num_heads,
                    mlp_hidden_dim=vision_config.intermediate_size,
                    act_fn=_ACTIVATION_REGISTRY[vision_config.hidden_act],
                    norm_layer=norm_layer,
                    quant_config=quant_config,
                    prefix=f"{prefix}.blocks.{layer_idx}",
                    use_data_parallel=use_data_parallel,
                    attn_backend=self.attn_backend,
                )
                for layer_idx in range(vision_config.depth)
            ]
        )
403
404
405
406
407
408
409
410
411

    @property
    def dtype(self) -> torch.dtype:
        return self.patch_embed.proj.weight.dtype

    @property
    def device(self) -> torch.device:
        return self.patch_embed.proj.weight.device

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
    @staticmethod
    @lru_cache(maxsize=1024)
    def rot_pos_ids(h: int, w: int, spatial_merge_size: int) -> torch.Tensor:
        hpos_ids = np.broadcast_to(np.arange(h).reshape(h, 1), (h, w))
        h_div = h // spatial_merge_size
        w_div = w // spatial_merge_size
        hpos_ids = hpos_ids.reshape(
            h_div,
            spatial_merge_size,
            w_div,
            spatial_merge_size,
        )
        hpos_ids = hpos_ids.transpose(0, 2, 1, 3)
        hpos_ids = hpos_ids.flatten()

        wpos_ids = np.broadcast_to(np.arange(w).reshape(1, w), (h, w))
        wpos_ids = wpos_ids.reshape(
            h_div,
            spatial_merge_size,
            w_div,
            spatial_merge_size,
        )
        wpos_ids = wpos_ids.transpose(0, 2, 1, 3)
        wpos_ids = wpos_ids.flatten()

        return torch.from_numpy(np.stack([hpos_ids, wpos_ids], axis=-1))

439
440
    def rot_pos_emb(self, grid_thw: list[list[int]]):
        max_grid_size = max(max(h, w) for _, h, w in grid_thw)
441
442
443
444
445
446
        pos_ids = [
            self.rot_pos_ids(h, w, self.spatial_merge_size)
            if t == 1
            else self.rot_pos_ids(h, w, self.spatial_merge_size).repeat(t, 1)
            for t, h, w in grid_thw
        ]
447
        pos_ids = torch.cat(pos_ids, dim=0).to(self.device, non_blocking=True)
448
449
450
451

        # Use pre-computed cos_sin_cache from RotaryEmbedding
        cos, sin = self.rotary_pos_emb.get_cos_sin(max_grid_size)

452
453
        cos_combined = cos[pos_ids].flatten(1)
        sin_combined = sin[pos_ids].flatten(1)
454
455

        return cos_combined, sin_combined
456

457
    def fast_pos_embed_interpolate(self, grid_thw: list[list[int]]) -> torch.Tensor:
458
459
460
        num_grid_per_side = self.num_grid_per_side
        m_size = self.spatial_merge_size
        hidden_dim = self.pos_embed.embedding_dim
461

462
        outputs = []
463
        for t, h, w in grid_thw:
464
465
466
467
468
469
            h_idxs = torch.linspace(
                0, num_grid_per_side - 1, h, dtype=torch.float32, device=self.device
            )
            w_idxs = torch.linspace(
                0, num_grid_per_side - 1, w, dtype=torch.float32, device=self.device
            )
470
471
472
473
474
475
476
477
478

            h_floor = h_idxs.to(torch.long)
            w_floor = w_idxs.to(torch.long)
            h_ceil = torch.clamp(h_floor + 1, max=num_grid_per_side - 1)
            w_ceil = torch.clamp(w_floor + 1, max=num_grid_per_side - 1)

            dh = h_idxs - h_floor
            dw = w_idxs - w_floor

479
            # Create meshgrid view for all h, w vars
480
481
482
            dh_grid, dw_grid = torch.meshgrid(dh, dw, indexing="ij")
            h_floor_grid, w_floor_grid = torch.meshgrid(h_floor, w_floor, indexing="ij")
            h_ceil_grid, w_ceil_grid = torch.meshgrid(h_ceil, w_ceil, indexing="ij")
483
484
485
486
487
488
489
490
491
492
493

            # original computation of weights
            # w00 = (1 - dh_grid) * (1 - dw_grid)
            # w01 = (1 - dh_grid) * dw_grid
            # w10 = dh_grid * (1 - dw_grid)
            # w11 = dh_grid * dw_grid
            # we reuse w11 here to avoid duplicate
            # dh_grid * dw_grid computation
            w11 = dh_grid * dw_grid
            w10 = dh_grid - w11
            w01 = dw_grid - w11
494
            w00 = 1 - dh_grid - w01
495

496
497
498
            h_grid = torch.stack([h_floor_grid, h_floor_grid, h_ceil_grid, h_ceil_grid])
            w_grid = torch.stack([w_floor_grid, w_ceil_grid, w_floor_grid, w_ceil_grid])
            h_grid_idx = h_grid * num_grid_per_side
499

500
            indices = (h_grid_idx + w_grid).reshape(4, -1)
501
            weights = torch.stack([w00, w01, w10, w11], dim=0).reshape(4, -1, 1)
502
            weights = weights.to(dtype=self.dtype)
503
504

            embeds = self.pos_embed(indices)
505
506
            embeds *= weights
            combined = embeds.sum(dim=0)
507

508
509
            combined = combined.reshape(
                h // m_size, m_size, w // m_size, m_size, hidden_dim
510
            )
511
512
            combined = combined.permute(0, 2, 1, 3, 4).reshape(1, -1, hidden_dim)
            repeated = combined.expand(t, -1, -1).reshape(-1, hidden_dim)
513
514
515
            outputs.append(repeated)

        return torch.cat(outputs, dim=0)
516
517
518
519

    def compute_attn_mask_seqlen(
        self,
        cu_seqlens: torch.Tensor,
520
    ) -> torch.Tensor:
521
        max_seqlen = torch.zeros([], device=cu_seqlens.device)
522
        if (
523
524
            self.attn_backend == AttentionBackendEnum.FLASH_ATTN
            or self.attn_backend == AttentionBackendEnum.ROCM_AITER_FA
525
        ):
526
            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
527
        return max_seqlen
528
529
530
531

    def forward(
        self,
        x: torch.Tensor,
532
        grid_thw: torch.Tensor | list[list[int]],
533
    ) -> torch.Tensor:
534
        hidden_states = x.to(device=self.device, dtype=self.dtype, non_blocking=True)
535
536
        hidden_states = self.patch_embed(hidden_states)

537
538
        if isinstance(grid_thw, list):
            grid_thw_list = grid_thw
539
            grid_thw = np.array(grid_thw, dtype=np.int32)
540
541
        else:
            grid_thw_list = grid_thw.tolist()
542
            grid_thw = grid_thw.numpy()
543
544

        pos_embeds = self.fast_pos_embed_interpolate(grid_thw_list)
545
        hidden_states = hidden_states + pos_embeds
546
        rotary_pos_emb_cos, rotary_pos_emb_sin = self.rot_pos_emb(grid_thw_list)
547

548
549
550
551
552
        cu_seqlens = np.repeat(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
            axis=0, dtype=np.int32
        )
        cu_seqlens = np.concatenate([np.zeros(1, dtype=np.int32), cu_seqlens])
        cu_seqlens = torch.from_numpy(cu_seqlens)
553
554

        hidden_states = hidden_states.unsqueeze(1)
555
        max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens)
556
        cu_seqlens = cu_seqlens.to(self.device, non_blocking=True)
557
558
559

        deepstack_feature_lists = []
        for layer_num, blk in enumerate(self.blocks):
560
561
562
            hidden_states = blk(
                hidden_states,
                cu_seqlens=cu_seqlens,
563
564
                rotary_pos_emb_cos=rotary_pos_emb_cos,
                rotary_pos_emb_sin=rotary_pos_emb_sin,
565
566
                max_seqlen=max_seqlen,
            )
567
            if layer_num in self.deepstack_visual_indexes:
568
569
570
571
                deepstack_merger_idx = self.deepstack_visual_indexes.index(layer_num)
                deepstack_feature = self.deepstack_merger_list[deepstack_merger_idx](
                    hidden_states
                )
572
573
574
                deepstack_feature_lists.append(deepstack_feature)
        hidden_states = self.merger(hidden_states)
        hidden_states = torch.cat(
575
576
            [hidden_states] + deepstack_feature_lists, dim=1
        )  # [seq_len, hidden_size * (1 + depth_of_deepstack)]
577
578
        return hidden_states

579
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
580
581
582
583
584
585
586
587
588
589
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("attn.qkv.", "attn.q.", "q"),
            ("attn.qkv.", "attn.k.", "k"),
            ("attn.qkv.", "attn.v.", "v"),
        ]
        params_dict = dict(self.named_parameters(remove_duplicate=False))
        loaded_params: set[str] = set()

        for name, loaded_weight in weights:
590
            for param_name, weight_name, shard_id in stacked_params_mapping:
591
592
593
594
595
596
597
598
599
600
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)

                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                param = params_dict[name]
601
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params


class Qwen3VLProcessingInfo(Qwen2VLProcessingInfo):
    def get_hf_config(self):
        return self.ctx.get_hf_config(Qwen3VLConfig)

    def get_hf_processor(self, **kwargs: object) -> Qwen3VLProcessor:
        return self.ctx.get_hf_processor(
            Qwen3VLProcessor,
            use_fast=kwargs.pop("use_fast", True),
            **kwargs,
        )

618
    def get_image_processor(self, **kwargs: object) -> Qwen2VLImageProcessorFast:
619
620
621
622
623
624
625
626
627
628
629
630
        return self.get_hf_processor(**kwargs).image_processor

    def get_video_processor(self, **kwargs: object) -> Qwen3VLVideoProcessor:
        return self.get_hf_processor(**kwargs).video_processor

    def _get_vision_info(
        self,
        *,
        image_width: int,
        image_height: int,
        num_frames: int = 2,
        do_resize: bool = True,
631
        image_processor: Qwen2VLImageProcessorFast | Qwen3VLVideoProcessor | None,
632
    ) -> tuple[ImageSize, int]:
633
634
635
        if image_processor is None and num_frames > 1:
            image_processor = self.get_video_processor()
        elif image_processor is None:
636
637
            image_processor = self.get_image_processor()

638
639
        is_video = isinstance(image_processor, Qwen3VLVideoProcessor)

640
641
642
643
644
645
646
        hf_config = self.get_hf_config()
        vision_config = hf_config.vision_config
        patch_size = vision_config.patch_size
        merge_size = vision_config.spatial_merge_size
        temporal_patch_size = vision_config.temporal_patch_size

        if do_resize:
647
648
649
650
            if is_video:
                smart_resize = video_smart_resize
                extra_kwargs = {
                    "num_frames": num_frames,
651
                    "temporal_factor": temporal_patch_size,
652
653
654
655
                }
            else:
                smart_resize = image_smart_resize
                extra_kwargs = {}
656
657
658
659
660
661
            resized_height, resized_width = smart_resize(
                height=image_height,
                width=image_width,
                factor=patch_size * merge_size,
                min_pixels=image_processor.size["shortest_edge"],
                max_pixels=image_processor.size["longest_edge"],
662
                **extra_kwargs,
663
            )
664
            preprocessed_size = ImageSize(width=resized_width, height=resized_height)
665
        else:
666
            preprocessed_size = ImageSize(width=image_width, height=image_height)
667
668
669
670
671
672
673
674
675
676
677
678

        padded_num_frames = num_frames + num_frames % temporal_patch_size

        grid_t = max(padded_num_frames // temporal_patch_size, 1)
        grid_h = preprocessed_size.height // patch_size
        grid_w = preprocessed_size.width // patch_size

        num_patches = grid_t * grid_h * grid_w
        num_vision_tokens = num_patches // (merge_size**2)

        return preprocessed_size, num_vision_tokens

679
680
681
682
    def _get_max_video_frames(self, max_tokens: int, start_num_frames: int = 2) -> int:
        return super()._get_max_video_frames(
            max_tokens, start_num_frames=start_num_frames
        )
683
684
685
686
687
688
689

    def get_num_frames_with_most_features(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> int:
        return super().get_num_frames_with_most_features(
690
691
            seq_len, mm_counts, max_frames_per_video=_MAX_FRAMES_PER_VIDEO
        )
692
693
694
695
696
697
698
699
700
701

    def get_max_video_tokens(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> int:
        target_width, target_height = self.get_image_size_with_most_features()
        video_soft_tokens = self.get_num_video_tokens(
            image_width=target_width,
            image_height=target_height,
702
            num_frames=self.get_num_frames_with_most_features(seq_len, mm_counts),
703
704
705
706
707
708
709
710
            image_processor=None,
        )

        # NOTE: By default in Qwen3-VL, one video token is converted to
        # "<{timestamp} seconds>" (on average 9.5 tokens) + vision_start_token + video_token + vision_end_token # noqa: E501
        formatted_video_soft_tokens = video_soft_tokens * 12.5
        return int(formatted_video_soft_tokens)

711
712
713
    def _calculate_timestamps(
        self, indices: list[int] | torch.Tensor, video_fps: float, merge_size: int
    ):
714
715
716
717
        if not isinstance(indices, list):
            indices = indices.tolist()
        if len(indices) % merge_size != 0:
            # don't update metadata's frames_indices directly
718
            indices = indices + [indices[-1]] * (merge_size - len(indices) % merge_size)
719
        timestamps = [idx / video_fps for idx in indices]
720
721
722
723
        timestamps = [
            (timestamps[i] + timestamps[i + merge_size - 1]) / 2
            for i in range(0, len(timestamps), merge_size)
        ]
724
725
726
        return timestamps

    def _get_video_second_idx(
727
728
729
        self,
        metadata: dict[str, Any],
        out_item: MultiModalKwargsItem,
730
731
        do_sample_frames: bool | None = None,
        sampled_fps: float | None = None,
732
    ) -> list[int]:
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
        video_processor = self.get_video_processor()
        merge_size = video_processor.merge_size
        indices = metadata["frames_indices"]

        # metadata["fps"] refers to the true fps of the input video.
        video_fps = metadata["fps"]
        if do_sample_frames is None:
            do_sample_frames = metadata.get("do_sample_frames", False)

        # If video frames are sampled in HF processor (instead of vLLM
        # video loader), we need to re-calculate the indices from original
        # metadata.
        if do_sample_frames:
            # here video_fps is the fps of the sampled video, and
            # metadata["fps"] refers to the fps of the original video.
748
            sampled_fps = sampled_fps if sampled_fps else video_processor.fps
749
            total_num_frames = metadata["total_num_frames"]
750
            num_frames = int(total_num_frames / metadata["fps"] * sampled_fps)
751
            num_frames = min(
752
753
754
755
756
757
758
759
760
761
762
763
                min(
                    max(num_frames, video_processor.min_frames),
                    video_processor.max_frames,
                ),
                total_num_frames,
            )
            indices = (
                np.linspace(0, total_num_frames - 1, num_frames)
                .round()
                .astype(int)
                .tolist()
            )
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
        timestamps = self._calculate_timestamps(indices, video_fps, merge_size)
        return timestamps


class Qwen3VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3VLProcessingInfo]):
    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        num_images = mm_counts.get("image", 0)
        num_videos = mm_counts.get("video", 0)

        image_token = "<|vision_start|><|image_pad|><|vision_end|>"
        video_token = "<|vision_start|><|video_pad|><|vision_end|>"

        return image_token * num_images + video_token * num_videos

    def get_dummy_mm_data(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
782
        mm_options: Mapping[str, BaseDummyOptions] | None = None,
783
784
785
    ) -> MultiModalDataDict:
        num_images = mm_counts.get("image", 0)
        num_videos = mm_counts.get("video", 0)
786
787
        image_overrides = mm_options.get("image") if mm_options else None
        video_overrides = mm_options.get("video") if mm_options else None
788

789
        target_width, target_height = self.info.get_image_size_with_most_features()
790
        target_num_frames = self.info.get_num_frames_with_most_features(
791
792
            seq_len, mm_counts
        )
793
794
795
796
797
798
799
800
801

        if video_overrides:
            assert isinstance(video_overrides, VideoDummyOptions)
            num_frames_override = video_overrides.num_frames
            if num_frames_override:
                if num_frames_override > target_num_frames:
                    logger.warning(
                        "video.num_frames override (%d) exceeds model's "
                        "maximum number of frames (%d), will be ignored",
802
803
804
                        num_frames_override,
                        target_num_frames,
                    )
805
806
807
                if num_frames_override < 2:
                    logger.warning(
                        "video.num_frames override (%d) cannot be less "
808
809
810
                        "than 2, will be ignored",
                        num_frames_override,
                    )
811
812
813
                target_num_frames = min(target_num_frames, num_frames_override)
        target_num_frames = max(target_num_frames, 2)

814
815
816
817
818
819
        target_video_size, _ = self.info._get_vision_info(
            image_width=target_width,
            image_height=target_height,
            num_frames=target_num_frames,
            image_processor=self.info.get_video_processor(),
        )
820
821
822
823
824
825
826
827
828
829
        # NOTE: we need to do this check here since Qwen3-VL resizes video
        # frames depending on how many frames there are.
        width, height = target_video_size.width, target_video_size.height
        if video_overrides:
            assert isinstance(video_overrides, VideoDummyOptions)
            width_override = video_overrides.width
            if width_override:
                if width_override > width:
                    logger.warning(
                        "video.width override (%d) exceeds model's "
830
831
832
833
                        "maximum width (%d), will be ignored",
                        width_override,
                        width,
                    )
834
835
836
837
838
839
840
                width = min(width, width_override)
            height_override = video_overrides.height
            if height_override:
                if height_override > height:
                    logger.warning(
                        "video.height override (%d) exceeds model's "
                        "maximum height (%d), will be ignored",
841
842
843
                        height_override,
                        height,
                    )
844
                height = min(height, height_override)
845

846
        return {
847
848
849
850
851
852
853
            "image": self._get_dummy_images(
                width=target_width,
                height=target_height,
                num_images=num_images,
                overrides=image_overrides,
            ),
            "video": self._get_dummy_videos(
854
855
                width=width,
                height=height,
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
                num_frames=target_num_frames,
                num_videos=num_videos,
            ),
        }

    def _get_dummy_videos(
        self,
        *,
        width: int,
        height: int,
        num_frames: int,
        num_videos: int,
    ) -> list[VideoItem]:
        video = np.full((num_frames, width, height, 3), 255, dtype=np.uint8)
        video_items = []
        for i in range(num_videos):
            video_metadata = {
                "fps": 2.0,
                "duration": num_frames / 2.0,
                "total_num_frames": num_frames,
                "frames_indices": [i for i in range(num_frames)],
                "video_backend": "opencv",
                "do_sample_frames": False,
            }
            video_item = (video.copy(), video_metadata)
            video_items.append(video_item)
        return video_items


885
class Qwen3VLMultiModalProcessor(BaseMultiModalProcessor[Qwen3VLProcessingInfo]):
886
    def _get_data_parser(self) -> MultiModalDataParser:
887
888
889
890
        return Qwen2VLMultiModalDataParser(
            self.info.get_hf_config().vision_config.spatial_merge_size,
            video_needs_metadata=True,
        )
891
892
893
894
895
896
897
898
899
900
901
902

    def _call_hf_processor(
        self,
        prompt: str,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
        tok_kwargs: Mapping[str, object],
    ) -> BatchFeature:
        mm_data = dict(mm_data)
        processor = self.info.get_hf_processor(**mm_kwargs)

        # Separate video processing from image processing. Because the videos
903
904
        # are processed into several image patches
        if videos := mm_data.pop("videos", []):
905
906
907
            video_grid_thw_lst = []
            pixel_values_videos_lst = []

908
            for item in videos:
909
910
911
912
913
914
915
916
917
918
919
920
921
922
                video_array, metadata = item

                # NOTE: @JJJYmmm new attr metadata.frames_indices indicates
                # the sampled frames indices of pre-sampled videos, which is
                # used to calculate the timestamps. Make sure that
                # do_sample_frames in mm_kwargs is false for presampled videos.

                # NOTE: a copy of is created to update do_sample_frames,
                # otherwise mm_hash for the object will be incorrect.
                video_mm_kwargs = dict(**mm_kwargs)
                if "do_sample_frames" not in video_mm_kwargs:
                    # qwen_vl_utils already has "do_sample_frames" in
                    # mm_kwargs, don't overwrite it.
                    video_mm_kwargs["do_sample_frames"] = metadata.get(
923
924
                        "do_sample_frames", False
                    )
925

926
927
928
                metadata = VideoMetadata(
                    **{k: metadata[k] for k in metadata if k != "do_sample_frames"}
                )
929
930
931
932
933
934
935
936
937
938
939
940

                video_mm_data = dict()
                video_mm_data["videos"] = [[video_array]]
                video_mm_data["video_metadata"] = [[metadata]]

                video_outputs = super()._call_hf_processor(
                    prompt="<|vision_start|><|video_pad|><|vision_end|>",
                    mm_data=video_mm_data,
                    mm_kwargs=video_mm_kwargs,
                    tok_kwargs=tok_kwargs,
                )
                input_ids = video_outputs.pop("input_ids")
941
                video_placeholder = processor.tokenizer.batch_decode(input_ids)[0]
942
943
944
945
946
947
948
                prompt = prompt.replace(
                    "<|vision_start|><|video_pad|><|vision_end|>",
                    video_placeholder,
                    1,
                )

                video_grid_thw_lst.append(video_outputs["video_grid_thw"])
949
                pixel_values_videos_lst.append(video_outputs["pixel_values_videos"])
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
            video_outputs = dict(
                pixel_values_videos=torch.cat(pixel_values_videos_lst),
                video_grid_thw=torch.cat(video_grid_thw_lst),
            )
        else:
            video_outputs = dict()

        processed_outputs = super()._call_hf_processor(
            prompt=prompt,
            mm_data=mm_data,
            mm_kwargs=mm_kwargs,
            tok_kwargs=tok_kwargs,
        )
        combined_outputs = dict(
            processed_outputs,
            **video_outputs,
        )
        return BatchFeature(combined_outputs)

    def _get_mm_fields_config(
        self,
        hf_inputs: BatchFeature,
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
        image_grid_thw = hf_inputs.get("image_grid_thw", torch.empty((0, 3)))
        image_grid_sizes = image_grid_thw.prod(-1)

        video_grid_thw = hf_inputs.get("video_grid_thw", torch.empty((0, 3)))
        video_grid_sizes = video_grid_thw.prod(-1)

        return dict(
            pixel_values=MultiModalFieldConfig.flat_from_sizes(
982
983
                "image", image_grid_sizes
            ),
984
            image_embeds=MultiModalFieldConfig.flat_from_sizes(
985
986
                "image", image_grid_sizes
            ),
987
            image_grid_thw=MultiModalFieldConfig.batched("image", keep_on_cpu=True),
988
            pixel_values_videos=MultiModalFieldConfig.flat_from_sizes(
989
990
                "video", video_grid_sizes
            ),
991
            video_embeds=MultiModalFieldConfig.flat_from_sizes(
992
993
                "video", video_grid_sizes
            ),
994
            video_grid_thw=MultiModalFieldConfig.batched("video", keep_on_cpu=True),
995
996
997
998
999
1000
1001
1002
1003
        )

    def _get_prompt_updates(
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, Any],
        out_mm_kwargs: MultiModalKwargsItems,
    ) -> Sequence[PromptUpdate]:
        hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
1004
        image_processor = self.info.get_image_processor(**hf_processor_mm_kwargs)
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
        tokenizer = self.info.get_tokenizer()
        hf_config = self.info.get_hf_config()

        video_token_id = hf_config.video_token_id
        vision_start_token_id = hf_config.vision_start_token_id
        vision_end_token_id = hf_config.vision_end_token_id

        merge_length = image_processor.merge_size**2

        def get_image_replacement_qwen3vl(item_idx: int):
            out_item = out_mm_kwargs["image"][item_idx]
            grid_thw = out_item["image_grid_thw"].data
            assert isinstance(grid_thw, torch.Tensor)

            num_tokens = int(grid_thw.prod()) // merge_length
            return [hf_processor.image_token_id] * num_tokens

        def get_video_replacement_qwen3vl(item_idx: int):
            out_item = out_mm_kwargs["video"][item_idx]
            grid_thw = out_item["video_grid_thw"].data
            assert isinstance(grid_thw, torch.Tensor)

            video, metadata = mm_items["video"][item_idx]
            do_sample_frames = hf_processor_mm_kwargs.get("do_sample_frames")
            sampled_fps = hf_processor_mm_kwargs.get("fps")
            if is_list_of(sampled_fps, float):
                sampled_fps = sampled_fps[item_idx]
            timestamps = self.info._get_video_second_idx(
1033
1034
                metadata, out_item, do_sample_frames, sampled_fps
            )
1035
1036
1037

            assert len(timestamps) == grid_thw[0], (
                f"The timestamps length({len(timestamps)}) should be equal "
1038
1039
                f"video length ({grid_thw[0]})."
            )
1040
1041

            frames_idx_token = [
1042
                tokenizer.encode(f"<{curr_time:.1f} seconds>", add_special_tokens=False)
1043
1044
1045
1046
1047
1048
                for curr_time in timestamps
            ]
            num_tokens_per_frame = int(grid_thw[1:].prod()) // merge_length
            placeholder = []
            for frame_idx in frames_idx_token:
                placeholder.extend(frame_idx)
1049
1050
1051
1052
1053
1054
                placeholder.extend(
                    [vision_start_token_id]
                    + [video_token_id] * num_tokens_per_frame
                    + [vision_end_token_id]
                )
            return PromptUpdateDetails.select_token_id(placeholder, video_token_id)
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080

        return [
            PromptReplacement(
                modality="image",
                target=hf_processor.image_token,
                replacement=get_image_replacement_qwen3vl,
            ),
            # NOTE: We match string on purpose since searching sequence of
            # token ids takes more time.
            PromptReplacement(
                modality="video",
                target="<|vision_start|><|video_pad|><|vision_end|>",
                replacement=get_video_replacement_qwen3vl,
            ),
        ]


@support_torch_compile(
    dynamic_arg_dims={
        "input_ids": 0,
        # positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl,
        # otherwise (seq_len, ).
        "positions": -1,
        "intermediate_tensors": 0,
        "inputs_embeds": 0,
        # the same shape as input_embeds
1081
1082
1083
        "deepstack_input_embeds": 0,
    }
)
1084
1085
1086
1087
1088
class Qwen3LLMModel(Qwen3Model):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__(vllm_config=vllm_config, prefix=prefix)
        if not get_pp_group().is_first_rank:
            assert self.start_layer >= len(
1089
1090
1091
1092
1093
                vllm_config.model_config.hf_config.vision_config.deepstack_visual_indexes
            ), (
                "start_layer should be greater than or equal to "
                "len(deepstack_visual_indexes)"
            )
1094
1095
1096
1097
1098

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1099
1100
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
1101
        # args for deepstack
1102
1103
        deepstack_input_embeds: IntermediateTensors | None = None,
    ) -> torch.Tensor | IntermediateTensors:
1104
1105
1106
1107
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
1108
                hidden_states = self.embed_input_ids(input_ids)
1109
1110
1111
1112
1113
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]
1114

1115
        aux_hidden_states = []
1116
1117
        for layer_idx, layer in islice(
            enumerate(self.layers), self.start_layer, self.end_layer
1118
        ):
1119
1120
            if layer_idx in self.aux_hidden_state_layers:
                aux_hidden_states.append(hidden_states + residual)
1121

1122
1123
1124
1125
1126
1127
            hidden_states, residual = layer(
                positions,
                hidden_states,
                residual,
            )

1128
1129
1130
1131
1132
1133
1134
            if deepstack_input_embeds is not None and layer_idx in range(
                0, len(deepstack_input_embeds)
            ):
                hidden_states = (
                    hidden_states
                    + deepstack_input_embeds[f"deepstack_input_embeds_{layer_idx}"]
                )
1135
1136

        if not get_pp_group().is_last_rank:
1137
1138
1139
            return IntermediateTensors(
                {"hidden_states": hidden_states, "residual": residual}
            )
1140
        hidden_states, _ = self.norm(hidden_states, residual)
1141
1142
1143

        if len(aux_hidden_states) > 0:
            return hidden_states, aux_hidden_states
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
        return hidden_states


class Qwen3LLMForCausalLM(Qwen3ForCausalLM):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super(Qwen3ForCausalLM, self).__init__()
        config = vllm_config.model_config.hf_config.text_config
        quant_config = vllm_config.quant_config

        self.config = config

        self.quant_config = quant_config
1156
1157
1158
        self.model = Qwen3LLMModel(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
1159
1160
1161
1162
1163

        if get_pp_group().is_last_rank:
            if config.tie_word_embeddings:
                self.lm_head = self.model.embed_tokens
            else:
1164
1165
1166
1167
1168
1169
                self.lm_head = ParallelLMHead(
                    config.vocab_size,
                    config.hidden_size,
                    quant_config=quant_config,
                    prefix="lm_head",
                )
1170
1171
1172
1173
1174
1175
        else:
            self.lm_head = PPMissingLayer()

        self.logits_processor = LogitsProcessor(config.vocab_size)

        self.make_empty_intermediate_tensors = (
1176
1177
            self.model.make_empty_intermediate_tensors
        )
1178
1179


1180
1181
1182
1183
1184
1185
@MULTIMODAL_REGISTRY.register_processor(
    Qwen3VLMultiModalProcessor,
    info=Qwen3VLProcessingInfo,
    dummy_inputs=Qwen3VLDummyInputsBuilder,
)
class Qwen3VLForConditionalGeneration(
1186
1187
1188
1189
1190
1191
    nn.Module,
    SupportsMultiModal,
    SupportsLoRA,
    SupportsPP,
    SupportsMRoPE,
    SupportsEagle3,
1192
):
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }
1204
1205
1206

    supports_encoder_tp_data = True

1207
1208
1209
1210
1211
1212
    # To ensure correct weight loading and mapping.
    hf_to_vllm_mapper = WeightsMapper(
        orig_to_new_prefix={
            "model.visual.": "visual.",
            "lm_head.": "language_model.lm_head.",
            "model.language_model.": "language_model.model.",
1213
1214
        }
    )
1215
1216

    @classmethod
1217
    def get_placeholder_str(cls, modality: str, i: int) -> str | None:
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
        if modality.startswith("image"):
            return "<|vision_start|><|image_pad|><|vision_end|>"
        if modality.startswith("video"):
            return "<|vision_start|><|video_pad|><|vision_end|>"

        raise ValueError("Only image or video modality is supported")

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "model"):
        super().__init__()
        config: Qwen3VLConfig = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        multimodal_config = vllm_config.model_config.multimodal_config

        self.config = config
        self.multimodal_config = multimodal_config
1233
        self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
1234
1235
1236
        if not multimodal_config.get_limit_per_prompt(
            "image"
        ) and not multimodal_config.get_limit_per_prompt("video"):
1237
1238
            self.visual = None
        else:
1239
1240
1241
1242
1243
            attn_backend_override = (
                multimodal_config.mm_encoder_attn_backend
                if multimodal_config is not None
                else None
            )
1244
1245
1246
1247
1248
1249
            self.visual = Qwen3_VisionTransformer(
                config.vision_config,
                norm_eps=getattr(config, "rms_norm_eps", 1e-6),
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "visual"),
                use_data_parallel=self.use_data_parallel,
1250
                attn_backend_override=attn_backend_override,
1251
            )
1252

1253
1254
1255
        self.language_model = Qwen3LLMForCausalLM(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "language_model")
        )
1256
1257

        self.make_empty_intermediate_tensors = (
1258
1259
            self.language_model.make_empty_intermediate_tensors
        )
1260

1261
1262
1263
1264
1265
1266
        self.use_deepstack = hasattr(config.vision_config, "deepstack_visual_indexes")
        self.deepstack_num_level = (
            len(config.vision_config.deepstack_visual_indexes)
            if self.use_deepstack
            else 0
        )
1267
        # register buffer for deepstack
1268
1269
1270
1271
        if self.use_deepstack and self.visual is not None:
            self.deepstack_input_embeds = [
                torch.zeros(
                    vllm_config.scheduler_config.max_num_batched_tokens,
1272
1273
                    config.text_config.hidden_size,
                )
1274
1275
1276
1277
                for _ in range(self.deepstack_num_level)
            ]
        else:
            self.deepstack_input_embeds = None
1278
1279
        self.visual_dim = config.vision_config.out_hidden_size
        self.multiscale_dim = self.visual_dim * self.deepstack_num_level
1280

1281
1282
1283
1284
1285
1286
1287
    def set_aux_hidden_state_layers(self, layers: tuple[int, ...]) -> None:
        self.language_model.model.aux_hidden_state_layers = layers

    def get_eagle3_aux_hidden_state_layers(self) -> tuple[int, ...]:
        num_layers = len(self.language_model.model.layers)
        return (2, num_layers // 2, num_layers - 3)

1288
    def _get_deepstack_input_embeds(self, num_tokens: int) -> IntermediateTensors:
1289
        # get deepstack_input_embeds from buffer, and clear the buffer
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
        return IntermediateTensors(
            {
                f"deepstack_input_embeds_{idx}": self.deepstack_input_embeds[idx][
                    :num_tokens
                ]
                for idx in range(self.deepstack_num_level)
            }
        )

    def _set_deepstack_input_embeds(self, deepstack_input_embeds: torch.Tensor) -> None:
1300
1301
1302
1303
        # set deepstack_input_embeds to buffer
        num_tokens = deepstack_input_embeds.size(1)
        if num_tokens > self.deepstack_input_embeds[0].size(0):
            self.deepstack_input_embeds = [
1304
1305
1306
1307
1308
1309
                torch.zeros(
                    num_tokens,
                    self.config.text_config.hidden_size,
                    device=self.deepstack_input_embeds[0].device,
                    dtype=self.deepstack_input_embeds[0].dtype,
                )
1310
1311
1312
1313
                for _ in range(self.deepstack_num_level)
            ]
        for idx in range(self.deepstack_num_level):
            self.deepstack_input_embeds[idx][:num_tokens].copy_(
1314
1315
                deepstack_input_embeds[idx]
            )
1316
1317
1318
1319
1320
1321
1322
1323

    def _clear_deepstack_input_embeds(self, num_tokens: int) -> None:
        # clear deepstack_input_embeds in buffer
        if num_tokens > 0:
            for idx in range(self.deepstack_num_level):
                self.deepstack_input_embeds[idx][:num_tokens].zero_()

    def _parse_and_validate_image_input(
1324
        self, **kwargs: object
1325
    ) -> Qwen2_5_VLImageInputs | None:
1326
1327
1328
1329
1330
1331
1332
1333
        pixel_values = kwargs.pop("pixel_values", None)
        image_embeds = kwargs.pop("image_embeds", None)
        image_grid_thw = kwargs.pop("image_grid_thw", None)

        if pixel_values is None and image_embeds is None:
            return None

        if pixel_values is not None:
1334
1335
1336
1337
1338
            return Qwen2_5_VLImagePixelInputs(
                type="pixel_values",
                pixel_values=pixel_values,
                image_grid_thw=image_grid_thw,
            )
1339
1340
1341
1342
1343

        if image_embeds is not None:
            return Qwen2_5_VLImageEmbeddingInputs(
                type="image_embeds",
                image_embeds=image_embeds,
1344
1345
                image_grid_thw=image_grid_thw,
            )
1346
1347

    def _parse_and_validate_video_input(
1348
        self, **kwargs: object
1349
    ) -> Qwen2_5_VLVideoInputs | None:
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
        pixel_values_videos = kwargs.pop("pixel_values_videos", None)
        video_embeds = kwargs.pop("video_embeds", None)
        video_grid_thw = kwargs.pop("video_grid_thw", None)
        second_per_grid_ts = kwargs.pop("second_per_grid_ts", None)

        if pixel_values_videos is None and video_embeds is None:
            return None

        if pixel_values_videos is not None:
            return Qwen2_5_VLVideoPixelInputs(
                type="pixel_values_videos",
                pixel_values_videos=pixel_values_videos,
                video_grid_thw=video_grid_thw,
                second_per_grid_ts=second_per_grid_ts,
            )

        if video_embeds is not None:
            return Qwen2_5_VLVideoEmbeddingInputs(
                type="video_embeds",
                video_embeds=video_embeds,
1370
1371
                video_grid_thw=video_grid_thw,
            )
1372
1373

    def _process_image_input(
1374
1375
        self, image_input: Qwen2_5_VLImageInputs
    ) -> tuple[torch.Tensor, ...]:
1376
1377
1378
1379
1380
1381
1382
        grid_thw = image_input["image_grid_thw"]
        assert grid_thw.ndim == 2

        if image_input["type"] == "image_embeds":
            image_embeds = image_input["image_embeds"].type(self.visual.dtype)
        else:
            pixel_values = image_input["pixel_values"].type(self.visual.dtype)
1383
            if self.use_data_parallel:
1384
                return run_dp_sharded_mrope_vision_model(
1385
                    self.visual, pixel_values, grid_thw.tolist(), rope_type="rope_3d"
1386
                )
1387
            else:
1388
                image_embeds = self.visual(pixel_values, grid_thw=grid_thw)
1389
1390
1391

        # Split concatenated embeddings for each image item.
        merge_size = self.visual.spatial_merge_size
1392
        sizes = (grid_thw.prod(-1) // merge_size // merge_size).tolist()
1393
1394
1395
        return image_embeds.split(sizes)

    def _process_video_input(
1396
1397
        self, video_input: Qwen2_5_VLVideoInputs
    ) -> tuple[torch.Tensor, ...]:
1398
1399
1400
1401
1402
1403
1404
        grid_thw = video_input["video_grid_thw"]
        assert grid_thw.ndim == 2

        if video_input["type"] == "video_embeds":
            video_embeds = video_input["video_embeds"].type(self.visual.dtype)
        else:
            pixel_values_videos = video_input["pixel_values_videos"].type(
1405
1406
                self.visual.dtype
            )
1407
            if self.use_data_parallel:
1408
                grid_thw_list = grid_thw.tolist()
1409
1410
1411
                return run_dp_sharded_mrope_vision_model(
                    self.visual, pixel_values_videos, grid_thw_list, rope_type="rope_3d"
                )
1412
            else:
1413
                video_embeds = self.visual(pixel_values_videos, grid_thw=grid_thw)
1414
1415
1416

        # Split concatenated embeddings for each video item.
        merge_size = self.visual.spatial_merge_size
1417
        sizes = (grid_thw.prod(-1) // merge_size // merge_size).tolist()
1418
1419
1420
1421
1422
        return video_embeds.split(sizes)

    def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
        mm_input_by_modality = {}
        for input_key in kwargs:
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
            if (
                input_key in ("pixel_values", "image_embeds")
                and "image" not in mm_input_by_modality
            ):
                mm_input_by_modality["image"] = self._parse_and_validate_image_input(
                    **kwargs
                )
            if (
                input_key in ("pixel_values_videos", "video_embeds")
                and "video" not in mm_input_by_modality
            ):
                mm_input_by_modality["video"] = self._parse_and_validate_video_input(
                    **kwargs
                )
1437
1438
        return mm_input_by_modality

1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
    def iter_mm_grid_hw(
        self, input_tokens: list[int], mm_features: list[MultiModalFeatureSpec]
    ) -> Iterator[tuple[int, int, int]]:
        video_token_id = self.config.video_token_id
        spatial_merge_size = self.config.vision_config.spatial_merge_size
        for mm_feature in sorted(mm_features, key=lambda f: f.mm_position.offset):
            offset = mm_feature.mm_position.offset
            if mm_feature.modality == "image":
                t, h, w = mm_feature.data["image_grid_thw"].data.tolist()
                assert t == 1, f"Image must have 1 frame, got {t}"
                yield offset, h // spatial_merge_size, w // spatial_merge_size
            elif mm_feature.modality == "video":
                t, h, w = mm_feature.data["video_grid_thw"].data.tolist()
                llm_grid_h = h // spatial_merge_size
                llm_grid_w = w // spatial_merge_size
                for _ in range(t):
                    offset = input_tokens.index(video_token_id, offset)
                    yield offset, llm_grid_h, llm_grid_w
                    offset += llm_grid_h * llm_grid_w
            else:
                raise ValueError(f"Unsupported modality: {mm_feature.modality}")

1461
    def get_mrope_input_positions(
1462
        self,
1463
        input_tokens: list[int],
1464
        mm_features: list[MultiModalFeatureSpec],
1465
    ) -> tuple[torch.Tensor, int]:
1466
        llm_pos_ids_list = []
1467
        st = 0
1468
1469
1470
1471
        for offset, llm_grid_h, llm_grid_w in self.iter_mm_grid_hw(
            input_tokens, mm_features
        ):
            text_len = offset - st
1472
1473
            st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
            llm_pos_ids_list.append(
1474
                np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx
1475
1476
            )

1477
1478
1479
            grid_indices = np.indices((1, llm_grid_h, llm_grid_w)).reshape(3, -1)
            llm_pos_ids_list.append(grid_indices + text_len + st_idx)
            st = offset + llm_grid_h * llm_grid_w
1480
1481
1482
1483
1484

        if st < len(input_tokens):
            st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
            text_len = len(input_tokens) - st
            llm_pos_ids_list.append(
1485
                np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx
1486
1487
            )

1488
        llm_positions = np.concatenate(llm_pos_ids_list, axis=1).reshape(3, -1)
1489
        mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()
1490
        return torch.from_numpy(llm_positions), mrope_position_delta
1491

1492
1493
1494
    def get_language_model(self) -> torch.nn.Module:
        return self.language_model

1495
    def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings | None:
1496
        mm_input_by_modality = self._parse_and_validate_multimodal_inputs(**kwargs)
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
        if not mm_input_by_modality:
            return None

        # The result multimodal_embeddings is tuple of tensors, with each
        # tensor correspoending to a multimodal data item (image or video).
        multimodal_embeddings: tuple[torch.Tensor, ...] = ()

        # NOTE: It is important to iterate over the keys in this dictionary
        # to preserve the order of the modalities.
        for modality in mm_input_by_modality:
            multimodal_input = mm_input_by_modality[modality]
            if modality == "image":
1509
1510
                image_embeddings = self._process_image_input(multimodal_input)
                multimodal_embeddings += tuple(image_embeddings)
1511
1512
            if modality == "video":
                video_embeddings = self._process_video_input(multimodal_input)
1513
                multimodal_embeddings += tuple(video_embeddings)
1514
1515
1516
        return multimodal_embeddings

    def _compute_deepstack_embeds(
1517
1518
1519
1520
1521
1522
        self,
        inputs_embeds: torch.Tensor,
        multimodal_embeddings: MultiModalEmbeddings,
        is_multimodal: torch.Tensor,
    ) -> tuple[torch.Tensor, MultiModalEmbeddings]:
        visual_lens = [len(x) for x in multimodal_embeddings]
1523
1524
        multimodal_embeddings_cat = torch.cat(multimodal_embeddings, dim=0)

1525
1526
1527
1528
1529
1530
1531
1532
        (
            multimodal_embeddings_main,
            multimodal_embeddings_multiscale,
        ) = torch.split(
            multimodal_embeddings_cat,
            [self.visual_dim, self.multiscale_dim],
            dim=-1,
        )
1533

1534
1535
1536
        multimodal_embeddings = torch.split(
            multimodal_embeddings_main, visual_lens, dim=0
        )
1537
        multimodal_embeddings_multiscale = torch.split(
1538
1539
            multimodal_embeddings_multiscale, visual_lens, dim=0
        )
1540
1541

        deepstack_input_embeds = inputs_embeds.new_zeros(
1542
1543
            inputs_embeds.size(0), self.deepstack_num_level * inputs_embeds.size(1)
        )
1544

1545
1546
1547
1548
        deepstack_input_embeds = _merge_multimodal_embeddings(
            inputs_embeds=deepstack_input_embeds,
            multimodal_embeddings=multimodal_embeddings_multiscale,
            is_multimodal=is_multimodal,
1549
1550
        )
        deepstack_input_embeds = deepstack_input_embeds.view(
1551
1552
            inputs_embeds.shape[0], self.deepstack_num_level, self.visual_dim
        )
1553
        deepstack_input_embeds = deepstack_input_embeds.permute(1, 0, 2)
1554

1555
1556
        return deepstack_input_embeds, multimodal_embeddings

1557
    def embed_input_ids(
1558
1559
        self,
        input_ids: torch.Tensor,
1560
        multimodal_embeddings: MultiModalEmbeddings | None = None,
1561
        *,
1562
        is_multimodal: torch.Tensor | None = None,
1563
        handle_oov_mm_token: bool = False,
1564
    ) -> torch.Tensor:
1565
        inputs_embeds = self._embed_text_input_ids(
1566
            input_ids,
1567
            self.language_model.embed_input_ids,
1568
1569
1570
1571
1572
1573
1574
1575
1576
            is_multimodal=is_multimodal,
            handle_oov_mm_token=handle_oov_mm_token,
        )

        if multimodal_embeddings is None or len(multimodal_embeddings) == 0:
            return inputs_embeds

        if is_multimodal is None:
            raise ValueError(
1577
                "`embed_input_ids` now requires `is_multimodal` arg, "
1578
                "please update your model runner according to "
1579
1580
                "https://github.com/vllm-project/vllm/pull/16229."
            )
1581
1582

        if self.use_deepstack:
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
            (
                deepstack_input_embeds,
                multimodal_embeddings,
            ) = self._compute_deepstack_embeds(
                inputs_embeds=inputs_embeds,
                multimodal_embeddings=multimodal_embeddings,
                is_multimodal=is_multimodal,
            )
        else:
            deepstack_input_embeds = None

        inputs_embeds = _merge_multimodal_embeddings(
            inputs_embeds=inputs_embeds,
            multimodal_embeddings=multimodal_embeddings,
            is_multimodal=is_multimodal,
        )

        if deepstack_input_embeds is not None:
1601
1602
1603
1604
1605
1606
1607
1608
            self._set_deepstack_input_embeds(deepstack_input_embeds)

        return inputs_embeds

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
1609
1610
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
1611
        **kwargs: object,
1612
    ) -> torch.Tensor | IntermediateTensors:
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
        """Run forward pass for Qwen3VL.

        Args:
            input_ids: Flattened (concatenated) input_ids corresponding to a
                batch.
            positions: Flattened (concatenated) position ids corresponding to a
                batch.
                **NOTE**: If mrope is enabled (default setting for Qwen3VL
                opensource models), the shape will be `(3, seq_len)`,
                otherwise it will be `(seq_len,).
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
            intermediate_tensors: Intermediate tensors from previous pipeline
                stages.
            inputs_embeds: Pre-computed input embeddings.
            **kwargs: Additional keyword arguments including:
                - pixel_values: Pixel values to be fed to a model.
                    `None` if no images are passed.
                - image_grid_thw: Tensor `(n_images, 3)` of image 3D grid in
                    LLM. `None` if no images are passed.
                - pixel_values_videos: Pixel values of videos to be fed to a
                    model. `None` if no videos are passed.
                - video_grid_thw: Tensor `(n_videos, 3)` of video 3D grid in
                    LLM. `None` if no videos are passed.
1635
1636
1637
1638
1639
        """

        if intermediate_tensors is not None:
            inputs_embeds = None

1640
1641
1642
1643
1644
        if (
            self.use_deepstack
            and inputs_embeds is not None
            and get_pp_group().is_first_rank
        ):
1645
            deepstack_input_embeds = self._get_deepstack_input_embeds(
1646
1647
                inputs_embeds.size(0)
            )
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
        else:
            deepstack_input_embeds = None

        hidden_states = self.language_model.model(
            input_ids=input_ids,
            positions=positions,
            intermediate_tensors=intermediate_tensors,
            inputs_embeds=inputs_embeds,
            # args for deepstack
            deepstack_input_embeds=deepstack_input_embeds,
        )

        if inputs_embeds is not None and get_pp_group().is_first_rank:
            self._clear_deepstack_input_embeds(inputs_embeds.size(0))

        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
1668
    ) -> torch.Tensor | None:
1669
        return self.language_model.compute_logits(hidden_states)
1670

1671
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
1672
1673
1674
1675
        skip_prefixes = []
        if self.visual is None:
            skip_prefixes.extend(["visual."])
        loader = AutoWeightsLoader(self, skip_prefixes=skip_prefixes)
1676
1677
1678
1679
1680
1681
1682
1683
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

    def get_mm_mapping(self) -> MultiModelKeys:
        """
        Get the module prefix in multimodal models
        """
        return MultiModelKeys.from_string_field(
            language_model="language_model",
1684
1685
            connector="visual.merger",
            tower_model="visual.",
1686
        )