qwen2_vl.py 52.7 KB
Newer Older
1
2
# SPDX-License-Identifier: Apache-2.0

3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
# Adapted from
# https://github.com/huggingface/transformers/blob/19e6e80e10118f855137b90740936c0b11ac397f/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py
# Copyright 2024 The Qwen team.
# Copyright 2023 The vLLM team.
# Copyright 2022 EleutherAI and 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 Qwen2-VL model compatible with HuggingFace weights."""
26
from collections.abc import Iterable, Mapping, Sequence
27
from functools import cached_property, partial
28
29
from typing import (Any, Callable, Literal, Optional, Set, Tuple, TypedDict,
                    Union)
30
31
32
33
34

import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange, repeat
35
36
37
from transformers import BatchFeature
from transformers.models.qwen2_vl import (Qwen2VLImageProcessor,
                                          Qwen2VLProcessor)
38
39
from transformers.models.qwen2_vl.configuration_qwen2_vl import (
    Qwen2VLConfig, Qwen2VLVisionConfig)
40
from transformers.models.qwen2_vl.image_processing_qwen2_vl import smart_resize
41

42
from vllm.config import VllmConfig
43
from vllm.distributed import parallel_state, tensor_model_parallel_all_gather
44
45
46
47
48
49
from vllm.distributed import utils as dist_utils
from vllm.logger import init_logger
from vllm.model_executor import SamplingMetadata
from vllm.model_executor.layers.activation import QuickGELU
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                               RowParallelLinear)
50
51
52
53
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.quantization.gptq import GPTQConfig
from vllm.model_executor.layers.quantization.gptq_marlin import (
    GPTQMarlinConfig)
Joe Runde's avatar
Joe Runde committed
54
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
55
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
56
from vllm.model_executor.models.module_mapping import MultiModelKeys
57
from vllm.multimodal import MULTIMODAL_REGISTRY
58
from vllm.multimodal.inputs import (ImageItem, ModalityData,
59
                                    MultiModalFieldConfig, MultiModalKwargs,
60
                                    VideoItem)
61
62
63
from vllm.multimodal.parse import (DictEmbeddingItems, ImageSize,
                                   ModalityDataItems, MultiModalDataItems,
                                   MultiModalDataParser)
64
from vllm.multimodal.processing import (BaseMultiModalProcessor,
65
66
                                        BaseProcessingInfo, PromptReplacement,
                                        PromptUpdate)
67
from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
68
from vllm.platforms import _Backend
69
from vllm.sequence import IntermediateTensors
70
from vllm.transformers_utils.config import uses_mrope
71
72
from vllm.transformers_utils.processor import (
    cached_image_processor_from_config)
73

74
75
from .interfaces import (MultiModalEmbeddings, SupportsLoRA,
                         SupportsMultiModal, SupportsPP)
76
from .utils import (AutoWeightsLoader, WeightsMapper,
77
78
                    init_vllm_registered_model, maybe_prefix,
                    merge_multimodal_embeddings)
79
from .vision import get_vit_attn_backend
80

81
82
logger = init_logger(__name__)

83
84
85
# For profile run
_MAX_FRAMES_PER_VIDEO = 16

86
87
88
# === Vision Inputs === #


89
90
class Qwen2VLImagePixelInputs(TypedDict):
    type: Literal["pixel_values"]
91
    pixel_values: torch.Tensor
92
    """Shape:
93
94
95
96
97
98
99
100
101
    `(num_patches, num_channels * patch_size * patch_size)`
    """

    image_grid_thw: torch.Tensor
    """Shape: `(num_images, 3)`
    This should be in `(grid_t, grid_h, grid_w)` format.
    """


102
103
class Qwen2VLImageEmbeddingInputs(TypedDict):
    type: Literal["image_embeds"]
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
    image_embeds: torch.Tensor
    """Supported types:
    - List[`torch.Tensor`]: A list of tensors holding all images' features.
        Each tensor holds an image's features.
    - `torch.Tensor`: A tensor holding all images' features
        (concatenation of all images' feature tensors).
    
    Tensor shape: `(num_image_features, hidden_size)`
    - `num_image_features` varies based on
        the number and resolution of the images.
    - `hidden_size` must match the hidden size of language model backbone.
    """

    image_grid_thw: torch.Tensor
    """Shape: `(num_images, 3)`
    This should be in `(grid_t, grid_h, grid_w)` format.
120
121
122
123
124
125
126
    """


Qwen2VLImageInputs = Union[Qwen2VLImagePixelInputs,
                           Qwen2VLImageEmbeddingInputs]


127
128
class Qwen2VLVideoPixelInputs(TypedDict):
    type: Literal["pixel_values_videos"]
129
    pixel_values_videos: torch.Tensor
130
131
    """Shape:
    `(num_patches,
132
133
134
135
136
      num_channels * temporal_patch_size * patch_size * patch_size)`
    """

    video_grid_thw: torch.Tensor
    """Shape: `(num_videos, 3)`
137

138
139
140
141
    This should be in `(grid_t, grid_h, grid_w)` format.
    """


142
143
144
145
146
147
148
class Qwen2VLVideoEmbeddingInputs(TypedDict):
    type: Literal["video_embeds"]
    video_embeds: torch.Tensor
    """Supported types:
    - List[`torch.Tensor`]: A list of tensors holding all videos' features.
        Each tensor holds an video's features.
    - `torch.Tensor`: A tensor holding all videos' features
149
        (concatenation of all videos' feature tensors).
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
    
    Tensor shape: `(num_image_features, hidden_size)`
    - `num_image_features` varies based on 
        the number and resolution of the videos.
    - `hidden_size` must match the hidden size of language model backbone.
    """

    video_grid_thw: torch.Tensor
    """Shape: `(num_videos, 3)`
    This should be in `(grid_t, grid_h, grid_w)` format.
    """


Qwen2VLVideoInputs = Union[Qwen2VLVideoPixelInputs,
                           Qwen2VLVideoEmbeddingInputs]

166
167
168
169
170
171
172
173
# === Vision Encoder === #


class Qwen2VisionMLP(nn.Module):

    def __init__(
        self,
        in_features: int,
174
        hidden_features: int,
175
        act_layer: type[nn.Module] = QuickGELU,
176
        quant_config: Optional[QuantizationConfig] = None,
177
        prefix: str = "",
178
179
180
181
    ):
        super().__init__()
        self.fc1 = ColumnParallelLinear(in_features,
                                        hidden_features,
182
183
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.fc1")
184
185
186
        self.act = act_layer()
        self.fc2 = RowParallelLinear(hidden_features,
                                     in_features,
187
188
                                     quant_config=quant_config,
                                     prefix=f"{prefix}.fc2")
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

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x_parallel, _ = self.fc1(x)
        x_parallel = self.act(x_parallel)
        x, _ = self.fc2(x_parallel)
        return x


def rotate_half(x: torch.Tensor, interleaved: bool = False) -> torch.Tensor:
    if not interleaved:
        x1, x2 = x.chunk(2, dim=-1)
        return torch.cat((-x2, x1), dim=-1)
    else:
        x1, x2 = x[..., ::2], x[..., 1::2]
        return rearrange(torch.stack((-x2, x1), dim=-1),
                         "... d two -> ... (d two)",
                         two=2)


def apply_rotary_emb_torch(x: torch.Tensor,
                           cos: torch.Tensor,
                           sin: torch.Tensor,
                           interleaved: bool = False) -> torch.Tensor:
    """
    x: (batch_size, seqlen, nheads, headdim)
    cos, sin: (seqlen, rotary_dim / 2) or (batch_size, seqlen, rotary_dim / 2)
    """
    ro_dim = cos.shape[-1] * 2
    assert ro_dim <= x.shape[-1]
    cos = repeat(
        cos,
        "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
    sin = repeat(
        sin,
        "... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
    return torch.cat(
        [
            x[..., :ro_dim] * cos +
            rotate_half(x[..., :ro_dim], interleaved) * sin, x[..., ro_dim:]
        ],
        dim=-1,
    )


def apply_rotary_pos_emb_vision(t: torch.Tensor,
燃's avatar
committed
234
235
                                freqs: torch.Tensor,
                                use_flash_attn=False) -> torch.Tensor:
236
237
238
    t_ = t.float()
    cos = freqs.cos()
    sin = freqs.sin()
燃's avatar
committed
239
240
241
242
    apply_rotary_emb = apply_rotary_emb_torch
    if use_flash_attn:
        from flash_attn.layers.rotary import apply_rotary_emb
    output = apply_rotary_emb(t_, cos, sin).type_as(t)
243
244
245
246
247
248
249
    return output


class Qwen2VisionAttention(nn.Module):

    def __init__(
        self,
250
251
252
        embed_dim: int,
        num_heads: int,
        projection_size: int,
253
        quant_config: Optional[QuantizationConfig] = None,
254
        prefix: str = "",
255
256
257
258
    ) -> None:
        super().__init__()
        # Per attention head and per partition values.
        world_size = parallel_state.get_tensor_model_parallel_world_size()
259
260
        self.tp_size = world_size
        self.tp_rank = parallel_state.get_tensor_model_parallel_rank()
261
262
263
264
265
266
267
        self.hidden_size_per_attention_head = dist_utils.divide(
            projection_size, num_heads)
        self.num_attention_heads_per_partition = dist_utils.divide(
            num_heads, world_size)

        self.qkv = ColumnParallelLinear(input_size=embed_dim,
                                        output_size=3 * projection_size,
268
269
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.qkv")
270
271
        self.proj = RowParallelLinear(input_size=projection_size,
                                      output_size=embed_dim,
272
273
                                      quant_config=quant_config,
                                      prefix=f"{prefix}.proj")
274
275

        # Detect attention implementation.
276
        self.attn_backend: _Backend = get_vit_attn_backend(support_fa=True)
277
278
279
280
281
        if self.attn_backend not in {
                _Backend.FLASH_ATTN, _Backend.TORCH_SDPA, _Backend.XFORMERS
        }:
            raise RuntimeError(
                f"Qwen2-VL does not support {self.attn_backend} backend now.")
282

283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
    def split_qkv(self, qkv: torch.Tensor) -> tuple[torch.Tensor, ...]:
        # [s, b, 3 * head * head_dim]
        seq_len, bs, _ = qkv.shape
        if self.tp_size > 1:
            qkv = tensor_model_parallel_all_gather(qkv)

        # [s, b, 3 * head * head_dim] -> 3 * [s, b, head * head_dim]
        q, k, v = qkv.chunk(3, dim=2)

        # 3 * [s, b, head * head_dim]
        if self.tp_size > 1:
            splitter = partial(dist_utils.split_tensor_along_last_dim,
                               num_partitions=self.tp_size)
            q = splitter(q)[self.tp_rank]
            k = splitter(k)[self.tp_rank]
            v = splitter(v)[self.tp_rank]

        # 3 * [s, b, head * head_dim] -> 3 * [s, b, head, head_dim]
        new_shape = (seq_len, bs, self.num_attention_heads_per_partition,
                     self.hidden_size_per_attention_head)
        q, k, v = (x.view(*new_shape) for x in (q, k, v))
        return q, k, v

306
    def forward(
307
308
309
310
311
312
            self,
            x: torch.Tensor,
            cu_seqlens: torch.Tensor,
            rotary_pos_emb: torch.Tensor,
            max_seqlen: Optional[int] = None,  # Only used for Flash Attention
            seqlens: Optional[list[int]] = None,  # Only used for xFormers
313
314
    ) -> torch.Tensor:

315
316
        # [s, b, c] --> [s, b, 3 * head * head_dim]
        x, _ = self.qkv(x)
317

318
319
        # [s, b, 3 * head * head_dim] -> 3 * [s, b, head, head_dim]
        q, k, v = self.split_qkv(x)
320
321
        batch_size = q.shape[1]

322
323
        q, k, v = (rearrange(x, "s b ... -> b s ...").contiguous()
                   for x in (q, k, v))
324
325
326
327
        if rotary_pos_emb is not None:
            q = apply_rotary_pos_emb_vision(q, rotary_pos_emb)
            k = apply_rotary_pos_emb_vision(k, rotary_pos_emb)

328
        if self.attn_backend == _Backend.FLASH_ATTN:
329
330
331
332
            # from vllm_flash_attn.flash_attn_interface import (
            #   flash_attn_varlen_func)
            from flash_attn import flash_attn_varlen_func

333
            q, k, v = (rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v])
334
335
336
337
338
339
340
341
342
343
344
345
346
347

            output = flash_attn_varlen_func(q,
                                            k,
                                            v,
                                            cu_seqlens_q=cu_seqlens,
                                            cu_seqlens_k=cu_seqlens,
                                            max_seqlen_q=max_seqlen,
                                            max_seqlen_k=max_seqlen,
                                            dropout_p=0,
                                            causal=False)

            context_layer = rearrange(output,
                                      "(b s) ... -> b s ...",
                                      b=batch_size)
348
        elif self.attn_backend == _Backend.TORCH_SDPA:
燃's avatar
committed
349
350
            # Execute attention entry by entry for speed & less VRAM.
            outputs = []
351
            for i in range(1, len(cu_seqlens)):
燃's avatar
committed
352
353
354
355
356
357
358
359
360
361
362
363
364
365
                start_idx = cu_seqlens[i - 1]
                end_idx = cu_seqlens[i]
                q_i = q[:, start_idx:end_idx]
                k_i = k[:, start_idx:end_idx]
                v_i = v[:, start_idx:end_idx]
                q_i, k_i, v_i = (rearrange(x, "b s h d -> b h s d")
                                 for x in [q_i, k_i, v_i])
                output_i = F.scaled_dot_product_attention(q_i,
                                                          k_i,
                                                          v_i,
                                                          dropout_p=0.0)
                output_i = rearrange(output_i, "b h s d -> b s h d ")
                outputs.append(output_i)
            context_layer = torch.cat(outputs, dim=1)
366
        elif self.attn_backend == _Backend.XFORMERS:
367
368
369
370
            from xformers import ops as xops
            from xformers.ops.fmha.attn_bias import BlockDiagonalMask

            attn_bias = BlockDiagonalMask.from_seqlens(q_seqlen=seqlens,
371
372
                                                       kv_seqlen=None,
                                                       device=q.device)
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389

            context_layer = xops.memory_efficient_attention_forward(
                q, k, v, attn_bias=attn_bias, p=0, scale=None)
        context_layer = rearrange(context_layer,
                                  "b s h d -> s b (h d)").contiguous()

        output, _ = self.proj(context_layer)
        return output


class Qwen2VisionBlock(nn.Module):

    def __init__(
        self,
        dim: int,
        num_heads: int,
        mlp_ratio: float,
390
        act_layer: type[nn.Module] = QuickGELU,
391
        norm_layer: Optional[Callable[[int], nn.Module]] = None,
392
        quant_config: Optional[QuantizationConfig] = None,
393
        prefix: str = "",
394
395
396
397
398
399
400
401
402
403
404
    ) -> 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)
        mlp_hidden_dim = int(dim * mlp_ratio)

        self.attn = Qwen2VisionAttention(embed_dim=dim,
                                         num_heads=num_heads,
                                         projection_size=dim,
405
406
                                         quant_config=quant_config,
                                         prefix=f"{prefix}.attn")
407
408
409
        self.mlp = Qwen2VisionMLP(dim,
                                  mlp_hidden_dim,
                                  act_layer=act_layer,
410
411
                                  quant_config=quant_config,
                                  prefix=f"{prefix}.mlp")
412

413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
    def forward(
            self,
            x: torch.Tensor,
            cu_seqlens: torch.Tensor,
            rotary_pos_emb: torch.Tensor,
            max_seqlen: Optional[int] = None,  # Only used for Flash Attention
            seqlens: Optional[list[int]] = None,  # Only used for xFormers
    ) -> torch.Tensor:
        x = x + self.attn(
            self.norm1(x),
            cu_seqlens=cu_seqlens,
            rotary_pos_emb=rotary_pos_emb,
            max_seqlen=max_seqlen,
            seqlens=seqlens,
        )

429
430
431
432
433
434
435
436
437
438
        x = x + self.mlp(self.norm2(x))
        return x


class Qwen2VisionPatchEmbed(nn.Module):

    def __init__(
        self,
        patch_size: int = 14,
        temporal_patch_size: int = 2,
439
        in_channels: int = 3,
440
441
442
443
444
445
446
        embed_dim: int = 1152,
    ) -> None:
        super().__init__()
        self.patch_size = patch_size
        self.temporal_patch_size = temporal_patch_size
        self.embed_dim = embed_dim

447
448
        kernel_size = (temporal_patch_size, patch_size, patch_size)
        self.proj = nn.Conv3d(in_channels,
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
                              embed_dim,
                              kernel_size=kernel_size,
                              stride=kernel_size,
                              bias=False)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        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.embed_dim)
        return x


class Qwen2VisionPatchMerger(nn.Module):

    def __init__(
        self,
        d_model: int,
        context_dim: int,
468
        norm_layer: Optional[Callable[[int], nn.Module]] = None,
469
470
        spatial_merge_size: int = 2,
        quant_config: Optional[QuantizationConfig] = None,
471
        prefix: str = "",
472
473
474
475
476
477
478
479
480
481
    ) -> None:
        super().__init__()
        self.hidden_size = context_dim * (spatial_merge_size**2)
        if norm_layer is None:
            norm_layer = partial(nn.LayerNorm, eps=1e-6)
        self.ln_q = norm_layer(context_dim)
        self.mlp = nn.ModuleList([
            ColumnParallelLinear(self.hidden_size,
                                 self.hidden_size,
                                 bias=True,
482
483
                                 quant_config=quant_config,
                                 prefix=f"{prefix}.mlp.0"),
484
485
486
487
            nn.GELU(),
            RowParallelLinear(self.hidden_size,
                              d_model,
                              bias=True,
488
489
                              quant_config=quant_config,
                              prefix=f"{prefix}.mlp.2"),
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
        ])

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.ln_q(x)
        x = x.view(-1, self.hidden_size)

        mlp_fc1, mlp_act, mlp_fc2 = self.mlp
        x_parallel, _ = mlp_fc1(x)
        x_parallel = mlp_act(x_parallel)
        out, _ = mlp_fc2(x_parallel)
        return out


class Qwen2VisionRotaryEmbedding(nn.Module):

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

    def update_freqs_cache(self, seqlen: int) -> None:
        if seqlen > self._seq_len_cached:
            seqlen *= 2
            self._seq_len_cached = seqlen
            self.inv_freq = 1.0 / (self.theta**(torch.arange(
                0, self.dim, 2, dtype=torch.float, device=self.inv_freq.device)
                                                / self.dim))
            seq = torch.arange(seqlen,
                               device=self.inv_freq.device,
                               dtype=self.inv_freq.dtype)
            freqs = torch.outer(seq, self.inv_freq)
            self._freqs_cached = freqs

    def forward(self, seqlen: int) -> torch.Tensor:
        self.update_freqs_cache(seqlen)
        return self._freqs_cached[:seqlen]


class Qwen2VisionTransformer(nn.Module):

    def __init__(
        self,
        vision_config: Qwen2VLVisionConfig,
        norm_eps: float = 1e-6,
        quant_config: Optional[QuantizationConfig] = None,
540
        prefix: str = "",
541
542
543
    ) -> None:
        super().__init__()

544
545
546
547
548
549
550
551
552
        patch_size = vision_config.patch_size
        temporal_patch_size = vision_config.temporal_patch_size
        spatial_merge_size = vision_config.spatial_merge_size
        in_channels = vision_config.in_channels
        hidden_size = vision_config.hidden_size
        embed_dim = vision_config.embed_dim
        depth = vision_config.depth
        num_heads = vision_config.num_heads
        mlp_ratio = vision_config.mlp_ratio
553
554

        self.spatial_merge_size = spatial_merge_size
555
556
        self.num_heads = num_heads
        self.embed_dim = embed_dim
557
558
559
560

        self.patch_embed = Qwen2VisionPatchEmbed(
            patch_size=patch_size,
            temporal_patch_size=temporal_patch_size,
561
            in_channels=in_channels,
562
563
564
565
566
567
568
569
            embed_dim=embed_dim,
        )

        norm_layer = partial(nn.LayerNorm, eps=norm_eps)
        head_dim = embed_dim // num_heads
        self.rotary_pos_emb = Qwen2VisionRotaryEmbedding(head_dim // 2)

        self.blocks = nn.ModuleList([
570
571
572
573
574
575
576
            Qwen2VisionBlock(dim=embed_dim,
                             num_heads=num_heads,
                             mlp_ratio=mlp_ratio,
                             norm_layer=norm_layer,
                             quant_config=quant_config,
                             prefix=f"{prefix}.blocks.{layer_idx}")
            for layer_idx in range(depth)
577
578
579
580
581
582
        ])
        self.merger = Qwen2VisionPatchMerger(
            d_model=hidden_size,
            context_dim=embed_dim,
            norm_layer=norm_layer,
            quant_config=quant_config,
583
            prefix=f"{prefix}.merger",
584
        )
585
        self.attn_backend: _Backend = get_vit_attn_backend(support_fa=True)
586
587
588

    @property
    def dtype(self) -> torch.dtype:
589
        return self.patch_embed.proj.weight.dtype
590
591
592

    @property
    def device(self) -> torch.device:
593
        return self.patch_embed.proj.weight.device
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

    def rot_pos_emb(self, grid_thw: torch.Tensor) -> torch.Tensor:
        pos_ids = []
        for t, h, w in grid_thw:
            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
            hpos_ids = hpos_ids.reshape(
                h // self.spatial_merge_size,
                self.spatial_merge_size,
                w // self.spatial_merge_size,
                self.spatial_merge_size,
            ).permute(0, 2, 1, 3).flatten()
            wpos_ids = wpos_ids.reshape(
                h // self.spatial_merge_size,
                self.spatial_merge_size,
                w // self.spatial_merge_size,
                self.spatial_merge_size,
            ).permute(0, 2, 1, 3).flatten()
            pos_ids.append(
                torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
        pos_ids = torch.cat(pos_ids, dim=0)
        max_grid_size = grid_thw[:, 1:].max()
        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
        return rotary_pos_emb

620
621
622
623
624
625
626
627
628
629
    def compute_attn_mask_seqlen(
            self, cu_seqlens: torch.Tensor
    ) -> tuple[Optional[int], Optional[list[int]]]:
        max_seqlen, seqlens = None, None
        if self.attn_backend == _Backend.FLASH_ATTN:
            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
        elif self.attn_backend == _Backend.XFORMERS:
            seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
        return max_seqlen, seqlens

630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
    def forward(
        self,
        x: torch.Tensor,
        grid_thw: torch.Tensor,
    ) -> torch.Tensor:
        # patchify
        x = x.to(device=self.device, dtype=self.dtype)
        x = self.patch_embed(x)

        # compute position embedding
        rotary_pos_emb = self.rot_pos_emb(grid_thw)

        # compute cu_seqlens
        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2],
                                             grid_thw[:, 0]).cumsum(
                                                 dim=0, dtype=torch.int32)
        cu_seqlens = F.pad(cu_seqlens, (1, 0), "constant", 0)

        # transformers
        x = x.unsqueeze(1)
650

651
652
        # pre-compute seqlens for attn mask to reduce cuMemcpy operations
        max_seqlen, seqlens = self.compute_attn_mask_seqlen(cu_seqlens)
653
        for blk in self.blocks:
654
655
656
657
658
659
660
            x = blk(
                x,
                cu_seqlens=cu_seqlens,
                rotary_pos_emb=rotary_pos_emb,
                max_seqlen=max_seqlen,
                seqlens=seqlens,
            )
661
662
663

        # adapter
        x = self.merger(x)
664

665
666
        return x

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
    def load_weights(self, weights: Iterable[Tuple[str,
                                                   torch.Tensor]]) -> Set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]
        params_dict = dict(self.named_parameters(remove_duplicate=False))
        loaded_params: Set[str] = set()

        for name, loaded_weight in weights:
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                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]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params

696

697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
def _qwen2vl_field_config(hf_inputs: Mapping[str, torch.Tensor]):
    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(
            "image", image_grid_sizes),
        image_embeds=MultiModalFieldConfig.flat_from_sizes(
            "image", image_grid_sizes),
        image_grid_thw=MultiModalFieldConfig.batched("image"),
        pixel_values_videos=MultiModalFieldConfig.flat_from_sizes(
            "video", video_grid_sizes),
        video_embeds=MultiModalFieldConfig.flat_from_sizes(
            "video", video_grid_sizes),
        video_grid_thw=MultiModalFieldConfig.batched("video"),
    )
716

717

Roger Wang's avatar
Roger Wang committed
718
class Qwen2VLMultiModalDataParser(MultiModalDataParser):
719
720
721
722
723
724

    def _parse_image_data(
        self,
        data: Union[dict[str, torch.Tensor], ModalityData[ImageItem]],
    ) -> ModalityDataItems[Any, Any]:
        if isinstance(data, dict):
725
726
727
728
            return DictEmbeddingItems(
                data,
                modality="image",
                required_fields={"image_embeds", "image_grid_thw"},
729
                fields_factory=_qwen2vl_field_config,
730
            )
731
732
733
734

        return super()._parse_image_data(data)

    def _parse_video_data(
735
        self,
736
737
738
        data: Union[dict[str, torch.Tensor], ModalityData[VideoItem]],
    ) -> ModalityDataItems[Any, Any]:
        if isinstance(data, dict):
739
740
741
742
            return DictEmbeddingItems(
                data,
                modality="video",
                required_fields={"video_embeds", "video_grid_thw"},
743
                fields_factory=_qwen2vl_field_config,
744
            )
745
746
747
748

        return super()._parse_video_data(data)


749
class Qwen2VLProcessingInfo(BaseProcessingInfo):
750

751
    def get_hf_config(self):
752
753
        return self.ctx.get_hf_config(Qwen2VLConfig)

754
    def get_hf_processor(
755
756
757
758
        self,
        *,
        min_pixels: Optional[int] = None,
        max_pixels: Optional[int] = None,
759
760
        size: Optional[dict[str, int]] = None,
        **kwargs: object,
761
    ) -> Qwen2VLProcessor:
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
        return self.ctx.get_hf_processor(
            Qwen2VLProcessor,
            image_processor=self.get_image_processor(min_pixels=min_pixels,
                                                     max_pixels=max_pixels,
                                                     size=size),
            **kwargs,
        )

    def _get_image_processor_kwargs(
        self,
        *,
        min_pixels: Optional[int] = None,
        max_pixels: Optional[int] = None,
        size: Optional[dict[str, int]] = None,
        **kwargs: object,
    ):
        if self.ctx.model_config.mm_processor_kwargs:
            kwargs.update(self.ctx.model_config.mm_processor_kwargs)

        if min_pixels is not None:
            kwargs["min_pixels"] = min_pixels

            if size is None:
                size = {"shortest_edge": min_pixels}
            else:
                size["shortest_edge"] = min_pixels

        if max_pixels is not None:
            kwargs["max_pixels"] = max_pixels

            if size is None:
                size = {"longest_edge": max_pixels}
            else:
                size["longest_edge"] = max_pixels

        if size is not None:
            kwargs["size"] = size

        return kwargs
801

802
    def get_image_processor(
803
804
805
806
        self,
        *,
        min_pixels: Optional[int] = None,
        max_pixels: Optional[int] = None,
807
808
        size: Optional[dict[str, int]] = None,
        **kwargs: object,
809
    ) -> Qwen2VLImageProcessor:
810
811
812
813
814
815
816
        return cached_image_processor_from_config(
            self.ctx.model_config,
            **self._get_image_processor_kwargs(min_pixels=min_pixels,
                                               max_pixels=max_pixels,
                                               size=size,
                                               **kwargs),
        )
817

818
819
820
    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
        return {"image": None, "video": None}

821
822
823
824
825
826
827
    def _get_vision_info(
        self,
        *,
        image_width: int,
        image_height: int,
        num_frames: int = 1,
        do_resize: bool = True,
828
        image_processor: Optional[Qwen2VLImageProcessor],
829
    ) -> tuple[ImageSize, int]:
830
831
832
833
        if image_processor is None:
            image_processor = self.get_image_processor()

        hf_config = self.get_hf_config()
834
        vision_config = hf_config.vision_config
835
836
837
        patch_size = vision_config.patch_size
        merge_size = vision_config.spatial_merge_size
        temporal_patch_size = vision_config.temporal_patch_size
838

839
840
841
842
843
844
845
846
847
848
849
850
851
852
        if do_resize:
            resized_height, resized_width = smart_resize(
                height=image_height,
                width=image_width,
                factor=patch_size * merge_size,
                min_pixels=image_processor.min_pixels,
                max_pixels=image_processor.max_pixels,
            )
            preprocessed_size = ImageSize(width=resized_width,
                                          height=resized_height)
        else:
            preprocessed_size = ImageSize(width=image_width,
                                          height=image_height)

853
854
855
856
857
        # NOTE: Frames are padded to be divisible by `temporal_patch_size`
        # https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/qwen2_vl/image_processing_qwen2_vl.py#L294
        padded_num_frames = num_frames + num_frames % temporal_patch_size

        grid_t = max(padded_num_frames // temporal_patch_size, 1)
858
859
860
861
862
863
864
865
        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

866
    def get_num_image_tokens(
867
868
869
870
        self,
        *,
        image_width: int,
        image_height: int,
871
        image_processor: Optional[Qwen2VLImageProcessor],
872
873
874
875
    ) -> int:
        _, num_image_tokens = self._get_vision_info(
            image_width=image_width,
            image_height=image_height,
876
            image_processor=image_processor,
877
878
879
        )
        return num_image_tokens

880
    def get_num_video_tokens(
881
882
883
884
885
        self,
        *,
        image_width: int,
        image_height: int,
        num_frames: int,
886
        image_processor: Optional[Qwen2VLImageProcessor],
887
888
889
890
891
    ) -> int:
        _, num_video_tokens = self._get_vision_info(
            image_width=image_width,
            image_height=image_height,
            num_frames=num_frames,
892
            image_processor=image_processor,
893
894
895
        )
        return num_video_tokens

896
    def get_image_size_with_most_features(self) -> ImageSize:
897
898
899
        max_image_size, _ = self._get_vision_info(
            image_width=9999999,
            image_height=9999999,
900
            image_processor=None,
901
902
903
        )
        return max_image_size

904
905
    def get_max_image_tokens(self) -> int:
        target_width, target_height = self.get_image_size_with_most_features()
906

907
        return self.get_num_image_tokens(
908
909
            image_width=target_width,
            image_height=target_height,
910
            image_processor=None,
911
        )
912
913

    def _get_max_video_frames(self, max_tokens: int) -> int:
914
        target_width, target_height = self.get_image_size_with_most_features()
915

916
917
918
919
        num_frames = 0

        while True:
            next_num_frames = num_frames + 1
920
            next_max_tokens = self.get_num_video_tokens(
921
922
923
                image_width=target_width,
                image_height=target_height,
                num_frames=next_num_frames,
924
                image_processor=None,
925
            )
926

927
            if next_max_tokens > max_tokens:
928
929
930
931
932
933
                break

            num_frames = next_num_frames

        return num_frames

934
935
936
937
938
939
940
    def get_num_frames_with_most_features(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> int:
        max_images = mm_counts.get("image", 0)
        max_videos = mm_counts.get("video", 0)
941

942
        max_image_tokens = self.get_max_image_tokens() * max_images
943
944
        max_total_frames = self._get_max_video_frames(seq_len -
                                                      max_image_tokens)
945
946
        max_frames_per_video = min(max_total_frames // max(max_videos, 1),
                                   _MAX_FRAMES_PER_VIDEO)
947

948
        return max(max_frames_per_video, 1)
949

950
951
952
953
954
    def get_max_video_tokens(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> int:
955
        target_width, target_height = self.get_image_size_with_most_features()
956

957
        return self.get_num_video_tokens(
958
959
            image_width=target_width,
            image_height=target_height,
960
961
            num_frames=self.get_num_frames_with_most_features(
                seq_len, mm_counts),
962
            image_processor=None,
963
964
        )

965
966
967

class Qwen2VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen2VLProcessingInfo]):

968
969
970
971
972
973
974
975
    def get_dummy_processor_inputs(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> ProcessorInputs:
        num_images = mm_counts.get("image", 0)
        num_videos = mm_counts.get("video", 0)

976
        hf_processor = self.info.get_hf_processor()
977
978
        image_token: str = hf_processor.image_token
        video_token: str = hf_processor.video_token
979
980
981
982

        target_width, target_height = \
            self.info.get_image_size_with_most_features()
        target_num_frames = \
983
            self.info.get_num_frames_with_most_features(seq_len, mm_counts)
984
985
986
987
988
989
990
991
992
993

        mm_data = {
            "image":
            self._get_dummy_images(width=target_width,
                                   height=target_height,
                                   num_images=num_images),
            "video":
            self._get_dummy_videos(
                width=target_width,
                height=target_height,
994
                num_frames=target_num_frames,
995
996
                num_videos=num_videos,
            )
997
998
        }

999
1000
1001
1002
        return ProcessorInputs(
            prompt_text=image_token * num_images + video_token * num_videos,
            mm_data=mm_data,
        )
1003

1004

1005
1006
class Qwen2VLMultiModalProcessor(BaseMultiModalProcessor[Qwen2VLProcessingInfo]
                                 ):
1007

1008
    def _get_data_parser(self) -> MultiModalDataParser:
Roger Wang's avatar
Roger Wang committed
1009
        return Qwen2VLMultiModalDataParser()
1010

1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
    def _call_hf_processor(
        self,
        prompt: str,
        mm_data: Mapping[str, object],
        mm_kwargs: Mapping[str, object],
    ) -> BatchFeature:
        return self.info.ctx.call_hf_processor(
            self.info.get_hf_processor(**mm_kwargs),
            dict(text=prompt, **mm_data),
            self.info._get_image_processor_kwargs(**mm_kwargs),
        )

1023
    def _get_prompt_updates(
1024
1025
        self,
        mm_items: MultiModalDataItems,
1026
        hf_processor_mm_kwargs: Mapping[str, Any],
1027
        out_mm_kwargs: MultiModalKwargs,
1028
    ) -> Sequence[PromptUpdate]:
1029
1030
1031
        hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
        image_processor = self.info.get_image_processor(
            **hf_processor_mm_kwargs)
1032
1033
        tokenizer = self.info.get_tokenizer()
        vocab = tokenizer.get_vocab()
1034
1035

        placeholder = {
1036
1037
            "image": vocab[hf_processor.image_token],
            "video": vocab[hf_processor.video_token],
1038
        }
1039

1040
1041
1042
        merge_length = image_processor.merge_size**2

        def get_replacement_qwen2vl(item_idx: int, modality: str):
1043
1044
1045
            grid_thw = out_mm_kwargs[f"{modality}_grid_thw"][item_idx]
            assert isinstance(grid_thw, torch.Tensor)

1046
1047
            num_tokens = int(grid_thw.prod()) // merge_length
            return [placeholder[modality]] * num_tokens
1048
1049
1050
1051

        return [
            PromptReplacement(
                modality=modality,
1052
                target=[placeholder[modality]],
1053
1054
1055
1056
                replacement=partial(get_replacement_qwen2vl,
                                    modality=modality),
            ) for modality in ("image", "video")
        ]
1057

1058
1059
1060
1061
1062
    def _get_mm_fields_config(
        self,
        hf_inputs: BatchFeature,
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
1063
        return _qwen2vl_field_config(hf_inputs)
1064

1065

1066
1067
1068
@MULTIMODAL_REGISTRY.register_processor(Qwen2VLMultiModalProcessor,
                                        info=Qwen2VLProcessingInfo,
                                        dummy_inputs=Qwen2VLDummyInputsBuilder)
1069
class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
                                      SupportsLoRA, SupportsPP):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

1083
1084
1085
1086
1087
    # To ensure correct weight loading and mapping.
    hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={
        "lm_head.": "language_model.lm_head.",
        "model.": "language_model.model.",
    })
1088

1089
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
1090
        super().__init__()
1091
        config: Qwen2VLConfig = vllm_config.model_config.hf_config
1092
1093
        quant_config = vllm_config.quant_config
        multimodal_config = vllm_config.model_config.multimodal_config
1094
1095
1096
1097
1098
1099
1100

        self.config = config
        self.multimodal_config = multimodal_config

        self.visual = Qwen2VisionTransformer(
            config.vision_config,
            norm_eps=getattr(config, "rms_norm_eps", 1e-6),
1101
            quant_config=self._maybe_ignore_quant_config(quant_config),
1102
            prefix=maybe_prefix(prefix, "visual"),
1103
1104
        )

1105
1106
1107
1108
1109
        self.language_model = init_vllm_registered_model(
            vllm_config=vllm_config,
            prefix=maybe_prefix(prefix, "language_model"),
            architectures=["Qwen2ForCausalLM"],
        )
1110

1111
1112
        self.make_empty_intermediate_tensors = (
            self.language_model.make_empty_intermediate_tensors)
1113

1114
1115
1116
1117
    @cached_property
    def sampler(self):
        if hasattr(self.language_model, "sampler"):
            return self.language_model.sampler
1118

1119
        return get_sampler()
1120

1121
1122
1123
1124
1125
1126
1127
1128
    def _maybe_ignore_quant_config(self, quant_config: QuantizationConfig):
        # GPTQ configs do not have a list of ignored modules, however AutoGPTQ
        # seems to avoid vision encoder sections for some models.
        # See: https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct-GPTQ-Int4
        if isinstance(quant_config, (GPTQConfig, GPTQMarlinConfig)):
            return None
        return quant_config

1129
    def _validate_and_reshape_mm_tensor(self, mm_input: object,
1130
1131
1132
1133
1134
1135
1136
1137
1138
                                        name: str) -> torch.Tensor:
        if not isinstance(mm_input, (torch.Tensor, list)):
            raise ValueError(f"Incorrect type of {name}. "
                             f"Got type: {type(mm_input)}")
        if isinstance(mm_input, torch.Tensor):
            if mm_input.ndim == 2:
                return mm_input
            if mm_input.ndim != 3:
                raise ValueError(f"{name} should be 2D or batched 3D tensor. "
1139
1140
                                 f"Got ndim: {mm_input.ndim} "
                                 f"(shape={mm_input.shape})")
1141
1142
1143
1144
1145
1146
1147
            return torch.concat(list(mm_input))
        else:
            return torch.concat(mm_input)

    def _parse_and_validate_image_input(
            self, **kwargs: object) -> Optional[Qwen2VLImageInputs]:
        pixel_values = kwargs.pop("pixel_values", None)
1148
        image_embeds = kwargs.pop("image_embeds", None)
1149
1150
        image_grid_thw = kwargs.pop("image_grid_thw", None)

1151
        if pixel_values is None and image_embeds is None:
1152
1153
            return None

1154
1155
1156
1157
1158
        if pixel_values is not None:
            pixel_values = self._validate_and_reshape_mm_tensor(
                pixel_values, "image pixel values")
            image_grid_thw = self._validate_and_reshape_mm_tensor(
                image_grid_thw, "image grid_thw")
1159

1160
1161
1162
            if not isinstance(pixel_values, (torch.Tensor, list)):
                raise ValueError("Incorrect type of image pixel values. "
                                 f"Got type: {type(pixel_values)}")
1163

1164
            return Qwen2VLImagePixelInputs(type="pixel_values",
1165
                                           pixel_values=pixel_values,
1166
1167
1168
                                           image_grid_thw=image_grid_thw)

        if image_embeds is not None:
1169
1170
            image_embeds = self._validate_and_reshape_mm_tensor(
                image_embeds, "image embeds")
1171
1172
            image_grid_thw = self._validate_and_reshape_mm_tensor(
                image_grid_thw, "image grid_thw")
1173

1174
1175
1176
1177
            if not isinstance(image_embeds, torch.Tensor):
                raise ValueError("Incorrect type of image embeddings. "
                                 f"Got type: {type(image_embeds)}")
            return Qwen2VLImageEmbeddingInputs(type="image_embeds",
1178
1179
                                               image_embeds=image_embeds,
                                               image_grid_thw=image_grid_thw)
1180
1181
1182
1183

    def _parse_and_validate_video_input(
            self, **kwargs: object) -> Optional[Qwen2VLVideoInputs]:
        pixel_values_videos = kwargs.pop("pixel_values_videos", None)
1184
        video_embeds = kwargs.pop("video_embeds", None)
1185
1186
        video_grid_thw = kwargs.pop("video_grid_thw", None)

1187
        if pixel_values_videos is None and video_embeds is None:
1188
1189
            return None

1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
        if pixel_values_videos is not None:
            pixel_values_videos = self._validate_and_reshape_mm_tensor(
                pixel_values_videos, "video pixel values")
            video_grid_thw = self._validate_and_reshape_mm_tensor(
                video_grid_thw, "video grid_thw")

            return Qwen2VLVideoPixelInputs(
                type="pixel_values_videos",
                pixel_values_videos=pixel_values_videos,
                video_grid_thw=video_grid_thw,
            )

        if video_embeds is not None:
            video_embeds = self._validate_and_reshape_mm_tensor(
                video_embeds, "video embeds")
            video_grid_thw = self._validate_and_reshape_mm_tensor(
                video_grid_thw, "video grid_thw")

            if not isinstance(video_embeds, torch.Tensor):
                raise ValueError("Incorrect type of video embeddings. "
                                 f"Got type: {type(video_embeds)}")
            return Qwen2VLVideoEmbeddingInputs(type="video_embeds",
                                               video_embeds=video_embeds,
                                               video_grid_thw=video_grid_thw)
1214

1215
1216
1217
1218
1219
1220
    def _process_image_input(
            self, image_input: Qwen2VLImageInputs) -> tuple[torch.Tensor, ...]:

        grid_thw = image_input["image_grid_thw"]
        assert grid_thw.ndim == 2

1221
        if image_input["type"] == "image_embeds":
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
            image_embeds = image_input["image_embeds"].type(self.visual.dtype)
        else:
            pixel_values = image_input["pixel_values"].type(self.visual.dtype)
            image_embeds = self.visual(pixel_values, grid_thw=grid_thw)

        # Split concatenated embeddings for each image item.
        merge_size = self.visual.spatial_merge_size
        sizes = grid_thw.prod(-1) // merge_size // merge_size

        return image_embeds.split(sizes.tolist())

    def _process_video_input(
            self, video_input: Qwen2VLVideoInputs) -> tuple[torch.Tensor, ...]:
1235

1236
1237
        grid_thw = video_input["video_grid_thw"]
        assert grid_thw.ndim == 2
1238

1239
        if video_input["type"] == "video_embeds":
1240
1241
1242
1243
1244
            video_embeds = video_input["video_embeds"].type(self.visual.dtype)
        else:
            pixel_values_videos = video_input["pixel_values_videos"].type(
                self.visual.dtype)
            video_embeds = self.visual(pixel_values_videos, grid_thw=grid_thw)
1245

1246
1247
1248
        # Split concatenated embeddings for each video item.
        merge_size = self.visual.spatial_merge_size
        sizes = grid_thw.prod(-1) // merge_size // merge_size
1249

1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
        return video_embeds.split(sizes.tolist())

    def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict:
        modalities = {}

        # Preserve the order of modalities if there are multiple of them
        # from the order of kwargs.
        for input_key in kwargs:
            if input_key in ("pixel_values",
                             "image_embeds") and "images" not in modalities:
                modalities["images"] = self._parse_and_validate_image_input(
                    **kwargs)
            if input_key in ("pixel_values_videos",
                             "video_embeds") and "videos" not in modalities:
                modalities["videos"] = self._parse_and_validate_video_input(
                    **kwargs)

        return modalities
1268

1269
1270
1271
    def get_language_model(self) -> torch.nn.Module:
        return self.language_model

1272
    def get_multimodal_embeddings(
1273
            self, **kwargs: object) -> Optional[MultiModalEmbeddings]:
1274

1275
1276
        modalities = self._parse_and_validate_multimodal_inputs(**kwargs)
        if not modalities:
1277
1278
            return None

1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
        # 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 modalities:
            if modality == "images":
                image_input = modalities["images"]
                vision_embeddings = self._process_image_input(image_input)
                multimodal_embeddings += vision_embeddings
            if modality == "videos":
                video_input = modalities["videos"]
                video_embeddings = self._process_video_input(video_input)
                multimodal_embeddings += video_embeddings
1294
1295
1296
1297
1298
1299

        return multimodal_embeddings

    def get_input_embeddings(
        self,
        input_ids: torch.Tensor,
1300
        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
1301
    ) -> torch.Tensor:
1302
        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
1303
        if multimodal_embeddings is not None:
1304
1305
1306
            inputs_embeds = merge_multimodal_embeddings(
                input_ids, inputs_embeds, multimodal_embeddings,
                [self.config.image_token_id, self.config.video_token_id])
1307
1308
        return inputs_embeds

1309
1310
1311
    def get_input_embeddings_v0(
        self,
        input_ids: torch.Tensor,
1312
1313
        image_input: Optional[Qwen2VLImagePixelInputs] = None,
        video_input: Optional[Qwen2VLVideoPixelInputs] = None,
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
    ) -> torch.Tensor:
        inputs_embeds = self.get_input_embeddings(input_ids)
        if image_input is not None:
            image_embeds = self._process_image_input(image_input)
            inputs_embeds = merge_multimodal_embeddings(
                input_ids,
                inputs_embeds,
                image_embeds,
                placeholder_token_id=self.config.image_token_id,
            )

        if video_input is not None:
            video_embeds = self._process_video_input(video_input)
            inputs_embeds = merge_multimodal_embeddings(
                input_ids,
                inputs_embeds,
                video_embeds,
                placeholder_token_id=self.config.video_token_id,
            )
        return inputs_embeds

1335
1336
1337
1338
1339
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
1340
        inputs_embeds: Optional[torch.Tensor] = None,
1341
        **kwargs: object,
1342
    ) -> Union[torch.Tensor, IntermediateTensors]:
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
        """Run forward pass for Qwen2-VL.

        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 Qwen2-VL
                opensource models), the shape will be `(3, seq_len)`,
                otherwise it will be `(seq_len,).
            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.
        """
1362

1363
        if intermediate_tensors is not None:
1364
            inputs_embeds = None
1365

1366
1367
1368
        # NOTE: In v1, inputs_embeds is always generated at model runner from
        # `get_multimodal_embeddings` and `get_input_embeddings`, this
        # condition is only for v0 compatibility.
1369
        elif inputs_embeds is None:
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
            image_input = self._parse_and_validate_image_input(**kwargs)
            video_input = self._parse_and_validate_video_input(**kwargs)

            if image_input is None and video_input is None:
                inputs_embeds = None
            else:
                if uses_mrope(self.config):
                    assert positions.ndim == 2 and positions.size(0) == 3, (
                        "multimodal section rotary embedding requires "
                        f"(3, seq_len) positions, but got {positions.size()}")
                inputs_embeds = self.get_input_embeddings_v0(
                    input_ids,
                    image_input=image_input,
                    video_input=video_input)
                input_ids = None
1385

1386
        hidden_states = self.language_model.model(
1387
1388
            input_ids=input_ids,
            positions=positions,
1389
            intermediate_tensors=intermediate_tensors,
1390
1391
1392
1393
            inputs_embeds=inputs_embeds,
        )
        return hidden_states

1394
1395
1396
1397
1398
1399
1400
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        return self.language_model.compute_logits(hidden_states,
                                                  sampling_metadata)
1401
1402
1403
1404
1405
1406

    def sample(
        self,
        logits: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[SamplerOutput]:
1407
        return self.language_model.sample(logits, sampling_metadata)
1408

1409
1410
    def load_weights(self, weights: Iterable[Tuple[str,
                                                   torch.Tensor]]) -> Set[str]:
1411
1412

        loader = AutoWeightsLoader(self)
1413
        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
1414
1415
1416
1417
1418
1419
1420
1421
1422

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