glm4_1v.py 61.3 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

# Adapted from
# https://github.com/huggingface/transformers/blob/main/src/transformers/models/Glm4v/modeling_Glm4v.py
# Copyright 2025 The vLLM team.
# Copyright 2025 The ZhipuAI 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.
27
28
"""Inference-only GLM-4.1V & GLM-4.6V-Flash, AutoGLM-Phone-9B model
compatible with HuggingFace weights."""
29
30

import math
31
from collections.abc import Callable, Iterable, Iterator, Mapping, Sequence
32
from functools import partial
33
from typing import Annotated, Any, Literal, TypeAlias
34
35
36
37
38
39

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
40
from transformers import BatchFeature, Glm4vProcessor
41
42
43
44
from transformers.models.glm4v.configuration_glm4v import (
    Glm4vTextConfig,
    Glm4vVisionConfig,
)
45
from transformers.models.glm4v.image_processing_glm4v import (
46
47
48
49
    Glm4vImageProcessor,
    smart_resize,
)
from transformers.models.glm4v.video_processing_glm4v import Glm4vVideoProcessor
50
51
from transformers.video_utils import VideoMetadata

52
from vllm.config import VllmConfig
53
from vllm.config.multimodal import BaseDummyOptions, VideoDummyOptions
54
from vllm.distributed import get_tensor_model_parallel_world_size, parallel_state
55
from vllm.distributed import utils as dist_utils
56
from vllm.inputs import MultiModalDataDict
57
from vllm.logger import init_logger
58
from vllm.model_executor.layers.attention import (
59
60
    MMEncoderAttention,
)
61
from vllm.model_executor.layers.conv import Conv2dLayer, Conv3dLayer
62
from vllm.model_executor.layers.layernorm import RMSNorm
63
64
65
66
67
68
from vllm.model_executor.layers.linear import (
    ColumnParallelLinear,
    MergedColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
69
from vllm.model_executor.layers.quantization import QuantizationConfig
70
71
72
from vllm.model_executor.layers.quantization.compressed_tensors import (
    compressed_tensors,
)
73
from vllm.model_executor.layers.rotary_embedding import get_rope
74
75
76
from vllm.model_executor.layers.rotary_embedding.common import (
    ApplyRotaryEmb,
)
77
78
79
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
80
from vllm.multimodal.inputs import (
81
    MultiModalFeatureSpec,
82
83
84
85
86
87
    MultiModalFieldConfig,
    MultiModalKwargsItems,
    VideoItem,
)
from vllm.multimodal.parse import ImageSize, MultiModalDataItems, MultiModalDataParser
from vllm.multimodal.processing import (
88
    BaseDummyInputsBuilder,
89
90
91
92
93
94
    BaseMultiModalProcessor,
    BaseProcessingInfo,
    PromptReplacement,
    PromptUpdate,
    PromptUpdateDetails,
)
95
from vllm.sequence import IntermediateTensors
96
from vllm.utils.tensor_schema import TensorSchema, TensorShape
97
from vllm.v1.attention.backends.registry import AttentionBackendEnum
98
99

from ..layers.activation import SiluAndMul
100
101
102
from .interfaces import (
    MultiModalEmbeddings,
    SupportsLoRA,
103
    SupportsMRoPE,
104
105
106
    SupportsMultiModal,
    SupportsPP,
)
107
from .qwen2_vl import _create_qwen2vl_field_factory
108
109
110
111
112
113
from .utils import (
    AutoWeightsLoader,
    WeightsMapper,
    init_vllm_registered_model,
    maybe_prefix,
)
114
115
from .vision import (
    get_vit_attn_backend,
116
    is_vit_use_data_parallel,
117
118
    run_dp_sharded_mrope_vision_model,
)
119
120
121
122
123
124
125
126
127

logger = init_logger(__name__)

# For profile run
_MAX_FRAMES_PER_VIDEO = 600

# === Vision Inputs === #


128
class Glm4vImagePixelInputs(TensorSchema):
129
    """
130
131
132
133
134
    Dimensions:
        - np: Number of patches
        - cpp: Number of channels * patch_size * patch_size
        - ni: Number of images
        - g: Grid dimensions (3 for grid_t, grid_h, grid_w)
135
    """
136

137
    type: Literal["pixel_values"] = "pixel_values"
138

139
140
    pixel_values: Annotated[torch.Tensor, TensorShape("np", "cpp")]
    image_grid_thw: Annotated[torch.Tensor, TensorShape("ni", 3)]
141
142


143
class Glm4vImageEmbeddingInputs(TensorSchema):
144
    """
145
146
147
148
149
    Dimensions:
        - f: Number of image features (varies based on image resolution)
        - h: Hidden size (must match language model backbone)
        - n: Number of images
        - g: Grid dimensions (3 for grid_t, grid_h, grid_w)
150
    """
151

152
153
154
155
    type: Literal["image_embeds"] = "image_embeds"

    image_embeds: Annotated[torch.Tensor, TensorShape("f", "h")]
    image_grid_thw: Annotated[torch.Tensor, TensorShape("n", 3)]
156
157


158
Glm4vImageInputs: TypeAlias = Glm4vImagePixelInputs | Glm4vImageEmbeddingInputs
159
160


161
class Glm4vVideoPixelInputs(TensorSchema):
162
    """
163
164
165
166
167
    Dimensions:
        - np: Number of patches
        - ctpp: Number of channels * temporal_patch_size *
            patch_size * patch_size
        - f: Number of frames
168
        - g: Grid dimensions (3 for grid_t which is usually 1 for processed
169
          video, grid_h, grid_w)
170
    """
171

172
    type: Literal["pixel_values_videos"] = "pixel_values_videos"
173

174
    pixel_values_videos: Annotated[torch.Tensor, TensorShape("np", "ctpp")]
175
    video_grid_thw: Annotated[torch.Tensor, TensorShape("f", 3)]
176
177


178
class Glm4vVideoEmbeddingInputs(TensorSchema):
179
    """
180
181
182
    Dimensions:
        - p: Number of video patches across all frames
        - h: Hidden size (must match language model backbone)
183
        - f: Number of frames
184
        - g: Grid dimensions (3 for grid_t which is usually 1 for processed
185
          video, grid_h, grid_w)
186
    """
187

188
    type: Literal["video_embeds"] = "video_embeds"
189

190
    video_embeds: Annotated[torch.Tensor, TensorShape("p", "h")]
191
    video_grid_thw: Annotated[torch.Tensor, TensorShape("f", 3)]
192
193


194
Glm4vVideoInputs: TypeAlias = Glm4vVideoPixelInputs | Glm4vVideoEmbeddingInputs
195

196
# ==== Vision Encoder ==== #
197
198
199
200
201
202
203
204


class Glm4vVisionMLP(nn.Module):
    def __init__(
        self,
        in_features: int,
        hidden_features: int,
        bias: bool = False,
205
        quant_config: QuantizationConfig | None = None,
206
        prefix: str = "",
207
208
    ):
        super().__init__()
209
        use_data_parallel = is_vit_use_data_parallel()
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
        self.gate_up_proj = MergedColumnParallelLinear(
            input_size=in_features,
            output_sizes=[hidden_features] * 2,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
            disable_tp=use_data_parallel,
        )
        self.down_proj = RowParallelLinear(
            hidden_features,
            in_features,
            bias=bias,
            quant_config=quant_config,
            prefix=f"{prefix}.down_proj",
            disable_tp=use_data_parallel,
        )
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
        self.act_fn = SiluAndMul()

    def forward(self, x: torch.Tensor):
        x, _ = self.gate_up_proj(x)
        x = self.act_fn(x)
        x, _ = self.down_proj(x)
        return x


def all_gather_interleave(local_tensor, hidden_size: int, tp_size: int):
    """All-gather the input tensor interleavely across model parallel group."""
    import torch.distributed as dist

    gathered_tensors = [torch.zeros_like(local_tensor) for _ in range(tp_size)]
    dist.all_gather(
        gathered_tensors,
        local_tensor,
        group=parallel_state.get_tp_group().device_group,
    )

    gathered_tensors_split = [
247
        torch.split(tensor, hidden_size // tp_size, -1) for tensor in gathered_tensors
248
249
250
251
252
253
254
255
256
257
258
259
260
261
    ]
    ordered_tensors = [
        tensor for pair in zip(*gathered_tensors_split) for tensor in pair
    ]
    result_tensor = torch.cat(ordered_tensors, dim=-1)
    return result_tensor


class Glm4vVisionAttention(nn.Module):
    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        projection_size: int,
262
        quant_config: QuantizationConfig | None = None,
263
264
265
266
        prefix: str = "",
    ) -> None:
        super().__init__()
        # Per attention head and per partition values.
267
        use_data_parallel = is_vit_use_data_parallel()
268
269
270
271
272
273
        self.tp_size = (
            1 if use_data_parallel else get_tensor_model_parallel_world_size()
        )
        self.tp_rank = (
            0 if use_data_parallel else parallel_state.get_tensor_model_parallel_rank()
        )
274
        self.hidden_size_per_attention_head = dist_utils.divide(
275
276
            projection_size, num_heads
        )
277
        self.num_attention_heads_per_partition = dist_utils.divide(
278
279
            num_heads, self.tp_size
        )
280

281
282
283
284
285
286
287
288
        self.qkv = QKVParallelLinear(
            hidden_size=embed_dim,
            head_size=self.hidden_size_per_attention_head,
            total_num_heads=num_heads,
            total_num_kv_heads=num_heads,
            bias=False,
            quant_config=quant_config,
            # Change qkv prefix to align with GLM-4.5V-FP8 quantization cfg
289
290
291
            prefix=f"{prefix}.qkv_proj"
            if isinstance(quant_config, compressed_tensors.CompressedTensorsConfig)
            else f"{prefix}.qkv",
292
293
294
295
296
297
298
299
300
301
            disable_tp=use_data_parallel,
        )
        self.proj = RowParallelLinear(
            input_size=projection_size,
            output_size=embed_dim,
            quant_config=quant_config,
            prefix=f"{prefix}.proj",
            bias=False,
            disable_tp=use_data_parallel,
        )
302

303
304
        self.attn = MMEncoderAttention(
            num_heads=self.num_attention_heads_per_partition,
305
            head_size=self.hidden_size_per_attention_head,
306
            scale=self.hidden_size_per_attention_head**-0.5,
307
            prefix=f"{prefix}.attn",
308
        )
309

310
311
        self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True)

312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
    def split_qkv(self, qkv: torch.Tensor) -> tuple[torch.Tensor, ...]:
        # [s, b, 3 * head * head_dim]
        seq_len, bs, _ = qkv.shape

        # [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] -> 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

    def forward(
330
331
332
        self,
        x: torch.Tensor,
        cu_seqlens: torch.Tensor,
333
334
        rotary_pos_emb_cos: torch.Tensor,
        rotary_pos_emb_sin: torch.Tensor,
335
        max_seqlen: torch.Tensor | None = None,  # Only used for Flash Attention
336
337
338
339
340
341
342
    ) -> torch.Tensor:
        # [s, b, c] --> [s, b, head * 3 * head_dim]
        x, _ = self.qkv(x)

        # [s, b, 3 * head * head_dim] -> 3 * [s, b, head, head_dim]
        q, k, v = self.split_qkv(x)

343
        q, k, v = (rearrange(x, "s b ... -> b s ...").contiguous() for x in (q, k, v))
344
        if rotary_pos_emb_cos is not None and rotary_pos_emb_sin is not None:
345
346
            # [2 * b, s, heads, head_dim]
            qk_concat = torch.cat([q, k], dim=0)
347
348
349
350
            qk_rotated = self.apply_rotary_emb(
                qk_concat,
                rotary_pos_emb_cos,
                rotary_pos_emb_sin,
351
            )
352
            q, k = torch.chunk(qk_rotated, 2, dim=0)
353

354
355
356
357
358
359
360
361
        context_layer = self.attn(
            query=q,
            key=k,
            value=v,
            cu_seqlens=cu_seqlens,
            max_seqlen=max_seqlen,
        )
        context_layer = rearrange(context_layer, "b s h d -> s b (h d)").contiguous()
362
363
364
365
366
367
368
369
370
371
372

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


class Glm4vVisionBlock(nn.Module):
    def __init__(
        self,
        dim: int,
        num_heads: int,
        mlp_hidden_dim: int,
373
374
        norm_layer: Callable[[int], nn.Module] | None = None,
        quant_config: QuantizationConfig | None = None,
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
        prefix: str = "",
    ) -> 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)
        self.attn = Glm4vVisionAttention(
            embed_dim=dim,
            num_heads=num_heads,
            projection_size=dim,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
        )
        self.mlp = Glm4vVisionMLP(
            dim,
            mlp_hidden_dim,
            bias=False,
            quant_config=quant_config,
394
            prefix=f"{prefix}.mlp",
395
396
397
        )

    def forward(
398
399
400
        self,
        x: torch.Tensor,
        cu_seqlens: torch.Tensor,
401
402
        rotary_pos_emb_cos: torch.Tensor,
        rotary_pos_emb_sin: torch.Tensor,
403
        max_seqlen: int | None = None,  # Only used for Flash Attention
404
    ) -> torch.Tensor:
405
        x_attn = self.attn(
406
407
            self.norm1(x),
            cu_seqlens=cu_seqlens,
408
409
            rotary_pos_emb_cos=rotary_pos_emb_cos,
            rotary_pos_emb_sin=rotary_pos_emb_sin,
410
411
            max_seqlen=max_seqlen,
        )
412
413
        x_fused_norm, residual = self.norm2(x, residual=x_attn)
        x = residual + self.mlp(x_fused_norm)
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431

        return x


class Glm4vVisionPatchEmbed(nn.Module):
    def __init__(
        self,
        patch_size: int = 14,
        temporal_patch_size: int = 1,
        in_channels: int = 3,
        hidden_size: int = 1536,
    ) -> 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)
432
433
        self.proj = Conv3dLayer(
            in_channels,
434
            hidden_size,
435
436
            kernel_size=kernel_size,
            stride=kernel_size,
437
438
439
440
            bias=True,
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
441
442
443
        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)
444
445
446
447
448
449
450
451
        return x


class Glm4vPatchMerger(nn.Module):
    def __init__(
        self,
        d_model: int,
        context_dim: int,
452
        quant_config: QuantizationConfig | None = None,
453
        bias: bool = False,
454
        prefix: str = "",
455
456
    ) -> None:
        super().__init__()
457
        use_data_parallel = is_vit_use_data_parallel()
458
        self.hidden_size = d_model
459
460
461
462
463
464
465
466
467
        self.proj = ColumnParallelLinear(
            self.hidden_size,
            self.hidden_size,
            bias=bias,
            gather_output=True,
            quant_config=quant_config,
            prefix=f"{prefix}.proj",
            disable_tp=use_data_parallel,
        )
468
        self.post_projection_norm = nn.LayerNorm(self.hidden_size)
469
        self.gate_up_proj = MergedColumnParallelLinear(
470
471
472
473
            input_size=self.hidden_size,
            output_sizes=[context_dim] * 2,
            bias=bias,
            quant_config=quant_config,
474
            prefix=f"{prefix}.gate_up_proj",
475
            disable_tp=use_data_parallel,
476
        )
477
        self.down_proj = RowParallelLinear(
478
479
480
481
            context_dim,
            self.hidden_size,
            bias=bias,
            quant_config=quant_config,
482
            prefix=f"{prefix}.down_proj",
483
            disable_tp=use_data_parallel,
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
        )
        self.act_fn = SiluAndMul()
        self.extra_activation_func = nn.GELU()

    def forward(self, x: torch.Tensor):
        x, _ = self.proj(x)
        x = self.extra_activation_func(self.post_projection_norm(x))
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.down_proj(x)
        return x


class Glm4vVisionEmbeddings(nn.Module):
    def __init__(self, config: Glm4vVisionConfig):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.image_size = config.image_size
        self.patch_size = config.patch_size

505
        self.num_patches = (self.image_size // self.patch_size) ** 2
506
        self.num_positions = self.num_patches
507
        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
508
509
510
511
512
513
        self.register_buffer(
            "position_ids",
            torch.arange(self.num_positions).expand((1, -1)),
            persistent=False,
        )

514
515
516
    def forward(
        self, embeddings, lengths, image_shapes, h_coords, w_coords
    ) -> torch.Tensor:
517
518
519
520
521
522
523
524
525
526
        pos_embed_weight = self.position_embedding.weight
        hidden_size = pos_embed_weight.shape[1]
        total_seq = h_coords.shape[0]
        device = pos_embed_weight.device

        # Move coordinates to correct device
        h_coords, w_coords = h_coords.to(device), w_coords.to(device)

        # Handle empty sequence case
        if total_seq == 0:
527
528
529
            adapted_pos_embed = torch.empty(
                0, hidden_size, device=device, dtype=pos_embed_weight.dtype
            )
530
531
532
        else:
            # Convert inputs to tensors if needed
            if isinstance(lengths, list):
533
                lengths = torch.tensor(lengths, device=device, dtype=torch.long)
534
            if not isinstance(image_shapes, torch.Tensor):
535
536
537
                image_shapes = torch.tensor(
                    image_shapes, device=device, dtype=torch.long
                )
538
539
540
541

            # Prepare 2D position embedding
            orig_size_sq = pos_embed_weight.shape[0]
            orig_size = int(orig_size_sq**0.5)
542
543
544
545
546
547
            pos_embed_2d = (
                pos_embed_weight.view(orig_size, orig_size, hidden_size)
                .permute(2, 0, 1)
                .unsqueeze(0)
                .to(device=device, dtype=torch.float32)
            )
548
549

            # Calculate target dimensions for each patch
550
551
552
553
554
555
556
557
558
559
            # Add bounds checking for data parallel mode
            if len(lengths) > image_shapes.shape[0]:
                # In data parallel mode, some GPUs might not have all
                # image shapes
                # Use available image shapes, cycling if necessary
                target_h_list = []
                target_w_list = []
                for i in range(len(lengths)):
                    # Cycle through available shapes
                    shape_idx = i % image_shapes.shape[0]
560
561
562
563
564
565
566
567
                    target_h_list.append(image_shapes[shape_idx, 1].repeat(lengths[i]))
                    target_w_list.append(image_shapes[shape_idx, 2].repeat(lengths[i]))
                target_h = torch.cat(target_h_list).to(
                    device=device, dtype=torch.float32
                )
                target_w = torch.cat(target_w_list).to(
                    device=device, dtype=torch.float32
                )
568
            else:
569
570
571
572
573
574
                target_h = torch.cat(
                    [image_shapes[i, 1].repeat(lengths[i]) for i in range(len(lengths))]
                ).to(device=device, dtype=torch.float32)
                target_w = torch.cat(
                    [image_shapes[i, 2].repeat(lengths[i]) for i in range(len(lengths))]
                ).to(device=device, dtype=torch.float32)
575
576
577
578
579
580
581
582

            # Normalize coordinates to [-1, 1] range for grid_sample
            h_coords = h_coords.to(device=device, dtype=torch.float32)
            w_coords = w_coords.to(device=device, dtype=torch.float32)
            norm_w = ((w_coords + 0.5) / target_w) * 2 - 1
            norm_h = ((h_coords + 0.5) / target_h) * 2 - 1

            # Create sampling grid
583
            grid = torch.stack((norm_w, norm_h), dim=-1).unsqueeze(0).unsqueeze(2)
584
585
586
587
588
589
590
591
592
593
594
595

            # Perform bicubic interpolation
            interpolated_embed_fp32 = F.grid_sample(
                pos_embed_2d,
                grid,
                mode="bicubic",
                align_corners=False,
                padding_mode="border",
            )

            # Reshape and convert back to original dtype
            adapted_pos_embed_fp32 = (
596
597
598
599
600
                interpolated_embed_fp32.squeeze(0).squeeze(-1).permute(1, 0)
            )
            adapted_pos_embed = adapted_pos_embed_fp32.to(pos_embed_weight.dtype).to(
                embeddings.device
            )
601
602
603
604
605
606
607
608
609

        # Add adapted position encoding to embeddings
        embeddings = embeddings + adapted_pos_embed
        return embeddings


class Glm4vVisionTransformer(nn.Module):
    def __init__(
        self,
610
        text_config: Glm4vTextConfig,
611
612
        vision_config: Glm4vVisionConfig,
        norm_eps: float = 1e-6,
613
        quant_config: QuantizationConfig | None = None,
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
        prefix: str = "",
    ) -> None:
        super().__init__()

        patch_size = vision_config.patch_size
        temporal_patch_size = vision_config.temporal_patch_size
        in_channels = vision_config.in_channels
        depth = vision_config.depth
        self.hidden_size = vision_config.hidden_size
        self.num_heads = vision_config.num_heads

        self.patch_size = vision_config.patch_size
        self.spatial_merge_size = vision_config.spatial_merge_size
        self.out_hidden_size = vision_config.out_hidden_size

        self.patch_embed = Glm4vVisionPatchEmbed(
            patch_size=patch_size,
            temporal_patch_size=temporal_patch_size,
            in_channels=in_channels,
            hidden_size=self.hidden_size,
        )

        norm_layer = partial(RMSNorm, eps=norm_eps)
        head_dim = self.hidden_size // self.num_heads
638
639
640
641
        self.rotary_pos_emb = get_rope(
            head_size=head_dim,
            max_position=8192,
            is_neox_style=True,
642
            rope_parameters={"partial_rotary_factor": 0.5},
643
        )
644
645
646
647
648
649
650
651
652
653
654
655
656
        self.blocks = nn.ModuleList(
            [
                Glm4vVisionBlock(
                    dim=self.hidden_size,
                    num_heads=self.num_heads,
                    mlp_hidden_dim=vision_config.out_hidden_size,
                    norm_layer=norm_layer,
                    quant_config=quant_config,
                    prefix=f"{prefix}.blocks.{layer_idx}",
                )
                for layer_idx in range(depth)
            ]
        )
657
658
659
660
661
        self.merger = Glm4vPatchMerger(
            d_model=vision_config.out_hidden_size,
            context_dim=vision_config.intermediate_size,
            quant_config=quant_config,
            bias=False,
662
            prefix=f"{prefix}.merger",
663
664
665
        )
        self.embeddings = Glm4vVisionEmbeddings(vision_config)

666
667
668
        self.post_conv_layernorm = RMSNorm(
            vision_config.hidden_size, eps=vision_config.rms_norm_eps
        )
669
        self.downsample = Conv2dLayer(
670
671
672
673
674
            in_channels=vision_config.hidden_size,
            out_channels=vision_config.out_hidden_size,
            kernel_size=vision_config.spatial_merge_size,
            stride=vision_config.spatial_merge_size,
        )
675
676
677
        self.post_layernorm = RMSNorm(
            vision_config.hidden_size, eps=vision_config.rms_norm_eps
        )
678

679
        self.attn_backend = get_vit_attn_backend(
680
681
            head_size=head_dim,
            dtype=torch.get_default_dtype(),
682
        )
683
684
685
686
687
688
689
690
691

    @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

692
693
694
    def rot_pos_emb(
        self, grid_thw: torch.Tensor
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
695
696
697
698
        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)
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
            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))
720
721
        pos_ids = torch.cat(pos_ids, dim=0)
        max_grid_size = grid_thw[:, 1:].max()
722
723
724
725

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

726
727
        cos_combined = cos[pos_ids].flatten(1)
        sin_combined = sin[pos_ids].flatten(1)
728
        return cos_combined, sin_combined, pos_ids
729
730
731
732

    def compute_attn_mask_seqlen(
        self,
        cu_seqlens: torch.Tensor,
733
    ) -> torch.Tensor | None:
734
        max_seqlen = None
735
736
737
738
739
        if self.attn_backend in {
            AttentionBackendEnum.FLASH_ATTN,
            AttentionBackendEnum.ROCM_AITER_FA,
            AttentionBackendEnum.TRITON_ATTN,
        }:
740
            max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
741
        return max_seqlen
742
743
744
745

    def forward(
        self,
        x: torch.Tensor,
746
        grid_thw: torch.Tensor | list[list[int]],
747
    ) -> torch.Tensor:
748
749
        if isinstance(grid_thw, list):
            grid_thw = torch.tensor(grid_thw, dtype=torch.int32)
750

751
752
753
754
755
756
        # patchify
        x = x.to(device=self.device, dtype=self.dtype)
        x = self.patch_embed(x)
        x = self.post_conv_layernorm(x)

        # compute position embedding
757
758
759
        rotary_pos_emb_cos, rotary_pos_emb_sin, image_type_ids = self.rot_pos_emb(
            grid_thw
        )
760
        # compute cu_seqlens
761
762
763
        cu_seqlens = torch.repeat_interleave(
            grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
        ).cumsum(dim=0, dtype=torch.int32)
764
        cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
765
766
767
        # pre-compute max_seqlen for attn mask to reduce cuMemcpy operations
        max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens)
        seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
768
        cu_seqlens = cu_seqlens.to(self.device, non_blocking=True)
769
770
771
        x = self.embeddings(
            x, seqlens, grid_thw, image_type_ids[:, 0], image_type_ids[:, 1]
        )
772
773
774
775
776
777
778

        # transformers
        x = x.unsqueeze(1)
        for blk in self.blocks:
            x = blk(
                x,
                cu_seqlens=cu_seqlens,
779
780
                rotary_pos_emb_cos=rotary_pos_emb_cos,
                rotary_pos_emb_sin=rotary_pos_emb_sin,
781
782
783
784
785
786
                max_seqlen=max_seqlen,
            )

        # adapter
        x = self.post_layernorm(x)

787
        x = x.view(-1, self.spatial_merge_size, self.spatial_merge_size, x.shape[-1])
788
789
790
791
792
793
        x = x.permute(0, 3, 1, 2)
        x = self.downsample(x).view(-1, self.out_hidden_size)
        x = self.merger(x)

        return x

794
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("attn.qkv.", "attn.q.", "q"),
            ("attn.qkv.", "attn.k.", "k"),
            ("attn.qkv.", "attn.v.", "v"),
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]
        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]
818
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
819
820
821
822
823
824
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params


class Glm4vProcessingInfo(BaseProcessingInfo):
825
    def get_supported_mm_limits(self) -> Mapping[str, int | None]:
826
827
        return {"image": None, "video": 1}

828
829
    def get_image_processor(self, **kwargs: object) -> Glm4vImageProcessor:
        return self.get_hf_processor(**kwargs).image_processor
830

831
832
    def get_video_processor(self, **kwargs: object) -> Glm4vVideoProcessor:
        return self.get_hf_processor(**kwargs).video_processor
833

834
835
836
837
838
839
    def get_data_parser(self):
        return MultiModalDataParser(
            video_needs_metadata=True,
            expected_hidden_size=self._get_expected_hidden_size(),
        )

840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
    def _get_vision_info(
        self,
        *,
        image_width: int,
        image_height: int,
        num_frames: int = 16,
        do_resize: bool = True,
        max_image_pixels: int = 28 * 28 * 2 * 30000,
    ) -> tuple[ImageSize, int]:
        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:
            resized_height, resized_width = smart_resize(
                num_frames=num_frames
857
858
                if num_frames > temporal_patch_size
                else temporal_patch_size,
859
860
861
862
863
                height=image_height,
                width=image_width,
                factor=patch_size * merge_size,
                max_pixels=max_image_pixels,
            )
864
            preprocessed_size = ImageSize(width=resized_width, height=resized_height)
865
        else:
866
            preprocessed_size = ImageSize(width=image_width, height=image_height)
867
868
869
870
871
872
873
874
875
876
877
878
879
880

        # 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)
        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

881
882
883
884
885
886
887
888
889
890
    def _get_image_max_pixels(self) -> int:
        """Read max_pixels from the HF image processor config.

        Despite the name, ``longest_edge`` is a pixel **area** (total pixel
        count), not an edge length.  The HF processor passes it directly to
        ``smart_resize`` as the ``max_pixels`` argument, which constrains
        ``t_bar * h_bar * w_bar <= max_pixels``.
        """
        return self.get_image_processor().size["longest_edge"]

891
    def get_image_size_with_most_features(self) -> ImageSize:
892
893
894
895
896
897
        # Use num_frames=1 for single-image budget estimation.
        # _get_vision_info defaults to num_frames=16 (video), which
        # makes smart_resize constrain 16*H*W <= max_pixels, vastly
        # underestimating the spatial budget for a single image and
        # causing encoder cache overflow for large images
        # (see https://github.com/vllm-project/vllm/issues/34040).
898
        max_image_size, _ = self._get_vision_info(
899
900
901
902
            image_width=9999999,
            image_height=9999999,
            num_frames=1,
            max_image_pixels=self._get_image_max_pixels(),
903
        )
904
905
906
907
908
909
910
911
912
913
914
        return max_image_size

    def get_num_image_tokens(
        self,
        *,
        image_width: int,
        image_height: int,
    ) -> int:
        _, num_image_tokens = self._get_vision_info(
            image_width=image_width,
            image_height=image_height,
915
916
            num_frames=1,
            max_image_pixels=self._get_image_max_pixels(),
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
        )
        return num_image_tokens

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

        return self.get_num_image_tokens(
            image_width=target_width,
            image_height=target_height,
        )

    def get_num_video_tokens(
        self,
        *,
        image_width: int,
        image_height: int,
        num_frames: int,
    ) -> int:
        _, num_video_tokens = self._get_vision_info(
            image_width=image_width,
            image_height=image_height,
            num_frames=num_frames,
            max_image_pixels=28 * 28 * 2 * 30000,
        )
        return num_video_tokens

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

        num_frames = 0

        while True:
            next_num_frames = num_frames + 1
            next_max_tokens = self.get_num_video_tokens(
                image_width=target_width,
                image_height=target_height,
                num_frames=next_num_frames,
            )
            if next_max_tokens > max_tokens or next_max_tokens == 0:
                break

            num_frames = next_num_frames

        return num_frames

    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)

        max_image_tokens = self.get_max_image_tokens() * max_images
971
972
973
974
        max_total_frames = self._get_max_video_frames(seq_len - max_image_tokens)
        max_frames_per_video = min(
            max_total_frames // max(max_videos, 1), _MAX_FRAMES_PER_VIDEO
        )
975
976
977

        return max(max_frames_per_video, 1)

978
    def _get_video_second_idx_glm4v(
979
980
        self, metadata: dict[str, Any], total_frames: int
    ) -> list[int]:
981
982
        video_processor = self.get_video_processor()

983
        video_fps = metadata.get("fps", video_processor.fps)
984
985
        meta_frames = metadata.get("total_num_frames", total_frames)
        max_frame_idx = meta_frames - 1
986
        duration = metadata.get("duration", round(max_frame_idx / video_fps) + 1)
987
988
989
        do_sample_frames = metadata["do_sample_frames"]
        if not do_sample_frames:
            frame_indices = metadata["frames_indices"]
990
        else:
991
992
            if duration <= video_processor.max_duration:
                n = int(math.floor(duration * video_processor.fps))
993
                frame_indices = [
994
995
996
                    min(
                        max_frame_idx,
                        int(math.ceil(i * video_fps / video_processor.fps)),
997
998
                    )
                    for i in range(n)
999
                ]
1000
            else:
1001
                num_samples = int(video_processor.max_duration * video_processor.fps)
1002
1003
1004
                if num_samples >= meta_frames:
                    frame_indices = list(range(meta_frames))
                else:
1005
1006
1007
                    target_seconds = np.linspace(
                        0, duration, num_samples, endpoint=True
                    )
1008
1009
1010
1011
                    frame_indices = [
                        min(max_frame_idx, int(math.ceil(t * video_fps)))
                        for t in target_seconds
                    ]
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028

        seen, uniq = set(), []
        for idx in frame_indices:
            if idx not in seen:
                seen.add(idx)
                uniq.append(idx)
        if len(uniq) & 1:
            uniq.append(uniq[-1])
        frame_indices = uniq

        full_second_idxs = [int(idx / video_fps) for idx in frame_indices]
        timestamps_list = full_second_idxs[::2]
        selected_timestamps = []
        for idx in range(0, len(timestamps_list)):
            selected_timestamps.append(timestamps_list[idx])
        return selected_timestamps

1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
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
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
    def _get_video_second_idx_glm46v(
        self, metadata: dict[str, Any], total_frames: int
    ) -> list[int]:
        video_processor = self.get_video_processor()

        video_fps = metadata["fps"]
        meta_frames = metadata.get("total_num_frames", total_frames)
        max_frame_idx = meta_frames - 1
        duration = metadata.get("duration", round(max_frame_idx / video_fps) + 1)

        do_sample_frames = metadata.get("do_sample_frames", True)
        if not do_sample_frames:
            frame_indices = metadata["frames_indices"]
        else:
            DYNAMIC_FPS_THRES = {30: 3, 300: 1, 2400: 0.5}
            MAX_FRAME_COUNT_DYNAMIC = 640
            MAX_DURATION = 2400

            effective_duration = min(duration, MAX_DURATION)
            if effective_duration <= 30:
                target_fps = DYNAMIC_FPS_THRES[30]
            elif effective_duration <= 300:
                target_fps = DYNAMIC_FPS_THRES[300]
            else:
                target_fps = DYNAMIC_FPS_THRES[2400]

            temporal_patch_size = getattr(video_processor, "temporal_patch_size", 1)
            extract_t = int(effective_duration * target_fps * temporal_patch_size)
            extract_t = min(extract_t, MAX_FRAME_COUNT_DYNAMIC)

            duration_per_frame = 1 / video_fps
            timestamps = [i * duration_per_frame for i in range(meta_frames)]
            max_second = int(duration)

            if meta_frames < extract_t:
                frame_indices = np.linspace(
                    0, meta_frames - 1, extract_t, dtype=int
                ).tolist()
            else:
                frame_indices = []
                current_second = 0.0
                inv_fps = 1 / (temporal_patch_size * target_fps)
                for frame_index in range(meta_frames):
                    if timestamps[frame_index] >= current_second:
                        current_second += inv_fps
                        frame_indices.append(frame_index)
                        if current_second >= max_second:
                            break

            if len(frame_indices) < extract_t:
                if len(frame_indices) == 0:
                    start, end = 0, max(meta_frames - 1, 0)
                else:
                    start, end = frame_indices[0], frame_indices[-1]
                frame_indices = np.linspace(start, end, extract_t, dtype=int).tolist()
            elif len(frame_indices) > extract_t:
                frame_indices = np.linspace(
                    0, meta_frames - 1, extract_t, dtype=int
                ).tolist()

        seen, uniq = set(), []
        for idx in frame_indices:
            if idx not in seen:
                seen.add(idx)
                uniq.append(idx)

        if len(uniq) & 1:
            uniq.append(uniq[-1])

        frame_indices = uniq
        full_second_idxs = [int(idx / video_fps) for idx in frame_indices]
        timestamps_list = full_second_idxs[::2]
        selected_timestamps = []
        for idx in range(len(timestamps_list)):
            selected_timestamps.append(timestamps_list[idx])
        return selected_timestamps

1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
    def _construct_video_placeholder(
        self,
        video_array: np.ndarray,
        metadata: dict[str, Any],
        grid_thw: torch.Tensor,
    ) -> str:
        hf_processor = self.get_hf_processor()
        tokenizer = self.get_tokenizer()
        image_processor = hf_processor.image_processor

        hf_config = self.get_hf_config()
        boi_token_id = hf_config.image_start_token_id
        eoi_token_id = hf_config.image_end_token_id
        bov_token_id = hf_config.video_start_token_id
        eov_token_id = hf_config.video_end_token_id
        merge_length = image_processor.merge_size**2

        assert isinstance(grid_thw, torch.Tensor)
1124
1125
1126
1127
1128
1129
1130
1131
1132
        timestamps = (
            self._get_video_second_idx_glm4v(metadata, len(video_array))
            if isinstance(hf_processor, Glm4vProcessor)
            else self._get_video_second_idx_glm46v(metadata, len(video_array))
        )

        timestamp_format = (
            "{}" if isinstance(hf_processor, Glm4vProcessor) else "{:.1f} seconds"
        )
1133
        frames_idx_token = [
1134
1135
            tokenizer.encode(timestamp_format.format(i), add_special_tokens=False)
            for i in timestamps
1136
1137
1138
1139
1140
1141
1142
        ]
        T, H, W = grid_thw
        num_tokens_per_frame = int(H * W) // merge_length
        placeholder = []
        placeholder.append(bov_token_id)
        for frame_idx in frames_idx_token:
            placeholder.append(boi_token_id)
1143
            placeholder.extend([hf_processor.video_token_id] * num_tokens_per_frame)
1144
1145
1146
1147
1148
1149
            placeholder.append(eoi_token_id)
            placeholder.extend(frame_idx)
        placeholder.append(eov_token_id)

        return placeholder

1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173

class Glm4vDummyInputsBuilder(BaseDummyInputsBuilder[Glm4vProcessingInfo]):
    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)

        hf_config = self.info.get_hf_config()
        hf_processor = self.info.get_hf_processor()
        tokenizer = self.info.get_tokenizer()

        image_token: str = hf_processor.image_token
        video_token_ids = [
            hf_config.video_start_token_id,
            hf_processor.video_token_id,
            hf_config.video_end_token_id,
        ]
        video_token = tokenizer.decode(video_token_ids)

        return image_token * num_images + video_token * num_videos

    def get_dummy_mm_data(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
1174
        mm_options: Mapping[str, BaseDummyOptions],
1175
1176
1177
1178
    ) -> MultiModalDataDict:
        num_images = mm_counts.get("image", 0)
        num_videos = mm_counts.get("video", 0)

1179
        target_width, target_height = self.info.get_image_size_with_most_features()
1180
        target_num_frames = self.info.get_num_frames_with_most_features(
1181
1182
            seq_len, mm_counts
        )
1183

1184
1185
        image_overrides = mm_options.get("image")
        video_overrides = mm_options.get("video")
1186

1187
        return {
1188
1189
1190
1191
1192
1193
1194
            "image": self._get_dummy_images(
                width=target_width,
                height=target_height,
                num_images=num_images,
                overrides=image_overrides,
            ),
            "video": self._get_dummy_videos(
1195
1196
1197
1198
                width=target_width,
                height=target_height,
                num_frames=target_num_frames,
                num_videos=num_videos,
1199
                overrides=video_overrides,
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
            ),
        }

    def _get_dummy_videos(
        self,
        *,
        width: int,
        height: int,
        num_frames: int,
        num_videos: int,
1210
        overrides: VideoDummyOptions | None = None,
1211
    ) -> list[VideoItem]:
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
        # GLM 4.6V requires at least 2 frames
        num_frames = max(num_frames, 2)
        if overrides and overrides.num_frames:
            overrides.num_frames = max(overrides.num_frames, 2)

        videos = super()._get_dummy_videos(
            width=width,
            height=height,
            num_frames=num_frames,
            num_videos=num_videos,
            overrides=overrides,
        )
        videos = [v.copy() for v in videos]
1225

1226
        video_items = []
1227
1228
        for video in videos:
            video_num_frames = video.shape[0]
1229
1230
            video_metadata = {
                "fps": 2.0,
1231
1232
1233
                "duration": video_num_frames / 2.0,
                "total_num_frames": video_num_frames,
                "frames_indices": list(range(video_num_frames)),
1234
                "video_backend": "opencv",
1235
                "do_sample_frames": False,
1236
            }
1237
            video_items.append((video, video_metadata))
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255

        return video_items


class Glm4vMultiModalProcessor(BaseMultiModalProcessor[Glm4vProcessingInfo]):
    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)

        # GLM-4.1V use `image_token_id` as video placeholder, we need to
        # replace it with `video_token_id` for video processing. So we
        # separate video processing from image processing.
1256
1257
1258
1259
1260
        if (
            "videos" in mm_data
            and isinstance(mm_data["videos"], list)
            and len(mm_data["videos"]) > 0
        ):
1261
1262
1263
1264
1265
            video_grid_thw_lst = []
            pixel_values_videos_lst = []
            for item in mm_data.pop("videos", []):
                video_array, metadata = item

1266
1267
1268
                # don't update mm_kwargs inplace
                video_mm_kwargs = dict(**mm_kwargs)
                video_mm_kwargs["do_sample_frames"] = metadata.get(
1269
1270
                    "do_sample_frames", True
                )
1271
1272
1273

                video_mm_data = dict()
                video_mm_data["videos"] = [[video_array]]
1274
1275

                unuse_metadata = ["do_sample_frames"]
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
                video_mm_data["video_metadata"] = [
                    [
                        VideoMetadata(
                            **{
                                k: metadata[k]
                                for k in metadata
                                if k not in unuse_metadata
                            }
                        )
                    ]
                ]
1287
1288
1289
1290

                video_outputs = super()._call_hf_processor(
                    prompt="<|begin_of_video|><|video|><|end_of_video|>",
                    mm_data=video_mm_data,
1291
                    mm_kwargs=video_mm_kwargs,
1292
1293
                    tok_kwargs=tok_kwargs,
                )
1294
1295
1296
1297
1298
                input_ids = video_outputs.pop("input_ids")
                input_ids[input_ids == processor.image_token_id] = (
                    processor.video_token_id
                )
                video_placeholder = processor.tokenizer.batch_decode(input_ids)[0]
1299
1300
1301
                prompt = prompt.replace(
                    "<|begin_of_video|><|video|><|end_of_video|>",
                    video_placeholder,
1302
                    1,
1303
1304
                )

1305
                video_grid_thw_lst.append(video_outputs["video_grid_thw"])
1306
                pixel_values_videos_lst.append(video_outputs["pixel_values_videos"])
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
            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]:
1331
        return _create_qwen2vl_field_factory(
1332
1333
            self.info.get_hf_config().vision_config.spatial_merge_size
        )(hf_inputs)
1334
1335
1336
1337
1338

    def _get_prompt_updates(
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, Any],
1339
        out_mm_kwargs: MultiModalKwargsItems,
1340
1341
    ) -> Sequence[PromptUpdate]:
        hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
1342
        image_processor = self.info.get_image_processor(**hf_processor_mm_kwargs)
1343
1344
1345
1346

        merge_length = image_processor.merge_size**2

        def get_image_replacement_glm4v(item_idx: int):
1347
1348
            out_item = out_mm_kwargs["image"][item_idx]
            grid_thw = out_item["image_grid_thw"].data
1349
1350
1351
1352
1353
1354
            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_glm4v(item_idx: int):
1355
1356
            out_item = out_mm_kwargs["video"][item_idx]
            grid_thw = out_item["video_grid_thw"].data
1357
1358
1359
            assert isinstance(grid_thw, torch.Tensor)

            video, metadata = mm_items["video"][item_idx]
1360
            placeholder = self.info._construct_video_placeholder(
1361
1362
                video, metadata, grid_thw
            )
1363
1364
1365
1366
            return PromptUpdateDetails.select_token_id(
                placeholder,
                embed_token_id=hf_processor.video_token_id,
            )
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386

        return [
            PromptReplacement(
                modality="image",
                target=hf_processor.image_token,
                replacement=get_image_replacement_glm4v,
            ),
            PromptReplacement(
                modality="video",
                target="<|begin_of_video|><|video|><|end_of_video|>",
                replacement=get_video_replacement_glm4v,
            ),
        ]


@MULTIMODAL_REGISTRY.register_processor(
    Glm4vMultiModalProcessor,
    info=Glm4vProcessingInfo,
    dummy_inputs=Glm4vDummyInputsBuilder,
)
1387
class Glm4vForConditionalGeneration(
1388
    nn.Module, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE
1389
):
1390
1391
1392
1393
1394
1395
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
1396
        "gate_up_proj": ["gate_up_proj"],
1397
1398
1399
1400
1401
1402
1403
1404
    }

    # 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.": "language_model.model.",
            "model.visual.": "visual.",
1405
1406
        }
    )
1407

1408
1409
    supports_encoder_tp_data = True

1410
    @classmethod
1411
    def get_placeholder_str(cls, modality: str, i: int) -> str | None:
1412
1413
1414
1415
1416
1417
1418
        if modality.startswith("image"):
            return "<|begin_of_image|><|image|><|end_of_image|>"
        if modality.startswith("video"):
            return "<|begin_of_video|><|video|><|end_of_video|>"

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

1419
1420
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
Yuxuan Zhang's avatar
Yuxuan Zhang committed
1421
        config = vllm_config.model_config.hf_config
1422
1423
1424
1425
1426
        quant_config = vllm_config.quant_config
        multimodal_config = vllm_config.model_config.multimodal_config

        self.config = config
        self.multimodal_config = multimodal_config
1427
        self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
1428

1429
1430
        with self._mark_tower_model(vllm_config, {"image", "video"}):
            self.visual = Glm4vVisionTransformer(
1431
                config.text_config,
1432
1433
1434
1435
1436
                config.vision_config,
                norm_eps=getattr(config, "rms_norm_eps", 1e-5),
                quant_config=quant_config,
                prefix=maybe_prefix(prefix, "visual"),
            )
1437

1438
        if config.model_type in ("glm4v", "glm_ocr"):
Yuxuan Zhang's avatar
Yuxuan Zhang committed
1439
1440
1441
1442
1443
1444
            architectures = ["Glm4ForCausalLM"]
        elif config.model_type == "glm4v_moe":
            architectures = ["Glm4MoeForCausalLM"]
        else:
            architectures = None

1445
1446
1447
1448
1449
1450
1451
        with self._mark_language_model(vllm_config):
            self.language_model = init_vllm_registered_model(
                vllm_config=vllm_config,
                hf_config=config.text_config,
                prefix=maybe_prefix(prefix, "language_model"),
                architectures=architectures,
            )
1452
1453

        self.make_empty_intermediate_tensors = (
1454
1455
            self.language_model.make_empty_intermediate_tensors
        )
1456
1457

    def _parse_and_validate_image_input(
1458
        self, **kwargs: object
1459
    ) -> Glm4vImageInputs | None:
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
        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:
            return Glm4vImagePixelInputs(
                type="pixel_values",
                pixel_values=pixel_values,
                image_grid_thw=image_grid_thw,
            )

        if image_embeds is not None:
            return Glm4vImageEmbeddingInputs(
                type="image_embeds",
                image_embeds=image_embeds,
                image_grid_thw=image_grid_thw,
            )

    def _parse_and_validate_video_input(
1482
        self, **kwargs: object
1483
    ) -> Glm4vVideoInputs | None:
1484
1485
1486
        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)
1487

1488
1489
        if pixel_values_videos is None and video_embeds is None:
            return None
1490

1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
        if pixel_values_videos is not None:
            return Glm4vVideoPixelInputs(
                type="pixel_values_videos",
                pixel_values_videos=pixel_values_videos,
                video_grid_thw=video_grid_thw,
            )

        if video_embeds is not None:
            return Glm4vVideoEmbeddingInputs(
                type="video_embeds",
                video_embeds=video_embeds,
                video_grid_thw=video_grid_thw,
            )

    def _process_image_input(
1506
1507
        self, image_input: Glm4vImageInputs
    ) -> tuple[torch.Tensor, ...]:
1508
1509
1510
1511
1512
1513
1514
        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)
1515
            if self.use_data_parallel:
1516
1517
1518
                return run_dp_sharded_mrope_vision_model(
                    self.visual, pixel_values, grid_thw.tolist(), rope_type="rope_3d"
                )
1519
            else:
1520
1521
                image_embeds = self.visual(pixel_values, grid_thw=grid_thw)

1522
        merge_size = self.visual.spatial_merge_size
1523
        sizes = (grid_thw.prod(-1) // merge_size // merge_size).tolist()
1524
        return image_embeds.split(sizes)
1525
1526

    def _process_video_input(
1527
1528
        self, video_input: Glm4vVideoInputs
    ) -> tuple[torch.Tensor, ...]:
1529
1530
1531
1532
1533
1534
1535
        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(
1536
1537
                self.visual.dtype
            )
1538
            if self.use_data_parallel:
1539
1540
1541
1542
1543
1544
                return run_dp_sharded_mrope_vision_model(
                    self.visual,
                    pixel_values_videos,
                    grid_thw.tolist(),
                    rope_type="rope_3d",
                )
1545
            else:
1546
1547
                video_embeds = self.visual(pixel_values_videos, grid_thw=grid_thw)

1548
1549
        # Split concatenated embeddings for each video item.
        merge_size = self.visual.spatial_merge_size
1550
        sizes = (grid_thw.prod(-1) // merge_size // merge_size).tolist()
1551
        return video_embeds.split(sizes)
1552
1553
1554
1555
1556
1557
1558

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

        # Preserve the order of modalities if there are multiple of them
        # from the order of kwargs.
        for input_key in kwargs:
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
            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
                )
1573
1574
        return mm_input_by_modality

1575
    def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings | None:
1576
        mm_input_by_modality = self._parse_and_validate_multimodal_inputs(**kwargs)
1577
1578
1579
1580
        if not mm_input_by_modality:
            return None

        # The result multimodal_embeddings is tuple of tensors, with each
1581
        # tensor corresponding to a multimodal data item (image or video).
1582
1583
1584
1585
1586
1587
1588
        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":
1589
1590
                image_embeddings = self._process_image_input(multimodal_input)
                multimodal_embeddings += tuple(image_embeddings)
1591
1592
            if modality == "video":
                video_embeddings = self._process_video_input(multimodal_input)
1593
                multimodal_embeddings += tuple(video_embeddings)
1594
1595
        return multimodal_embeddings

1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
    def iter_mm_grid_thw(
        self, mm_features: list[MultiModalFeatureSpec]
    ) -> Iterator[tuple[int, int, int, int]]:
        hf_config = self.config
        spatial_merge_size = hf_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, t, h // spatial_merge_size, w // spatial_merge_size
            elif mm_feature.modality == "video":
                t, h, w = mm_feature.data["video_grid_thw"].data.tolist()
                yield (
                    offset,
                    t,
                    h // spatial_merge_size,
                    w // spatial_merge_size,
                )
            else:
                raise ValueError(f"Unsupported modality: {mm_feature.modality}")

1618
1619
1620
    def get_mrope_input_positions(
        self,
        input_tokens: list[int],
1621
        mm_features: list[MultiModalFeatureSpec],
1622
1623
    ) -> tuple[torch.Tensor, int]:
        llm_pos_ids_list: list = []
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
        st = 0
        for (
            offset,
            llm_grid_t,
            llm_grid_h,
            llm_grid_w,
        ) in self.iter_mm_grid_thw(mm_features):
            text_len = offset - st
            st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
            llm_pos_ids_list.append(
                np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx
            )
            grid_indices = np.indices((llm_grid_t, llm_grid_h, llm_grid_w)).reshape(
                3, -1
            )
            llm_pos_ids_list.append(grid_indices + text_len + st_idx)
            st = offset + llm_grid_t * llm_grid_h * llm_grid_w

        if st < len(input_tokens):
            text_len = len(input_tokens) - st
            st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
            llm_pos_ids_list.append(
                np.broadcast_to(np.arange(text_len), (3, text_len)) + st_idx
            )
1648

1649
        llm_positions = np.concatenate(llm_pos_ids_list, axis=1).reshape(3, -1)
1650
        mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item()
1651
        return torch.from_numpy(llm_positions), mrope_position_delta
1652

1653
1654
    def forward(
        self,
1655
        input_ids: torch.Tensor | None,
1656
        positions: torch.Tensor,
1657
1658
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
1659
        **kwargs: object,
1660
    ) -> torch.Tensor | IntermediateTensors:
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
        """Run forward pass for GLM-4V.

        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 GLM-4V
                opensource models), the shape will be `(3, seq_len)`,
                otherwise it will be `(seq_len,).
1671
1672
1673
1674
            intermediate_tensors: Optional intermediate tensors for pipeline
                parallelism.
            inputs_embeds: Optional pre-computed input embeddings.
            **kwargs: Additional keyword arguments.
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
        """
        if intermediate_tensors is not None:
            inputs_embeds = None

        hidden_states = self.language_model.model(
            input_ids=input_ids,
            positions=positions,
            intermediate_tensors=intermediate_tensors,
            inputs_embeds=inputs_embeds,
        )
        return hidden_states

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
1690
    ) -> torch.Tensor | None:
1691
        return self.language_model.compute_logits(hidden_states)
1692

1693
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
1694
1695
1696
1697
1698
1699
1700
1701
        loader = AutoWeightsLoader(self)
        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(
Jee Jee Li's avatar
Jee Jee Li committed
1702
            language_model="language_model.model",
1703
1704
1705
            connector="visual.merger.",
            tower_model="visual.",
        )
Jee Jee Li's avatar
Jee Jee Li committed
1706

1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
    def get_num_mm_encoder_tokens(
        self,
        num_image_tokens: int,
    ) -> int:
        merge_size = self.config.vision_config.spatial_merge_size
        return num_image_tokens * (merge_size**2)

    def get_num_mm_connector_tokens(
        self,
        num_vision_tokens: int,
    ) -> int:
        merge_size = self.config.vision_config.spatial_merge_size
        return num_vision_tokens // (merge_size**2)

Jee Jee Li's avatar
Jee Jee Li committed
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738

@MULTIMODAL_REGISTRY.register_processor(
    Glm4vMultiModalProcessor,
    info=Glm4vProcessingInfo,
    dummy_inputs=Glm4vDummyInputsBuilder,
)
class Glm4vMoeForConditionalGeneration(Glm4vForConditionalGeneration):
    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }