| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Adapted from |
| 5 | # https://github.com/huggingface/transformers/blob/main/src/transformers/models/Glm4v/modeling_Glm4v.py |
| 6 | # Copyright 2026 The ZhipuAI Team. |
| 7 | # Copyright 2026 The vLLM team. |
| 8 | # Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. |
| 9 | # All rights reserved. |
| 10 | # |
| 11 | # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX |
| 12 | # and OPT implementations in this library. It has been modified from its |
| 13 | # original forms to accommodate minor architectural differences compared |
| 14 | # to GPT-NeoX and OPT used by the Meta AI team that trained the model. |
| 15 | # |
| 16 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 17 | # you may not use this file except in compliance with the License. |
| 18 | # You may obtain a copy of the License at |
| 19 | # |
| 20 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 21 | # |
| 22 | # Unless required by applicable law or agreed to in writing, software |
| 23 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 24 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 25 | # See the License for the specific language governing permissions and |
| 26 | # limitations under the License. |
| 27 | """Inference-only GLM-OCR model compatible with HuggingFace weights.""" |
| 28 | |
| 29 | from collections.abc import Callable |
| 30 | from functools import partial |
| 31 | from typing import TYPE_CHECKING |
| 32 | |
| 33 | import torch |
| 34 | import torch.nn as nn |
| 35 | from einops import rearrange |
| 36 | |
| 37 | if TYPE_CHECKING: |
| 38 | from transformers.models.glm_ocr.configuration_glm_ocr import ( |
| 39 | GlmOcrTextConfig, |
| 40 | GlmOcrVisionConfig, |
| 41 | ) |
| 42 | |
| 43 | from vllm.config import VllmConfig |
| 44 | from vllm.distributed import get_tensor_model_parallel_world_size, parallel_state |
| 45 | from vllm.distributed import utils as dist_utils |
| 46 | from vllm.logger import init_logger |
| 47 | from vllm.model_executor.layers.attention.mm_encoder_attention import ( |
| 48 | MMEncoderAttention, |
| 49 | ) |
| 50 | from vllm.model_executor.layers.conv import Conv2dLayer |
| 51 | from vllm.model_executor.layers.fusion.mm_input_norm import build_mm_input_norm |
| 52 | from vllm.model_executor.layers.layernorm import RMSNorm |
| 53 | from vllm.model_executor.layers.linear import ( |
| 54 | QKVParallelLinear, |
| 55 | RowParallelLinear, |
| 56 | ) |
| 57 | from vllm.model_executor.layers.quantization import QuantizationConfig |
| 58 | from vllm.model_executor.layers.rotary_embedding import get_rope |
| 59 | from vllm.model_executor.layers.rotary_embedding.common import ( |
| 60 | ApplyRotaryEmb, |
| 61 | ) |
| 62 | from vllm.model_executor.models.glm4_1v import ( |
| 63 | Glm4vDummyInputsBuilder, |
| 64 | Glm4vForConditionalGeneration, |
| 65 | Glm4vMultiModalProcessor, |
| 66 | Glm4vPatchMerger, |
| 67 | Glm4vProcessingInfo, |
| 68 | Glm4vVisionBlock, |
| 69 | Glm4vVisionMLP, |
| 70 | Glm4vVisionPatchEmbed, |
| 71 | Glm4vVisionTransformer, |
| 72 | ) |
| 73 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 74 | from vllm.utils.torch_utils import PIN_MEMORY |
| 75 | |
| 76 | from .utils import ( |
| 77 | maybe_prefix, |
| 78 | ) |
| 79 | from .vision import ( |
| 80 | get_vit_attn_backend, |
| 81 | is_vit_use_data_parallel, |
| 82 | ) |
| 83 | |
| 84 | logger = init_logger(__name__) |
| 85 | |
| 86 | |
| 87 | class GlmOcrVisionMLP(Glm4vVisionMLP): |
| 88 | pass |
| 89 | |
| 90 | |
| 91 | class GlmOcrVisionAttention(nn.Module): |
| 92 | def __init__( |
| 93 | self, |
| 94 | embed_dim: int, |
| 95 | num_heads: int, |
| 96 | projection_size: int, |
| 97 | quant_config: QuantizationConfig | None = None, |
| 98 | prefix: str = "", |
| 99 | ) -> None: |
| 100 | super().__init__() |
| 101 | # Per attention head and per partition values. |
| 102 | use_data_parallel = is_vit_use_data_parallel() |
| 103 | self.tp_size = ( |
| 104 | 1 if use_data_parallel else get_tensor_model_parallel_world_size() |
| 105 | ) |
| 106 | self.tp_rank = ( |
| 107 | 0 if use_data_parallel else parallel_state.get_tensor_model_parallel_rank() |
| 108 | ) |
| 109 | self.hidden_size_per_attention_head = dist_utils.divide( |
| 110 | projection_size, num_heads |
| 111 | ) |
| 112 | self.num_attention_heads_per_partition = dist_utils.divide( |
| 113 | num_heads, self.tp_size |
| 114 | ) |
| 115 | |
| 116 | self.head_dim = embed_dim // num_heads |
| 117 | |
| 118 | self.q_norm = RMSNorm(self.head_dim, eps=1e-5) |
| 119 | self.k_norm = RMSNorm(self.head_dim, eps=1e-5) |
| 120 | |
| 121 | self.qkv = QKVParallelLinear( |
| 122 | hidden_size=embed_dim, |
| 123 | head_size=self.hidden_size_per_attention_head, |
| 124 | total_num_heads=num_heads, |
| 125 | total_num_kv_heads=num_heads, |
| 126 | bias=True, |
| 127 | quant_config=quant_config, |
| 128 | # Change qkv prefix to align with GLM-4.5V-FP8 quantization cfg |
| 129 | prefix=f"{prefix}.qkv_proj" if quant_config else f"{prefix}.qkv", |
| 130 | disable_tp=use_data_parallel, |
| 131 | ) |
| 132 | self.proj = RowParallelLinear( |
| 133 | input_size=projection_size, |
| 134 | output_size=embed_dim, |
| 135 | quant_config=quant_config, |
| 136 | prefix=f"{prefix}.proj", |
| 137 | bias=True, |
| 138 | disable_tp=use_data_parallel, |
| 139 | ) |
| 140 | |
| 141 | self.attn = MMEncoderAttention( |
| 142 | num_heads=self.num_attention_heads_per_partition, |
| 143 | head_size=self.hidden_size_per_attention_head, |
| 144 | scale=self.hidden_size_per_attention_head**-0.5, |
| 145 | prefix=f"{prefix}.attn", |
| 146 | ) |
| 147 | self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True) |
| 148 | |
| 149 | def split_qkv(self, qkv: torch.Tensor) -> tuple[torch.Tensor, ...]: |
| 150 | # [s, b, 3 * head * head_dim] |
| 151 | seq_len, bs, _ = qkv.shape |
| 152 | |
| 153 | # [s, b, 3 * head * head_dim] -> 3 * [s, b, head * head_dim] |
| 154 | q, k, v = qkv.chunk(3, dim=2) |
| 155 | |
| 156 | # 3 * [s, b, head * head_dim] -> 3 * [s, b, head, head_dim] |
| 157 | new_shape = ( |
| 158 | seq_len, |
| 159 | bs, |
| 160 | self.num_attention_heads_per_partition, |
| 161 | self.hidden_size_per_attention_head, |
| 162 | ) |
| 163 | q, k, v = (x.view(*new_shape) for x in (q, k, v)) |
| 164 | return q, k, v |
| 165 | |
| 166 | def forward( |
| 167 | self, |
| 168 | x: torch.Tensor, |
| 169 | cu_seqlens: torch.Tensor, |
| 170 | rotary_pos_emb_cos: torch.Tensor, |
| 171 | rotary_pos_emb_sin: torch.Tensor, |
| 172 | max_seqlen: torch.Tensor | None = None, # Only used for Flash Attention |
| 173 | ) -> torch.Tensor: |
| 174 | # [s, b, c] --> [s, b, head * 3 * head_dim] |
| 175 | x, _ = self.qkv(x) |
| 176 | |
| 177 | # [s, b, 3 * head * head_dim] -> 3 * [s, b, head, head_dim] |
| 178 | q, k, v = self.split_qkv(x) |
| 179 | |
| 180 | # RMSNorm on q, k |
| 181 | q_shape, k_shape = q.shape, k.shape |
| 182 | q = self.q_norm(q.reshape(-1, self.head_dim)).view(q_shape) |
| 183 | k = self.k_norm(k.reshape(-1, self.head_dim)).view(k_shape) |
| 184 | |
| 185 | q, k, v = (rearrange(x, "s b ... -> b s ...").contiguous() for x in (q, k, v)) |
| 186 | if rotary_pos_emb_cos is not None and rotary_pos_emb_sin is not None: |
| 187 | # [2 * b, s, heads, head_dim] |
| 188 | qk_concat = torch.cat([q, k], dim=0) |
| 189 | qk_rotated = self.apply_rotary_emb( |
| 190 | qk_concat, |
| 191 | rotary_pos_emb_cos, |
| 192 | rotary_pos_emb_sin, |
| 193 | ) |
| 194 | q, k = torch.chunk(qk_rotated, 2, dim=0) |
| 195 | |
| 196 | context_layer = self.attn( |
| 197 | query=q, |
| 198 | key=k, |
| 199 | value=v, |
| 200 | cu_seqlens=cu_seqlens, |
| 201 | max_seqlen=max_seqlen, |
| 202 | ) |
| 203 | context_layer = rearrange(context_layer, "b s h d -> s b (h d)").contiguous() |
| 204 | |
| 205 | output, _ = self.proj(context_layer) |
| 206 | return output |
| 207 | |
| 208 | |
| 209 | class GlmOcrVisionBlock(Glm4vVisionBlock): |
| 210 | def __init__( |
| 211 | self, |
| 212 | dim: int, |
| 213 | num_heads: int, |
| 214 | mlp_hidden_dim: int, |
| 215 | norm_layer: Callable[[int], nn.Module] | None = None, |
| 216 | quant_config: QuantizationConfig | None = None, |
| 217 | prefix: str = "", |
| 218 | ) -> None: |
| 219 | super().__init__( |
| 220 | dim, |
| 221 | num_heads, |
| 222 | mlp_hidden_dim, |
| 223 | norm_layer, |
| 224 | quant_config, |
| 225 | prefix, |
| 226 | ) |
| 227 | if norm_layer is None: |
| 228 | norm_layer = partial(nn.LayerNorm, eps=1e-6) |
| 229 | self.norm1 = norm_layer(dim) |
| 230 | self.norm2 = norm_layer(dim) |
| 231 | self.attn = GlmOcrVisionAttention( |
| 232 | embed_dim=dim, |
| 233 | num_heads=num_heads, |
| 234 | projection_size=dim, |
| 235 | quant_config=quant_config, |
| 236 | prefix=f"{prefix}.attn", |
| 237 | ) |
| 238 | self.mlp = GlmOcrVisionMLP( |
| 239 | dim, |
| 240 | mlp_hidden_dim, |
| 241 | bias=True, |
| 242 | quant_config=quant_config, |
| 243 | prefix=f"{prefix}.mlp", |
| 244 | ) |
| 245 | |
| 246 | |
| 247 | class GlmOcrPatchMerger(Glm4vPatchMerger): |
| 248 | pass |
| 249 | |
| 250 | |
| 251 | class GlmOcrVisionTransformer(Glm4vVisionTransformer): |
| 252 | def __init__( |
| 253 | self, |
| 254 | text_config: "GlmOcrTextConfig", |
| 255 | vision_config: "GlmOcrVisionConfig", |
| 256 | norm_eps: float = 1e-5, |
| 257 | quant_config: QuantizationConfig | None = None, |
| 258 | input_norm: nn.Module | None = None, |
| 259 | prefix: str = "", |
| 260 | ) -> None: |
| 261 | super().__init__( |
| 262 | text_config, |
| 263 | vision_config, |
| 264 | norm_eps, |
| 265 | quant_config, |
| 266 | input_norm, |
| 267 | prefix, |
| 268 | ) |
| 269 | |
| 270 | del self.post_conv_layernorm |
| 271 | del self.embeddings |
| 272 | |
| 273 | patch_size = vision_config.patch_size |
| 274 | temporal_patch_size = vision_config.temporal_patch_size |
| 275 | in_channels = vision_config.in_channels |
| 276 | depth = vision_config.depth |
| 277 | self.hidden_size = vision_config.hidden_size |
| 278 | self.num_heads = vision_config.num_heads |
| 279 | |
| 280 | self.patch_size = vision_config.patch_size |
| 281 | self.spatial_merge_size = vision_config.spatial_merge_size |
| 282 | self.out_hidden_size = vision_config.out_hidden_size |
| 283 | |
| 284 | self.patch_embed = Glm4vVisionPatchEmbed( |
| 285 | patch_size=patch_size, |
| 286 | temporal_patch_size=temporal_patch_size, |
| 287 | in_channels=in_channels, |
| 288 | hidden_size=self.hidden_size, |
| 289 | ) |
| 290 | |
| 291 | norm_layer = partial(RMSNorm, eps=norm_eps) |
| 292 | head_dim = self.hidden_size // self.num_heads |
| 293 | self.rotary_pos_emb = get_rope( |
| 294 | head_size=head_dim, |
| 295 | max_position=8192, |
| 296 | is_neox_style=True, |
| 297 | rope_parameters={"partial_rotary_factor": 0.5}, |
| 298 | ) |
| 299 | self.blocks = nn.ModuleList( |
| 300 | [ |
| 301 | GlmOcrVisionBlock( |
| 302 | dim=self.hidden_size, |
| 303 | num_heads=self.num_heads, |
| 304 | mlp_hidden_dim=vision_config.intermediate_size, |
| 305 | norm_layer=norm_layer, |
| 306 | quant_config=quant_config, |
| 307 | prefix=f"{prefix}.blocks.{layer_idx}", |
| 308 | ) |
| 309 | for layer_idx in range(depth) |
| 310 | ] |
| 311 | ) |
| 312 | self.merger = GlmOcrPatchMerger( |
| 313 | d_model=vision_config.out_hidden_size, |
| 314 | context_dim=text_config.intermediate_size, |
| 315 | quant_config=quant_config, |
| 316 | bias=False, |
| 317 | prefix=f"{prefix}.merger", |
| 318 | ) |
| 319 | |
| 320 | self.downsample = Conv2dLayer( |
| 321 | in_channels=vision_config.hidden_size, |
| 322 | out_channels=vision_config.out_hidden_size, |
| 323 | kernel_size=vision_config.spatial_merge_size, |
| 324 | stride=vision_config.spatial_merge_size, |
| 325 | ) |
| 326 | self.post_layernorm = RMSNorm( |
| 327 | vision_config.hidden_size, eps=vision_config.rms_norm_eps |
| 328 | ) |
| 329 | |
| 330 | self.attn_backend = get_vit_attn_backend( |
| 331 | head_size=head_dim, |
| 332 | dtype=torch.get_default_dtype(), |
| 333 | ) |
| 334 | |
| 335 | def forward( |
| 336 | self, |
| 337 | x: torch.Tensor, |
| 338 | grid_thw: torch.Tensor | list[list[int]], |
| 339 | encoder_metadata: dict[str, torch.Tensor] | None = None, |
| 340 | ) -> torch.Tensor: |
| 341 | del encoder_metadata |
| 342 | if isinstance(grid_thw, list): |
| 343 | grid_thw = torch.tensor(grid_thw, dtype=torch.int32) |
| 344 | |
| 345 | # patchify |
| 346 | x = self.input_norm(x.to(device=self.device), self.dtype) |
| 347 | x = self.patch_embed(x) |
| 348 | |
| 349 | # compute position embedding |
| 350 | rotary_pos_emb_cos, rotary_pos_emb_sin, image_type_ids = self.rot_pos_emb( |
| 351 | grid_thw |
| 352 | ) |
| 353 | # compute cu_seqlens |
| 354 | cu_seqlens = torch.repeat_interleave( |
| 355 | grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0] |
| 356 | ).cumsum(dim=0, dtype=torch.int32) |
| 357 | # cat straight into a pinned buffer so the H2D copy stays non-blocking. |
| 358 | cu_seqlens = torch.cat( |
| 359 | [cu_seqlens.new_zeros(1), cu_seqlens], |
| 360 | out=cu_seqlens.new_empty(cu_seqlens.numel() + 1, pin_memory=PIN_MEMORY), |
| 361 | ).to(self.device, non_blocking=True) |
| 362 | |
| 363 | # pre-compute max_seqlen for attn mask to reduce cuMemcpy operations |
| 364 | max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens) |
| 365 | |
| 366 | # transformers |
| 367 | x = x.unsqueeze(1) |
| 368 | for blk in self.blocks: |
| 369 | x = blk( |
| 370 | x, |
| 371 | cu_seqlens=cu_seqlens, |
| 372 | rotary_pos_emb_cos=rotary_pos_emb_cos, |
| 373 | rotary_pos_emb_sin=rotary_pos_emb_sin, |
| 374 | max_seqlen=max_seqlen, |
| 375 | ) |
| 376 | |
| 377 | # adapter |
| 378 | x = self.post_layernorm(x) |
| 379 | |
| 380 | x = x.view(-1, self.spatial_merge_size, self.spatial_merge_size, x.shape[-1]) |
| 381 | x = x.permute(0, 3, 1, 2) |
| 382 | x = self.downsample(x).view(-1, self.out_hidden_size) |
| 383 | x = self.merger(x) |
| 384 | |
| 385 | return x |
| 386 | |
| 387 | |
| 388 | @MULTIMODAL_REGISTRY.register_processor( |
| 389 | Glm4vMultiModalProcessor, |
| 390 | info=Glm4vProcessingInfo, |
| 391 | dummy_inputs=Glm4vDummyInputsBuilder, |
| 392 | ) |
| 393 | class GlmOcrForConditionalGeneration(Glm4vForConditionalGeneration): |
| 394 | def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 395 | super().__init__(vllm_config=vllm_config, prefix=prefix) |
| 396 | config = vllm_config.model_config.hf_config |
| 397 | quant_config = vllm_config.quant_config |
| 398 | |
| 399 | with self._mark_tower_model(vllm_config, {"image", "video"}): |
| 400 | self.visual = GlmOcrVisionTransformer( |
| 401 | config.text_config, |
| 402 | config.vision_config, |
| 403 | norm_eps=getattr(config, "rms_norm_eps", 1e-5), |
| 404 | quant_config=quant_config, |
| 405 | input_norm=build_mm_input_norm(self.model_config), |
| 406 | prefix=maybe_prefix(prefix, "visual"), |
| 407 | ) |