# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project # 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.""" import math from collections.abc import Callable, Iterable, Mapping, Sequence from functools import partial from typing import Annotated, Any, Literal, TypeAlias import numpy as np import torch import torch.nn as nn from einops import rearrange from transformers import BatchFeature from transformers.models.qwen2_vl import Qwen2VLImageProcessor, Qwen2VLProcessor from transformers.models.qwen2_vl.configuration_qwen2_vl import ( Qwen2VLConfig, Qwen2VLVisionConfig, ) from transformers.models.qwen2_vl.image_processing_qwen2_vl import smart_resize from transformers.models.qwen2_vl.video_processing_qwen2_vl import Qwen2VLVideoProcessor from vllm.attention.backends.registry import AttentionBackendEnum from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention from vllm.config import MultiModalConfig, VllmConfig from vllm.config.multimodal import BaseDummyOptions from vllm.distributed import parallel_state, tensor_model_parallel_all_gather from vllm.distributed import utils as dist_utils from vllm.logger import init_logger from vllm.model_executor.layers.activation import QuickGELU from vllm.model_executor.layers.conv import Conv3dLayer from vllm.model_executor.layers.linear import ( ColumnParallelLinear, RowParallelLinear, ) from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.rotary_embedding.common import ( ApplyRotaryEmb, ) 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 from vllm.multimodal.inputs import ( ImageItem, ModalityData, MultiModalDataDict, MultiModalFeatureSpec, MultiModalFieldConfig, MultiModalKwargsItems, VideoItem, ) from vllm.multimodal.parse import ( DictEmbeddingItems, ImageSize, ModalityDataItems, MultiModalDataItems, MultiModalDataParser, ) from vllm.multimodal.processing import ( BaseMultiModalProcessor, BaseProcessingInfo, PromptReplacement, PromptUpdate, ) from vllm.multimodal.profiling import BaseDummyInputsBuilder from vllm.sequence import IntermediateTensors from vllm.tokenizers import TokenizerLike from vllm.utils.tensor_schema import TensorSchema, TensorShape from .interfaces import ( MultiModalEmbeddings, SupportsLoRA, SupportsMRoPE, SupportsMultiModal, SupportsPP, ) from .utils import ( AutoWeightsLoader, WeightsMapper, init_vllm_registered_model, maybe_prefix, ) from .vision import ( get_vit_attn_backend, run_dp_sharded_mrope_vision_model, ) logger = init_logger(__name__) # For profile run _MAX_FRAMES_PER_VIDEO = 14 # === Vision Inputs === # class Qwen2VLImagePixelInputs(TensorSchema): """ Dimensions: - np: The total number of patches over each image over each prompt in the batch - ni: Number of images - cps: Number of channels * patch_size * patch_size Historical context: - pixel_values shape: (num_patches, num_channels * patch_size * patch_size) - image_grid_thw shape: (num_images, 3) in (grid_t, grid_h, grid_w) format """ type: Literal["pixel_values"] pixel_values: Annotated[ torch.Tensor, TensorShape("np", "cps"), ] image_grid_thw: Annotated[ torch.Tensor, TensorShape("ni", 3), ] class Qwen2VLImageEmbeddingInputs(TensorSchema): """ Dimensions: - nf: Number of image features - hs: Hidden size - ni: Number of images Historical context: - image_embeds 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 shape: (num_images, 3) in (grid_t, grid_h, grid_w) format """ type: Literal["image_embeds"] image_embeds: Annotated[ torch.Tensor, TensorShape("nf", "hs"), ] image_grid_thw: Annotated[ torch.Tensor, TensorShape("ni", 3), ] Qwen2VLImageInputs: TypeAlias = Qwen2VLImagePixelInputs | Qwen2VLImageEmbeddingInputs class Qwen2VLVideoPixelInputs(TensorSchema): """ Dimensions: - np: The total number of patches over each video over each prompt in the batch - ctps: Number of channels * temporal_patch_size * patch_size * patch_size - nv: Number of videos Historical context: - pixel_values_videos shape: (num_patches, num_channels * temporal_patch_size * patch_size * patch_size) - video_grid_thw shape: (num_videos, 3) in (grid_t, grid_h, grid_w) format """ type: Literal["pixel_values_videos"] pixel_values_videos: Annotated[ torch.Tensor, TensorShape("np", "ctps"), ] video_grid_thw: Annotated[ torch.Tensor, TensorShape("nv", 3), ] class Qwen2VLVideoEmbeddingInputs(TensorSchema): """ Dimensions: - nf: Number of video features - hs: Hidden size - nv: Number of videos Historical context: - video_embeds shape: (num_video_features, hidden_size) - num_video_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 shape: (num_videos, 3) in (grid_t, grid_h, grid_w) format """ type: Literal["video_embeds"] video_embeds: Annotated[ torch.Tensor, TensorShape("nf", "hs"), ] video_grid_thw: Annotated[ torch.Tensor, TensorShape("nv", 3), ] Qwen2VLVideoInputs: TypeAlias = Qwen2VLVideoPixelInputs | Qwen2VLVideoEmbeddingInputs # === Vision Encoder === # class Qwen2VisionMLP(nn.Module): def __init__( self, in_features: int, hidden_features: int, act_layer: type[nn.Module] = QuickGELU, quant_config: QuantizationConfig | None = None, multimodal_config: MultiModalConfig | None = None, prefix: str = "", ): super().__init__() use_data_parallel = ( multimodal_config.mm_encoder_tp_mode == "data" if multimodal_config else False ) self.fc1 = ColumnParallelLinear( in_features, hidden_features, quant_config=quant_config, prefix=f"{prefix}.fc1", disable_tp=use_data_parallel, ) self.act = act_layer() self.fc2 = RowParallelLinear( hidden_features, in_features, quant_config=quant_config, prefix=f"{prefix}.fc2", disable_tp=use_data_parallel, ) 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 class Qwen2VisionAttention(nn.Module): def __init__( self, embed_dim: int, num_heads: int, projection_size: int, quant_config: QuantizationConfig | None = None, multimodal_config: MultiModalConfig | None = None, prefix: str = "", ) -> None: super().__init__() # Per attention head and per partition values. use_data_parallel = ( multimodal_config.mm_encoder_tp_mode == "data" if multimodal_config else False ) self.tp_size = ( 1 if use_data_parallel else parallel_state.get_tensor_model_parallel_world_size() ) self.tp_rank = parallel_state.get_tensor_model_parallel_rank() self.hidden_size_per_attention_head = dist_utils.divide( projection_size, num_heads ) self.num_attention_heads_per_partition = dist_utils.divide( num_heads, self.tp_size ) self.qkv = ColumnParallelLinear( input_size=embed_dim, output_size=3 * projection_size, quant_config=quant_config, prefix=f"{prefix}.qkv", disable_tp=use_data_parallel, ) self.proj = RowParallelLinear( input_size=projection_size, output_size=embed_dim, quant_config=quant_config, prefix=f"{prefix}.proj", disable_tp=use_data_parallel, ) self.attn = MMEncoderAttention( num_heads=self.num_attention_heads_per_partition, head_size=self.hidden_size_per_attention_head, multimodal_config=multimodal_config, ) self.apply_rotary_emb = ApplyRotaryEmb(enforce_enable=True) 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 def forward( self, x: torch.Tensor, cu_seqlens: torch.Tensor, rotary_pos_emb_cos: torch.Tensor, rotary_pos_emb_sin: torch.Tensor, max_seqlen: int | None = None, # Only used for Flash Attention ) -> torch.Tensor: # [s, b, c] --> [s, b, 3 * head * 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) q, k, v = (rearrange(x, "s b ... -> b s ...") for x in (q, k, v)) # [2 * b, s, heads, head_dim] qk_concat = torch.cat([q, k], dim=0) qk_rotated = self.apply_rotary_emb( qk_concat, rotary_pos_emb_cos, rotary_pos_emb_sin, ) q, k = torch.chunk(qk_rotated, 2, dim=0) 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() output, _ = self.proj(context_layer) return output class Qwen2VisionBlock(nn.Module): def __init__( self, dim: int, num_heads: int, mlp_ratio: float, act_layer: type[nn.Module] = QuickGELU, norm_layer: Callable[[int], nn.Module] | None = None, quant_config: QuantizationConfig | None = None, multimodal_config: MultiModalConfig | None = None, 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) mlp_hidden_dim = int(dim * mlp_ratio) self.attn = Qwen2VisionAttention( embed_dim=dim, num_heads=num_heads, projection_size=dim, quant_config=quant_config, multimodal_config=multimodal_config, prefix=f"{prefix}.attn", ) self.mlp = Qwen2VisionMLP( dim, mlp_hidden_dim, act_layer=act_layer, quant_config=quant_config, multimodal_config=multimodal_config, prefix=f"{prefix}.mlp", ) def forward( self, x: torch.Tensor, cu_seqlens: torch.Tensor, rotary_pos_emb_cos: torch.Tensor, rotary_pos_emb_sin: torch.Tensor, max_seqlen: int | None = None, # Only used for Flash Attention ) -> torch.Tensor: x = x + self.attn( self.norm1(x), cu_seqlens=cu_seqlens, rotary_pos_emb_cos=rotary_pos_emb_cos, rotary_pos_emb_sin=rotary_pos_emb_sin, max_seqlen=max_seqlen, ) 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, in_channels: int = 3, embed_dim: int = 1152, ) -> None: super().__init__() self.patch_size = patch_size self.temporal_patch_size = temporal_patch_size self.embed_dim = embed_dim kernel_size = (temporal_patch_size, patch_size, patch_size) self.proj = Conv3dLayer( in_channels, 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, norm_layer: Callable[[int], nn.Module] | None = None, spatial_merge_size: int = 2, quant_config: QuantizationConfig | None = None, multimodal_config: MultiModalConfig | None = None, prefix: str = "", ) -> None: super().__init__() use_data_parallel = ( multimodal_config.mm_encoder_tp_mode == "data" if multimodal_config else False ) 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, quant_config=quant_config, prefix=f"{prefix}.mlp.0", disable_tp=use_data_parallel, ), nn.GELU(), RowParallelLinear( self.hidden_size, d_model, bias=True, quant_config=quant_config, prefix=f"{prefix}.mlp.2", disable_tp=use_data_parallel, ), ] ) 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 Qwen2VisionTransformer(nn.Module): def __init__( self, vision_config: Qwen2VLVisionConfig, norm_eps: float = 1e-6, quant_config: QuantizationConfig | None = None, multimodal_config: MultiModalConfig | None = None, prefix: str = "", ) -> None: super().__init__() 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 self.use_data_parallel = ( multimodal_config.mm_encoder_tp_mode == "data" if multimodal_config else False ) self.out_hidden_size = vision_config.hidden_size self.spatial_merge_size = spatial_merge_size self.num_heads = num_heads self.embed_dim = embed_dim self.patch_embed = Qwen2VisionPatchEmbed( patch_size=patch_size, temporal_patch_size=temporal_patch_size, in_channels=in_channels, embed_dim=embed_dim, ) norm_layer = partial(nn.LayerNorm, eps=norm_eps) head_dim = embed_dim // num_heads self.rotary_pos_emb = get_rope( head_size=head_dim, max_position=8192, is_neox_style=True, rope_parameters={"partial_rotary_factor": 0.5}, ) self.blocks = nn.ModuleList( [ 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}", multimodal_config=multimodal_config, ) for layer_idx in range(depth) ] ) self.merger = Qwen2VisionPatchMerger( d_model=hidden_size, context_dim=embed_dim, norm_layer=norm_layer, quant_config=quant_config, prefix=f"{prefix}.merger", multimodal_config=multimodal_config, ) attn_backend_override = ( multimodal_config.mm_encoder_attn_backend if multimodal_config else None ) self.attn_backend = get_vit_attn_backend( head_size=head_dim, dtype=torch.get_default_dtype(), attn_backend_override=attn_backend_override, ) @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 def rot_pos_emb( self, grid_thw: list[list[int]] ) -> tuple[torch.Tensor, torch.Tensor]: pos_ids = [] max_grid_size = 0 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)) max_grid_size = max(max_grid_size, h, w) pos_ids = torch.cat(pos_ids, dim=0) # Use pre-computed cos_sin_cache from RotaryEmbedding cos, sin = self.rotary_pos_emb.get_cos_sin(max_grid_size) cos_combined = cos[pos_ids].flatten(1) sin_combined = sin[pos_ids].flatten(1) return cos_combined, sin_combined def compute_attn_mask_seqlen(self, cu_seqlens: torch.Tensor) -> int | None: max_seqlen = None if self.attn_backend in { AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.ROCM_AITER_FA, }: max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max() return max_seqlen def forward( self, x: torch.Tensor, grid_thw: torch.Tensor | list[list[int]], ) -> torch.Tensor: # patchify x = x.to(device=self.device, dtype=self.dtype) x = self.patch_embed(x) if isinstance(grid_thw, list): grid_thw_list = grid_thw grid_thw = np.array(grid_thw, dtype=np.int32) else: grid_thw_list = grid_thw.tolist() grid_thw = grid_thw.numpy() # compute position embedding rotary_pos_emb_cos, rotary_pos_emb_sin = self.rot_pos_emb(grid_thw_list) # compute cu_seqlens cu_seqlens = np.repeat(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum( axis=0, dtype=np.int32 ) cu_seqlens = np.concatenate([np.zeros(1, dtype=np.int32), cu_seqlens]) cu_seqlens = torch.from_numpy(cu_seqlens) # transformers x = x.unsqueeze(1) # pre-compute seqlens for attn mask to reduce cuMemcpy operations max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens) cu_seqlens = cu_seqlens.to(self.device, non_blocking=True) for blk in self.blocks: x = blk( x, cu_seqlens=cu_seqlens, rotary_pos_emb_cos=rotary_pos_emb_cos, rotary_pos_emb_sin=rotary_pos_emb_sin, max_seqlen=max_seqlen, ) # adapter x = self.merger(x) return x 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 def _create_qwen2vl_field_factory( spatial_merge_size: int, ) -> Callable[ [Mapping[str, torch.Tensor]], Mapping[str, MultiModalFieldConfig], ]: def _qwen2vl_field_config(hf_inputs: Mapping[str, torch.Tensor]): image_grid_thw = hf_inputs.get("image_grid_thw", torch.empty((0, 3))) image_pixel_grid_sizes = image_grid_thw.prod(-1) image_embed_grid_sizes = ( image_pixel_grid_sizes // spatial_merge_size // spatial_merge_size ) video_grid_thw = hf_inputs.get("video_grid_thw", torch.empty((0, 3))) video_grid_sizes = video_grid_thw.prod(-1) video_embed_grid_sizes = ( video_grid_sizes // spatial_merge_size // spatial_merge_size ) return dict( pixel_values=MultiModalFieldConfig.flat_from_sizes( "image", image_pixel_grid_sizes ), image_embeds=MultiModalFieldConfig.flat_from_sizes( "image", image_embed_grid_sizes ), image_grid_thw=MultiModalFieldConfig.batched("image", keep_on_cpu=True), pixel_values_videos=MultiModalFieldConfig.flat_from_sizes( "video", video_grid_sizes ), video_embeds=MultiModalFieldConfig.flat_from_sizes( "video", video_embed_grid_sizes ), video_grid_thw=MultiModalFieldConfig.batched("video", keep_on_cpu=True), ) return _qwen2vl_field_config class Qwen2VLMultiModalDataParser(MultiModalDataParser): def __init__(self, spatial_merge_size: int, *args, **kwargs): self._spatial_merge_size = spatial_merge_size super().__init__(*args, **kwargs) def _parse_image_data( self, data: dict[str, torch.Tensor] | ModalityData[ImageItem], ) -> ModalityDataItems[Any, Any] | None: if isinstance(data, dict): return DictEmbeddingItems( data, modality="image", required_fields={"image_embeds", "image_grid_thw"}, fields_factory=_create_qwen2vl_field_factory(self._spatial_merge_size), ) return super()._parse_image_data(data) def _parse_video_data( self, data: dict[str, torch.Tensor] | ModalityData[VideoItem], ) -> ModalityDataItems[Any, Any] | None: if isinstance(data, dict): return DictEmbeddingItems( data, modality="video", required_fields={"video_embeds", "video_grid_thw"}, fields_factory=_create_qwen2vl_field_factory(self._spatial_merge_size), ) return super()._parse_video_data(data) class Qwen2VLProcessingInfo(BaseProcessingInfo): def get_hf_config(self): return self.ctx.get_hf_config(Qwen2VLConfig) def get_hf_processor(self, **kwargs: object) -> Qwen2VLProcessor: return self.ctx.get_hf_processor( Qwen2VLProcessor, use_fast=kwargs.pop("use_fast", True), **kwargs, ) def get_image_processor(self, **kwargs: object) -> Qwen2VLImageProcessor: return self.get_hf_processor(**kwargs).image_processor def get_supported_mm_limits(self) -> Mapping[str, int | None]: return {"image": None, "video": None} def get_mm_max_tokens_per_item( self, seq_len: int, mm_counts: Mapping[str, int], ) -> Mapping[str, int]: max_image_tokens = self.get_max_image_tokens() max_video_tokens = self.get_max_video_tokens(seq_len, mm_counts) return {"image": max_image_tokens, "video": max_video_tokens} def _get_vision_info( self, *, image_width: int, image_height: int, num_frames: int = 1, do_resize: bool = True, image_processor: Qwen2VLImageProcessor | None, ) -> tuple[ImageSize, int]: if image_processor is None: image_processor = self.get_image_processor() 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( 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) # 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 def get_num_image_tokens( self, *, image_width: int, image_height: int, image_processor: Qwen2VLImageProcessor | None, ) -> int: _, num_image_tokens = self._get_vision_info( image_width=image_width, image_height=image_height, num_frames=1, image_processor=image_processor, ) return num_image_tokens def get_num_video_tokens( self, *, image_width: int, image_height: int, num_frames: int, image_processor: Qwen2VLImageProcessor | None, ) -> int: _, num_video_tokens = self._get_vision_info( image_width=image_width, image_height=image_height, num_frames=num_frames, image_processor=image_processor, ) return num_video_tokens def get_image_size_with_most_features(self) -> ImageSize: # NOTE: Simply processing a huge size with _get_vision_info might not give a # size that maximizes the number of featrues, i.e., the number of (merged) # patches. This is because the number of patches limits the allowed aspect # ratios. For example, suppose the maximum number of patches is 1280. A square # image cannot be broken down into 1280 patches, so feeding a giant square image # into _get_vision_info will not yield a size that maximizes the number of # patches. Therefore, we directly factorize the maximum number of patches into # height and width. The tricky part is to avoid extreme aspect ratios (>200 for # qwen2-vl). If we can't find a suitable aspect ratio, we decrease the number of # patches and retry. This is safe because the processor does not accept extreme # aspect ratios, so there is no valid post-resize image with the number of # patches that yields extreme aspect ratios. 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 image_processor = self.get_image_processor() max_pixels = image_processor.max_pixels or image_processor.size["longest_edge"] unit = patch_size * merge_size max_seq_len = max_pixels // (unit * unit) def closest_factor_pair(n: int) -> tuple[int, int]: # left <= right for d in range(math.isqrt(n), 0, -1): if n % d == 0: return d, n // d return 1, n height_factor, width_factor = 1, max_seq_len for seq_len in range(max_seq_len, 0, -1): height_factor, width_factor = closest_factor_pair(seq_len) if width_factor / height_factor <= 200: break return ImageSize(width=unit * width_factor, height=unit * height_factor) 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, image_processor=None, ) def _get_max_video_frames(self, max_tokens: int, start_num_frames: int = 1) -> int: target_width, target_height = self.get_image_size_with_most_features() num_frames = start_num_frames 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, image_processor=None, ) if next_max_tokens > max_tokens: 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], max_frames_per_video: int = _MAX_FRAMES_PER_VIDEO, ) -> int: max_videos = mm_counts.get("video", 0) max_total_frames = self._get_max_video_frames(seq_len) max_frames_per_video = min( max_total_frames // max(max_videos, 1), max_frames_per_video ) return max(max_frames_per_video, 1) def get_max_video_tokens( self, seq_len: int, mm_counts: Mapping[str, int], ) -> int: target_width, target_height = self.get_image_size_with_most_features() return self.get_num_video_tokens( image_width=target_width, image_height=target_height, num_frames=self.get_num_frames_with_most_features(seq_len, mm_counts), image_processor=None, ) class Qwen2VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen2VLProcessingInfo]): 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_processor = self.info.get_hf_processor() image_token: str = hf_processor.image_token video_token: str = hf_processor.video_token return image_token * num_images + video_token * num_videos def get_dummy_mm_data( self, seq_len: int, mm_counts: Mapping[str, int], mm_options: Mapping[str, BaseDummyOptions] | None = None, ) -> MultiModalDataDict: num_images = mm_counts.get("image", 0) num_videos = mm_counts.get("video", 0) target_width, target_height = self.info.get_image_size_with_most_features() target_num_frames = self.info.get_num_frames_with_most_features( seq_len, mm_counts ) image_overrides = mm_options.get("image") if mm_options else None video_overrides = mm_options.get("video") if mm_options else None return { "image": self._get_dummy_images( width=target_width, height=target_height, num_images=num_images, overrides=image_overrides, ), "video": self._get_dummy_videos( width=target_width, height=target_height, num_frames=target_num_frames, num_videos=num_videos, overrides=video_overrides, ), } class Qwen2VLMultiModalProcessor(BaseMultiModalProcessor[Qwen2VLProcessingInfo]): def _get_data_parser(self) -> MultiModalDataParser: return Qwen2VLMultiModalDataParser( self.info.get_hf_config().vision_config.spatial_merge_size ) def _get_prompt_updates( self, mm_items: MultiModalDataItems, hf_processor_mm_kwargs: Mapping[str, Any], out_mm_kwargs: MultiModalKwargsItems, ) -> Sequence[PromptUpdate]: hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) image_processor = self.info.get_image_processor(**hf_processor_mm_kwargs) tokenizer = self.info.get_tokenizer() vocab = tokenizer.get_vocab() placeholder = { "image": vocab[hf_processor.image_token], "video": vocab[hf_processor.video_token], } merge_length = image_processor.merge_size**2 def get_replacement_qwen2vl(item_idx: int, modality: str): out_item = out_mm_kwargs[modality][item_idx] grid_thw = out_item[f"{modality}_grid_thw"].data assert isinstance(grid_thw, torch.Tensor) num_tokens = int(grid_thw.prod()) // merge_length return [placeholder[modality]] * num_tokens return [ PromptReplacement( modality=modality, target=[placeholder[modality]], replacement=partial(get_replacement_qwen2vl, modality=modality), ) for modality in ("image", "video") ] def _get_mm_fields_config( self, hf_inputs: BatchFeature, hf_processor_mm_kwargs: Mapping[str, object], ) -> Mapping[str, MultiModalFieldConfig]: return _create_qwen2vl_field_factory( self.info.get_hf_config().vision_config.spatial_merge_size )(hf_inputs) @MULTIMODAL_REGISTRY.register_processor( Qwen2VLMultiModalProcessor, info=Qwen2VLProcessingInfo, dummy_inputs=Qwen2VLDummyInputsBuilder, ) class Qwen2VLForConditionalGeneration( nn.Module, SupportsMultiModal, SupportsLoRA, SupportsPP, SupportsMRoPE ): # To ensure correct weight loading and mapping. hf_to_vllm_mapper = WeightsMapper( orig_to_new_prefix={ # mapping for new names in checkpoint saved after transformers v4.52 "model.language_model.": "language_model.model.", "model.visual.": "visual.", # mapping for original checkpoint "lm_head.": "language_model.lm_head.", "model.": "language_model.model.", } ) supports_encoder_tp_data = True def get_mrope_input_positions( self, input_tokens: list[int], mm_features: list[MultiModalFeatureSpec], ) -> tuple[torch.Tensor, int]: kwargs = MultiModalFeatureSpec.gather_kwargs( mm_features, {"image_grid_thw", "video_grid_thw", "second_per_grid_ts"}, ) image_grid_thw = [item.tolist() for item in kwargs.get("image_grid_thw", [])] video_grid_thw = [item.tolist() for item in kwargs.get("video_grid_thw", [])] second_per_grid_ts = kwargs.get("second_per_grid_ts", []) hf_config = self.config image_token_id = hf_config.image_token_id video_token_id = hf_config.video_token_id vision_start_token_id = hf_config.vision_start_token_id spatial_merge_size = hf_config.vision_config.spatial_merge_size tokens_per_second = getattr(hf_config.vision_config, "tokens_per_second", 1.0) input_tokens_tensor = torch.tensor(input_tokens) vision_start_indices = torch.argwhere( input_tokens_tensor == vision_start_token_id ).squeeze(1) vision_tokens = input_tokens_tensor[vision_start_indices + 1] image_nums = (vision_tokens == image_token_id).sum() video_nums = (vision_tokens == video_token_id).sum() llm_pos_ids_list: list = [] st = 0 remain_images, remain_videos = image_nums, video_nums image_index, video_index = 0, 0 for _ in range(image_nums + video_nums): video_second_per_grid_t = 0.0 if remain_images > 0: try: ed_image = input_tokens.index(image_token_id, st) except ValueError: ed_image = len(input_tokens) + 1 else: ed_image = len(input_tokens) + 1 if remain_videos > 0: try: ed_video = input_tokens.index(video_token_id, st) except ValueError: ed_video = len(input_tokens) + 1 else: ed_video = len(input_tokens) + 1 if ed_image < ed_video: t, h, w = image_grid_thw[image_index] image_index += 1 remain_images -= 1 ed = ed_image else: t, h, w = video_grid_thw[video_index] video_second_per_grid_t = 1.0 if second_per_grid_ts: video_second_per_grid_t = second_per_grid_ts[video_index] video_index += 1 remain_videos -= 1 ed = ed_video llm_grid_t, llm_grid_h, llm_grid_w = ( t, h // spatial_merge_size, w // spatial_merge_size, ) text_len = ed - st st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0 llm_pos_ids_list.append( torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx ) t_index = ( ( torch.arange(llm_grid_t) .view(-1, 1) .expand(-1, llm_grid_h * llm_grid_w) * video_second_per_grid_t * tokens_per_second ) .long() .flatten() ) h_index = ( torch.arange(llm_grid_h) .view(1, -1, 1) .expand(llm_grid_t, -1, llm_grid_w) .flatten() ) w_index = ( torch.arange(llm_grid_w) .view(1, 1, -1) .expand(llm_grid_t, llm_grid_h, -1) .flatten() ) llm_pos_ids_list.append( torch.stack([t_index, h_index, w_index]) + text_len + st_idx ) st = ed + llm_grid_t * llm_grid_h * llm_grid_w if st < len(input_tokens): st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0 text_len = len(input_tokens) - st llm_pos_ids_list.append( torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx ) llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1) mrope_position_delta = (llm_positions.max() + 1 - len(input_tokens)).item() return llm_positions, mrope_position_delta @classmethod def get_placeholder_str(cls, modality: str, i: int) -> str | None: if modality.startswith("image"): return "<|vision_start|><|image_pad|><|vision_end|>" if modality.startswith("video"): return "<|vision_start|><|video_pad|><|vision_end|>" raise ValueError("Only image or video modality is supported") def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config: Qwen2VLConfig = vllm_config.model_config.hf_config quant_config = vllm_config.quant_config multimodal_config = vllm_config.model_config.multimodal_config self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data" self.config = config self.multimodal_config = multimodal_config if multimodal_config.get_limit_per_prompt( "image" ) or multimodal_config.get_limit_per_prompt("video"): self.visual = Qwen2VisionTransformer( config.vision_config, norm_eps=getattr(config, "rms_norm_eps", 1e-6), quant_config=quant_config, multimodal_config=multimodal_config, prefix=maybe_prefix(prefix, "visual"), ) else: self.visual = None self.language_model = init_vllm_registered_model( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "language_model"), architectures=["Qwen2ForCausalLM"], ) self.make_empty_intermediate_tensors = ( self.language_model.make_empty_intermediate_tensors ) def _parse_and_validate_image_input( self, **kwargs: object ) -> Qwen2VLImageInputs | None: 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 Qwen2VLImagePixelInputs( type="pixel_values", pixel_values=pixel_values, image_grid_thw=image_grid_thw, ) if image_embeds is not None: return Qwen2VLImageEmbeddingInputs( type="image_embeds", image_embeds=image_embeds, image_grid_thw=image_grid_thw, ) def _parse_and_validate_video_input( self, **kwargs: object ) -> Qwen2VLVideoInputs | None: 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) if pixel_values_videos is None and video_embeds is None: return None if pixel_values_videos is not None: return Qwen2VLVideoPixelInputs( type="pixel_values_videos", pixel_values_videos=pixel_values_videos, video_grid_thw=video_grid_thw, ) if video_embeds is not None: return Qwen2VLVideoEmbeddingInputs( type="video_embeds", video_embeds=video_embeds, video_grid_thw=video_grid_thw, ) def _process_image_input( self, image_input: Qwen2VLImageInputs ) -> tuple[torch.Tensor, ...]: grid_thw = image_input["image_grid_thw"] assert grid_thw.ndim == 2 if image_input["type"] == "image_embeds": image_embeds = image_input["image_embeds"] else: pixel_values = image_input["pixel_values"] if self.use_data_parallel: return run_dp_sharded_mrope_vision_model( self.visual, pixel_values, grid_thw.tolist(), rope_type="rope_3d" ) else: 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).tolist() return image_embeds.split(sizes) def _process_video_input( self, video_input: Qwen2VLVideoInputs ) -> tuple[torch.Tensor, ...]: grid_thw = video_input["video_grid_thw"] assert grid_thw.ndim == 2 if video_input["type"] == "video_embeds": video_embeds = video_input["video_embeds"] else: pixel_values_videos = video_input["pixel_values_videos"] if self.use_data_parallel: return run_dp_sharded_mrope_vision_model( self.visual, pixel_values_videos, grid_thw.tolist(), rope_type="rope_3d", ) else: video_embeds = self.visual(pixel_values_videos, grid_thw=grid_thw) # Split concatenated embeddings for each video item. merge_size = self.visual.spatial_merge_size sizes = (grid_thw.prod(-1) // merge_size // merge_size).tolist() return video_embeds.split(sizes) 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 def get_language_model(self) -> torch.nn.Module: return self.language_model def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: modalities = self._parse_and_validate_multimodal_inputs(**kwargs) if not modalities: return [] # 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"] image_embeddings = self._process_image_input(image_input) multimodal_embeddings += tuple(image_embeddings) if modality == "videos": video_input = modalities["videos"] video_embeddings = self._process_video_input(video_input) multimodal_embeddings += tuple(video_embeddings) return multimodal_embeddings def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, intermediate_tensors: IntermediateTensors | None = None, inputs_embeds: torch.Tensor | None = None, **kwargs: object, ) -> torch.Tensor | IntermediateTensors: """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,)`. intermediate_tensors: Intermediate tensors from prior forward pass. inputs_embeds: Optional tensor of input embeddings. """ 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, ) -> torch.Tensor | None: return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: skip_prefixes = [] if self.visual is None: skip_prefixes.extend(["visual."]) loader = AutoWeightsLoader(self, skip_prefixes=skip_prefixes) return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) def get_mm_mapping(self) -> MultiModelKeys: """ Get the module prefix in multimodal models """ return MultiModelKeys.from_string_field( language_model="language_model", connector="visual.merger.", tower_model="visual.", ) class Tarsier2MultiModalProcessor(Qwen2VLMultiModalProcessor): pass class Tarsier2ImageProcessor(Qwen2VLImageProcessor): def __init__( self, size: dict[str, int] | None = None, **kwargs, ) -> None: if size is not None and "min_pixels" in size and "max_pixels" in size: # Remap if Tarsier2-specific format is provided remapped_size = { "shortest_edge": size["min_pixels"], "longest_edge": size["max_pixels"], } super().__init__(size=remapped_size, **kwargs) else: super().__init__(size=size, **kwargs) class Tarsier2Processor(Qwen2VLProcessor): def __init__( self, vision_config: dict, tokenizer: TokenizerLike, **kwargs, ): self.image_processor = Tarsier2ImageProcessor(**vision_config) super().__init__( image_processor=self.image_processor, tokenizer=tokenizer, video_processor=Qwen2VLVideoProcessor(**vision_config), chat_template=None, **kwargs, ) class Tarsier2ProcessingInfo(Qwen2VLProcessingInfo): def get_hf_config(self) -> Qwen2VLConfig: model_path = self.ctx.model_config.model correct_config = Qwen2VLConfig.from_pretrained(model_path) return correct_config def get_hf_processor(self, **kwargs: object) -> Tarsier2Processor: return Tarsier2Processor( vision_config=self.ctx.get_hf_image_processor_config(), tokenizer=self.get_tokenizer(), **kwargs, ) def get_image_processor(self) -> Tarsier2ImageProcessor: return Tarsier2ImageProcessor(**self.ctx.get_hf_image_processor_config()) @MULTIMODAL_REGISTRY.register_processor( Tarsier2MultiModalProcessor, info=Tarsier2ProcessingInfo, dummy_inputs=Qwen2VLDummyInputsBuilder, ) class Tarsier2ForConditionalGeneration(Qwen2VLForConditionalGeneration): hf_to_vllm_mapper = WeightsMapper( orig_to_new_prefix={ "vision_tower.": "visual.", } ) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: skip_prefixes = [] if self.visual is None: skip_prefixes.extend(["visual."]) loader = AutoWeightsLoader(self, skip_prefixes=skip_prefixes) return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)