sgl-project--sglang/python/sglang/srt/models/qwen2_5_vl.py

# coding=utf-8
# 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 logging
import re
from functools import partial
from typing import Iterable, List, Optional, Tuple, Type

import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
from transformers.activations import ACT2FN
from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import (
    Qwen2_5_VLConfig,
    Qwen2_5_VLVisionConfig,
)
from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
    Qwen2_5_VisionPatchEmbed,
    Qwen2_5_VisionRotaryEmbedding,
)

from sglang.srt.distributed.parallel_state import get_pp_group
from sglang.srt.environ import envs
from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.attention.vision import VisionAttention
from sglang.srt.layers.layernorm import RMSNorm
from sglang.srt.layers.linear import (
    ColumnParallelLinear,
    MergedColumnParallelLinear,
    RowParallelLinear,
)
from sglang.srt.layers.logits_processor import LogitsProcessor
from sglang.srt.layers.pooler import Pooler, PoolingType
from sglang.srt.layers.quantization.base_config import QuantizationConfig
from sglang.srt.layers.utils import PPMissingLayer, get_layer_id
from sglang.srt.layers.vocab_parallel_embedding import ParallelLMHead
from sglang.srt.managers.mm_utils import (
    MultiModalityDataPaddingPatternMultimodalTokens,
    general_mm_embed_routine,
)
from sglang.srt.managers.schedule_batch import (
    Modality,
    MultimodalDataItem,
    MultimodalInputs,
)
from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
from sglang.srt.model_loader.weight_utils import default_weight_loader
from sglang.srt.models.qwen2 import Qwen2Model
from sglang.srt.models.utils import RotaryPosMixin, WeightsMapper, permute_inv
from sglang.srt.multimodal.mm_utils import run_dp_sharded_mrope_vision_model
from sglang.srt.multimodal.vit_cuda_graph_runner import ViTCudaGraphRunner
from sglang.srt.runtime_context import get_parallel, get_server_args
from sglang.srt.utils import add_prefix, is_cpu, is_cuda, is_npu

_is_cuda = is_cuda()
_is_cpu = is_cpu()

logger = logging.getLogger(__name__)


class Qwen2_5_VLMLP(nn.Module):
    def __init__(
        self,
        in_features: int,
        hidden_features: int = None,
        bias: bool = True,
        hidden_act="silu",
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        use_data_parallel: bool = False,
    ):
        super().__init__()
        self.tp_size = 1 if use_data_parallel else get_parallel().tp_size
        self.tp_rank = 0 if use_data_parallel else get_parallel().tp_rank
        self.gate_up_proj = MergedColumnParallelLinear(
            input_size=in_features,
            output_sizes=[hidden_features] * 2,  # [gate_proj, up_proj]
            bias=bias,
            quant_config=quant_config,
            prefix=add_prefix("gate_up_proj", prefix),
            tp_size=self.tp_size,
            tp_rank=self.tp_rank,
        )
        self.down_proj = RowParallelLinear(
            hidden_features,
            in_features,
            bias=bias,
            quant_config=quant_config,
            prefix=add_prefix("down_proj", prefix),
            tp_size=self.tp_size,
            tp_rank=self.tp_rank,
        )
        self.hidden_act = hidden_act
        if self.hidden_act == "silu":
            self.act = SiluAndMul()
        else:
            base_act = ACT2FN[self.hidden_act]

            def _act_fn(x: torch.Tensor) -> torch.Tensor:
                gate, up = x.chunk(2, dim=-1)
                return base_act(gate) * up

            self.act = _act_fn

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        gate_up, _ = self.gate_up_proj(x)
        x = self.act(gate_up)
        x_down, _ = self.down_proj(x)
        return x_down


class Qwen2_5_VisionBlock(nn.Module):

    def __init__(
        self,
        dim: int,
        intermediate_dim: int,
        num_heads: int,
        head_size: int,
        hidden_act="silu",
        norm_layer: Type[nn.Module] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        num_dummy_heads: int = 0,
        rms_norm_eps: float = 1e-6,
        use_data_parallel: bool = False,
    ) -> None:
        super().__init__()
        self.norm1 = RMSNorm(dim, eps=rms_norm_eps)
        self.norm2 = RMSNorm(dim, eps=rms_norm_eps)

        self.attn = VisionAttention(
            embed_dim=dim,
            num_heads=num_heads,
            head_size=head_size,
            projection_size=num_heads * head_size,
            use_qkv_parallel=True,
            proj_bias=True,
            flatten_batch=True,
            quant_config=quant_config,
            prefix=add_prefix("attn", prefix),
            num_dummy_heads=num_dummy_heads,
            use_data_parallel=use_data_parallel,
        )
        self.mlp = Qwen2_5_VLMLP(
            dim,
            intermediate_dim,
            hidden_act=hidden_act,
            quant_config=quant_config,
            prefix=add_prefix("mlp", prefix),
            use_data_parallel=use_data_parallel,
        )

    def forward(
        self,
        x: torch.Tensor,
        cu_seqlens: torch.Tensor,
        position_embeddings: torch.Tensor,
        output_ws=None,
    ) -> torch.Tensor:
        S, B, H = x.shape
        # norm1: flatten to 2D -> [S*B, H], then reshape back
        x2d = x.reshape(-1, H)
        hidden_states = self.norm1(x2d).reshape(S, B, H)

        # Attention expects [B, S, H]
        hidden_states = rearrange(hidden_states, "s b h -> b s h")
        attn = self.attn(
            hidden_states,
            cu_seqlens=cu_seqlens,
            position_embeddings=position_embeddings,
            output_ws=output_ws,
        )
        attn = rearrange(attn, "b s h -> s b h")

        # norm2 with fused residual-add: also 2D
        attn2d = attn.reshape(-1, H)
        x_norm_2d, x_after_add_2d = self.norm2(x2d, residual=attn2d)
        x_norm = x_norm_2d.reshape(S, B, H)
        x_after_add = x_after_add_2d.reshape(S, B, H)

        # MLP and final residual
        mlp_out = self.mlp(x_norm)
        x = x_after_add + mlp_out
        return x


class Qwen2_5_VisionPatchMerger(nn.Module):

    def __init__(
        self,
        dim: int,
        context_dim: int,
        padded_context_dim: int,
        spatial_merge_size: int = 2,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        use_data_parallel: bool = False,
    ) -> None:
        super().__init__()
        self.hidden_size = context_dim * (spatial_merge_size**2)
        self.padded_context_dim = padded_context_dim * (spatial_merge_size**2)
        self.ln_q = RMSNorm(context_dim, eps=1e-6)
        tp_size = 1 if use_data_parallel else get_parallel().tp_size
        tp_rank = 0 if use_data_parallel else get_parallel().tp_rank
        self.mlp = nn.ModuleList(
            [
                ColumnParallelLinear(
                    self.hidden_size,
                    self.padded_context_dim,
                    bias=True,
                    quant_config=quant_config,
                    prefix=add_prefix("mlp.0", prefix),
                    tp_size=tp_size,
                    tp_rank=tp_rank,
                ),
                nn.GELU(),
                RowParallelLinear(
                    self.padded_context_dim,
                    dim,
                    bias=True,
                    quant_config=quant_config,
                    prefix=add_prefix("mlp.2", prefix),
                    tp_size=tp_size,
                    tp_rank=tp_rank,
                ),
            ]
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        # x expected shape: [S, B, context_dim]
        S, B, D = x.shape
        x2d = x.reshape(-1, D)
        x2d = self.ln_q(x2d)  # RMSNorm expects 2D
        x2d = x2d.view(-1, self.hidden_size)  # group into spatial_merge_unit
        mlp_fc1, mlp_act, mlp_fc2 = self.mlp
        x_parallel, _ = mlp_fc1(x2d)
        x_parallel = mlp_act(x_parallel)
        out, _ = mlp_fc2(x_parallel)
        return out


class Qwen2_5_VisionTransformer(nn.Module, RotaryPosMixin):

    def __init__(
        self,
        vision_config: Qwen2_5_VLVisionConfig,
        norm_eps: float = 1e-6,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
        use_data_parallel: bool = False,
        max_context_len: Optional[int] = None,
    ) -> None:
        super().__init__()

        patch_size: int = vision_config.patch_size
        temporal_patch_size: int = vision_config.temporal_patch_size
        spatial_merge_size: int = vision_config.spatial_merge_size
        self.spatial_merge_size = spatial_merge_size
        self.spatial_merge_unit: int = spatial_merge_size * spatial_merge_size
        in_channels: int = vision_config.in_channels
        hidden_size: int = vision_config.hidden_size
        depth: int = vision_config.depth
        num_heads: int = vision_config.num_heads
        self.fullatt_block_indexes = vision_config.fullatt_block_indexes
        self.window_size = vision_config.window_size
        self.patch_size = vision_config.patch_size
        mlp_hidden_size: int = ((vision_config.intermediate_size + 7) // 8) * 8
        self.use_data_parallel = use_data_parallel
        self.out_hidden_size = vision_config.out_hidden_size
        self.patch_embed = Qwen2_5_VisionPatchEmbed(
            patch_size=patch_size,
            temporal_patch_size=temporal_patch_size,
            in_channels=in_channels,
            embed_dim=hidden_size,
        )

        norm_layer = partial(nn.LayerNorm, eps=norm_eps)
        if _is_cpu and hasattr(vision_config, "original_num_heads"):
            head_dim = hidden_size // vision_config.original_num_heads
        else:
            head_dim = hidden_size // num_heads
        self.rotary_pos_emb = Qwen2_5_VisionRotaryEmbedding(head_dim // 2)
        self.blocks = nn.ModuleList(
            [
                Qwen2_5_VisionBlock(
                    dim=hidden_size,
                    intermediate_dim=mlp_hidden_size,
                    num_heads=num_heads,
                    head_size=head_dim,
                    hidden_act=vision_config.hidden_act,
                    norm_layer=norm_layer,
                    quant_config=quant_config,
                    prefix=add_prefix(f"blocks.{i}", prefix),
                    use_data_parallel=use_data_parallel,
                )
                for i in range(depth)
            ]
        )
        self.merger = Qwen2_5_VisionPatchMerger(
            dim=vision_config.out_hidden_size,
            context_dim=hidden_size,
            padded_context_dim=num_heads * head_dim,
            spatial_merge_size=spatial_merge_size,
            quant_config=quant_config,
            prefix=add_prefix("merger", prefix),
            use_data_parallel=use_data_parallel,
        )

        # Resource prepared for vit cuda graph
        self.tp_size = 1 if use_data_parallel else get_parallel().tp_size
        self.max_context_len = max_context_len
        self.enable_cg = _is_cuda and envs.SGLANG_VIT_ENABLE_CUDA_GRAPH.get()

        self.cuda_graph_runner: Optional[ViTCudaGraphRunner] = None
        if self.enable_cg:
            self.cuda_graph_runner = ViTCudaGraphRunner(self)

    def get_window_index(self, grid_thw):
        cu_window_seqlens: list = [0]
        window_index_id = 0
        vit_merger_window_size = (
            self.window_size // self.spatial_merge_size // self.patch_size
        )
        window_index: list = []
        for grid_t, grid_h, grid_w in grid_thw:
            llm_grid_h, llm_grid_w = (
                grid_h // self.spatial_merge_size,
                grid_w // self.spatial_merge_size,
            )
            index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(
                grid_t, llm_grid_h, llm_grid_w
            )
            pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
            pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
            num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
            num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
            index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
            index_padded = index_padded.reshape(
                grid_t,
                num_windows_h,
                vit_merger_window_size,
                num_windows_w,
                vit_merger_window_size,
            )
            index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
                grid_t,
                num_windows_h * num_windows_w,
                vit_merger_window_size,
                vit_merger_window_size,
            )
            seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
            index_padded = index_padded.reshape(-1)
            index_new = index_padded[index_padded != -100]
            window_index.append(index_new + window_index_id)
            cu_seqlens_tmp = (
                seqlens.cumsum(0) * self.spatial_merge_unit + cu_window_seqlens[-1]
            )
            cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
            window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()
        window_index = torch.cat(window_index, dim=0)
        return window_index, cu_window_seqlens

    @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: torch.Tensor) -> torch.Tensor:
        pos_ids = []
        for t, h, w in grid_thw:
            base = self.rot_pos_ids(h, w, self.spatial_merge_size)
            pos_ids.append(base if t == 1 else base.repeat(t, 1))

        pos_ids = torch.cat(pos_ids, dim=0)
        max_grid_size = int(grid_thw[:, 1:].max())
        # transformers 5.12's rotary forward takes 1-D position_ids on the input device (grid_thw is CPU).
        rotary_pos_emb_full = self.rotary_pos_emb(
            torch.arange(max_grid_size, device=self.device)
        )
        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
        return rotary_pos_emb

    def forward(
        self,
        x: torch.Tensor,
        grid_thw: torch.Tensor,
    ) -> torch.Tensor:
        if self.enable_cg:
            return self.forward_with_cuda_graph(x, grid_thw)

        # patchify
        x = x.to(device=self.device, dtype=self.dtype)
        x = self.patch_embed(x)

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

        window_index, cu_window_seqlens = self.get_window_index(grid_thw)
        cu_window_seqlens = torch.tensor(
            cu_window_seqlens,
            device=x.device,
            dtype=torch.int32,
        )
        cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)

        # Move window_index to the same device as x before using it to index x
        window_index = window_index.to(device=x.device)
        reverse_indices = permute_inv(window_index)

        # Ensure rotary_pos_emb is on the same device/dtype as x
        rotary_pos_emb = rotary_pos_emb.to(device=x.device, dtype=x.dtype)

        seq_len, _ = x.size()

        x = x.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        x = x[window_index, :, :]
        x = x.reshape(seq_len, -1)
        rotary_pos_emb = rotary_pos_emb.reshape(
            seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1
        )
        rotary_pos_emb = rotary_pos_emb[window_index, :, :]
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
        position_embeddings = (emb.cos(), emb.sin())
        # After building position_embeddings, make sure both cos and sin are on the same device/dtype as the attention input
        position_embeddings = (
            position_embeddings[0].to(x.device, x.dtype),
            position_embeddings[1].to(x.device, x.dtype),
        )

        # compute cu_seqlens - move cu_seqlens to GPU and make it int32
        cu_seqlens = torch.cat(
            [
                torch.tensor([0], device=x.device, dtype=torch.int32),
                (grid_thw[:, 0] * grid_thw[:, 1] * grid_thw[:, 2])
                .cumsum(dim=0)
                .to(device=x.device, dtype=torch.int32),
            ]
        )
        cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
        # cu_seqlens must be on cpu because of npu_flash_attention_unpad operator restriction
        if is_npu():
            cu_seqlens = cu_seqlens.to("cpu")
            cu_window_seqlens = cu_window_seqlens.to("cpu")
        # transformers
        x = x.unsqueeze(1)
        for layer_num, blk in enumerate(self.blocks):
            fullatt_indexes = self.fullatt_block_indexes
            if isinstance(fullatt_indexes, torch.Tensor):
                fullatt_indexes = fullatt_indexes.tolist()
            if layer_num in fullatt_indexes:
                cu_seqlens_now = cu_seqlens
            else:
                cu_seqlens_now = cu_window_seqlens
            x = blk(
                x, cu_seqlens=cu_seqlens_now, position_embeddings=position_embeddings
            )

        # adapter
        x = self.merger(x)
        x = x[reverse_indices, :]

        return x

    def forward_with_cuda_graph(
        self,
        x: torch.Tensor,
        grid_thw: torch.Tensor,
    ) -> torch.Tensor:
        # patchify
        x = x.to(device=self.device, dtype=self.dtype)
        x = self.patch_embed(x)

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

        window_index, cu_window_seqlens = self.get_window_index(grid_thw)
        cu_window_seqlens = torch.tensor(
            cu_window_seqlens,
            device=x.device,
            dtype=torch.int32,
        )
        cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)

        window_index = window_index.to(device=x.device)
        reverse_indices = permute_inv(window_index)
        rotary_pos_emb = rotary_pos_emb.to(device=x.device, dtype=x.dtype)

        # patch token num
        seq_len, _ = x.size()

        # [G, M, hidden]
        x = x.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
        x = x[window_index, :, :]  # [G, M, hidden]
        x = x.reshape(seq_len, -1)  # [seq_len, hidden]

        rotary_pos_emb = rotary_pos_emb.reshape(
            seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1
        )
        rotary_pos_emb = rotary_pos_emb[window_index, :, :]
        rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)

        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
        position_embeddings = (emb.cos(), emb.sin())
        # After building position_embeddings, make sure both cos and sin are on
        # the same device/dtype as the attention input
        position_embeddings = (
            position_embeddings[0].to(x.device, x.dtype),
            position_embeddings[1].to(x.device, x.dtype),
        )

        # compute cu_seqlens - move cu_seqlens to GPU and make it int32
        cu_seqlens = torch.cat(
            [
                torch.tensor([0], device=x.device, dtype=torch.int32),
                (grid_thw[:, 0] * grid_thw[:, 1] * grid_thw[:, 2])
                .cumsum(dim=0)
                .to(device=x.device, dtype=torch.int32),
            ]
        )
        cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])

        return self.cuda_graph_runner.run(
            x=x,
            position_embeddings=position_embeddings,
            cu_seqlens=cu_seqlens,
            cu_window_seqlens=cu_window_seqlens,
            output_indices=reverse_indices,
        )


class Qwen2_5_VLForConditionalGeneration(nn.Module):
    # BitandBytes specific attributes
    default_bitsandbytes_target_modules = [
        ".gate_up_proj.",
        ".down_proj.",
        ".q_proj.",
        ".k_proj.",
        ".v_proj.",
        ".o_proj.",
    ]
    bitsandbytes_stacked_params_mapping = {
        # shard_name, weight_name, index
        "q_proj": ("qkv_proj", 0),
        "k_proj": ("qkv_proj", 1),
        "v_proj": ("qkv_proj", 2),
        "gate_proj": ("gate_up_proj", 0),
        "up_proj": ("gate_up_proj", 1),
    }

    packed_modules_mapping = {
        "gate_up_proj": ["gate_proj", "up_proj"],
    }
    # To ensure correct weight loading and mapping.
    hf_to_sglang_mapper = WeightsMapper(
        orig_to_new_substr={
            "attn.qkv": "attn.qkv_proj",
        },
        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.",
        },
    )

    def __init__(
        self,
        config: Qwen2_5_VLConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()

        self.pp_group = get_pp_group()
        self.config = config
        self.use_data_parallel = get_server_args().mm_enable_dp_encoder

        if not self.config.encoder_only:
            self.model = Qwen2Model(
                config,
                quant_config,
                prefix=add_prefix("model", prefix),
            )

            if self.pp_group.is_last_rank:
                if self.pp_group.world_size == 1 and self.config.tie_word_embeddings:
                    self.lm_head = self.model.embed_tokens
                else:
                    self.lm_head = ParallelLMHead(
                        self.config.vocab_size,
                        self.config.hidden_size,
                        quant_config=quant_config,
                        prefix=add_prefix("lm_head", prefix),
                    )
            else:
                # ranks other than the last rank will have a placeholder layer
                self.lm_head = PPMissingLayer()
        else:
            # encoder_only mode: no language model, so no lm_head needed
            self.lm_head = None

        self.visual = Qwen2_5_VisionTransformer(
            config.vision_config,
            norm_eps=getattr(config, "rms_norm_eps", 1e-6),
            # NOTE: Qwen2_5-VL vision encoder currently supports BitsAndBytes 4-bit quantization.
            # Other quantization methods (e.g., GPTQ, AWQ) are untested and may not be supported.
            quant_config=quant_config,
            prefix=add_prefix("visual", prefix),
            use_data_parallel=self.use_data_parallel,
            max_context_len=self.config.max_position_embeddings,
        )

        self.is_mrope_enabled = "mrope_section" in self.config.rope_scaling

        self.logits_processor = LogitsProcessor(config)
        self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)

        # For EAGLE3 support
        self.capture_aux_hidden_states = False

    def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
        pattern = MultiModalityDataPaddingPatternMultimodalTokens()
        return pattern.pad_input_tokens(input_ids, mm_inputs)

    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
        # in qwen-vl, last dim is the same
        pixel_values = torch.cat([item.feature for item in items], dim=0).type(
            self.visual.dtype
        )
        image_grid_thw = torch.concat([item.image_grid_thw for item in items], dim=0)

        expected_dim = getattr(self.visual, "embed_dim", -1)

        if expected_dim == -1:
            vision_conf = self.config.vision_config
            expected_dim = getattr(
                vision_conf, "embed_dim", getattr(vision_conf, "hidden_size", -1)
            )

        raw_patch_dim = 1176

        if pixel_values.dim() == 2:
            current_dim = pixel_values.shape[-1]
            if current_dim == expected_dim:
                return pixel_values
            if current_dim != raw_patch_dim:

                return pixel_values

        assert pixel_values.dim() == 2, pixel_values.dim()
        assert image_grid_thw.dim() == 2, image_grid_thw.dim()
        if self.use_data_parallel:
            return run_dp_sharded_mrope_vision_model(
                self.visual, pixel_values, image_grid_thw.tolist(), rope_type="rope_3d"
            )
        else:
            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
        return image_embeds

    _lora_pattern = re.compile(
        r"^model\.layers\.(\d+)\.(?:self_attn|mlp)\.(?:qkv_proj|o_proj|down_proj|gate_up_proj)$"
    )

    def should_apply_lora(self, module_name: str) -> bool:
        return bool(self._lora_pattern.match(module_name))

    def get_video_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
        # in qwen-vl, last dim is the same
        pixel_values = torch.cat([item.feature for item in items], dim=0).type(
            self.visual.dtype
        )
        video_grid_thw = torch.concat([item.video_grid_thw for item in items], dim=0)
        assert pixel_values.dim() == 2, pixel_values.dim()
        assert video_grid_thw.dim() == 2, video_grid_thw.dim()
        if self.use_data_parallel:
            return run_dp_sharded_mrope_vision_model(
                self.visual, pixel_values, video_grid_thw.tolist(), rope_type="rope_3d"
            )
        else:
            video_embeds = self.visual(pixel_values, grid_thw=video_grid_thw)
        return video_embeds

    def post_process(
        self,
        inputs_embeds,
        modalities: List[Modality],
        embeddings: List[torch.Tensor],
        indices: List[torch.Tensor],
        forward_batch: ForwardBatch,
    ) -> torch.Tensor:
        # Placeholder for post_process
        new_embeddings = []
        for i, (modality, embedding, index) in enumerate(
            zip(modalities, embeddings, indices)
        ):
            if embedding is None or index is None:
                continue

            new_embeddings.append(embedding)
        return new_embeddings, forward_batch

    def get_input_embeddings(self):
        return self.model.embed_tokens

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds=None,
        get_embedding: bool = False,
        pp_proxy_tensors: Optional[PPProxyTensors] = None,
    ):
        """Run forward pass for Qwen2_5-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,).
                (Use input_metadata.mrope_positions to replace it)
        """
        if self.is_mrope_enabled:
            positions = forward_batch.mrope_positions

        if not (
            forward_batch.forward_mode.is_decode()
            or not forward_batch.contains_image_inputs()
        ):
            if self.is_mrope_enabled:
                assert positions.ndim == 2 and positions.size(0) == 3, (
                    "multimodal section rotary embedding requires "
                    f"(3, seq_len) positions, but got {positions.size()}"
                )

        hidden_states = general_mm_embed_routine(
            input_ids=input_ids,
            forward_batch=forward_batch,
            language_model=self.model,
            multimodal_model=self,
            positions=positions,
            pp_proxy_tensors=pp_proxy_tensors,
        )

        aux_hidden_states = None
        if self.capture_aux_hidden_states:
            hidden_states, aux_hidden_states = hidden_states

        if self.pp_group.is_last_rank:
            if not get_embedding:
                return self.logits_processor(
                    input_ids,
                    hidden_states,
                    self.lm_head,
                    forward_batch,
                    aux_hidden_states,
                )
            else:
                return self.pooler(hidden_states, forward_batch)
        else:
            return hidden_states

    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            (".qkv_proj", ".q_proj", "q"),
            (".qkv_proj", ".k_proj", "k"),
            (".qkv_proj", ".v_proj", "v"),
            ("gate_up_proj", "up_proj", 1),
            ("gate_up_proj", "gate_proj", 0),
        ]
        params_dict = dict(self.named_parameters(remove_duplicate=False))
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue

            if (
                self.config.tie_word_embeddings
                and self.pp_group.is_last_rank
                and "model.embed_tokens.weight" in name
            ):
                if "lm_head.weight" in params_dict:
                    lm_head_param = params_dict["lm_head.weight"]
                    weight_loader = getattr(
                        lm_head_param, "weight_loader", default_weight_loader
                    )
                    weight_loader(lm_head_param, loaded_weight)

            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    continue
                if (
                    "visual" in name
                    and "up_proj" not in name
                    and "gate_proj" not in name
                ):
                    continue
                name = name.replace(weight_name, param_name)
                layer_id = get_layer_id(name)
                if (
                    layer_id is not None
                    and hasattr(self, "model")
                    and hasattr(self.model, "start_layer")
                    and (
                        layer_id < self.model.start_layer
                        or layer_id >= self.model.end_layer
                    )
                ):
                    continue

                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # Skip loading visual/language model weights
                if (
                    self.config.encoder_only or self.config.language_only
                ) and name not in params_dict:
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                if "visual" in name:
                    # adapt to VisionAttention
                    name = name.replace(r"attn.qkv.", r"attn.qkv_proj.")

                try:
                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue
                    if name in params_dict.keys():
                        param = params_dict[name]
                    else:
                        continue

                except KeyError:
                    print(params_dict.keys())
                    raise

                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                weight_loader(param, loaded_weight)

    def get_embed_and_head(self):
        return self.model.embed_tokens.weight, self.lm_head.weight

    def set_eagle3_layers_to_capture(self, layer_ids: Optional[List[int]] = None):
        self.capture_aux_hidden_states = True
        self.model.capture_aux_hidden_states = True
        if layer_ids is None:
            num_layers = self.config.num_hidden_layers
            self.model.layers_to_capture = [
                2,
                num_layers // 2,
                num_layers - 3,
            ]  # Specific layers for EAGLE3 support
        else:
            self.model.layers_to_capture = [val + 1 for val in layer_ids]


EntryClass = [Qwen2_5_VLForConditionalGeneration]