[Model] Add Ernie4.5 VL model support (#15679)

Signed-off-by: CSWYF3634076 <wangyafeng@baidu.com>
Signed-off-by: wangyafeng <wangyafeng@baidu.com>

docs/supported_models/multimodal_language_models.md

@@ -46,6 +46,7 @@ in the GitHub search bar.
 | **DotsVLM-OCR**            | `rednote-hilab/dots.ocr`                   | Specialized OCR variant of DotsVLM optimized for optical character recognition tasks with enhanced text extraction and document understanding capabilities. | Don't use `--trust-remote-code` |
 | **NVILA** (8B, 15B, Lite-2B, Lite-8B, Lite-15B) | `Efficient-Large-Model/NVILA-8B` | `chatml` | NVILA explores the full stack efficiency of multi-modal design, achieving cheaper training, faster deployment and better performance. |
 | **NVIDIA Nemotron Nano 2.0 VL** | `nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16` | NVIDIA Nemotron Nano v2 VL enables multi-image reasoning and video understanding, along with strong document intelligence, visual Q&A and summarization capabilities. It builds on Nemotron Nano V2, a hybrid Mamba-Transformer LLM, in order to achieve higher inference throughput in long document and video scenarios. | Use `--trust-remote-code`. You may need to adjust `--max-mamba-cache-size` [default is 512] to fit memory constraints. |
+| **Ernie4.5-VL** | `baidu/ERNIE-4.5-VL-28B-A3B-PT`              | Baidu's vision-language models(28B,424B). Support image and video comprehension, and also support thinking.                                                                     |  |
 | **JetVLM** |  | JetVLM is an vision-language model designed for high-performance multimodal understanding and generation tasks built upon Jet-Nemotron. | Coming soon |
 
 ## Video Input Support

python/sglang/srt/configs/model_config.py

@@ -1123,6 +1123,7 @@ def is_generation_model(model_architectures: List[str], is_embedding: bool = Fal
 multimodal_model_archs = [
     "CLIPModel",
     "DeepseekVL2ForCausalLM",
+    "Ernie4_5_VLMoeForConditionalGeneration",
     "Gemma3ForConditionalGeneration",
     "Gemma3nForConditionalGeneration",
     "Glm4vForConditionalGeneration",

python/sglang/srt/layers/rotary_embedding.py

@@ -2285,6 +2285,177 @@ class MRotaryEmbedding(RotaryEmbedding):
 
             return position_ids, mrope_position_deltas
 
+    @staticmethod
+    def get_rope_index_ernie45(
+        input_ids: torch.Tensor,
+        hf_config: Any,
+        image_grid_thw: Union[list[list[int]], torch.Tensor],
+        video_grid_thw: Union[list[list[int]], torch.Tensor],
+        **kwargs,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Get mrope input positions and delta value for Ernie VL."""
+
+        image_token_id = hf_config.im_patch_id
+        video_start_token_id = hf_config.video_start_token_id
+        video_end_token_id = hf_config.video_end_token_id
+        spatial_conv_size = hf_config.spatial_conv_size
+        temporal_conv_size = hf_config.temporal_conv_size
+
+        mrope_position_deltas = []
+        if input_ids is not None and (
+            image_grid_thw is not None or video_grid_thw is not None
+        ):
+            total_input_ids = input_ids
+            position_ids = torch.ones(
+                3,
+                input_ids.shape[0],
+                input_ids.shape[1],
+                dtype=input_ids.dtype,
+                device=input_ids.device,
+            )
+            image_index, video_index = 0, 0
+            for i, input_ids in enumerate(total_input_ids):
+                input_tokens = input_ids.tolist()
+
+                input_token_type = []
+                video_check_flg = False
+                for token in input_tokens:
+                    if token == video_start_token_id:
+                        video_check_flg = True
+                    elif token == video_end_token_id:
+                        video_check_flg = False
+
+                    if token == image_token_id and not video_check_flg:
+                        input_token_type.append("image")
+                    elif token == image_token_id and video_check_flg:
+                        input_token_type.append("video")
+                    else:
+                        input_token_type.append("text")
+
+                input_type_group = []
+                for key, group in itertools.groupby(
+                    enumerate(input_token_type), lambda x: x[1]
+                ):
+                    group = list(group)
+                    start_index = group[0][0]
+                    end_index = group[-1][0] + 1
+                    input_type_group.append((key, start_index, end_index))
+
+                llm_pos_ids_list = []
+                video_frame_num = 1
+                for modality_type, start_idx, end_idx in input_type_group:
+                    st_idx = (
+                        llm_pos_ids_list[-1].max() + 1
+                        if len(llm_pos_ids_list) > 0
+                        else 0
+                    )
+
+                    if modality_type == "image":
+                        t, h, w = (
+                            image_grid_thw[image_index][0],
+                            image_grid_thw[image_index][1],
+                            image_grid_thw[image_index][2],
+                        )
+                        llm_grid_t, llm_grid_h, llm_grid_w = (
+                            t.item(),
+                            h.item() // spatial_conv_size,
+                            w.item() // spatial_conv_size,
+                        )
+
+                        t_index = (
+                            torch.arange(llm_grid_t)
+                            .view(-1, 1)
+                            .expand(-1, llm_grid_h * llm_grid_w)
+                            .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]) + st_idx
+                        )
+
+                        image_index += 1
+                        video_frame_num = 1
+
+                    elif modality_type == "video":
+                        t, h, w = (
+                            video_grid_thw[video_index][0],
+                            video_grid_thw[video_index][1],
+                            video_grid_thw[video_index][2],
+                        )
+
+                        llm_grid_t, llm_grid_h, llm_grid_w = (
+                            t.item() // temporal_conv_size,
+                            h.item() // spatial_conv_size,
+                            w.item() // spatial_conv_size,
+                        )
+
+                        for t_idx in range(llm_grid_t):
+                            t_index = (
+                                torch.tensor(t_idx)
+                                .view(-1, 1)
+                                .expand(-1, llm_grid_h * llm_grid_w)
+                                .flatten()
+                            )
+
+                            h_index = (
+                                torch.arange(llm_grid_h)
+                                .view(1, -1, 1)
+                                .expand(1, -1, llm_grid_w)
+                                .flatten()
+                            )
+                            w_index = (
+                                torch.arange(llm_grid_w)
+                                .view(1, 1, -1)
+                                .expand(1, llm_grid_h, -1)
+                                .flatten()
+                            )
+                            llm_pos_ids_list.append(
+                                torch.stack([t_index, h_index, w_index]) + st_idx
+                            )
+
+                        video_index += 1
+                        video_frame_num += 1
+
+                    else:
+                        text_len = end_idx - start_idx
+                        llm_pos_ids_list.append(
+                            torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx
+                        )
+
+                        video_frame_num = 1
+
+                llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
+                position_ids[..., i, :] = llm_positions.to(position_ids.device)
+                mrope_position_deltas.append(
+                    llm_positions.max() + 1 - len(total_input_ids[i])
+                )
+            mrope_position_deltas = torch.tensor(
+                mrope_position_deltas, device=input_ids.device
+            ).unsqueeze(1)
+            return position_ids, mrope_position_deltas
+        else:
+            s = input_ids.shape[1]
+            position_ids = torch.arange(s)
+            position_ids = (
+                position_ids.unsqueeze(0).expand(3, -1, -1).to(input_ids.device)
+            )
+            max_position_ids = position_ids.max(0, keepdim=False)[0].max(
+                -1, keepdim=True
+            )[0]
+            mrope_position_deltas = max_position_ids + 1 - s
+            return position_ids, mrope_position_deltas
+
     # For qwen3-omni
     @staticmethod
     def _get_feat_extract_output_lengths(input_lengths):
@@ -2324,6 +2495,91 @@ class MRotaryEmbedding(RotaryEmbedding):
         return llm_pos_ids
 
 
+class Ernie4_5_VLRotaryEmbedding(MRotaryEmbedding):
+    """3D rotary positional embedding. [h w h w h w h w... t t t...]"""
+
+    def forward_native(  # type: ignore[override]
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor | None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        assert positions.ndim == 1 or positions.ndim == 2
+        assert key is not None
+
+        num_tokens = positions.shape[-1]
+        cos_sin = self.cos_sin_cache[positions]
+        cos, sin = cos_sin.chunk(2, dim=-1)
+        if positions.ndim == 2:
+            assert self.mrope_section
+
+            section_h = self.mrope_section[0]  # 22
+            section_w = self.mrope_section[1]  # 22
+            section_t = self.mrope_section[2]  # 20
+            assert section_h == section_w
+            # Split according to [h w h w h w h w... t t t...]
+            section_cos_t = cos[..., -section_t:]
+            section_cos_h = cos[..., : section_h + section_w : 2]
+            section_cos_w = cos[..., 1 : section_h + section_w : 2]
+
+            cos_t, cos_h, cos_w = section_cos_t[0], section_cos_h[1], section_cos_w[2]
+            cos_hw = torch.stack([cos_h, cos_w], dim=-1).reshape(
+                cos_h.shape[:-1] + (cos_h.shape[-1] * 2,)
+            )
+            cos = torch.cat([cos_hw, cos_t], dim=-1)
+
+            section_sin_t = sin[..., -section_t:]
+            section_sin_h = sin[..., : section_h + section_w : 2]
+            section_sin_w = sin[..., 1 : section_h + section_w : 2]
+
+            sin_t, sin_h, sin_w = section_sin_t[0], section_sin_h[1], section_sin_w[2]
+            sin_hw = torch.stack([sin_h, sin_w], dim=-1).reshape(
+                sin_h.shape[:-1] + (sin_h.shape[-1] * 2,)
+            )
+            sin = torch.cat([sin_hw, sin_t], dim=-1)
+
+        query_shape = query.shape
+        query = query.view(num_tokens, -1, self.head_size)
+        query_rot = query[..., : self.rotary_dim]
+        query_pass = query[..., self.rotary_dim :]
+        query_rot = _apply_rotary_emb(query_rot, cos, sin, self.is_neox_style)
+        query = torch.cat((query_rot, query_pass), dim=-1).reshape(query_shape)
+
+        key_shape = key.shape
+        key = key.view(num_tokens, -1, self.head_size)
+        key_rot = key[..., : self.rotary_dim]
+        key_pass = key[..., self.rotary_dim :]
+        key_rot = _apply_rotary_emb(key_rot, cos, sin, self.is_neox_style)
+        key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
+        return query, key
+
+    def forward_cuda(  # type: ignore[override]
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor | None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor | None]:
+        return self.forward_native(positions, query, key)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        fused_set_kv_buffer_arg: Optional[FusedSetKVBufferArg] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Forward pass with optional Triton kernel acceleration.
+        Args:
+            positions:
+                [num_tokens,] (text only) or
+                [3, num_tokens] (T/H/W positions with multimodal inputs)
+            query: [num_tokens, num_heads * head_size]
+            key: [num_tokens, num_kv_heads * head_size]
+        """
+        assert positions.ndim == 1 or positions.ndim == 2
+        return self.forward_native(positions, query, key)
+
+
 class DualChunkRotaryEmbedding(MultiPlatformOp):
     """Rotary positional embedding for Dual Chunk Attention."""
 

python/sglang/srt/models/ernie45_moe_vl.py

@@ -0,0 +1,552 @@
+# Copyright 2023-2025 SGLang Team
+# 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 Ernie4.5 VL model compatible with baidu/ERNIE-4.5-VL-*-PT weights. """
+
+import logging
+from itertools import islice
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+
+from sglang.srt.distributed import (
+    get_pp_group,
+    get_tensor_model_parallel_world_size,
+    tensor_model_parallel_all_reduce,
+)
+from sglang.srt.layers.dp_attention import is_dp_attention_enabled
+from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import (
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
+from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
+from sglang.srt.layers.moe.topk import TopK
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.rotary_embedding import Ernie4_5_VLRotaryEmbedding
+from sglang.srt.layers.utils import PPMissingLayer
+from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
+from sglang.srt.models.deepseek_v2 import DeepseekV2MLP as Ernie4_5_VLMoeMLP
+from sglang.srt.utils import add_prefix, make_layers
+
+logger = logging.getLogger(__name__)
+
+
+class Ernie4_5_VLMoeAttention(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        rope_is_neox_style: bool = True,
+        freq_allocation: int = 20,
+        max_position_embeddings: int = 8192,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        bias: bool = False,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
+        self.head_dim = getattr(
+            config, "head_dim", self.hidden_size // self.total_num_heads
+        )
+        partial_rotary_factor = getattr(config, "partial_rotary_factor", 1)
+        self.rotary_dim = int(partial_rotary_factor * self.head_dim)
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=add_prefix("qkv_proj", prefix),
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=add_prefix("o_proj", prefix),
+        )
+
+        # 3D rope
+        t_rope = freq_allocation
+        h_rope = (self.head_dim // 2 - freq_allocation) // 2
+        w_rope = (self.head_dim // 2 - freq_allocation) // 2
+
+        self.rotary_emb = Ernie4_5_VLRotaryEmbedding(
+            head_size=self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position_embeddings=max_position_embeddings,
+            base=rope_theta,
+            is_neox_style=False,
+            dtype=torch.get_default_dtype(),
+            mrope_section=[h_rope, w_rope, t_rope],
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+            quant_config=quant_config,
+            prefix=add_prefix("attn", prefix),
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, forward_batch)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class Ernie4_5_VLMoeMoE(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        layer_id: int,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.layer_id = layer_id
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.moe_num_shared_experts = getattr(config, "moe_num_shared_experts", 0)
+        self.hidden_size = config.hidden_size
+
+        moe_num_experts = config.moe_num_experts
+        max_moe_num_experts = max(moe_num_experts)
+
+        if self.tp_size > max_moe_num_experts:
+            raise ValueError(
+                f"Tensor parallel size {self.tp_size} is greater than "
+                f"the number of experts {moe_num_experts}."
+            )
+
+        moe_layer_start_index = config.moe_layer_start_index
+        text_moe_layer_start_index = moe_layer_start_index[0]
+        vision_moe_layer_start_index = moe_layer_start_index[1]
+        moe_layer_end_index = config.moe_layer_end_index
+        moe_layer_end_index = getattr(
+            config,
+            "moe_layer_end_index",
+            [config.num_hidden_layers - 1, config.num_hidden_layers - 1],
+        )
+        text_moe_layer_end_index = moe_layer_end_index[0]
+        vision_moe_layer_end_index = moe_layer_end_index[1]
+
+        assert config.moe_num_experts[0] == config.moe_num_experts[1]
+        self.e_score_correction_bias = nn.Parameter(
+            torch.empty(2, config.moe_num_experts[0], dtype=torch.float32)
+        )
+
+        assert text_moe_layer_start_index <= text_moe_layer_end_index
+
+        if (
+            layer_id >= text_moe_layer_start_index
+            and layer_id <= text_moe_layer_end_index
+        ):
+            self.text_experts_gate = ReplicatedLinear(
+                config.hidden_size,
+                config.moe_num_experts[0],
+                bias=False,
+                params_dtype=torch.float32,
+                quant_config=quant_config,
+                prefix=add_prefix("text_experts_gate", prefix),
+            )
+
+            self.text_experts_topk = TopK(
+                top_k=config.moe_k,
+                renormalize=True,
+                use_grouped_topk=False,
+                correction_bias=self.e_score_correction_bias[0],
+            )
+
+            self.text_experts = get_moe_impl_class(quant_config)(
+                num_experts=config.moe_num_experts[0],
+                top_k=config.moe_k,
+                hidden_size=config.hidden_size,
+                intermediate_size=config.moe_intermediate_size[0],
+                layer_id=self.layer_id,
+                quant_config=quant_config,
+                prefix=add_prefix("text_experts", prefix),
+            )
+
+        assert vision_moe_layer_start_index <= vision_moe_layer_end_index
+        if (
+            layer_id >= vision_moe_layer_start_index
+            and layer_id <= vision_moe_layer_end_index
+        ):
+
+            self.vision_experts_gate = ReplicatedLinear(
+                config.hidden_size,
+                config.moe_num_experts[1],
+                bias=False,
+                params_dtype=torch.float32,
+                quant_config=quant_config,
+                prefix=add_prefix("vision_experts_gate", prefix),
+            )
+
+            self.vision_experts_topk = TopK(
+                top_k=config.moe_k,
+                renormalize=True,
+                use_grouped_topk=False,
+                correction_bias=self.e_score_correction_bias[1],
+            )
+
+            self.vision_experts = get_moe_impl_class(quant_config)(
+                num_experts=config.moe_num_experts[1],
+                top_k=config.moe_k,
+                hidden_size=config.hidden_size,
+                intermediate_size=config.moe_intermediate_size[1],
+                layer_id=self.layer_id,
+                quant_config=quant_config,
+                prefix=add_prefix("vision_experts", prefix),
+            )
+
+        if self.moe_num_shared_experts > 0:
+            intermediate_size = (
+                config.moe_intermediate_size[0] * config.moe_num_shared_experts
+            )
+            self.shared_experts = Ernie4_5_VLMoeMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                reduce_results=False,
+                prefix=add_prefix("shared_experts", prefix),
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        visual_token_mask: torch.Tensor,
+        **kwargs: object,
+    ) -> torch.Tensor:
+        shared_output = (
+            self.shared_experts(hidden_states)
+            if self.moe_num_shared_experts > 0
+            else None
+        )
+
+        orig_shape = hidden_states.shape
+        hidden_dim = hidden_states.shape[-1]
+        hidden_states = hidden_states.view(-1, hidden_dim)
+
+        capturing = torch.cuda.is_current_stream_capturing()
+
+        if visual_token_mask is not None and not capturing:
+            all_visual = visual_token_mask.all()
+            any_visual = visual_token_mask.any()
+        else:
+            # During CUDA Graph capture, all set false
+            all_visual = False
+            any_visual = False
+
+        if all_visual:
+            # vision modal input processing directly
+            vision_router_logits, _ = self.vision_experts_gate(
+                hidden_states.to(dtype=torch.float32)
+            )
+            vision_topk_output = self.vision_experts_topk(
+                hidden_states, vision_router_logits
+            )
+            final_hidden_states = self.vision_experts(
+                hidden_states=hidden_states, topk_output=vision_topk_output
+            )
+        elif any_visual:
+            visual_token_mask = visual_token_mask.repeat(1, self.hidden_size).bool()
+            text_token_mask = ~visual_token_mask
+            final_hidden_states = torch.zeros_like(hidden_states)
+
+            text_hidden_states = hidden_states[text_token_mask].reshape(
+                -1, self.hidden_size
+            )
+            vision_hidden_states = hidden_states[visual_token_mask].reshape(
+                -1, self.hidden_size
+            )
+
+            text_router_logits, _ = self.text_experts_gate(
+                text_hidden_states.to(dtype=torch.float32)
+            )
+            text_topk_output = self.text_experts_topk(
+                text_hidden_states, text_router_logits
+            )
+            final_hidden_states[text_token_mask] = self.text_experts(
+                hidden_states=text_hidden_states, topk_output=text_topk_output
+            ).flatten()
+
+            vision_router_logits, _ = self.vision_experts_gate(
+                vision_hidden_states.to(dtype=torch.float32)
+            )
+            vision_topk_output = self.vision_experts_topk(
+                vision_hidden_states, vision_router_logits
+            )
+            final_hidden_states[visual_token_mask] = self.vision_experts(
+                hidden_states=vision_hidden_states, topk_output=vision_topk_output
+            ).flatten()
+
+        else:
+            # text modal input processing directly
+            text_router_logits, _ = self.text_experts_gate(
+                hidden_states.to(dtype=torch.float32)
+            )
+            topk_output = self.text_experts_topk(hidden_states, text_router_logits)
+            final_hidden_states = self.text_experts(
+                hidden_states=hidden_states, topk_output=topk_output
+            )
+
+        if shared_output is not None:
+            final_hidden_states = final_hidden_states + shared_output
+
+        if self.tp_size > 1:
+            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
+
+        return final_hidden_states.view(orig_shape)
+
+
+class Ernie4_5_VLMoeDecoderLayer(nn.Module):
+    """A single transformer layer.
+
+    Transformer layer takes input with size [s, b, h] and returns an
+    output of the same size.
+    """
+
+    def __init__(
+        self,
+        config,
+        layer_id: int,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        rope_theta = getattr(config, "rope_theta", 500000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        rope_is_neox_style = getattr(config, "rope_is_neox_style", False)
+        freq_allocation = getattr(config, "freq_allocation", 20)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 131072)
+        # Self attention.
+        self.self_attn = Ernie4_5_VLMoeAttention(
+            config=config,
+            hidden_size=config.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            rope_is_neox_style=rope_is_neox_style,
+            freq_allocation=freq_allocation,
+            max_position_embeddings=config.max_position_embeddings,
+            quant_config=quant_config,
+            prefix=add_prefix("self_attn", prefix),
+            bias=config.use_bias,
+        )
+
+        # MoE
+        moe_layer_start_index = config.moe_layer_start_index
+        min_moe_layer_start_index = min(moe_layer_start_index)
+        moe_layer_end_index = getattr(
+            config,
+            "moe_layer_end_index",
+            [config.num_hidden_layers - 1, config.num_hidden_layers - 1],
+        )
+        max_moe_layer_end_index = max(moe_layer_end_index)
+        assert min_moe_layer_start_index <= max_moe_layer_end_index
+        moe_num_experts = config.moe_num_experts
+        max_moe_num_experts = max(moe_num_experts)
+        moe_layer_interval = getattr(config, "moe_layer_interval", 1)
+        use_moe = getattr(config, "use_moe", max_moe_num_experts > 0)
+        # MLP
+        if (
+            use_moe
+            and ((layer_id + 1) % moe_layer_interval == 0)
+            and layer_id >= min_moe_layer_start_index
+            and layer_id <= max_moe_layer_end_index
+        ):
+            self.mlp = Ernie4_5_VLMoeMoE(
+                config=config,
+                layer_id=layer_id,
+                quant_config=quant_config,
+                prefix=add_prefix("mlp", prefix),
+            )
+        else:
+            self.mlp = Ernie4_5_VLMoeMLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=config.intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                prefix=add_prefix("mlp", prefix),
+            )
+
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+        residual: Optional[torch.Tensor],
+        visual_token_mask: torch.Tensor | None,
+        **kwargs: object,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            forward_batch=forward_batch,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        if isinstance(self.mlp, Ernie4_5_VLMoeMoE):
+            hidden_states = self.mlp(hidden_states, visual_token_mask, **kwargs)
+        else:
+            hidden_states = self.mlp(hidden_states)
+
+        return hidden_states, residual
+
+
+# only used as text backbone for ernie4.5 vl
+class Ernie4_5_VLMoeModel(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.pp_group = get_pp_group()
+
+        if self.pp_group.is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+                enable_tp=not is_dp_attention_enabled(),
+                prefix=add_prefix("embed_tokens", prefix),
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
+        self.layers, self.start_layer, self.end_layer = make_layers(
+            config.num_hidden_layers,
+            lambda idx, prefix: Ernie4_5_VLMoeDecoderLayer(
+                layer_id=idx,
+                config=config,
+                quant_config=quant_config,
+                prefix=prefix,
+            ),
+            pp_rank=self.pp_group.rank_in_group,
+            pp_size=self.pp_group.world_size,
+            prefix=add_prefix("layers", prefix),
+        )
+        if self.pp_group.is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer(return_tuple=True)
+
+    def get_input_embeddings(self) -> torch.Tensor:
+        return self.embed_tokens
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+        pp_proxy_tensors: Optional[PPProxyTensors] = None,
+        visual_token_mask: torch.Tensor | None = None,
+    ) -> Union[torch.Tensor, PPProxyTensors]:
+
+        if self.pp_group.is_first_rank:
+            if input_embeds is None:
+                hidden_states = self.embed_tokens(input_ids)
+            else:
+                hidden_states = input_embeds
+            residual = None
+        else:
+            assert pp_proxy_tensors is not None
+            hidden_states = pp_proxy_tensors["hidden_states"]
+            residual = pp_proxy_tensors["residual"]
+
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                forward_batch,
+                residual,
+                visual_token_mask,
+            )
+
+        if not self.pp_group.is_last_rank:
+            return PPProxyTensors(
+                {
+                    "hidden_states": hidden_states,
+                    "residual": residual,
+                }
+            )
+
+        if hidden_states.shape[0] != 0:
+            if residual is None:
+                hidden_states = self.norm(hidden_states)
+            else:
+                hidden_states, _ = self.norm(hidden_states, residual)
+
+        return hidden_states

python/sglang/srt/models/ernie45_vl.py

@@ -0,0 +1,845 @@
+# Copyright 2023-2025 SGLang Team
+# 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 Ernie45-VL model compatible with HuggingFace weights."""
+import logging
+from functools import lru_cache, partial
+from typing import Iterable, List, Optional, Tuple, Type
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from transformers import PretrainedConfig
+
+from sglang.srt.layers.activation import QuickGELU
+from sglang.srt.layers.attention.vision import VisionAttention
+from sglang.srt.layers.layernorm import RMSNorm
+from sglang.srt.layers.linear import ColumnParallelLinear, RowParallelLinear
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+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 MultimodalDataItem, MultimodalInputs
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.models.ernie45_moe_vl import Ernie4_5_VLMoeModel
+from sglang.srt.utils import add_prefix
+from sglang.srt.utils.hf_transformers_utils import get_processor
+
+logger = logging.getLogger(__name__)
+
+
+# === Vision Encoder === #
+
+
+class Ernie4_5_VisionMLP(nn.Module):
+
+    def __init__(
+        self,
+        in_features: int,
+        hidden_features: int = None,
+        act_layer: Type[nn.Module] = QuickGELU,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.fc1 = ColumnParallelLinear(
+            in_features,
+            hidden_features,
+            quant_config=quant_config,
+            prefix=add_prefix("fc1", prefix),
+        )
+        self.act = act_layer()
+        self.fc2 = RowParallelLinear(
+            hidden_features,
+            in_features,
+            quant_config=quant_config,
+            prefix=add_prefix("fc2", prefix),
+        )
+
+    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 Ernie4_5_VisionBlock(nn.Module):
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float,
+        act_layer: Type[nn.Module] = QuickGELU,
+        norm_layer: Type[nn.Module] = None,
+        quant_config: Optional[QuantizationConfig] = 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 = VisionAttention(
+            embed_dim=dim,
+            num_heads=num_heads,
+            projection_size=dim,
+            use_qkv_parallel=True,
+            flatten_batch=True,
+            quant_config=quant_config,
+            prefix=add_prefix("attn", prefix),
+        )
+        self.mlp = Ernie4_5_VisionMLP(
+            dim,
+            mlp_hidden_dim,
+            act_layer=act_layer,
+            quant_config=quant_config,
+            prefix=add_prefix("mlp", prefix),
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        position_embeddings: torch.Tensor,
+    ) -> torch.Tensor:
+        hidden_states = self.norm1(x)
+        hidden_states = rearrange(hidden_states, "s b ... -> b s ...")
+        attn = self.attn(
+            hidden_states,
+            cu_seqlens=cu_seqlens,
+            position_embeddings=position_embeddings,
+        )
+        attn = rearrange(attn, "b s ... -> s b ...")
+        x = x + attn
+        x = x + self.mlp(self.norm2(x))
+        return x
+
+
+class Ernie4_5_VisionPatchEmbed(nn.Module):
+
+    def __init__(
+        self,
+        patch_size: int = 14,
+        in_chans: int = 3,
+        embed_dim: int = 1280,
+    ) -> None:
+        super().__init__()
+        self.patch_size = patch_size
+        self.in_channels = in_chans
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Linear(in_chans * patch_size * patch_size, embed_dim, bias=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        target_dtype = self.proj.weight.dtype
+        hidden_states = hidden_states.to(target_dtype)
+        hidden_states = self.proj(hidden_states)
+
+        return hidden_states
+
+
+class VariableResolutionResamplerModel(nn.Module):
+    def __init__(
+        self,
+        in_dim,
+        out_dim,
+        spatial_conv_size,
+        temporal_conv_size,
+        config,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+        self.config = config
+        self.spatial_conv_size = spatial_conv_size
+        self.temporal_conv_size = temporal_conv_size
+        self.use_temporal_conv = config.use_temporal_conv
+
+        # compress 2d conv(picture) to 1d
+        self.spatial_dim = self.in_dim * self.spatial_conv_size * self.spatial_conv_size
+        # compress 3d conv(video) to 1d
+        self.temporal_dim = (
+            self.in_dim
+            * self.spatial_conv_size
+            * self.spatial_conv_size
+            * self.temporal_conv_size
+        )
+
+        self.spatial_linear1 = ColumnParallelLinear(
+            self.spatial_dim,
+            self.spatial_dim,
+            bias=True,
+            gather_output=True,
+            quant_config=getattr(config, "quant_config", None),
+            prefix=f"{prefix}.spatial_linear1",
+        )
+
+        self.spatial_gelu = nn.GELU()
+
+        self.spatial_linear2 = ColumnParallelLinear(
+            self.spatial_dim,
+            self.spatial_dim,
+            bias=True,
+            gather_output=True,
+            quant_config=getattr(config, "quant_config", None),
+            prefix=f"{prefix}.spatial_linear2",
+        )
+
+        self.spatial_norm = nn.LayerNorm(self.spatial_dim, eps=1e-6)
+
+        if self.use_temporal_conv:
+            self.temporal_linear1 = ColumnParallelLinear(
+                self.temporal_dim,
+                self.spatial_dim,
+                bias=True,
+                gather_output=True,
+                quant_config=getattr(config, "quant_config", None),
+                prefix=f"{prefix}.temporal_linear1",
+            )
+
+            self.temporal_gelu = nn.GELU()
+
+            self.temporal_linear2 = ColumnParallelLinear(
+                self.spatial_dim,
+                self.spatial_dim,
+                bias=True,
+                gather_output=True,
+                quant_config=getattr(config, "quant_config", None),
+                prefix=f"{prefix}.temporal_linear2",
+            )
+
+            self.temporal_norm = nn.LayerNorm(self.spatial_dim, eps=1e-6)
+
+        self.mlp = ColumnParallelLinear(
+            self.spatial_dim,
+            self.out_dim,
+            bias=True,
+            gather_output=True,
+            quant_config=getattr(config, "quant_config", None),
+            prefix=f"{prefix}.mlp",
+        )
+
+        self.after_norm = RMSNorm(
+            hidden_size=out_dim, eps=getattr(config, "rms_norm_eps", 1e-6)
+        )
+
+    def spatial_conv_reshape(self, x, spatial_conv_size):
+        S, C = x.shape
+        x = x.reshape([-1, C * (spatial_conv_size**2)])
+        return x
+
+    def forward(self, x, grid_thw):
+        def fwd_spatial(x):
+            x = self.spatial_conv_reshape(x, self.spatial_conv_size)
+
+            x, _ = self.spatial_linear1(x)
+            x = self.spatial_gelu(x)
+            x, _ = self.spatial_linear2(x)
+            x = self.spatial_norm(x)
+
+            return x
+
+        def fwd_placeholder(x, grid_thw, to_tensor=False):
+            grid_thw_cpu = grid_thw.cpu().numpy()
+            grid_t, grid_hw = grid_thw_cpu[:, 0], grid_thw_cpu[:, 1:]
+            grid_hw_after_conv = grid_hw.prod(-1) // (self.spatial_conv_size**2)
+
+            tokens_per_img_or_vid = grid_thw_cpu.prod(-1) // (self.spatial_conv_size**2)
+            batch_offset = np.empty(
+                tokens_per_img_or_vid.size, dtype=tokens_per_img_or_vid.dtype
+            )
+            batch_offset[0] = 0
+            batch_offset[1:] = tokens_per_img_or_vid.cumsum()[:-1]
+
+            slice_offsets = []
+            for temporoal_size, spatial_size, b_offset in zip(
+                grid_t, grid_hw_after_conv, batch_offset
+            ):
+                for temp_offset in range(0, temporoal_size, 2):
+                    slice_offsets.append(
+                        np.arange(
+                            b_offset + (temp_offset) * spatial_size,
+                            b_offset + (temp_offset + 1) * spatial_size,
+                        )
+                    )
+            slice_offsets = torch.tensor(np.concatenate(slice_offsets, axis=-1)).to(
+                x.device
+            )
+
+            slice_offsets2 = []
+            for temporoal_size, spatial_size, b_offset in zip(
+                grid_t, grid_hw_after_conv, batch_offset
+            ):
+                for temp_offset in range(
+                    1 if temporoal_size > 1 else 0, temporoal_size, 2
+                ):
+                    slice_offsets2.append(
+                        np.arange(
+                            b_offset + (temp_offset) * spatial_size,
+                            b_offset + (temp_offset + 1) * spatial_size,
+                        )
+                    )
+            slice_offsets2 = torch.tensor(np.concatenate(slice_offsets2, axis=-1)).to(
+                x.device
+            )
+
+            x_timestep_1 = torch.index_select(x, dim=0, index=slice_offsets)
+            x_timestep_2 = torch.index_select(x, dim=0, index=slice_offsets2)
+            x = torch.concat([x_timestep_1, x_timestep_2], dim=-1)
+            return x
+
+        def fwd_temporal(x):
+            x, _ = self.temporal_linear1(x)
+            x = self.temporal_gelu(x)
+            x, _ = self.temporal_linear2(x)
+            x = self.temporal_norm(x)
+            return x
+
+        def fwd_mlp(x):
+            x, _ = self.mlp(x)
+            x = self.after_norm(x)
+            return x
+
+        x = fwd_spatial(x)
+        if self.use_temporal_conv:
+            x = fwd_placeholder(x, grid_thw)
+            x = fwd_temporal(x)
+        x = fwd_mlp(x)
+        return x
+
+    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+        loaded_params: set[str] = set()
+
+        for name, loaded_weight in weights:
+            if name not in params_dict:
+                continue
+            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
+
+
+class Ernie4_5_VisionRotaryEmbedding(nn.Module):
+
+    def __init__(self, dim: int, theta: float = 10000.0) -> None:
+        super().__init__()
+        self.inv_freq = 1.0 / theta ** (
+            torch.arange(start=0, end=dim, step=2, dtype=torch.float32) / dim
+        )
+
+    def forward(self, seqlen: int) -> torch.Tensor:
+        seq = torch.arange(
+            seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype
+        )
+        freqs = torch.outer(input=seq, vec2=self.inv_freq)
+        return freqs
+
+
+class Ernie4_5_VisionTransformer(nn.Module):
+
+    def __init__(
+        self,
+        vision_config: PretrainedConfig,
+        norm_eps: float = 1e-6,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        patch_size: int = vision_config.patch_size
+        spatial_merge_size: int = vision_config.spatial_merge_size
+        in_chans: int = vision_config.in_chans
+        hidden_size: int = vision_config.hidden_size
+        embed_dim: int = vision_config.embed_dim
+        depth: int = vision_config.depth
+        num_heads: int = vision_config.num_heads
+        mlp_ratio: float = vision_config.mlp_ratio
+
+        self.spatial_merge_size = spatial_merge_size
+
+        self.patch_embed = Ernie4_5_VisionPatchEmbed(
+            patch_size=patch_size,
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+        )
+
+        norm_layer = partial(nn.LayerNorm, eps=norm_eps)
+        head_dim = embed_dim // num_heads
+        self.rotary_pos_emb = Ernie4_5_VisionRotaryEmbedding(head_dim // 2)
+        self.blocks = nn.ModuleList(
+            [
+                Ernie4_5_VisionBlock(
+                    dim=embed_dim,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    norm_layer=norm_layer,
+                    quant_config=quant_config,
+                    prefix=add_prefix(f"blocks.{i}", prefix),
+                )
+                for i in range(depth)
+            ]
+        )
+
+        self.ln = nn.LayerNorm(hidden_size, eps=1e-6)
+
+    @property
+    def dtype(self) -> torch.dtype:
+        return self.patch_embed.proj.weight.dtype
+
+    @property
+    def device(self) -> torch.device:
+        return self.blocks[0].mlp.fc2.weight.device
+
+    def rot_pos_emb(self, grid_thw: torch.Tensor) -> torch.Tensor:
+        pos_ids = []
+        for i in range(grid_thw.size(0)):
+            t, h, w = grid_thw[i].tolist()
+            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
+            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
+            hpos_ids = (
+                hpos_ids.reshape(
+                    h // self.spatial_merge_size,
+                    self.spatial_merge_size,
+                    w // self.spatial_merge_size,
+                    self.spatial_merge_size,
+                )
+                .permute(0, 2, 1, 3)
+                .flatten()
+            )
+            wpos_ids = (
+                wpos_ids.reshape(
+                    h // self.spatial_merge_size,
+                    self.spatial_merge_size,
+                    w // self.spatial_merge_size,
+                    self.spatial_merge_size,
+                )
+                .permute(0, 2, 1, 3)
+                .flatten()
+            )
+            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+        pos_ids = torch.cat(pos_ids, dim=0)
+        max_grid_size = grid_thw[:, 1:].max()
+        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+        return rotary_pos_emb
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        grid_thw: torch.Tensor,
+    ) -> torch.Tensor:
+        # patchify
+        x = x.to(device=self.device, dtype=self.dtype)
+        x = self.patch_embed(x)
+
+        # compute position embedding
+        rotary_pos_emb = self.rot_pos_emb(grid_thw)
+        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
+        position_embeddings = (emb.cos(), emb.sin())
+        # compute cu_seqlens
+        cu_seqlens = torch.repeat_interleave(
+            grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]
+        ).cumsum(dim=0, dtype=torch.int32)
+        cu_seqlens = torch.cat([cu_seqlens.new_zeros(1), cu_seqlens])
+
+        # transformers
+        x = x.unsqueeze(1)
+        for blk in self.blocks:
+            x = blk(x, cu_seqlens=cu_seqlens, position_embeddings=position_embeddings)
+
+        final_output = self.ln(x)
+
+        if final_output.ndim == 3:
+            final_output = final_output.squeeze(dim=1)
+
+        return final_output
+
+
+cached_get_processor = lru_cache(get_processor)
+
+
+class Ernie4_5_VLMoeForConditionalGeneration(nn.Module):
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        ".gate_proj.",
+        ".down_proj.",
+        ".up_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),
+    }
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+        self.vision_model = Ernie4_5_VisionTransformer(
+            config.vision_config,
+            norm_eps=getattr(config, "rms_norm_eps", 1e-6),
+            quant_config=quant_config,
+            prefix=add_prefix("vision_model", prefix),
+        )
+
+        self.model = Ernie4_5_VLMoeModel(
+            config, quant_config, prefix=add_prefix("model", prefix)
+        )
+
+        self.resampler_model = VariableResolutionResamplerModel(
+            self.config.pixel_hidden_size,
+            self.config.hidden_size,
+            self.config.spatial_conv_size,
+            self.config.temporal_conv_size,
+            config=self.config,
+            prefix=add_prefix("resampler_model", prefix),
+        )
+
+        if config.tie_word_embeddings:
+            self.lm_head = self.model.embed_tokens
+        else:
+            self.lm_head = ParallelLMHead(
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                prefix=add_prefix("lm_head", prefix),
+            )
+
+        self.is_mrope_enabled = "mrope_section" in self.config.rope_scaling
+        self.logits_processor = LogitsProcessor(config)
+
+        if getattr(self.config, "im_patch_id", None):
+            visual_token_ids = [
+                token_id
+                for token_id in [
+                    self.config.im_patch_id,
+                    getattr(self.config, "image_start_token_id", None),
+                    getattr(self.config, "image_end_token_id", None),
+                    getattr(self.config, "video_start_token_id", None),
+                    getattr(self.config, "video_end_token_id", None),
+                ]
+                if token_id is not None
+            ]
+            self._visual_token_ids_tensor_cache = torch.tensor(
+                visual_token_ids, dtype=torch.long
+            )
+        else:
+            self._visual_token_ids_tensor_cache = None
+
+    def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
+        pattern = MultiModalityDataPaddingPatternMultimodalTokens()
+        return pattern.pad_input_tokens(input_ids, mm_inputs)
+
+    def _vision_forward(
+        self,
+        pixel_values: torch.Tensor,
+        grid_thw: torch.Tensor,
+    ) -> torch.Tensor:
+        if grid_thw is not None:
+            grid_thw = grid_thw[grid_thw > 0]
+            if grid_thw.numel() % 3 != 0:
+                raise ValueError(
+                    f"grid_thw has {grid_thw.numel()} elements after filtering,"
+                    "which is not divisible by 3."
+                )
+            grid_thw = grid_thw.reshape(-1, 3)
+            # example: [[1,64,64],[2,80,80]] -> [[1,64,64],[1,80,80],[1,80,80]]
+            grid_thw = F.pad(
+                torch.repeat_interleave(grid_thw[:, 1:], grid_thw[:, 0], 0),
+                [1, 0, 0, 0],
+                value=1,
+            )
+        image_features = self.vision_model(pixel_values, grid_thw)
+        return image_features
+
+    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.vision_model.dtype
+        )
+        image_grid_thw = torch.concat([item.image_grid_thw for item in items], dim=0)
+        assert pixel_values.dim() == 2, pixel_values.dim()
+        assert image_grid_thw.dim() == 2, image_grid_thw.dim()
+        image_feature = self._vision_forward(pixel_values, grid_thw=image_grid_thw)
+        image_embeds = self.resampler_model(image_feature, image_grid_thw)
+        return image_embeds
+
+    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.vision_model.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()
+        video_feature = self._vision_forward(pixel_values, grid_thw=video_grid_thw)
+        video_embeds = self.resampler_model(video_feature, video_grid_thw)
+        return video_embeds
+
+    def _set_visual_token_mask(
+        self, input_ids: torch.Tensor, forward_batch: ForwardBatch
+    ) -> None:
+        """Set mask for visual tokens (image/video patches and delimiters)."""
+        if self._visual_token_ids_tensor_cache is None:
+            self.visual_token_mask = None
+            return
+        # Create tensor on the correct device
+        visual_token_ids_tensor = self._visual_token_ids_tensor_cache.to(
+            device=input_ids.device,
+            dtype=input_ids.dtype,
+        )
+
+        pad_values = []
+        if hasattr(forward_batch, "mm_inputs") and forward_batch.mm_inputs is not None:
+            for mm_input in forward_batch.mm_inputs:
+                if mm_input is None:
+                    continue
+                for item in mm_input.mm_items:
+                    pad_values.append(item.pad_value)
+        placeholder_tensor = torch.as_tensor(
+            pad_values,
+            device=input_ids.device,
+        )
+        pad_visual_token_ids_tensor = torch.cat(
+            [visual_token_ids_tensor, placeholder_tensor], dim=0
+        )
+        self.visual_token_mask = torch.isin(
+            input_ids, pad_visual_token_ids_tensor
+        ).reshape(-1, 1)
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def should_apply_lora(self, module_name: str) -> bool:
+        # skip vision_model
+        return not module_name.startswith("vision_model")
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        get_embedding: bool = False,
+    ):
+        """Run forward pass for Ernie45-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()}"
+                )
+
+        self._set_visual_token_mask(input_ids, forward_batch)
+
+        assert (
+            input_ids.numel() == positions.shape[-1]
+        ), f"input_ids {input_ids.shape} and position_ids {positions.shape} should have the same length"
+
+        hidden_states = general_mm_embed_routine(
+            input_ids=input_ids,
+            forward_batch=forward_batch,
+            language_model=self.model,
+            multimodal_model=self,
+            positions=positions,
+            visual_token_mask=self.visual_token_mask,
+        )
+
+        self.visual_token_mask = None
+
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head, forward_batch
+        )
+
+    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),
+        ]
+
+        # resampler_weight_mappings
+        resampler_weight_mapping = {
+            "spatial_linear.0.": "spatial_linear1.",
+            "spatial_linear.2.": "spatial_linear2.",
+            "spatial_linear.3.": "spatial_norm.",
+            "temporal_linear.0.": "temporal_linear1.",
+            "temporal_linear.2.": "temporal_linear2.",
+            "temporal_linear.3.": "temporal_norm.",
+        }
+
+        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+            ckpt_gate_proj_name="gate_proj",
+            ckpt_down_proj_name="down_proj",
+            ckpt_up_proj_name="up_proj",
+            num_experts=max(self.config.moe_num_experts),
+        )
+        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 "lm_head.weight" in name:
+                continue
+
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+
+                if ("mlp.experts." in name) and name not in params_dict:
+                    continue
+                name = name.replace(weight_name, param_name)
+
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") 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 "vision_model" in name:
+                    # adapt to VisionAttention
+                    name = name.replace(r"attn.qkv.", r"attn.qkv_proj.")
+                if name.startswith("model.resampler_model"):
+                    name = name.replace("model.resampler_model", "resampler_model")
+
+                for (
+                    old_weight_name,
+                    new_weight_name,
+                ) in resampler_weight_mapping.items():
+                    if old_weight_name in name:
+                        name = name.replace(old_weight_name, new_weight_name, 1)
+                        break
+
+                # Distinguish between vision experts and text experts
+                if "mlp.experts" in name:
+                    moe_offset = int(name.split(".")[-3])
+                    vision_expert_start_idx = self.config.moe_num_experts[0]
+                    is_text_expert = moe_offset <= vision_expert_start_idx - 1
+                    if is_text_expert:
+                        name = name.replace(".experts.", ".text_experts.")
+                    else:
+                        name = name.replace(
+                            f".experts.{moe_offset}",
+                            f".vision_experts.{moe_offset - vision_expert_start_idx}",
+                        )
+
+                for mapping in expert_params_mapping:
+                    param_name, weight_name, expert_id, shard_id = mapping
+                    if weight_name not in name:
+                        continue
+
+                    # Distinguish between vision experts and text experts
+                    moe_offset = int(name.split(".")[-3])
+                    is_text_expert = moe_offset <= self.config.moe_num_experts[0] - 1
+
+                    name = name.replace(weight_name, param_name)
+                    if is_text_expert:
+                        name = name.replace(".experts.", ".text_experts.")
+                    else:
+                        name = name.replace(".experts.", ".vision_experts.")
+
+                    # Skip loading extra bias for GPTQ models.
+                    if (
+                        name.endswith(".bias") or name.endswith("_bias")
+                    ) and name not in params_dict:
+                        continue
+
+                    if name in params_dict.keys():
+                        param = params_dict[name]
+                        weight_loader = param.weight_loader
+                        weight_loader(
+                            param,
+                            loaded_weight,
+                            name,
+                            shard_id=shard_id,
+                            expert_id=expert_id,
+                        )
+                    else:
+                        logger.warning(f"Parameter {name} not found in params_dict")
+                    break
+                else:
+                    # Distinguish between vision expert gate
+                    # and text expert gate
+                    if name.endswith("mlp.gate.weight"):
+                        name = name.replace("gate.weight", "text_experts_gate.weight")
+                        loaded_weight = loaded_weight.T
+                    elif name.endswith("mlp.gate.weight_1"):
+                        name = name.replace(
+                            "gate.weight_1", "vision_experts_gate.weight"
+                        )
+                        loaded_weight = loaded_weight.T
+
+                    if "e_score_correction_bias" in name:
+                        name = name.replace(".moe_statics.", ".")
+
+                    # Skip loading extra bias for GPTQ models.
+                    if (
+                        name.endswith(".bias") or name.endswith("_bias")
+                    ) and name not in params_dict:
+                        continue
+
+                    if name in params_dict.keys():
+                        param = params_dict[name]
+                        weight_loader = getattr(
+                            param, "weight_loader", default_weight_loader
+                        )
+                        weight_loader(param, loaded_weight)
+                    else:
+                        logger.warning(f"Parameter {name} not found in params_dict")
+
+
+EntryClass = [Ernie4_5_VLMoeForConditionalGeneration]

python/sglang/srt/multimodal/processors/ernie45_vl.py

@@ -0,0 +1,417 @@
+import math
+import os
+from typing import List, Union
+
+import numpy as np
+import torch
+import torchvision
+from PIL import Image
+from torchvision.transforms import InterpolationMode
+from transformers import BaseImageProcessorFast
+
+from sglang.srt.environ import envs
+from sglang.srt.layers.rotary_embedding import MRotaryEmbedding
+from sglang.srt.models.ernie45_vl import Ernie4_5_VLMoeForConditionalGeneration
+from sglang.srt.multimodal.processors.base_processor import (
+    BaseMultimodalProcessor as SGLangBaseProcessor,
+)
+from sglang.srt.multimodal.processors.base_processor import MultimodalSpecialTokens
+from sglang.srt.utils import get_bool_env_var, is_npu, logger
+
+_is_npu = is_npu()
+
+SGL_USE_CUDA_IPC = get_bool_env_var("SGLANG_USE_CUDA_IPC_TRANSPORT")
+
+
+IMAGE_FACTOR = 28
+MIN_PIXELS = 4 * 28 * 28
+# MAX_PIXELS = envs.SGLANG_IMAGE_MAX_PIXELS.get()
+MAX_PIXELS = 16384 * 28 * 28
+MAX_RATIO = 200
+RESIZE_RESAMPLE = getattr(Image, envs.SGLANG_RESIZE_RESAMPLE.get(), None)
+if envs.SGLANG_RESIZE_RESAMPLE.is_set() and RESIZE_RESAMPLE is None:
+    logger.warning(
+        f"Invalid RESIZE_RESAMPLE value: '{envs.SGLANG_RESIZE_RESAMPLE.get()}'. "
+        f"Ignoring and using default."
+    )
+VIDEO_TOTAL_PIXELS = int(
+    float(os.environ.get("VIDEO_MAX_PIXELS", 128000 * 28 * 28 * 0.9))
+)
+
+VIDEO_MIN_PIXELS = 299 * 28 * 28
+VIDEO_MAX_PIXELS = 1196 * 28 * 28
+FRAME_FACTOR = 2
+FPS = 2.0
+FPS_MIN_FRAMES = 16
+FPS_MAX_FRAMES = 180
+
+
+def smart_resize(
+    height: int,
+    width: int,
+    factor: int = IMAGE_FACTOR,
+    min_pixels: int = MIN_PIXELS,
+    max_pixels: int = MAX_PIXELS,
+):
+    if max(height, width) / min(height, width) > MAX_RATIO:
+        if height > width:
+            new_width = max(factor, round_by_factor(width, factor))
+            new_height = floor_by_factor(new_width * MAX_RATIO, factor)
+        else:
+            new_height = max(factor, round_by_factor(height, factor))
+            new_width = floor_by_factor(new_height * MAX_RATIO, factor)
+
+        height = new_height
+        width = new_width
+
+    h_bar = max(factor, round_by_factor(height, factor))
+    w_bar = max(factor, round_by_factor(width, factor))
+    if h_bar * w_bar > max_pixels:
+        beta = math.sqrt((height * width) / max_pixels)
+        h_bar = floor_by_factor(height / beta, factor)
+        w_bar = floor_by_factor(width / beta, factor)
+    elif h_bar * w_bar < min_pixels:
+        beta = math.sqrt(min_pixels / (height * width))
+        h_bar = ceil_by_factor(height * beta, factor)
+        w_bar = ceil_by_factor(width * beta, factor)
+
+    if min_pixels > h_bar * w_bar or h_bar * w_bar > max_pixels:
+        raise ValueError(f"encounter invalid h_bar: {h_bar}, w_bar: {w_bar}")
+
+    return h_bar, w_bar
+
+
+def resize_image(
+    image,
+    min_pixels: int = MIN_PIXELS,
+    max_pixels: int = MAX_PIXELS,
+    size_factor: int = IMAGE_FACTOR,
+) -> Image.Image:
+    width, height = image.size
+    min_pixels = min_pixels
+    max_pixels = max_pixels
+    resized_height, resized_width = smart_resize(
+        height,
+        width,
+        factor=size_factor,
+        min_pixels=min_pixels,
+        max_pixels=max_pixels,
+    )
+    image = image.resize((resized_width, resized_height), resample=RESIZE_RESAMPLE)
+    return image
+
+
+def round_by_factor(number: int | float, factor: int) -> int:
+    return round(number / factor) * factor
+
+
+def ceil_by_factor(number: int | float, factor: int) -> int:
+    return math.ceil(number / factor) * factor
+
+
+def floor_by_factor(number: int | float, factor: int) -> int:
+    return math.floor(number / factor) * factor
+
+
+async def resize_image_async(
+    image,
+    min_pixels: int = MIN_PIXELS,
+    max_pixels: int = MAX_PIXELS,
+    size_factor: int = IMAGE_FACTOR,
+):
+    return resize_image(image, min_pixels, max_pixels, size_factor)
+
+
+def smart_nframes(
+    ele: dict,
+    total_frames: int,
+    video_fps: int | float,
+) -> int:
+    """calculate the number of frames for video used for model inputs.
+
+    Args:
+        ele (dict): a dict contains the configuration of video.
+            support either `fps` or `nframes`:
+                - nframes: the number of frames to extract for model inputs.
+                - fps: the fps to extract frames for model inputs.
+                    - min_frames: the minimum number of frames of the video, only used when fps is provided.
+                    - max_frames: the maximum number of frames of the video, only used when fps is provided.
+        total_frames (int): the original total number of frames of the video.
+        video_fps (int | float): the original fps of the video.
+
+    Raises:
+        ValueError: nframes should in interval [FRAME_FACTOR, total_frames].
+
+    Returns:
+        int: the number of frames for video used for model inputs.
+    """
+    assert not (
+        "fps" in ele and "nframes" in ele
+    ), "Only accept either `fps` or `nframes`"
+    if "nframes" in ele:
+        nframes = round_by_factor(ele["nframes"], FRAME_FACTOR)
+    else:
+        fps = ele.get("fps", FPS)
+        min_frames = ceil_by_factor(ele.get("min_frames", FPS_MIN_FRAMES), FRAME_FACTOR)
+        max_frames = floor_by_factor(
+            ele.get("max_frames", min(FPS_MAX_FRAMES, total_frames)), FRAME_FACTOR
+        )
+        nframes = total_frames / video_fps * fps
+        if nframes > total_frames:
+            logger.warning(
+                f"smart_nframes: nframes[{nframes}] > total_frames[{total_frames}]"
+            )
+        nframes = min(min(max(nframes, min_frames), max_frames), total_frames)
+        nframes = floor_by_factor(nframes, FRAME_FACTOR)
+    if not (FRAME_FACTOR <= nframes and nframes <= total_frames):
+        raise ValueError(
+            f"nframes should in interval [{FRAME_FACTOR}, {total_frames}], but got {nframes}."
+        )
+    return nframes
+
+
+# process video, qwen-specific
+async def preprocess_video(
+    vr,
+    image_factor: int = IMAGE_FACTOR,
+) -> torch.Tensor:
+
+    total_frames, video_fps = len(vr), vr.get_avg_fps()
+    nframes = smart_nframes({}, total_frames=total_frames, video_fps=video_fps)
+    idx = np.linspace(0, total_frames - 1, num=nframes, dtype=np.int64)
+    idx = np.unique(idx)
+    video_np = vr.get_batch(idx).asnumpy()
+    video = torch.from_numpy(video_np).pin_memory()
+    video = video.permute(0, 3, 1, 2)  # Convert to TCHW format
+    nframes, _, height, width = video.shape
+    min_pixels = VIDEO_MIN_PIXELS
+    total_pixels = VIDEO_TOTAL_PIXELS
+    max_pixels = max(
+        min(VIDEO_MAX_PIXELS, total_pixels / nframes * FRAME_FACTOR),
+        int(min_pixels * 1.05),
+    )
+
+    resized_height, resized_width = smart_resize(
+        height,
+        width,
+        factor=image_factor,
+        min_pixels=min_pixels,
+        max_pixels=max_pixels,
+    )
+    video = torchvision.transforms.functional.resize(
+        video,
+        [resized_height, resized_width],
+        interpolation=InterpolationMode.BILINEAR,
+    )
+
+    video = video.permute(0, 2, 3, 1)
+    video = video.pin_memory()
+    video_metadata = {
+        "fps": video_fps,
+        "duration": total_frames / video_fps,
+        "total_num_frames": total_frames,
+        "frames_indices": idx,
+        "video_backend": "torchvision",
+    }
+
+    return video, video_metadata
+
+
+# Compatible with Ernie-VL Series
+class Ernie4_5_VLImageProcessor(SGLangBaseProcessor):
+    models = [Ernie4_5_VLMoeForConditionalGeneration]
+
+    def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
+        super().__init__(hf_config, server_args, _processor, *args, **kwargs)
+        self.hf_config = hf_config
+        self.model_type = hf_config.model_type
+        self.image_start_token_id = hf_config.image_start_token_id
+        self.image_end_token_id = hf_config.image_end_token_id
+        self.video_start_token_id = hf_config.video_start_token_id
+        self.video_end_token_id = hf_config.video_end_token_id
+
+        self.IMAGE_FACTOR = 28
+        self.MIN_PIXELS = 4 * 28 * 28
+        self.MAX_PIXELS = 16384 * 28 * 28
+        self.MAX_RATIO = 200
+        self.mm_tokens = MultimodalSpecialTokens(
+            image_token="<|IMAGE_START|><|image@placeholder|><|IMAGE_END|>",
+            video_token="<|VIDEO_START|><|video@placeholder|><|VIDEO_END|>",
+            image_token_id=hf_config.im_patch_id,
+            video_token_id=hf_config.im_patch_id,  # image and video use the same token_id
+        ).build(_processor)
+
+        self.tokenizer = self._processor.tokenizer
+        self.image_processor = self._processor.image_processor
+
+    def _pixel_values_norm(
+        self,
+        pixel_values: torch.Tensor,
+        mm_kwargs: object,
+    ) -> torch.Tensor:
+        hf_config = self.hf_config
+        vision_config = hf_config.vision_config
+        image_processor = self.image_processor
+        image_mean_tensor = torch.tensor(
+            image_processor.image_mean, dtype=torch.float32
+        ).reshape([1, 3, 1, 1])
+        image_std_tensor = torch.tensor(
+            image_processor.image_std, dtype=torch.float32
+        ).reshape([1, 3, 1, 1])
+        rescale_factor = torch.tensor(
+            image_processor.rescale_factor, dtype=torch.float32
+        )
+        patch_size_squared = vision_config.patch_size**2
+
+        image_mean_tensor = image_mean_tensor.squeeze([-2, -1]).repeat_interleave(
+            patch_size_squared, -1
+        )
+        image_std_tensor = image_std_tensor.squeeze([-2, -1]).repeat_interleave(
+            patch_size_squared, -1
+        )
+
+        if not image_mean_tensor.is_contiguous():
+            image_mean_tensor = image_mean_tensor.contiguous()
+        if not image_std_tensor.is_contiguous():
+            image_std_tensor = image_std_tensor.contiguous()
+
+        pixel_values = (
+            rescale_factor * pixel_values.to(torch.float32) - image_mean_tensor
+        ) / image_std_tensor
+        pixel_values = pixel_values.to(hf_config.dtype)
+        return pixel_values
+
+    def process_mm_data(
+        self, input_text, images=None, videos=None, audios=None, **kwargs
+    ) -> dict:
+        """
+        process multimodal data with transformers AutoProcessor
+        """
+        if images:
+            kwargs["images"] = images
+        if videos:
+            kwargs["videos"] = videos
+
+        processor = self._processor
+        if (
+            hasattr(processor, "image_processor")
+            and isinstance(processor.image_processor, BaseImageProcessorFast)
+            and not self.server_args.disable_fast_image_processor
+        ):
+            if not _is_npu:
+                kwargs["device"] = "cuda"
+
+        result = processor.__call__(
+            text=[input_text],
+            padding=True,
+            return_tensors="pt",
+            **kwargs,
+        )
+
+        # Divide the processor_output into two modalities: image and video.
+        if result is not None:
+            pixel_values = result["images"]
+            if pixel_values is not None:
+                result["images"] = self._pixel_values_norm(pixel_values, kwargs)
+            for key in list(result.keys()):
+                if result[key] is None:
+                    del result[key]
+                    continue
+                if key == "grid_thw":
+                    grid_thw = result["grid_thw"]
+                    pixel_values_all = result["images"]
+                    # Identify elements where the first
+                    # dimension is greater than 1 and
+                    # treat them as the video modality
+                    mask = grid_thw[:, 0] > 1
+                    result["video_grid_thw"] = grid_thw[mask]
+                    result["image_grid_thw"] = grid_thw[~mask]
+                    image_patch_num = result["image_grid_thw"].prod(dim=1).sum()
+                    result["pixel_values"] = pixel_values_all[:image_patch_num]
+                    result["pixel_values_videos"] = pixel_values_all[image_patch_num:]
+                    del result["images"]
+                    del result["grid_thw"]
+
+                    # del empty result
+                    if result["image_grid_thw"].numel() == 0:
+                        del result["image_grid_thw"]
+                    if result["pixel_values"].numel() == 0:
+                        del result["pixel_values"]
+                    if result["video_grid_thw"].numel() == 0:
+                        del result["video_grid_thw"]
+                    if result["pixel_values_videos"].numel() == 0:
+                        del result["pixel_values_videos"]
+
+        if not self.server_args.keep_mm_feature_on_device:
+            # move feature tensors to cpu
+            for feature_name in self.FEATURE_NAMES:
+                if SGL_USE_CUDA_IPC:
+                    pass
+                else:
+                    if feature_name in result and isinstance(
+                        result[feature_name], torch.Tensor
+                    ):
+                        result[feature_name] = result[feature_name].to("cpu")
+
+        return result
+
+    async def process_mm_data_async(
+        self,
+        image_data: List[Union[str, bytes]],
+        input_text,
+        request_obj,
+        *args,
+        **kwargs,
+    ):
+        base_output = self.load_mm_data(
+            prompt=input_text,
+            image_data=image_data,
+            video_data=request_obj.video_data,
+            audio_data=request_obj.audio_data,
+            multimodal_tokens=self.mm_tokens,
+        )
+
+        # resize images if they are raw Image objects
+        resized_images = []
+        if base_output.images and isinstance(base_output.images[0], Image.Image):
+            for image in base_output.images:
+                resized_image = resize_image(image)
+                resized_images.append(resized_image)
+            base_output.images = resized_images
+
+        if base_output.videos:
+            videos_processed = [
+                await preprocess_video(video) for video in base_output.videos
+            ]
+            base_output.videos, _ = map(list, zip(*videos_processed))
+
+        mm_items, input_ids, ret = self.process_and_combine_mm_data(
+            base_output, self.mm_tokens
+        )
+
+        input_ids = input_ids.flatten()
+
+        mrope_positions, mrope_position_delta = MRotaryEmbedding.get_rope_index_ernie45(
+            input_ids=input_ids.unsqueeze(0),
+            hf_config=self.hf_config,
+            image_grid_thw=getattr(ret, "image_grid_thw", None),
+            video_grid_thw=getattr(ret, "video_grid_thw", None),
+        )
+        mrope_positions = mrope_positions.squeeze(1)
+
+        assert (
+            input_ids.shape[0] == mrope_positions.shape[-1]
+        ), "input_ids and mrope_positions should have the same length"
+
+        mm_inputs = {
+            "input_ids": input_ids.tolist(),
+            "mm_items": mm_items,
+            "im_start_id": self.image_start_token_id,
+            "im_end_id": self.image_end_token_id,
+            "im_token_id": self.mm_tokens.image_token_id,
+            "video_token_id": self.mm_tokens.video_token_id,
+            "mrope_positions": mrope_positions,
+            "mrope_position_delta": mrope_position_delta,
+        }
+
+        return mm_inputs