sglang 0.5.3rc0__py3-none-any.whl → 0.5.3rc2__py3-none-any.whl

sglang/srt/models/starcoder2.py

@@ -0,0 +1,357 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Copyright 2024 BigCode 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.
+# Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/starcoder2.py
+""" PyTorch Starcoder2 model."""
+from collections.abc import Iterable
+from typing import Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import Starcoder2Config
+
+from sglang.srt.distributed import get_pp_group, get_tensor_model_parallel_world_size
+from sglang.srt.layers.activation import get_act_fn
+from sglang.srt.layers.linear import (
+    ColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.layers.rotary_embedding import get_rope
+from sglang.srt.layers.vocab_parallel_embedding import (
+    DEFAULT_VOCAB_PADDING_SIZE,
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import default_weight_loader
+from sglang.srt.utils import add_prefix, make_layers
+
+
+class Starcoder2Attention(nn.Module):
+
+    def __init__(
+        self,
+        config: Starcoder2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        layer_id: int = 0,
+    ):
+        super().__init__()
+        self.config = config
+
+        self.hidden_size = config.hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = config.num_attention_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = config.num_key_value_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)
+        self.head_dim = self.hidden_size // self.total_num_heads
+        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 = config.rope_theta
+        self.max_position_embeddings = config.max_position_embeddings
+        self.use_bias = config.use_bias
+
+        self.qkv_proj = QKVParallelLinear(
+            self.hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=self.use_bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            self.hidden_size,
+            bias=self.use_bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.o_proj",
+        )
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=self.max_position_embeddings,
+            base=int(self.rope_theta),
+            is_neox_style=True,
+        )
+        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=f"{prefix}.attn",
+        )
+
+    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 Starcoder2MLP(nn.Module):
+
+    def __init__(
+        self,
+        config: Starcoder2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.c_fc = ColumnParallelLinear(
+            config.hidden_size,
+            config.intermediate_size,
+            bias=config.use_bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.c_fc",
+        )
+        self.c_proj = RowParallelLinear(
+            config.intermediate_size,
+            config.hidden_size,
+            bias=config.use_bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.c_proj",
+        )
+        self.act = get_act_fn(config.hidden_act)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+        hidden_states, _ = self.c_fc(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states, _ = self.c_proj(hidden_states)
+        return hidden_states
+
+
+class Starcoder2DecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: Starcoder2Config,
+        layer_id: int,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = Starcoder2Attention(
+            config=config,
+            layer_id=layer_id,
+            quant_config=quant_config,
+            prefix=f"{prefix}.self_attn",
+        )
+        self.mlp = Starcoder2MLP(
+            config, quant_config=quant_config, prefix=f"{prefix}.mlp"
+        )
+        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.norm_epsilon)
+        self.post_attention_layernorm = nn.LayerNorm(
+            config.hidden_size, eps=config.norm_epsilon
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> torch.Tensor:
+        # Self Attention
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            forward_batch=forward_batch,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class Starcoder2Model(nn.Module):
+
+    def __init__(
+        self,
+        config: Starcoder2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+
+        self.config = config
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+            quant_config=quant_config,
+            prefix=f"{prefix}.embed_tokens",
+        )
+
+        pp_group = get_pp_group()
+        pp_size = pp_group.world_size
+        pp_rank = pp_group.rank
+        self.start_layer = pp_rank * config.num_hidden_layers // pp_size
+        self.end_layer = (pp_rank + 1) * config.num_hidden_layers // pp_size
+
+        self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda idx, prefix: Starcoder2DecoderLayer(
+                config=config, quant_config=quant_config, layer_id=idx, prefix=prefix
+            ),
+            prefix=f"{prefix}.layers",
+        )
+        self.norm = nn.LayerNorm(config.hidden_size, eps=config.norm_epsilon)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if inputs_embeds is None:
+            hidden_states = self.embed_tokens(input_ids)
+        else:
+            hidden_states = inputs_embeds
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            hidden_states = layer(
+                positions,
+                hidden_states,
+                forward_batch,
+            )
+        hidden_states = self.norm(hidden_states)
+        return hidden_states
+
+
+class Starcoder2ForCausalLM(nn.Module):
+
+    def __init__(
+        self,
+        config: Starcoder2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.config = config
+        self.model = Starcoder2Model(
+            config, quant_config, prefix=add_prefix("model", prefix)
+        )
+        self.vocab_size = config.vocab_size
+        self.unpadded_vocab_size = config.vocab_size
+        if config.tie_word_embeddings:
+            self.lm_head = self.model.embed_tokens
+        else:
+            self.unpadded_vocab_size = config.vocab_size
+            self.lm_head = ParallelLMHead(
+                self.unpadded_vocab_size,
+                config.hidden_size,
+                org_num_embeddings=config.vocab_size,
+                padding_size=DEFAULT_VOCAB_PADDING_SIZE,
+                quant_config=quant_config,
+                prefix=f"{prefix}.lm_head",
+            )
+        self.logits_processor = LogitsProcessor(config=config)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        hidden_states = self.model(
+            input_ids=input_ids,
+            positions=positions,
+            forward_batch=forward_batch,
+            inputs_embeds=inputs_embeds,
+        )
+        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"),
+        ]
+        params_dict = dict(self.named_parameters())
+
+        for name, loaded_weight in weights:
+            if "rotary_emb.inv_freqs" in name:
+                continue
+
+            is_stacked = False
+            for param_name, weight_name, shard_id in stacked_params_mapping:
+                if weight_name in name:
+                    name = name.replace(weight_name, param_name)
+                    param = params_dict[name]
+                    weight_loader = getattr(
+                        param, "weight_loader", default_weight_loader
+                    )
+                    weight_loader(param, loaded_weight, shard_id)
+                    is_stacked = True
+                    break
+            if is_stacked:
+                continue
+
+            param = params_dict.get(name)
+            if param is None:
+                continue
+
+            weight_loader = getattr(param, "weight_loader", default_weight_loader)
+            weight_loader(param, loaded_weight)
+
+
+EntryClass = Starcoder2ForCausalLM

sglang/srt/models/torch_native_llama.py

@@ -66,8 +66,8 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix
 
-tp_size = get_tensor_model_parallel_world_size()
-tp_rank = get_tensor_model_parallel_rank()
+tp_size: Optional[int] = None
+tp_rank: Optional[int] = None
 
 
 def gate_up_proj_weight_loader(
@@ -341,6 +341,13 @@ class LlamaModel(nn.Module):
         quant_config: Optional[QuantizationConfig] = None,
     ) -> None:
         super().__init__()
+
+        global tp_size, tp_rank
+        if tp_size is None:
+            tp_size = get_tensor_model_parallel_world_size()
+        if tp_rank is None:
+            tp_rank = get_tensor_model_parallel_rank()
+
         self.config = config
         self.padding_idx = config.pad_token_id
         self.vocab_size = config.vocab_size

sglang/srt/models/utils.py

@@ -0,0 +1,51 @@
+# 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.
+# ==============================================================================
+
+import torch
+
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.utils import is_cuda
+
+_is_cuda = is_cuda()
+
+
+if _is_cuda:
+    from sgl_kernel import FusedSetKVBufferArg
+
+
+def enable_fused_set_kv_buffer(forward_batch: ForwardBatch):
+    """Enable fused set_kv_buffer only on CUDA with bfloat16 KV cache."""
+    return _is_cuda and forward_batch.token_to_kv_pool.dtype == torch.bfloat16
+
+
+def create_fused_set_kv_buffer_arg(
+    value: torch.Tensor,
+    layer: RadixAttention,
+    forward_batch: ForwardBatch,
+):
+    layer_id = layer.layer_id
+    token_to_kv_pool = forward_batch.token_to_kv_pool
+
+    k_buffer = token_to_kv_pool.get_key_buffer(layer_id)
+    v_buffer = token_to_kv_pool.get_value_buffer(layer_id)
+
+    return FusedSetKVBufferArg(
+        value=value,
+        k_buffer=k_buffer.view(k_buffer.shape[0], -1),
+        v_buffer=v_buffer.view(v_buffer.shape[0], -1),
+        k_scale=layer.k_scale,
+        v_scale=layer.v_scale,
+        cache_loc=forward_batch.out_cache_loc,
+    )

sglang/srt/multimodal/processors/base_processor.py

@@ -234,19 +234,27 @@ class BaseMultimodalProcessor(ABC):
             and isinstance(processor.image_processor, BaseImageProcessorFast)
             and not self.server_args.disable_fast_image_processor
         ):
-            kwargs["device"] = "cuda" if not _is_npu else "npu"
+            if not _is_npu:
+                kwargs["device"] = "cuda"
+            elif processor.__class__.__name__ not in {
+                "Qwen2_5_VLProcessor",
+                "Qwen3VLProcessor",
+            }:
+                # Note: for qwen-vl, processor has some reshape issue because of dims restriction on Ascend.
+                kwargs["device"] = "npu"
         result = processor.__call__(
             text=[input_text],
             padding=True,
             return_tensors="pt",
             **kwargs,
         )
-        # move feature tensors to cpu
-        for feature_name in self.FEATURE_NAMES:
-            if feature_name in result and isinstance(
-                result[feature_name], torch.Tensor
-            ):
-                result[feature_name] = result[feature_name].to("cpu")
+        if not self.server_args.keep_mm_feature_on_device:
+            # move feature tensors to cpu
+            for feature_name in self.FEATURE_NAMES:
+                if feature_name in result and isinstance(
+                    result[feature_name], torch.Tensor
+                ):
+                    result[feature_name] = result[feature_name].to("cpu")
 
         return result
 

sglang/srt/multimodal/processors/dots_vlm.py

@@ -5,6 +5,7 @@ from typing import Dict, List, Union
 
 from PIL import Image
 
+from sglang.srt.models.dots_ocr import DotsOCRForCausalLM
 from sglang.srt.models.dots_vlm import DotsVLMForCausalLM
 from sglang.srt.multimodal.processors.base_processor import (
     BaseMultimodalProcessor,
@@ -14,7 +15,7 @@ from sglang.srt.multimodal.processors.qwen_vl import resize_image_async
 
 
 class DotsVLMImageProcessor(BaseMultimodalProcessor):
-    models = [DotsVLMForCausalLM]
+    models = [DotsVLMForCausalLM, DotsOCRForCausalLM]
 
     def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
         super().__init__(hf_config, server_args, _processor, *args, **kwargs)
@@ -82,11 +83,9 @@ class DotsVLMImageProcessor(BaseMultimodalProcessor):
                 for image in base_output.images
             ]
             base_output.images = await asyncio.gather(*resize_tasks)
-
         combined_mm_item, input_ids, _ = self.process_and_combine_mm_data(
             base_output, self.mm_tokens
         )
-
         if combined_mm_item is None:
             return None
 

sglang/srt/multimodal/processors/internvl.py

@@ -1,5 +1,7 @@
 # Adapted from https://huggingface.co/OpenGVLab/InternVL2-4B/blob/main/modeling_intern_vit.py
 
+from functools import lru_cache
+
 import numpy as np
 import torch
 import torchvision.transforms as T
@@ -19,6 +21,20 @@ from sglang.srt.multimodal.processors.base_processor import (
 class InternVLImageProcessor(BaseMultimodalProcessor):
     models = [InternVLChatModel, InternS1ForConditionalGeneration]
 
+    IMAGENET_MEAN = [0.485, 0.456, 0.406]
+    IMAGENET_STD = [0.229, 0.224, 0.225]
+
+    @staticmethod
+    @lru_cache(maxsize=1)
+    def _get_normalize_tensors(device="cuda", dtype=torch.float32):
+        mean = torch.tensor(
+            InternVLImageProcessor.IMAGENET_MEAN, device=device, dtype=dtype
+        ).view(-1, 1, 1)
+        std = torch.tensor(
+            InternVLImageProcessor.IMAGENET_STD, device=device, dtype=dtype
+        ).view(-1, 1, 1)
+        return mean, std
+
     def __init__(self, hf_config, server_args, _image_processor, *args, **kwargs):
         super().__init__(hf_config, server_args, _image_processor, *args, **kwargs)
         image_size = (
@@ -88,6 +104,8 @@ class InternVLImageProcessor(BaseMultimodalProcessor):
             bound, fps, max_frame, first_idx=0, num_segments=num_segments
         )
 
+        mean, std = InternVLImageProcessor._get_normalize_tensors(device="cuda")
+
         for frame_index in frame_indices:
             # Load frame
             frame = vr[frame_index]
@@ -97,10 +115,6 @@ class InternVLImageProcessor(BaseMultimodalProcessor):
                 img_np = frame.asnumpy()
                 img = torch.from_numpy(img_np).permute(2, 0, 1).cuda().float() / 255.0
 
-            # Using the mean and variance of the ImageNet dataset for all input images can lead to accuracy issues, while using the mean and variance of each input image is a more accurate choice.
-            mean = img.mean(dim=[1, 2], keepdim=True)
-            # Prevent division by zero; clamp to minimum value of 1e-6
-            std = img.std(dim=[1, 2], keepdim=True).clamp(min=1e-6)
             img = (img - mean) / std
 
             tiles = InternVLImageProcessor.dynamic_preprocess(
@@ -188,6 +202,8 @@ class InternVLImageProcessor(BaseMultimodalProcessor):
         num_patches_list = []
         pixel_values = []
 
+        mean, std = InternVLImageProcessor._get_normalize_tensors(device="cuda")
+
         # Process each input with allocated frames
         for image_index, image in enumerate(base_output.images):
             try:
@@ -201,10 +217,6 @@ class InternVLImageProcessor(BaseMultimodalProcessor):
                 else:
                     tensor = image.cuda()  # assume already tensor
 
-                # Using the mean and variance of the ImageNet dataset for all input images can lead to accuracy issues, while using the mean and variance of each input image is a more accurate choice.
-                mean = tensor.mean(dim=[1, 2], keepdim=True)
-                # Prevent division by zero; clamp to minimum value of 1e-6
-                std = tensor.std(dim=[1, 2], keepdim=True).clamp(min=1e-6)
                 tensor = (tensor - mean) / std
                 tiles = self.dynamic_preprocess(
                     tensor, image_size=448, max_num=12, use_thumbnail=True

sglang/srt/multimodal/processors/qwen_vl.py

@@ -12,6 +12,8 @@ from torchvision.transforms import InterpolationMode
 from sglang.srt.layers.rotary_embedding import MRotaryEmbedding
 from sglang.srt.models.qwen2_5_vl import Qwen2_5_VLForConditionalGeneration
 from sglang.srt.models.qwen2_vl import Qwen2VLForConditionalGeneration
+from sglang.srt.models.qwen3_vl import Qwen3VLForConditionalGeneration
+from sglang.srt.models.qwen3_vl_moe import Qwen3VLMoeForConditionalGeneration
 from sglang.srt.multimodal.processors.base_processor import (
     BaseMultimodalProcessor as SGLangBaseProcessor,
 )
@@ -209,7 +211,12 @@ async def preprocess_video(
 
 # Compatible with Qwen2VL and Qwen2_5VL
 class Qwen2_5VLImageProcessor(SGLangBaseProcessor):
-    models = [Qwen2VLForConditionalGeneration, Qwen2_5_VLForConditionalGeneration]
+    models = [
+        Qwen2VLForConditionalGeneration,
+        Qwen2_5_VLForConditionalGeneration,
+        Qwen3VLForConditionalGeneration,
+        Qwen3VLMoeForConditionalGeneration,
+    ]
 
     def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
         super().__init__(hf_config, server_args, _processor, *args, **kwargs)

sglang/srt/multimodal/processors/sarashina2_vision.py

@@ -0,0 +1,81 @@
+from typing import List, Union
+
+from sglang.srt.models.sarashina2_vision import Sarashina2VisionForCausalLM
+from sglang.srt.multimodal.processors.base_processor import (
+    BaseMultimodalProcessor,
+    MultimodalSpecialTokens,
+)
+
+
+class Sarashina2VisionProcessor(BaseMultimodalProcessor):
+    models = [Sarashina2VisionForCausalLM]
+
+    def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
+        super().__init__(hf_config, server_args, _processor, *args, **kwargs)
+
+        # Sarashina2Vision specific tokens (default is <|file|>)
+        self.IMAGE_TOKEN = "<|file|>"
+        self.IM_TOKEN_ID = getattr(hf_config, "image_token_index", 14)
+        self.IM_START_ID = getattr(hf_config, "start_image_token_index", 102397)
+        self.IM_END_ID = getattr(hf_config, "end_image_token_index", 102398)
+
+        self.mm_tokens = MultimodalSpecialTokens(
+            image_token=self.IMAGE_TOKEN,
+            image_token_id=self.IM_TOKEN_ID,
+        ).build(_processor)
+
+        # Patch the processor's image processor to handle parameter compatibility
+        if hasattr(_processor, "image_processor") and hasattr(
+            _processor.image_processor, "_preprocess"
+        ):
+            original_preprocess = _processor.image_processor._preprocess
+
+            def patched_preprocess(*args, **kwargs):
+                # Filter kwargs to only include parameters that the custom _preprocess method accepts
+                # Based on Sarashina2VisionImageProcessor._preprocess signature
+                allowed_params = {
+                    "do_resize",
+                    "resample",
+                    "do_rescale",
+                    "rescale_factor",
+                    "do_normalize",
+                    "image_mean",
+                    "image_std",
+                    "do_convert_rgb",
+                    "data_format",
+                    "input_data_format",
+                }
+                filtered_kwargs = {
+                    k: v for k, v in kwargs.items() if k in allowed_params
+                }
+                return original_preprocess(*args, **filtered_kwargs)
+
+            _processor.image_processor._preprocess = patched_preprocess
+
+    async def process_mm_data_async(
+        self,
+        image_data: List[Union[str, bytes]],
+        input_text,
+        request_obj,
+        *args,
+        **kwargs,
+    ):
+        """Process image data for Sarashina2Vision model using standard SGLang pattern."""
+        base_output = self.load_mm_data(
+            prompt=input_text,
+            image_data=image_data,
+            multimodal_tokens=self.mm_tokens,
+        )
+
+        mm_items, input_ids, ret = self.process_and_combine_mm_data(
+            base_output=base_output,
+            mm_tokens=self.mm_tokens,
+        )
+
+        return {
+            "mm_items": mm_items,
+            "input_ids": input_ids.tolist(),
+            "im_token_id": self.mm_tokens.image_token_id,
+            "im_start_id": self.IM_START_ID,
+            "im_end_id": self.IM_END_ID,
+        }

sglang/srt/parser/jinja_template_utils.py

@@ -89,6 +89,12 @@ def detect_jinja_template_content_format(chat_template: str) -> str:
     - If template has loops like {%- for content in message['content'] -%} → 'openai'
     - Otherwise → 'string'
     """
+    # Shortcut for multimodal templates
+    if any(
+        keyword in chat_template for keyword in ["image", "audio", "video", "vision"]
+    ):
+        return "openai"
+
     jinja_ast = _try_extract_ast(chat_template)
     if jinja_ast is None:
         return "string"