Urgent model support: support gemma-3-it (#4424)

python/sglang/lang/chat_template.py

@@ -520,6 +520,14 @@ def match_granite_instruct(model_path: str):
         return get_chat_template("granite-3-instruct")
 
 
+@register_chat_template_matching_function
+def match_gemma3_instruct(model_path: str):
+    model_path = model_path.lower()
+    if "gemma-3" in model_path and "1b" not in model_path:
+        # gemma-3-1b-it is completion model
+        return get_chat_template("gemma-it")
+
+
 if __name__ == "__main__":
     messages = [
         {"role": "system", "content": None},  # None means default

python/sglang/srt/configs/__init__.py

@@ -1,6 +1,7 @@
 from sglang.srt.configs.chatglm import ChatGLMConfig
 from sglang.srt.configs.dbrx import DbrxConfig
 from sglang.srt.configs.exaone import ExaoneConfig
+from sglang.srt.configs.gemma3 import Gemma3Config, Gemma3TextConfig
 from sglang.srt.configs.janus_pro import MultiModalityConfig
 from sglang.srt.configs.qwen2_5_vl_config import (
     Qwen2_5_VLConfig,
@@ -14,4 +15,6 @@ __all__ = [
     "Qwen2_5_VLConfig",
     "Qwen2_5_VLVisionConfig",
     "MultiModalityConfig",
+    "Gemma3Config",
+    "Gemma3TextConfig",
 ]

python/sglang/srt/configs/model_config.py

@@ -391,9 +391,13 @@ def _get_and_verify_dtype(
         dtype = dtype.lower()
         if dtype == "auto":
             if config_dtype == torch.float32:
-                if config.model_type == "gemma2":
+                if config.model_type.startswith("gemma"):
+                    if config.model_type == "gemma":
+                        gemma_version = ""
+                    else:
+                        gemma_version = config.model_type[5]
                     logger.info(
-                        "For Gemma 2, we downcast float32 to bfloat16 instead "
+                        f"For Gemma {gemma_version}, we downcast float32 to bfloat16 instead "
                         "of float16 by default. Please specify `dtype` if you "
                         "want to use float16."
                     )
@@ -453,6 +457,7 @@ multimodal_model_archs = [
     "LlavaQwenForCausalLM",
     "LlavaMistralForCausalLM",
     "LlavaVidForCausalLM",
+    "Gemma3ForConditionalGeneration",
     "Grok1VForCausalLM",
     "Grok1AForCausalLM",
     "MllamaForConditionalGeneration",

python/sglang/srt/conversation.py

@@ -45,6 +45,7 @@ class SeparatorStyle(IntEnum):
     DEEPSEEK_CHAT = auto()
     METAMATH = auto()
     QWEN2_VL_EMBED = auto()
+    GEMMA3 = auto()
 
 
 @dataclasses.dataclass
@@ -285,6 +286,18 @@ class Conversation:
                 else:
                     ret += role + ":"
             return ret
+        elif self.sep_style == SeparatorStyle.GEMMA3:
+            ret = system_prompt
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    if i == 0:
+                        ret += message + self.sep
+                    else:
+                        ret += role + message + self.sep
+                else:
+                    ret += role
+            return ret
+
         else:
             raise ValueError(f"Invalid style: {self.sep_style}")
 
@@ -604,6 +617,20 @@ register_conv_template(
     )
 )
 
+# Reference: https://huggingface.co/google/gemma-3-4b-it/blob/main/config.json
+register_conv_template(
+    Conversation(
+        name="gemma-it",
+        system_message="You are a helpful assistant.",
+        system_template="<bos><start_of_turn>user{system_message}\n\n",
+        roles=("<start_of_turn>user\n", "<start_of_turn>model\n"),
+        sep="<end_of_turn>\n",
+        sep_style=SeparatorStyle.GEMMA3,
+        stop_str=["<end_of_turn>"],
+        image_token="<start_of_image>",
+    )
+)
+
 # Reference: https://huggingface.co/Alibaba-NLP/gme-Qwen2-VL-2B-Instruct#usage
 register_conv_template(
     Conversation(

python/sglang/srt/hf_transformers_utils.py

@@ -34,6 +34,8 @@ from sglang.srt.configs import (
     ChatGLMConfig,
     DbrxConfig,
     ExaoneConfig,
+    Gemma3Config,
+    Gemma3TextConfig,
     MultiModalityConfig,
     Qwen2_5_VLConfig,
 )
@@ -46,6 +48,8 @@ _CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = {
     ExaoneConfig.model_type: ExaoneConfig,
     Qwen2_5_VLConfig.model_type: Qwen2_5_VLConfig,
     MultiModalityConfig.model_type: MultiModalityConfig,
+    Gemma3Config.model_type: Gemma3Config,
+    Gemma3TextConfig.model_type: Gemma3TextConfig,
 }
 
 for name, cls in _CONFIG_REGISTRY.items():

python/sglang/srt/layers/attention/vision.py

@@ -19,34 +19,10 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.quantization import QuantizationConfig
+from sglang.srt.layers.rotary_embedding import apply_rotary_pos_emb, rotate_half
 from sglang.srt.utils import add_prefix
 
 
-# Copied from transformers, modeling_qwen2_vl.py
-def rotate_half(x):
-    """Rotates half the hidden dims of the input."""
-    x1 = x[..., : x.shape[-1] // 2]
-    x2 = x[..., x.shape[-1] // 2 :]
-    return torch.cat((-x2, x1), dim=-1)
-
-
-def apply_rotary_pos_emb_vision(
-    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
-) -> Tuple[torch.Tensor, torch.Tensor]:
-    orig_q_dtype = q.dtype
-    orig_k_dtype = k.dtype
-    q, k = q.float(), k.float()
-
-    cos, sin = cos.unsqueeze(-2).float(), sin.unsqueeze(-2).float()
-    q_embed = (q * cos) + (rotate_half(q) * sin)
-    k_embed = (k * cos) + (rotate_half(k) * sin)
-
-    q_embed = q_embed.to(orig_q_dtype)
-    k_embed = k_embed.to(orig_k_dtype)
-
-    return q_embed, k_embed
-
-
 class VisionAttention(nn.Module):
     r"""
         Multi-headed attention without any cache, mostly used for ViT.
@@ -168,7 +144,7 @@ class VisionAttention(nn.Module):
             cos, sin = position_embeddings
             original_shape = q.shape
             q, k = q.view(s, head, -1), k.view(s, head, -1)
-            q, k = apply_rotary_pos_emb_vision(q, k, cos, sin)
+            q, k = apply_rotary_pos_emb(q, k, cos, sin)
             q, k = q.reshape(original_shape), k.reshape(original_shape)
 
         if self.use_qkv_parallel:

python/sglang/srt/layers/layernorm.py

@@ -119,6 +119,26 @@ class GemmaRMSNorm(CustomOp):
         return out
 
 
+class Gemma3RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.zeros(dim))
+
+    def _norm(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x):
+        output = self._norm(x.float())
+        # Llama does x.to(float16) * w whilst Gemma3 is (x * w).to(float16)
+        # See https://github.com/huggingface/transformers/pull/29402
+        output = output * (1.0 + self.weight.float())
+        return output.type_as(x)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+
 if not _is_cuda:
     logger.info(
         "sgl-kernel is not available on Non-NV platforms. Fallback to other kernel libraries."

python/sglang/srt/layers/rotary_embedding.py

@@ -1173,6 +1173,37 @@ def get_rope(
     return rotary_emb
 
 
+# Copied from transformers
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    unsqueeze_dim=1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    orig_q_dtype = q.dtype
+    orig_k_dtype = k.dtype
+    q, k = q.float(), k.float()
+
+    # embedding is performed in float
+    cos = cos.unsqueeze(unsqueeze_dim).float()
+    sin = sin.unsqueeze(unsqueeze_dim).float()
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+
+    q_embed = q_embed.to(orig_q_dtype)
+    k_embed = k_embed.to(orig_k_dtype)
+
+    return q_embed, k_embed
+
+
 def get_rope_cpu(
     head_size: int,
     rotary_dim: int,

python/sglang/srt/managers/image_processors/base_image_processor.py

@@ -111,7 +111,7 @@ class BaseImageProcessor(ABC):
 
     def load_images(
         self,
-        input_ids: list,
+        input_ids: list[int],
         image_data,
         image_token: str,
         max_req_input_len: int,
@@ -122,22 +122,21 @@ class BaseImageProcessor(ABC):
         Each frame of video/image will be replaced by a single image token
 
         Args:
-
             discard_alpha_channel: if True, discards the alpha channel in the returned images
 
         """
-        image_hashes, image_sizes = [], []
-        all_frames = []
-        new_text_parts = []
 
         if isinstance(input_ids, list) and return_text:
             assert len(input_ids) and isinstance(input_ids[0], int)
             input_text = self._processor.tokenizer.decode(input_ids)
         else:
             input_text = input_ids
-
         if return_text:
-            text_parts = input_text.split(image_token)
+            import re
+
+            pattern = "(" + "|".join(re.escape(sep) for sep in [image_token]) + ")"
+            # split text into list of normal text and special tokens
+            text_parts = re.split(pattern, input_text)
 
         # TODO(mick): load from server_args, env, or sampling_params
         MAX_NUM_FRAMES = 30
@@ -145,53 +144,65 @@ class BaseImageProcessor(ABC):
         total_frame_count = sum(estimated_frames_list)
         # a heuristic value, suggesting the maximum fraction of frames to embed from all visual inputs.
         # e.g., 0.1 suggests that 1 frame out of 10 input frames should be used
-        scaling_factor = min(1.0, MAX_NUM_FRAMES / total_frame_count)
+        _scaling_factor = min(1.0, MAX_NUM_FRAMES / max(1, total_frame_count))
 
         assert len(image_data) == len(estimated_frames_list)
 
-        # Process each input with allocated frames
-        for image_index, (image, estimated_frames) in enumerate(
-            zip(image_data, estimated_frames_list)
-        ):
-            if len(all_frames) >= MAX_NUM_FRAMES:
-                max_frames_to_process = 0
-            else:
-                max_frames_to_process = max(1, int(estimated_frames * scaling_factor))
-
-            if max_frames_to_process == 0:
-                frames = []
-            else:
-                try:
-                    if isinstance(image, str) and image.startswith("video:"):
-                        path = image[len("video:") :]
-                        frames = BaseImageProcessor.encode_video(
-                            path, frame_count_limit=max_frames_to_process
-                        )
+        image_index, audio_index = 0, 0
+        hashes, image_sizes, images, audios = [], [], [], []
+        new_text = ""
+        for index, text_part in enumerate(text_parts):
+            try:
+                if text_part == image_token:
+                    # load as image
+                    frames_to_process = estimated_frames_list[image_index]
+                    if frames_to_process == 0:
+                        frames = []
                     else:
-                        raw_image, _size = load_image(image)
-                        if discard_alpha_channel:
-                            raw_image = raw_image.convert("RGB")
-                        frames = [raw_image]
-                    assert len(frames) != 0
-                except FileNotFoundError as e:
-                    print(e)
-                    return None
+                        image_file = image_data[image_index]
+                        if isinstance(image_file, str) and image_file.startswith(
+                            "video:"
+                        ):
+                            # video
+                            path = image_file[len("video:") :]
+                            frames = self.encode_video(
+                                path, frame_count_limit=frames_to_process
+                            )
+                        else:
+                            # image
+                            raw_image, _size = load_image(image_file)
+                            if discard_alpha_channel:
+                                raw_image = raw_image.convert("RGB")
+                            frames = [raw_image]
+                        if len(frames) == 0:
+                            continue
 
-                image_sizes += [frames[0].size] * len(frames)
-                image_hashes += [hash(image)] * len(frames)
-                all_frames += frames
+                    image_sizes += frames[0].size * len(frames)
+                    hashes += [hash(image_file)] * len(frames)
+                    images += frames
+                    image_index += 1
+                    if frames_to_process != 0:
+                        new_text += image_token * len(frames)
+                    assert frames_to_process == len(frames)
+                else:
+                    # TODO(mick): handle video
+                    # normal text
+                    new_text += text_part
 
-            if return_text:
-                new_text_parts.append(text_parts[image_index])
-            if max_frames_to_process != 0:
-                new_text_parts.append(image_token * len(frames))
-            assert max_frames_to_process >= len(frames)
-        if return_text:
-            new_text_parts.append(text_parts[-1])
+            except Exception as e:
+                import openai
+
+                logger.error(f"An exception occurred while loading images: {e}")
+                raise BadRequestError(
+                    f"An exception occurred while loading images: {e}"
+                )
+                continue
 
-        input_text = "".join(new_text_parts)
         return BaseImageProcessorOutput(
-            image_hashes, image_sizes, all_frames, input_text
+            image_hashes=hashes,
+            image_sizes=image_sizes,
+            all_frames=images,
+            input_text=new_text,
         )
 
 

python/sglang/srt/managers/image_processors/gemma3.py

@@ -0,0 +1,100 @@
+import asyncio
+from typing import List, Union
+
+from transformers.utils import logging
+
+from sglang.srt.managers.image_processor import (
+    BaseImageProcessor as SGLangBaseImageProcessor,
+)
+from sglang.srt.managers.image_processors.base_image_processor import (
+    get_global_processor,
+)
+from sglang.srt.models.gemma3_mm import Gemma3ForConditionalGeneration
+
+# Copied from: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma3/image_processing_gemma3_fast.py
+# will be removed in the future
+logger = logging.get_logger(__name__)
+
+
+class Gemma3SGLangImageProcessor(SGLangBaseImageProcessor):
+    def __init__(self, hf_config, server_args, _processor):
+        super().__init__(hf_config, server_args, _processor)
+        self.IMAGE_TOKEN = "<start_of_image>"
+        self.IM_START_TOKEN_ID = hf_config.boi_token_index
+        self.IM_END_TOKEN_ID = hf_config.eoi_token_index
+
+    @staticmethod
+    def _process_images_task(images, input_text, _hf_config):
+        if isinstance(images, list) and len(images) == 0:
+            images = None
+        processor = get_global_processor()
+        result = processor.__call__(
+            text=[input_text],
+            images=images,
+            padding=True,
+            return_tensors="pt",
+            # if RGBA, this needs to be set
+            # images_kwargs={
+            #     "input_data_format": ChannelDimension.FIRST
+            # }
+        )
+
+        pixel_values = getattr(result, "pixel_values", None)
+
+        return {
+            "input_ids": result.input_ids,
+            "pixel_values": pixel_values,
+        }
+
+    async def _process_images(self, images, input_text) -> dict:
+        if self.executor is not None:
+            loop = asyncio.get_event_loop()
+            return await loop.run_in_executor(
+                self.executor,
+                Gemma3SGLangImageProcessor._process_images_task,
+                images,
+                input_text,
+                self.hf_config,
+            )
+        else:
+            return self._process_images_task(images, input_text, self.hf_config)
+
+    async def process_images_async(
+        self,
+        image_data: List[Union[str, bytes]],
+        input_ids,
+        request_obj,
+        max_req_input_len,
+        *args,
+        **kwargs,
+    ):
+        if not image_data:
+            return None
+        if isinstance(image_data, str):
+            image_data = [image_data]
+
+        image_token = self.IMAGE_TOKEN
+        base_output = self.load_images(
+            input_ids=input_ids,
+            image_data=image_data,
+            image_token=image_token,
+            max_req_input_len=max_req_input_len,
+            discard_alpha_channel=True,
+        )
+
+        ret = await self._process_images(
+            input_text=base_output.input_text, images=base_output.all_frames
+        )
+
+        return {
+            "input_ids": ret["input_ids"].flatten().tolist(),
+            "pixel_values": ret["pixel_values"],
+            "image_hashes": base_output.image_hashes,
+            "im_start_id": self.IM_START_TOKEN_ID,
+            "im_end_id": self.IM_END_TOKEN_ID,
+        }
+
+
+ImageProcessorMapping = {
+    Gemma3ForConditionalGeneration: Gemma3SGLangImageProcessor,
+}

python/sglang/srt/managers/image_processors/janus_pro.py

@@ -60,7 +60,10 @@ class JanusProProcessor(SGLangBaseImageProcessor):
             image_data = [image_data]
 
         base_out = self.load_images(
-            input_ids, image_data, "<image_placeholder>", max_req_input_len
+            input_ids=input_ids,
+            image_data=image_data,
+            image_token="<image_placeholder>",
+            max_req_input_len=max_req_input_len,
         )
         images = base_out.all_frames
         res = await self._process_images(images=images, input_text=base_out.input_text)

python/sglang/srt/managers/image_processors/minicpmv.py

@@ -52,7 +52,10 @@ class MiniCPMVImageProcessor(BaseImageProcessor):
             image_data = [image_data]
 
         base_output = self.load_images(
-            input_ids, image_data, self.IMAGE_TOKEN, max_req_input_len
+            input_ids=input_ids,
+            image_data=image_data,
+            image_token=self.IMAGE_TOKEN,
+            max_req_input_len=max_req_input_len,
         )
         if base_output is None:
             return None

python/sglang/srt/managers/image_processors/qwen_vl.py

@@ -72,10 +72,10 @@ class Qwen2_5VLImageProcessor(BaseImageProcessor):
 
         image_token = self.IMAGE_TOKEN
         base_output = self.load_images(
-            input_ids,
-            image_data,
-            image_token,
-            max_req_input_len,
+            input_ids=input_ids,
+            image_data=image_data,
+            image_token=image_token,
+            max_req_input_len=max_req_input_len,
         )
 
         def smart_resize(

python/sglang/srt/managers/schedule_batch.py

@@ -49,7 +49,7 @@ from sglang.srt.model_executor.forward_batch_info import CaptureHiddenMode, Forw
 from sglang.srt.sampling.sampling_batch_info import SamplingBatchInfo
 from sglang.srt.sampling.sampling_params import SamplingParams
 from sglang.srt.server_args import ServerArgs
-from sglang.srt.utils import get_compiler_backend, next_power_of_2
+from sglang.srt.utils import get_compiler_backend
 
 if TYPE_CHECKING:
     from sglang.srt.speculative.eagle_utils import EagleDraftInput, EagleVerifyInput
@@ -207,6 +207,9 @@ class ImageInputs:
         return ret
 
     def merge(self, other):
+        """
+        merge image inputs when requests are being merged
+        """
         assert self.pixel_values.shape[1:] == other.pixel_values.shape[1:]
         self.pixel_values = np.concatenate([self.pixel_values, other.pixel_values])
 

python/sglang/srt/model_executor/forward_batch_info.py

@@ -33,6 +33,7 @@ from dataclasses import dataclass
 from enum import IntEnum, auto
 from typing import TYPE_CHECKING, List, Optional, Union
 
+import numpy as np
 import torch
 import triton
 import triton.language as tl
@@ -331,6 +332,32 @@ class ForwardBatch:
 
         return ret
 
+    def get_merged_image_inputs(self) -> Optional[ImageInputs]:
+        """
+        Merge all image inputs in the batch into a single ImageInputs object.
+
+        Returns:
+            if none, current batch contains no image input
+
+        """
+        if not self.image_inputs or all(x is None for x in self.image_inputs):
+            return None
+
+        # Filter out None values
+        valid_inputs = [x for x in self.image_inputs if x is not None]
+
+        # Start with the first valid image input
+        merged = valid_inputs[0]
+
+        # Merge remaining inputs
+        for img_input in valid_inputs[1:]:
+            merged.merge(img_input)
+
+        if isinstance(merged.pixel_values, np.ndarray):
+            merged.pixel_values = torch.from_numpy(merged.pixel_values)
+
+        return merged
+
     def _compute_mrope_positions(
         self, model_runner: ModelRunner, batch: ModelWorkerBatch
     ):

python/sglang/srt/models/gemma3_causal.py

@@ -0,0 +1,687 @@
+# Copyright 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 copy
+from typing import Iterable, Optional, Set, Tuple
+
+import einops
+import torch
+import torch.nn.functional as F
+from torch import nn
+from transformers import (
+    ROPE_INIT_FUNCTIONS,
+    AutoModel,
+    PretrainedConfig,
+    PreTrainedModel,
+)
+
+from sglang.srt.configs.gemma3 import Gemma3TextConfig
+from sglang.srt.distributed import get_tensor_model_parallel_world_size
+from sglang.srt.layers.activation import GeluAndMul
+from sglang.srt.layers.layernorm import Gemma3RMSNorm
+from sglang.srt.layers.linear import (
+    MergedColumnParallelLinear,
+    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 apply_rotary_pos_emb, get_rope
+from sglang.srt.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import (
+    default_weight_loader,
+    maybe_remap_kv_scale_name,
+)
+from sglang.srt.utils import add_prefix, make_layers
+
+
+# Adapted from:
+# https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/gemma3.py
+def extract_layer_index(prefix: str) -> int:
+    """Extract the layer index from a prefix string."""
+    parts = prefix.split(".")
+    for part in parts:
+        if part.startswith("layers."):
+            layer_str = part.split(".")[-1]
+            try:
+                return int(layer_str)
+            except ValueError:
+                continue
+    return -1
+
+
+class Gemma3MLP(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_activation: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size,
+            [intermediate_size] * 2,
+            bias=False,
+            quant_config=quant_config,
+            prefix=add_prefix("gate_up_proj", prefix),
+        )
+        self.down_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            prefix=add_prefix("down_proj", prefix),
+        )
+        if hidden_activation != "gelu_pytorch_tanh":
+            raise ValueError(
+                "Gemma3 uses `gelu_pytorch_tanh` as the hidden activation "
+                "function. Please set `hidden_activation` to "
+                "`gelu_pytorch_tanh`."
+            )
+        self.act_fn = GeluAndMul()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class Gemma3Attention(nn.Module):
+    def __init__(
+        self,
+        layer_id: int,
+        config: Gemma3TextConfig,
+        max_position_embeddings: int,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.layer_id = layer_id
+        self.config = config
+        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
+
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+
+        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
+
+        hidden_size = config.hidden_size
+
+        head_dim = getattr(
+            config, "head_dim", hidden_size // config.num_attention_heads
+        )
+        self.head_dim = head_dim
+
+        self.q_size = self.num_heads * self.head_dim
+
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = config.query_pre_attn_scalar**-0.5
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=config.attention_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=config.attention_bias,
+            quant_config=quant_config,
+            prefix=add_prefix("o_proj", prefix),
+        )
+
+        # Determine if layer uses sliding window based on pattern
+        self.is_sliding = bool((layer_id + 1) % config.sliding_window_pattern)
+
+        # Initialize the rotary embedding.
+        if self.is_sliding:
+            # Local attention. Override the values in config.json.
+            self.rope_theta = config.rope_local_base_freq
+            self.rope_scaling = {"rope_type": "default"}
+            # FIXME(mick): idk why vllm does this
+            # self.sliding_window = config.interleaved_sliding_window
+            self.sliding_window = config.sliding_window
+        else:
+            # Global attention. Use the values in config.json.
+            self.rope_theta = config.rope_theta
+            self.rope_scaling = config.rope_scaling
+            self.sliding_window = None
+
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+            logit_cap=getattr(self.config, "attn_logit_softcapping", None),
+            sliding_window_size=self.sliding_window,
+            prefix=add_prefix("attn", prefix),
+        )
+
+        # Gemma3 adds normalization for q and k
+        self.q_norm = Gemma3RMSNorm(dim=config.head_dim, eps=config.rms_norm_eps)
+        self.k_norm = Gemma3RMSNorm(dim=config.head_dim, eps=config.rms_norm_eps)
+
+    def naive_attn_with_masks(
+        self,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        out: torch.Tensor,
+        **kwargs,
+    ) -> torch.Tensor:
+        q = q.view(-1, self.num_heads, self.head_dim)
+        # Expand the key and value to handle GQA.
+        num_queries_per_kv = self.num_heads // self.num_kv_heads
+        k = k.view(-1, self.num_kv_heads, self.head_dim)
+        k = k.repeat_interleave(num_queries_per_kv, dim=-2)
+        v = v.view(-1, self.num_kv_heads, self.head_dim)
+        v = v.repeat_interleave(num_queries_per_kv, dim=-2)
+
+        if self.is_sliding:
+            attn_masks = kwargs["local_attn_masks"]
+        else:
+            attn_masks = kwargs["global_attn_masks"]
+
+        seq_lens = kwargs["seq_lens"]
+        start_idx = 0
+        for seq_len, attn_mask in zip(seq_lens, attn_masks):
+            end_idx = start_idx + seq_len
+            query = q[start_idx:end_idx].unsqueeze(0)
+            key = k[start_idx:end_idx].unsqueeze(0)
+            value = v[start_idx:end_idx].unsqueeze(0)
+
+            # Transpose.
+            query = query.transpose(1, 2)
+            key = key.transpose(1, 2)
+            value = value.transpose(1, 2)
+
+            output = F.scaled_dot_product_attention(
+                query,
+                key,
+                value,
+                attn_mask,
+                self.scaling,
+            )
+            output = output.transpose(1, 2).flatten(-2, -1)
+            out[start_idx:end_idx] = output
+            start_idx = end_idx
+        return out
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        forward_batch: ForwardBatch,
+        **kwargs,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        # [s, h * head_dim]
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+
+        # [s, h, head_dim]
+        q = q.unflatten(-1, (self.num_heads, self.head_dim))
+        # -> [h, s, head_dim]
+        q = q.transpose(0, 1).unsqueeze(0)
+        q = self.q_norm(q)
+        k = k.unflatten(-1, (self.num_kv_heads, self.head_dim))
+        # -> [h, s, head_dim]
+        k = k.transpose(0, 1).unsqueeze(0)
+        k = self.k_norm(k)
+
+        # q, k = self.rotary_emb(positions, q, k)
+        cos, sin = position_embeddings
+        q, k = apply_rotary_pos_emb(q, k, cos, sin)
+
+        # [b, h, s, head_dim] ->  [b, s, h, head_dim]
+        q = q.permute(0, 2, 1, 3)
+        k = k.permute(0, 2, 1, 3)
+
+        attn_output = self.attn(q, k, v, forward_batch=forward_batch)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class Gemma3DecoderLayer(nn.Module):
+    def __init__(
+        self,
+        layer_id: int,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = Gemma3Attention(
+            layer_id=layer_id,
+            config=config,
+            max_position_embeddings=config.max_position_embeddings,
+            quant_config=quant_config,
+            prefix=add_prefix("self_attn", prefix),
+        )
+        self.hidden_size = config.hidden_size
+        self.mlp = Gemma3MLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_activation=config.hidden_activation,
+            quant_config=quant_config,
+            prefix=add_prefix("mlp", prefix),
+        )
+        self.input_layernorm = Gemma3RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.post_attention_layernorm = Gemma3RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.pre_feedforward_layernorm = Gemma3RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.post_feedforward_layernorm = Gemma3RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.is_sliding = self.self_attn.is_sliding
+        self.layer_id = layer_id
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        position_embeddings_global: torch.Tensor,
+        position_embeddings_local: torch.Tensor,
+        forward_batch: ForwardBatch,
+        **kwargs,
+    ) -> tuple[
+        torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]
+    ]:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # apply global RoPE to non-sliding layer only
+        if self.self_attn.is_sliding:
+            position_embeddings = position_embeddings_local
+        else:
+            position_embeddings = position_embeddings_global
+
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            position_embeddings=position_embeddings,
+            forward_batch=forward_batch,
+            **kwargs,
+        )
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.pre_feedforward_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = self.post_feedforward_layernorm(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        return outputs
+
+
+class Gemma3RotaryEmbedding(nn.Module):
+    def __init__(self, config: Gemma3TextConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get(
+                "rope_type", config.rope_scaling.get("type")
+            )
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(
+                self.config, device, seq_len=seq_len
+            )
+            self.register_buffer(
+                "inv_freq", inv_freq, persistent=False
+            )  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if (
+            seq_len < self.original_max_seq_len
+            and self.max_seq_len_cached > self.original_max_seq_len
+        ):  # reset
+            # This .to() is needed if the model has been moved to a device after being initialized (because
+            # the buffer is automatically moved, but not the original copy)
+            self.original_inv_freq = self.original_inv_freq.to(device)
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block
+        inv_freq_expanded = (
+            self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        )
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = (
+            device_type
+            if isinstance(device_type, str) and device_type != "mps"
+            else "cpu"
+        )
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (
+                inv_freq_expanded.float().to(x.device) @ position_ids_expanded.float()
+            ).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class Gemma3TextScaledWordEmbedding(nn.Embedding):
+    """
+    This module overrides nn.Embeddings' forward by multiplying with embeddings scale.
+    """
+
+    def __init__(
+        self,
+        num_embeddings: int,
+        embedding_dim: int,
+        padding_idx: int,
+        embed_scale: Optional[float] = 1.0,
+    ):
+        super().__init__(num_embeddings, embedding_dim, padding_idx)
+        self.embed_scale = embed_scale
+
+    def forward(self, input_ids: torch.Tensor):
+        return super().forward(input_ids) * self.embed_scale
+
+
+class Gemma3TextModel(PreTrainedModel):
+    def __init__(
+        self,
+        config: Gemma3TextConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__(config=config)
+        self.config = config
+        self.quant_config = quant_config
+
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        # Gemma3 downcasts the below to float16, causing sqrt(3072)=55.4256 to become 55.5. See https://github.com/huggingface/transformers/pull/29402
+        self.embed_tokens = Gemma3TextScaledWordEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+            self.padding_idx,
+            embed_scale=self.config.hidden_size**0.5,
+        )
+
+        self.norm = Gemma3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Gemma3RotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # when we want to create a local RoPE layer. Config defaults should hold values for global RoPE
+        config = copy.deepcopy(config)
+        config.rope_theta = config.rope_local_base_freq
+        config.rope_scaling = {"rope_type": "default"}
+        self.rotary_emb_local = Gemma3RotaryEmbedding(config=config)
+
+        self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda idx, prefix: Gemma3DecoderLayer(
+                layer_id=idx,
+                config=config,
+                quant_config=quant_config,
+                prefix=prefix,
+            ),
+            prefix=add_prefix("layers", prefix),
+        )
+        self.norm = Gemma3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_init()
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        if input_embeds is None:
+            hidden_states = self.embed_tokens(input_ids)
+        else:
+            hidden_states = input_embeds
+
+        if len(positions.shape) == 1:
+            positions = einops.rearrange(positions, "s -> 1 s")
+
+        position_embeddings_global = self.rotary_emb(hidden_states, positions)
+        position_embeddings_local = self.rotary_emb_local(hidden_states, positions)
+        for layer in self.layers:
+            layer_outputs = layer(
+                positions=positions,
+                position_embeddings_global=position_embeddings_global,
+                position_embeddings_local=position_embeddings_local,
+                hidden_states=hidden_states,
+                forward_batch=forward_batch,
+                **kwargs,
+            )
+            hidden_states = layer_outputs[0]
+
+        hidden_states = self.norm(hidden_states)
+
+        return hidden_states
+
+
+class Gemma3ForCausalLM(PreTrainedModel):
+    config_class = Gemma3TextConfig
+
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+    config_class = Gemma3TextConfig
+    base_model_prefix = "language_model"
+
+    # 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),
+    }
+
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+
+    # LoRA specific attributes
+    supported_lora_modules = [
+        "qkv_proj",
+        "o_proj",
+        "gate_up_proj",
+        "down_proj",
+    ]
+    # Gemma does not apply LoRA to the embedding layer.
+    embedding_modules = {}
+    embedding_padding_modules = []
+    supports_lora = True
+
+    def __init__(
+        self,
+        config: Gemma3TextConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__(config=config)
+        self.config = config
+        self.quant_config = quant_config
+        self.model = Gemma3TextModel(
+            config, quant_config, prefix=add_prefix("model", prefix)
+        )
+        self.logits_processor = LogitsProcessor(config)
+
+        if self.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.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def dtype(self) -> torch.dtype:
+        return self.model.layers[0].mlp.gate_up_proj.weight.dtype
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+        **kwargs,
+    ) -> LogitsProcessor:
+
+        hidden_states = self.model(
+            input_ids, positions, forward_batch, input_embeds, **kwargs
+        )
+
+        return self.logits_processor(
+            input_ids, hidden_states, self.model.embed_tokens, 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", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+        params_dict = dict(self.named_parameters())
+        loaded_params: Set[str] = set()
+        for name, loaded_weight in weights:
+            for param_name, shard_name, shard_id in stacked_params_mapping:
+                # if param_name in name:
+                # print(f"{param_name} is already in {name}")
+                if shard_name not in name:
+                    continue
+                name = name.replace(shard_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:
+                # lm_head is not used in vllm as it is tied with embed_token.
+                # To prevent errors, skip loading lm_head.weight.
+                if "lm_head.weight" in name:
+                    continue
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
+
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        # unloaded_params = params_dict.keys() - loaded_params
+        # if unloaded_params:
+        #     logger.warning(
+        #         "Some weights are not initialized from checkpoints: %s", unloaded_params
+        #     )
+        return loaded_params
+
+
+EntryClass = Gemma3ForCausalLM
+AutoModel.register(Gemma3TextConfig, Gemma3ForCausalLM, exist_ok=True)

python/sglang/srt/models/gemma3_mm.py

@@ -0,0 +1,462 @@
+# Copyright 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.
+# ==============================================================================
+
+# Adapted from:
+# https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/gemma3_mm.py
+
+import logging
+from functools import lru_cache
+from typing import Dict, Iterable, List, Optional, Set, Tuple, TypedDict
+
+import torch
+from torch import nn
+from transformers import AutoModel, PreTrainedModel
+
+from sglang.srt.configs import Gemma3Config
+from sglang.srt.hf_transformers_utils import get_processor
+from sglang.srt.layers.layernorm import Gemma3RMSNorm
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.managers.multi_modality_padding import (
+    MultiModalityDataPaddingPatternTokenPairs,
+)
+from sglang.srt.managers.schedule_batch import ImageInputs
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+from sglang.srt.model_loader.weight_utils import (
+    default_weight_loader,
+    maybe_remap_kv_scale_name,
+)
+from sglang.srt.models.gemma3_causal import Gemma3ForCausalLM
+from sglang.srt.utils import add_prefix
+
+logger = logging.getLogger(__name__)
+
+cached_get_processor = lru_cache(get_processor)
+
+
+class Gemma3ImagePixelInputs(TypedDict):
+    pixel_values: torch.Tensor
+    """Shape: `(batch_size * num_images, num_channels, height, width)`"""
+
+
+class Gemma3MultiModalProjector(nn.Module):
+    """Projector for Gemma3 multimodal."""
+
+    def __init__(self, config: Gemma3Config):
+        super().__init__()
+
+        self.mm_input_projection_weight = nn.Parameter(
+            torch.zeros(
+                config.vision_config.hidden_size, config.text_config.hidden_size
+            )
+        )
+
+        self.mm_soft_emb_norm = Gemma3RMSNorm(
+            config.vision_config.hidden_size, eps=config.vision_config.layer_norm_eps
+        )
+
+        self.patches_per_image = int(
+            config.vision_config.image_size // config.vision_config.patch_size
+        )
+        self.tokens_per_side = int(config.mm_tokens_per_image**0.5)
+        self.kernel_size = self.patches_per_image // self.tokens_per_side
+        self.avg_pool = nn.AvgPool2d(
+            kernel_size=self.kernel_size, stride=self.kernel_size
+        )
+
+    def forward(self, vision_outputs: torch.Tensor) -> torch.Tensor:
+        batch_size, seq_length, hidden_size = vision_outputs.shape
+
+        # Reshape for pooling
+        reshaped_vision_outputs = vision_outputs.transpose(1, 2)
+        reshaped_vision_outputs = reshaped_vision_outputs.reshape(
+            batch_size, hidden_size, self.patches_per_image, self.patches_per_image
+        )
+        reshaped_vision_outputs = reshaped_vision_outputs.contiguous()
+
+        # Apply pooling
+        pooled_vision_outputs = self.avg_pool(reshaped_vision_outputs)
+        pooled_vision_outputs = pooled_vision_outputs.flatten(2)
+        pooled_vision_outputs = pooled_vision_outputs.transpose(1, 2)
+
+        # Apply normalization
+        normed_vision_outputs = self.mm_soft_emb_norm(pooled_vision_outputs)
+
+        # Project to text embedding space
+        projected_vision_outputs = torch.matmul(
+            normed_vision_outputs, self.mm_input_projection_weight
+        )
+
+        return projected_vision_outputs.type_as(vision_outputs)
+
+
+class Gemma3ForConditionalGeneration(PreTrainedModel):
+    config_class = Gemma3Config
+    """Gemma3 multimodal model for conditional generation."""
+
+    # 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),
+    }
+
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+
+    # LoRA specific attributes
+    supported_lora_modules = [
+        "qkv_proj",
+        "o_proj",
+        "gate_up_proj",
+        "down_proj",
+    ]
+    # Gemma does not apply LoRA to the embedding layer.
+    embedding_modules = {}
+    embedding_padding_modules = []
+    supports_lora = True
+
+    def __init__(
+        self,
+        config: Gemma3Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__(config=config)
+        self.config = config
+        self.quant_config = quant_config
+        # Vision components
+        # TODO: replace with vision attention
+        # self.vision_tower = SiglipVisionModel(
+        #     config.vision_config,
+        #     quant_config,
+        #     prefix=add_prefix("vision_tower", prefix),
+        # )
+        self.vision_tower = AutoModel.from_config(config=config.vision_config)
+        self.multi_modal_projector = Gemma3MultiModalProjector(config)
+        self.vocab_size = config.text_config.vocab_size
+
+        # Text model
+        self.language_model = Gemma3ForCausalLM(
+            config.text_config, quant_config, prefix=add_prefix("model", prefix)
+        )
+        if self.language_model.logits_processor.logit_scale:
+            logit_scale = getattr(config, "logit_scale", 1.0)
+            self.language_model.logits_processor.logit_scale *= logit_scale
+        self.post_init()
+
+    def pad_input_ids(
+        self, input_ids: List[int], image_inputs: ImageInputs
+    ) -> List[int]:
+        """Pad input IDs with image tokens."""
+        # Get special token IDs
+        im_start_id: int = image_inputs.im_start_id
+        im_end_id: int = image_inputs.im_end_id
+
+        media_token_pairs = [(im_start_id, im_end_id)]
+        pattern = MultiModalityDataPaddingPatternTokenPairs(media_token_pairs)
+        ids = pattern.pad_input_tokens(input_ids, image_inputs)
+        return ids
+
+    def prepare_attn_masks(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        mask_dtype: torch.dtype,
+        **kwargs,
+    ) -> Dict:
+        """Prepare attention masks for multimodal inputs."""
+        kwargs["has_images"] = True
+
+        # Distinguish sequences by position id 0
+        start_indices = (positions == 0).cpu().nonzero()
+        num_seqs = len(start_indices)
+        seq_lens = []
+
+        for i in range(num_seqs):
+            start_idx = start_indices[i].item()
+            if i < num_seqs - 1:
+                end_idx = start_indices[i + 1].item()
+            else:
+                end_idx = len(input_ids)
+            seq_lens.append(end_idx - start_idx)
+
+        kwargs["seq_lens"] = seq_lens
+
+        # Create attention masks
+        global_attn_masks = []
+        local_attn_masks = []
+        sliding_window = self.config.text_config.interleaved_sliding_window
+
+        start_idx = 0
+        for seq_len in seq_lens:
+            end_idx = start_idx + seq_len
+            input_token_ids = input_ids[start_idx:end_idx]
+            start_idx = end_idx
+
+            # Create global causal mask
+            global_attn_mask = torch.empty(
+                1,
+                1,
+                seq_len,
+                seq_len,
+                dtype=mask_dtype,
+                device=input_ids.device,
+            )
+            global_attn_mask.fill_(float("-inf"))
+            global_attn_mask = global_attn_mask.triu(diagonal=1)
+
+            # Consider bidirectional attention between image tokens
+            img_mask = torch.zeros_like(global_attn_mask)
+            img_pos = input_token_ids == self.config.image_token_index
+            img_mask[:, :, :, img_pos] += 1
+            img_mask[:, :, img_pos, :] += 1
+            global_attn_mask = torch.where(img_mask == 2, 0, global_attn_mask)
+            global_attn_masks.append(global_attn_mask)
+
+            # Create local causal mask with sliding window
+            local_attn_mask = torch.ones_like(global_attn_mask)
+            local_attn_mask = torch.tril(local_attn_mask, diagonal=-sliding_window)
+            local_attn_mask = torch.where(
+                local_attn_mask == 0, global_attn_mask, float("-inf")
+            )
+            local_attn_masks.append(local_attn_mask)
+
+        kwargs["global_attn_masks"] = global_attn_masks
+        kwargs["local_attn_masks"] = local_attn_masks
+        return kwargs
+
+    def get_input_embeddings(self):
+        return self.language_model.get_input_embeddings()
+
+    def get_image_features(self, pixel_values: torch.Tensor):
+        """
+        Projects the last hidden state from the vision model into language model space.
+
+        Args:
+            pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`)
+               The tensors corresponding to the input images.
+        Returns:
+            image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`).
+        """
+        pixel_values = pixel_values.to("cuda")
+        pixel_values = pixel_values.to(dtype=self.language_model.dtype())
+
+        vision_outputs = self.vision_tower(pixel_values=pixel_values).last_hidden_state
+        image_features = self.multi_modal_projector(vision_outputs)
+        return image_features
+
+    def embed_image_inputs(
+        self,
+        input_ids: torch.Tensor,
+        forward_batch: ForwardBatch,
+        image_input: ImageInputs,
+    ) -> torch.Tensor:
+        if input_ids is None:
+            raise ValueError("Unimplemented")
+        # boolean-masking image tokens
+        special_image_mask = torch.isin(
+            input_ids,
+            torch.tensor(image_input.pad_values, device=input_ids.device),
+        ).unsqueeze(-1)
+        num_image_tokens_in_input_ids = special_image_mask.sum()
+
+        inputs_embeds = None
+        if num_image_tokens_in_input_ids == 0:
+            inputs_embeds = self.get_input_embeddings()(input_ids)
+            return inputs_embeds
+        else:
+            # print(f"image tokens from input_ids: {inputs_embeds[special_image_mask].numel()}")
+            image_features = self.get_image_features(image_input.pixel_values)
+
+            # print(f"image tokens from image embeddings: {image_features.numel()}")
+            num_image_tokens_in_embedding = (
+                image_features.shape[0] * image_features.shape[1]
+            )
+
+            if num_image_tokens_in_input_ids != num_image_tokens_in_embedding:
+                num_image = num_image_tokens_in_input_ids // image_features.shape[1]
+                image_features = image_features[:num_image, :]
+                logger.warning(
+                    f"Number of images does not match number of special image tokens in the input text. "
+                    f"Got {num_image_tokens_in_input_ids} image tokens in the text but {num_image_tokens_in_embedding} "
+                    "tokens from image embeddings."
+                )
+
+            # Important: clamp after extracting original image boundaries
+            input_ids.clamp_(min=0, max=self.vocab_size - 1)
+
+            inputs_embeds = self.get_input_embeddings()(input_ids)
+
+            special_image_mask = special_image_mask.expand_as(inputs_embeds).to(
+                inputs_embeds.device
+            )
+
+            image_features = image_features.to(
+                inputs_embeds.device, inputs_embeds.dtype
+            )
+            inputs_embeds = inputs_embeds.masked_scatter(
+                special_image_mask, image_features
+            )
+
+        return inputs_embeds
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        positions: torch.Tensor,
+        forward_batch: ForwardBatch,
+        input_embeds: torch.Tensor = None,
+        **kwargs: object,
+    ) -> LogitsProcessor:
+        r"""
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.text_config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.text_config.vocab_size]`.
+
+            logits_to_keep (`int` or `torch.Tensor`, *optional*):
+                If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+                If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+                This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import AutoProcessor, Gemma3ForConditionalGeneration
+
+        >>> model = Gemma3ForConditionalGeneration.from_pretrained("google/Gemma3-test-224px-hf")
+        >>> processor = AutoProcessor.from_pretrained("google/Gemma3-test-224px-hf")
+
+        >>> prompt = "answer en Where is the cow standing?"
+        >>> url = "https://huggingface.co/gv-hf/Gemma3-test-224px-hf/resolve/main/cow_beach_1.png"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, text=prompt,  return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(**inputs, max_length=30)
+        >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "answer en Where is the cow standing?\nbeach"
+        ```"""
+
+        # Important: position_ids in Gemma3 are 1-indexed
+        # This really does cost me sometime
+        positions += 1
+
+        # Replace image id with PAD if the image token if OOV, to avoid index-errors
+        if input_ids is not None and self.config.image_token_index >= self.vocab_size:
+            special_image_mask = input_ids == self.config.image_token_index
+            llm_input_ids = input_ids.clone()
+            llm_input_ids[special_image_mask] = 0
+        else:
+            llm_input_ids = input_ids
+
+        merged_image_input = forward_batch.get_merged_image_inputs()
+
+        if (
+            not forward_batch.forward_mode.is_decode()
+            and merged_image_input is not None
+        ):
+            inputs_embeds = self.embed_image_inputs(
+                input_ids=llm_input_ids,
+                forward_batch=forward_batch,
+                image_input=merged_image_input,
+            )
+        else:
+            llm_input_ids.clamp_(min=0, max=self.vocab_size - 1)
+            inputs_embeds = self.get_input_embeddings()(llm_input_ids)
+
+        outputs = self.language_model(
+            input_ids=None,
+            positions=positions,
+            forward_batch=forward_batch,
+            input_embeds=inputs_embeds,
+            **kwargs,
+        )
+
+        return outputs
+
+    def tie_weights(self):
+        return self.language_model.tie_weights()
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        """Load weights for the model."""
+        params_dict = dict(self.named_parameters())
+        loaded_params: Set[str] = set()
+
+        for name, loaded_weight in weights:
+            if "language_model" in name:
+                # Gemma3ForCausalLM.load_weights(self, [(name.replace("language_model.", ""), loaded_weight)])
+                causal_loaded_params = Gemma3ForCausalLM.load_weights(
+                    self, [(name, loaded_weight)]
+                )
+                loaded_params.update(causal_loaded_params)
+                continue
+            else:
+                # Skip lm_head.weight as it's tied with embed_tokens
+                if "lm_head.weight" in name:
+                    continue
+
+                # Skip loading extra bias for GPTQ models
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+
+                # Remapping the name of FP8 kv-scale
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                weight_loader(param, loaded_weight)
+                loaded_params.add(name)
+        unloaded_params = params_dict.keys() - loaded_params
+        if unloaded_params:
+            pass
+            # raise RuntimeError(
+            #     f"Some weights are not initialized from checkpoints: {unloaded_params}")
+        return loaded_params
+
+
+EntryClass = Gemma3ForConditionalGeneration
+
+AutoModel.register(Gemma3Config, Gemma3ForConditionalGeneration, exist_ok=True)

python/sglang/srt/utils.py

@@ -41,7 +41,6 @@ from functools import lru_cache
 from importlib.metadata import PackageNotFoundError, version
 from importlib.util import find_spec
 from io import BytesIO
-from multiprocessing import Pool
 from multiprocessing.reduction import ForkingPickler
 from pathlib import Path
 from typing import Any, Callable, Dict, List, Optional, Protocol, Set, Tuple, Union
@@ -454,8 +453,9 @@ def load_image(image_file: Union[str, bytes]):
         image = Image.open(BytesIO(image_file))
     elif image_file.startswith("http://") or image_file.startswith("https://"):
         timeout = int(os.getenv("REQUEST_TIMEOUT", "3"))
-        response = requests.get(image_file, timeout=timeout)
-        image = Image.open(BytesIO(response.content))
+        response = requests.get(image_file, stream=True, timeout=timeout).raw
+        image = Image.open(response)
+        response.close()
     elif image_file.lower().endswith(("png", "jpg", "jpeg", "webp", "gif")):
         image = Image.open(image_file)
     elif image_file.startswith("data:"):