model(vlm): mistral 3.1 (#5099)

Co-authored-by: KivenChen <sleigh-queue-0y@icloud.com>

docs/supported_models/vision_language_models.md

+# Vision Language Models
+
+These models accept multi-modal inputs (e.g., images and text) and generate text output. They augment language models with visual encoders and require a specific chat template for handling vision prompts.
+
+## Example launch Command
+
+```shell
+python3 -m sglang.launch_server \
+  --model-path meta-llama/Llama-3.2-11B-Vision-Instruct \  # example HF/local path
+  --host 0.0.0.0 \
+  --port 30000 \
+```
+
+## Supporting Matrixs
+
+| Model Family (Variants)        | Example HuggingFace Identifier                     | Chat Template        | Description                                                                            |
+|--------------------------------|--------------------------------------------------|----------------------|----------------------------------------------------------------------------------------|
+| **Qwen-VL** (Qwen2 series)     | `Qwen/Qwen2.5-VL-7B-Instruct`                    | `qwen2-vl`           | Alibaba’s vision-language extension of Qwen; for example, Qwen2.5-VL (7B and larger variants) can analyze and converse about image content. |
+| **DeepSeek-VL2**               | `deepseek-ai/deepseek-vl2`                       | `deepseek-vl2`       | Vision-language variant of DeepSeek (with a dedicated image processor), enabling advanced multimodal reasoning on image and text inputs. |
+| **Janus-Pro** (1B, 7B)      | `deepseek-ai/Janus-Pro-7B`                     | `janus-pro`       | DeepSeek’s open-source multimodal model capable of both image understanding and generation. Janus-Pro employs a decoupled architecture for separate visual encoding paths, enhancing performance in both tasks. |
+| **MiniCPM-V / MiniCPM-o**      | `openbmb/MiniCPM-V-2_6`                          | `minicpmv`           | MiniCPM-V (2.6, ~8B) supports image inputs, and MiniCPM-o adds audio/video; these multimodal LLMs are optimized for end-side deployment on mobile/edge devices. |
+| **Llama 3.2 Vision** (11B)     | `meta-llama/Llama-3.2-11B-Vision-Instruct`        | `llama_3_vision`     | Vision-enabled variant of Llama 3 (11B) that accepts image inputs for visual question answering and other multimodal tasks. |
+| **Pixtral** (12B, 124B)              | `mistral-community/pixtral-12b`                           | `mistral`     | Pixtral is a vision-language model from Mistral AI that can process both text and images. |
+| **LLaVA** (v1.5 & v1.6)        | *e.g.* `liuhaotian/llava-v1.5-13b`               | `vicuna_v1.1`        | Open vision-chat models that add an image encoder to LLaMA/Vicuna (e.g. LLaMA2 13B) for following multimodal instruction prompts. |
+| **LLaVA-NeXT** (8B, 72B)       | `lmms-lab/llava-next-72b`                        | `chatml-llava`       | Improved LLaVA models (with an 8B Llama3 version and a 72B version) offering enhanced visual instruction-following and accuracy on multimodal benchmarks. |
+| **LLaVA-OneVision**            | `lmms-lab/llava-onevision-qwen2-7b-ov`           | `chatml-llava`       | Enhanced LLaVA variant integrating Qwen as the backbone; supports multiple images (and even video frames) as inputs via an OpenAI Vision API-compatible format. |
+| **Gemma 3 (Multimodal)**       | `google/gemma-3-4b-it`                           | `gemma-it`           | Gemma 3’s larger models (4B, 12B, 27B) accept images (each image encoded as 256 tokens) alongside text in a combined 128K-token context. |
+| **Kimi-VL** (A3B)              | `moonshotai/Kimi-VL-A3B-Instruct`                | `kimi-vl`            | Kimi-VL is a multimodal model that can understand and generate text from images. |
+| **Mistral-Small-3.1-24B**      | `mistralai/Mistral-Small-3.1-24B-Instruct-2503`    | `mistral`            | Mistral 3.1 is a multimodal model that can generate text from text or images input. It also supports tool calling and structured output. |

python/sglang/srt/configs/model_config.py

@@ -549,6 +549,7 @@ multimodal_model_archs = [
     "LlavaVidForCausalLM",
     "MiniCPMO",
     "MiniCPMV",
+    "Mistral3ForConditionalGeneration",
     "MultiModalityCausalLM",
     "MllamaForConditionalGeneration",
     "Qwen2VLForConditionalGeneration",

python/sglang/srt/managers/multimodal_processors/llava.py

@@ -20,6 +20,7 @@ from sglang.srt.models.llava import (
     LlavaQwenForCausalLM,
 )
 from sglang.srt.models.llavavid import LlavaVidForCausalLM
+from sglang.srt.models.mistral import Mistral3ForConditionalGeneration
 from sglang.srt.utils import load_image, logger
 from sglang.utils import get_exception_traceback
 
@@ -176,10 +177,10 @@ class LlavaImageProcessor(BaseMultimodalProcessor):
 
 class LlavaMultimodalProcessor(BaseMultimodalProcessor):
     """
-    This is a wrapper class used to identify the multimodal processor for Llava architecture models.
+    This is a wrapper class used to identify the multimodal processor for Llava architectures' vision model.
     """
 
-    models = [LlavaForConditionalGeneration]
+    models = [LlavaForConditionalGeneration, Mistral3ForConditionalGeneration]
 
     def _get_sgl_processor_cls(self, model_type: str):
         if hf_name := HF_MAPPING_NAMES.get(model_type):

python/sglang/srt/models/llava.py

@@ -135,7 +135,6 @@ class LlavaBaseForCausalLM(nn.Module):
         """
         image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
         # NOTE: This is not memory efficient. (output_hidden_states=True) will save all the hidden stated.
-
         selected_image_feature = image_outputs.hidden_states[self.vision_feature_layer]
         if self.vision_feature_select_strategy in ["default", "patch"]:
             selected_image_feature = selected_image_feature[:, 1:]
@@ -146,7 +145,6 @@ class LlavaBaseForCausalLM(nn.Module):
                 f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}"
             )
         image_features = self.multi_modal_projector(selected_image_feature)
-
         return image_features
 
     @torch.no_grad()
@@ -613,6 +611,10 @@ class LlavaForConditionalGeneration(LlavaBaseForCausalLM):
 
     MULTIMODAL_PROJECTOR_TYPE = LlavaMultiModalProjector
 
+    @property
+    def dtype(self):
+        return self.torch_dtype
+
     def pad_input_ids(self, input_ids: List[int], image_inputs: MultimodalInputs):
         if hasattr(self.vision_tower, "pad_input_ids"):
             return self.vision_tower.pad_input_ids(input_ids, image_inputs)
@@ -672,11 +674,17 @@ class LlavaForConditionalGeneration(LlavaBaseForCausalLM):
         assert hasattr(config, "text_config")
         assert hasattr(config, "vision_config")
         self.config = config
-        self.text_config = config.text_config
-        self.vision_config = config.vision_config
+        self.text_config = self.config.text_config
+        self.vision_config = self.config.vision_config
+        self.torch_dtype = getattr(self.config, "torch_dtype")
+
+        if not getattr(self.text_config, "torch_dtype"):
+            self.text_config.torch_dtype = self.torch_dtype
+        if not getattr(self.vision_config, "torch_dtype"):
+            self.vision_config.torch_dtype = self.torch_dtype
 
         if not hasattr(self.config, "vocab_size"):
-            self.config.vocab_size = self.config.text_config.vocab_size
+            self.config.vocab_size = self.text_config.vocab_size
         if not hasattr(self.config, "image_aspect_ratio"):
             self.config.image_aspect_ratio = "anyres"
         if not hasattr(self.config, "image_grid_pinpoints"):
@@ -697,39 +705,39 @@ class LlavaForConditionalGeneration(LlavaBaseForCausalLM):
         if not hasattr(self.config, "projector_hidden_act"):
             self.config.projector_hidden_act = "gelu"
 
-        self.vision_feature_layer = getattr(config, "vision_feature_layer", -1)
+        self.vision_feature_layer = getattr(self.config, "vision_feature_layer", -1)
         self.vision_feature_select_strategy = getattr(
-            config, "vision_feature_select_strategy", "full"
+            self.config, "vision_feature_select_strategy", "full"
         )
-        self.image_size = self.config.vision_config.image_size
-        self.patch_size = self.config.vision_config.patch_size
+        self.image_size = self.vision_config.image_size
+        self.patch_size = self.vision_config.patch_size
 
-        self.mm_patch_merge_type = config.mm_patch_merge_type
-        self.image_aspect_ratio = config.image_aspect_ratio
-        self.image_grid_pinpoints = config.image_grid_pinpoints
+        self.mm_patch_merge_type = self.config.mm_patch_merge_type
+        self.image_aspect_ratio = self.config.image_aspect_ratio
+        self.image_grid_pinpoints = self.config.image_grid_pinpoints
 
         self.image_feature_len = int((self.image_size // self.patch_size) ** 2)
 
         self.multi_modal_projector = self.MULTIMODAL_PROJECTOR_TYPE(config)
 
         language_model_cls = self._get_sgl_model_cls(
-            config.text_config, AutoModelForCausalLM
+            self.text_config, AutoModelForCausalLM
         )
-        vision_model_cls = self._get_sgl_model_cls(config.vision_config, AutoModel)
+        vision_model_cls = self._get_sgl_model_cls(self.vision_config, AutoModel)
         self.language_model = language_model_cls(
-            config.text_config,
+            self.text_config,
             quant_config=quant_config,
             prefix=add_prefix("language_model", prefix),
         )
         self.vision_tower = vision_model_cls(
-            config.vision_config,
+            self.vision_config,
             quant_config=quant_config,
             prefix=add_prefix("vision_tower", prefix),
         )
 
-        if "unpad" in getattr(config, "mm_patch_merge_type", ""):
+        if "unpad" in getattr(self.config, "mm_patch_merge_type", ""):
             self.language_model.model.image_newline = nn.Parameter(
-                torch.empty(config.text_config.hidden_size, dtype=torch.float16)
+                torch.empty(self.text_config.hidden_size, dtype=self.torch_dtype)
             )
 
     def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:

python/sglang/srt/models/mistral.py

@@ -13,6 +13,12 @@
 # ==============================================================================
 """Inference-only Mistral model."""
 
+from typing import List, Union
+
+import torch
+from transformers.models.mistral3.modeling_mistral3 import Mistral3MultiModalProjector
+
+from sglang.srt.managers.schedule_batch import MultimodalDataItem
 from sglang.srt.models.llama import LlamaForCausalLM
 
 
@@ -20,4 +26,68 @@ class MistralForCausalLM(LlamaForCausalLM):
     pass
 
 
-EntryClass = MistralForCausalLM
+class Mistral3ForConditionalGeneration:
+    MULTIMODAL_PROJECTOR_TYPE = Mistral3MultiModalProjector
+
+    def __init__(self, **kwargs):
+        # lazy load inner class
+        # to bypass circular import
+        from sglang.srt.models.llava import LlavaForConditionalGeneration
+
+        # override config: mistral's projector adds patchmerger that doesn't require padding
+        kwargs["config"].vision_config.pad_image_border = False
+
+        self.inner = LlavaForConditionalGeneration(**kwargs)
+        self.inner.multi_modal_projector = self.MULTIMODAL_PROJECTOR_TYPE(
+            kwargs["config"]
+        )
+        self.inner.get_image_feature = self.get_image_feature
+
+    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        """Extract features from image inputs.
+
+        Args:
+            items: List of MultimodalDataItem objects containing image data
+                Note that an item can be either "image" or "multi-images"
+
+        Returns:
+            torch.Tensor: features from image inputs, concatenated
+        """
+        features = []
+        for item in items:
+            # in each item, we assume pixel_values is always batched
+            pixel_values, image_sizes = item.pixel_values, item.image_sizes
+            image_outputs = self.vision_tower(
+                pixel_values, image_sizes, output_hidden_states=True
+            )
+            selected_image_feature = image_outputs.hidden_states[
+                self.vision_feature_layer
+            ]
+
+            if self.vision_feature_select_strategy in ["default", "patch"]:
+                selected_image_feature = selected_image_feature[:, 1:]
+            elif self.vision_feature_select_strategy == "full":
+                selected_image_feature = selected_image_feature
+            else:
+                raise ValueError(
+                    f"Unexpected select feature: {self.vision_feature_select_strategy}"
+                )
+            features.append(
+                self.multi_modal_projector(
+                    selected_image_feature.squeeze(0), image_sizes
+                )
+            )
+        ret = torch.cat(features, dim=0)
+        return ret
+
+    def __getattr__(self, name):
+        return getattr(self.inner, name)
+
+    def __hasattr__(self, name):
+        return hasattr(self.inner, name)
+
+    def __call__(self, *args, **kwargs):
+        return self.inner(*args, **kwargs)
+
+
+EntryClass = [MistralForCausalLM, Mistral3ForConditionalGeneration]

test/srt/test_vision_openai_server.py

@@ -664,6 +664,28 @@ class TestPixtralServer(TestOpenAIVisionServer):
         pass
 
 
+class TestMistral3_1Server(TestOpenAIVisionServer):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "unsloth/Mistral-Small-3.1-24B-Instruct-2503"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.api_key = "sk-123456"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--trust-remote-code",
+                "--mem-fraction-static",
+                "0.8",
+            ],
+        )
+        cls.base_url += "/v1"
+
+    def test_video_chat_completion(self):
+        pass
+
+
 class TestDeepseekVL2Server(TestOpenAIVisionServer):
     @classmethod
     def setUpClass(cls):