Support Intern-S2-Preview (#24875)

python/sglang/srt/configs/__init__.py

@@ -8,6 +8,7 @@ from sglang.srt.configs.dots_vlm import DotsVLMConfig
 from sglang.srt.configs.exaone import ExaoneConfig
 from sglang.srt.configs.falcon_h1 import FalconH1Config
 from sglang.srt.configs.granitemoehybrid import GraniteMoeHybridConfig
+from sglang.srt.configs.interns2preview import InternS2PreviewConfig
 from sglang.srt.configs.janus_pro import MultiModalityConfig
 from sglang.srt.configs.jet_nemotron import JetNemotronConfig
 from sglang.srt.configs.jet_vlm import JetVLMConfig
@@ -57,6 +58,7 @@ __all__ = [
     "Qwen3NextConfig",
     "Qwen3_5Config",
     "Qwen3_5MoeConfig",
+    "InternS2PreviewConfig",
     "DotsVLMConfig",
     "DotsOCRConfig",
     "FalconH1Config",

python/sglang/srt/configs/interns2preview.py

@@ -0,0 +1,23 @@
+from sglang.srt.configs.qwen3_5 import (
+    Qwen3_5MoeConfig,
+    Qwen3_5MoeTextConfig,
+    Qwen3_5MoeVisionConfig,
+)
+
+
+class InternS2PreviewVisionConfig(Qwen3_5MoeVisionConfig):
+    model_type = "intern_s2_preview"
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+
+class InternS2PreviewConfig(Qwen3_5MoeConfig):
+    model_type = "intern_s2_preview"
+    sub_configs = {
+        "vision_config": InternS2PreviewVisionConfig,
+        "text_config": Qwen3_5MoeTextConfig,
+    }
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)

python/sglang/srt/configs/model_config.py

@@ -466,6 +466,7 @@ class ModelConfig:
         if is_draft_model and self.hf_config.architectures[0] in [
             "Qwen3_5ForConditionalGeneration",
             "Qwen3_5MoeForConditionalGeneration",
+            "InternS2PreviewForConditionalGeneration",
         ]:
             self.hf_config.architectures[0] = "Qwen3_5ForCausalLMMTP"
             self.hf_config.num_nextn_predict_layers = 1
@@ -1522,6 +1523,7 @@ multimodal_model_archs = [
     "Qwen3VLMoeForConditionalGeneration",
     "Qwen3_5ForConditionalGeneration",
     "Qwen3_5MoeForConditionalGeneration",
+    "InternS2PreviewForConditionalGeneration",
     "Qwen3ASRForConditionalGeneration",
     "Qwen3OmniMoeForConditionalGeneration",
     "KimiVLForConditionalGeneration",

python/sglang/srt/disaggregation/encode_server.py

@@ -970,7 +970,13 @@ class MMEncoder:
             # Get additional video metadata
             if (
                 self.model_type
-                in ["qwen3_vl", "qwen3_vl_moe", "qwen3_5", "qwen3_5_moe"]
+                in [
+                    "qwen3_vl",
+                    "qwen3_vl_moe",
+                    "qwen3_5",
+                    "qwen3_5_moe",
+                    "intern_s2_preview",
+                ]
                 and video_processor_kwargs.get("video_metadata", None) is not None
             ):
                 # For qwen3-vl/qwen3.5 models, we need to store the video timestamps

python/sglang/srt/layers/rotary_embedding/mrope_rope_index.py

@@ -159,6 +159,7 @@ def get_rope_index(
                     "qwen3_vl_moe",
                     "qwen3_5",
                     "qwen3_5_moe",
+                    "intern_s2_preview",
                 ):
                     t_index = (
                         torch.arange(llm_grid_t, device=position_ids.device)

python/sglang/srt/model_executor/model_runner.py

@@ -40,6 +40,7 @@ from sglang.srt.configs import (
     BailingHybridConfig,
     FalconH1Config,
     GraniteMoeHybridConfig,
+    InternS2PreviewConfig,
     JetNemotronConfig,
     JetVLMConfig,
     KimiLinearConfig,
@@ -2193,6 +2194,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
             Qwen3NextConfig
             | Qwen3_5Config
             | Qwen3_5MoeConfig
+            | InternS2PreviewConfig
             | JetNemotronConfig
             | JetVLMConfig,
         ):

python/sglang/srt/models/interns2preview.py

@@ -0,0 +1,9 @@
+# Models
+from sglang.srt.models.qwen3_5 import Qwen3_5MoeForConditionalGeneration
+
+
+class InternS2PreviewForConditionalGeneration(Qwen3_5MoeForConditionalGeneration):
+    """InternS2Preview Vision-Language Model."""
+
+
+EntryClass = [InternS2PreviewForConditionalGeneration]

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

@@ -17,6 +17,7 @@ from sglang.srt.managers.schedule_batch import (
     MultimodalDataItem,
     MultimodalProcessorOutput,
 )
+from sglang.srt.models.interns2preview import InternS2PreviewForConditionalGeneration
 from sglang.srt.models.qwen2_5_vl import Qwen2_5_VLForConditionalGeneration
 from sglang.srt.models.qwen2_vl import Qwen2VLForConditionalGeneration
 from sglang.srt.models.qwen3_5 import (
@@ -246,6 +247,7 @@ class QwenVLImageProcessor(SGLangBaseProcessor):
         Qwen3VLMoeForConditionalGeneration,
         Qwen3_5ForConditionalGeneration,
         Qwen3_5MoeForConditionalGeneration,
+        InternS2PreviewForConditionalGeneration,
         Qwen3OmniMoeForConditionalGeneration,
     ]
 
@@ -419,7 +421,14 @@ class QwenVLImageProcessor(SGLangBaseProcessor):
                 audio_seq_lens = (audio_seq_lens - 2) // 2 + 1
 
         if (
-            self.model_type in ["qwen3_vl", "qwen3_vl_moe", "qwen3_5", "qwen3_5_moe"]
+            self.model_type
+            in [
+                "qwen3_vl",
+                "qwen3_vl_moe",
+                "qwen3_5",
+                "qwen3_5_moe",
+                "intern_s2_preview",
+            ]
             and video_timestamps is not None
         ):
             input_ids, offsets, modality_list = self.build_input_ids_with_timestamps(
@@ -522,6 +531,7 @@ class QwenVLImageProcessor(SGLangBaseProcessor):
             "qwen3_vl_moe",
             "qwen3_5",
             "qwen3_5_moe",
+            "intern_s2_preview",
         ):
             mm_items, input_ids, ret = self.process_and_combine_mm_data(
                 base_output,

python/sglang/srt/server_args.py

@@ -2231,6 +2231,7 @@ class ServerArgs:
             "Qwen3VLMoeForConditionalGeneration",
             "Qwen3NextForCausalLM",
             "Qwen3_5MoeForConditionalGeneration",
+            "InternS2PreviewForConditionalGeneration",
             "Qwen3_5ForConditionalGeneration",
         ]:
             if is_sm100_supported():
@@ -2258,6 +2259,7 @@ class ServerArgs:
             if model_arch in [
                 "Qwen3NextForCausalLM",
                 "Qwen3_5MoeForConditionalGeneration",
+                "InternS2PreviewForConditionalGeneration",
                 "Qwen3_5ForConditionalGeneration",
             ]:
                 sm100_default_attn_backend = "triton"
@@ -2386,6 +2388,7 @@ class ServerArgs:
                 "Qwen3NextForCausalLM",
                 "KimiK25ForConditionalGeneration",
                 "Qwen3_5MoeForConditionalGeneration",
+                "InternS2PreviewForConditionalGeneration",
                 "Qwen3_5ForConditionalGeneration",
             ]
             and (is_sm90_supported() or is_sm100_supported())
@@ -3901,6 +3904,7 @@ class ServerArgs:
             "Qwen3VLMoeForConditionalGeneration",
             "Qwen3_5ForConditionalGeneration",
             "Qwen3_5MoeForConditionalGeneration",
+            "InternS2PreviewForConditionalGeneration",
             "Qwen3OmniMoeForConditionalGeneration",
             "Qwen2AudioForConditionalGeneration",
             "Qwen2_5OmniForConditionalGeneration",

python/sglang/srt/utils/hf_transformers/common.py

@@ -32,6 +32,7 @@ from sglang.srt.configs import (
     ExaoneConfig,
     FalconH1Config,
     GraniteMoeHybridConfig,
+    InternS2PreviewConfig,
     JetNemotronConfig,
     JetVLMConfig,
     KimiK25Config,
@@ -96,6 +97,7 @@ _CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = {
         DeepseekVLV2Config,
         Qwen3_5Config,
         Qwen3_5MoeConfig,
+        InternS2PreviewConfig,
         JetNemotronConfig,
         JetVLMConfig,
         KimiK25Config,