model: support Qwen3.5 (#18489)

Co-authored-by: 瑀澈 <yuche.lz@alibaba-inc.com>

python/sglang/srt/configs/__init__.py

@@ -18,6 +18,7 @@ from sglang.srt.configs.longcat_flash import LongcatFlashConfig
 from sglang.srt.configs.nano_nemotron_vl import NemotronH_Nano_VL_V2_Config
 from sglang.srt.configs.nemotron_h import NemotronHConfig
 from sglang.srt.configs.olmo3 import Olmo3Config
+from sglang.srt.configs.qwen3_5 import Qwen3_5Config, Qwen3_5MoeConfig
 from sglang.srt.configs.qwen3_next import Qwen3NextConfig
 from sglang.srt.configs.step3_vl import (
     Step3TextConfig,
@@ -43,6 +44,8 @@ __all__ = [
     "KimiLinearConfig",
     "KimiK25Config",
     "Qwen3NextConfig",
+    "Qwen3_5Config",
+    "Qwen3_5MoeConfig",
     "DotsVLMConfig",
     "DotsOCRConfig",
     "FalconH1Config",

python/sglang/srt/configs/model_config.py

@@ -319,6 +319,13 @@ class ModelConfig:
             self.hf_config.architectures[0] = "Qwen3NextForCausalLMMTP"
             self.hf_config.num_nextn_predict_layers = 1
 
+        if is_draft_model and self.hf_config.architectures[0] in [
+            "Qwen3_5ForConditionalGeneration",
+            "Qwen3_5MoeForConditionalGeneration",
+        ]:
+            self.hf_config.architectures[0] = "Qwen3_5ForCausalLMMTP"
+            self.hf_config.num_nextn_predict_layers = 1
+
         if is_draft_model and self.hf_config.architectures[0] == "ExaoneMoEForCausalLM":
             self.hf_config.architectures[0] = "ExaoneMoEForCausalLMMTP"
             self.hf_config.num_nextn_predict_layers = 1
@@ -1193,6 +1200,8 @@ multimodal_model_archs = [
     "Qwen2_5_VLForConditionalGeneration",
     "Qwen3VLForConditionalGeneration",
     "Qwen3VLMoeForConditionalGeneration",
+    "Qwen3_5ForConditionalGeneration",
+    "Qwen3_5MoeForConditionalGeneration",
     "Qwen3OmniMoeForConditionalGeneration",
     "KimiVLForConditionalGeneration",
     "InternVLChatModel",

python/sglang/srt/configs/qwen3_5.py

@@ -0,0 +1,113 @@
+from transformers import PretrainedConfig
+
+from sglang.srt.configs.qwen3_next import Qwen3NextConfig
+from sglang.srt.configs.qwen3_vl import Qwen3VLVisionConfig
+
+
+class Qwen3_5VisionConfig(Qwen3VLVisionConfig):
+    model_type = "qwen3_5"
+    base_config_key = "vision_config"
+
+
+class Qwen3_5TextConfig(Qwen3NextConfig):
+    model_type = "qwen3_5_text"
+    base_config_key = "text_config"
+
+    def __init__(
+        self,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        if self.rope_scaling is None:
+            self.rope_scaling = {}
+
+
+class Qwen3_5Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen3_5Model`]. It is used to instantiate a
+    Qwen3.5 model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen3.5.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        text_config (`Union[PreTrainedConfig, dict]`, *optional*, defaults to `Qwen3_5TextConfig`):
+            The config object or dictionary of the text backbone.
+        vision_config (`Union[PreTrainedConfig, dict]`,  *optional*, defaults to `Qwen3_5VisionConfig`):
+            The config object or dictionary of the vision backbone.
+        image_token_id (`int`, *optional*, defaults to 151655):
+            The image token index to encode the image prompt.
+        video_token_id (`int`, *optional*, defaults to 151656):
+            The video token index to encode the image prompt.
+        vision_start_token_id (`int`, *optional*, defaults to 151652):
+            The start token index to encode the image prompt.
+        vision_end_token_id (`int`, *optional*, defaults to 151653):
+            The end token index to encode the image prompt.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie the word embeddings.
+
+    ```python
+    >>> from transformers import Qwen3_5ForConditionalGeneration, Qwen3_5Config
+
+    >>> # Initializing a Qwen3.5 style configuration
+    >>> configuration = Qwen3_5Config()
+
+    >>> # Initializing a model from the Qwen3.5 style configuration
+    >>> model = Qwen3_5ForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "qwen3_5"
+    sub_configs = {
+        "vision_config": Qwen3_5VisionConfig,
+        "text_config": Qwen3_5TextConfig,
+    }
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        text_config=None,
+        vision_config=None,
+        image_token_id=151655,
+        video_token_id=151656,
+        vision_start_token_id=151652,
+        vision_end_token_id=151653,
+        tie_word_embeddings=False,
+        **kwargs,
+    ):
+        if isinstance(vision_config, dict):
+            self.vision_config = self.sub_configs["vision_config"](**vision_config)
+        elif vision_config is None:
+            self.vision_config = self.sub_configs["vision_config"]()
+
+        if isinstance(text_config, dict):
+            self.text_config = self.sub_configs["text_config"](**text_config)
+        elif text_config is None:
+            self.text_config = self.sub_configs["text_config"]()
+
+        self.image_token_id = image_token_id
+        self.video_token_id = video_token_id
+        self.vision_start_token_id = vision_start_token_id
+        self.vision_end_token_id = vision_end_token_id
+        super().__init__(**kwargs, tie_word_embeddings=tie_word_embeddings)
+
+
+class Qwen3_5MoeVisionConfig(Qwen3_5VisionConfig):
+    model_type = "qwen3_5_moe"
+
+
+class Qwen3_5MoeTextConfig(Qwen3_5TextConfig):
+    model_type = "qwen3_5_moe_text"
+
+
+class Qwen3_5MoeConfig(Qwen3_5Config):
+    model_type = "qwen3_5_moe"
+    sub_configs = {
+        "vision_config": Qwen3_5MoeVisionConfig,
+        "text_config": Qwen3_5MoeTextConfig,
+    }