[VLM]: allow Qwen3.5 models for encoder disaggregation (#21849)

python/sglang/srt/disaggregation/encode_server.py

@@ -867,10 +867,11 @@ class MMEncoder:
             )
             # Get additional video metadata
             if (
-                self.model_type in ["qwen3_vl", "qwen3_vl_moe"]
+                self.model_type
+                in ["qwen3_vl", "qwen3_vl_moe", "qwen3_5", "qwen3_5_moe"]
                 and video_processor_kwargs.get("video_metadata", None) is not None
             ):
-                # For qwen3-vl models, we need to store the video timestamps
+                # For qwen3-vl/qwen3.5 models, we need to store the video timestamps
                 video_metadata = video_processor_kwargs["video_metadata"]
                 try:
                     merge_size = (

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

@@ -422,7 +422,7 @@ class QwenVLImageProcessor(SGLangBaseProcessor):
                 audio_seq_lens = (audio_seq_lens - 2) // 2 + 1
 
         if (
-            self.model_type in ["qwen3_vl", "qwen3_vl_moe"]
+            self.model_type in ["qwen3_vl", "qwen3_vl_moe", "qwen3_5", "qwen3_5_moe"]
             and video_timestamps is not None
         ):
             input_ids, offsets, modality_list = self.build_input_ids_with_timestamps(

python/sglang/srt/server_args.py

@@ -3326,6 +3326,8 @@ class ServerArgs:
             "Qwen3VLForConditionalGeneration",
             "Qwen2_5_VLForConditionalGeneration",
             "Qwen3VLMoeForConditionalGeneration",
+            "Qwen3_5ForConditionalGeneration",
+            "Qwen3_5MoeForConditionalGeneration",
             "Qwen3OmniMoeForConditionalGeneration",
             "Qwen2AudioForConditionalGeneration",
             "Qwen2_5OmniForConditionalGeneration",