model: support Qwen3.5 (#18489)

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

python/sglang/srt/model_executor/forward_batch_info.py

@@ -350,6 +350,7 @@ class ForwardBatch(ForwardBatchDeepSeekMHAMixin):
     # Speculative decoding
     spec_info: Optional[SpecInput] = None
     spec_algorithm: SpeculativeAlgorithm = None
+    mm_input_embeds: Optional[torch.Tensor] = None
     capture_hidden_mode: CaptureHiddenMode = None
 
     # For padding

python/sglang/srt/model_executor/model_runner.py

@@ -38,6 +38,8 @@ from sglang.srt.configs import (
     Lfm2Config,
     NemotronH_Nano_VL_V2_Config,
     NemotronHConfig,
+    Qwen3_5Config,
+    Qwen3_5MoeConfig,
     Qwen3NextConfig,
 )
 from sglang.srt.configs.device_config import DeviceConfig
@@ -1548,8 +1550,15 @@ class ModelRunner(ModelRunnerKVCacheMixin):
 
     @property
     def hybrid_gdn_config(self):
-        config = self.model_config.hf_config
-        if isinstance(config, Qwen3NextConfig | JetNemotronConfig | JetVLMConfig):
+        config = self.model_config.hf_config.get_text_config()
+        if isinstance(
+            config,
+            Qwen3NextConfig
+            | Qwen3_5Config
+            | Qwen3_5MoeConfig
+            | JetNemotronConfig
+            | JetVLMConfig,
+        ):
             return config
         return None
 
@@ -2532,7 +2541,9 @@ class ModelRunner(ModelRunnerKVCacheMixin):
     def model_is_mrope(self) -> bool:
         """Detect if the model has "mrope" rope_scaling type.
         mrope requires keep "rope_deltas" between prompt and decoding phases."""
-        rope_scaling = getattr(self.model_config.hf_text_config, "rope_scaling", {})
+        rope_scaling = getattr(
+            self.model_config.hf_text_config, "rope_parameters", None
+        ) or getattr(self.model_config.hf_text_config, "rope_scaling", {})
         if rope_scaling is None:
             return False
         is_mrope_enabled = "mrope_section" in rope_scaling