--- src/sglang-pr/python/sglang/srt/configs/model_config.py	2026-10-30 14:12:31.930788808 -0700
+++ overlay/python/sglang/srt/configs/model_config.py	2026-10-30 14:31:57.333705000 -0700
@@ -1166,6 +1166,26 @@
             self.hf_config.architectures[0] = "Qwen3MoeForCausalLMMTP"
             self.hf_config.num_nextn_predict_layers = 1
 
+        if (
+            is_draft_model
+            and self.hf_config.architectures[0]
+            == "SenseNovaLoopForConditionalGeneration"
+        ):
+            # MTP patch: the bundled mtp.* head is a plain Qwen3.5 MTP layer (one
+            # full-attention layer, BF16). Shrink the looped execution layout to
+            # that single layer; deep-copy so the target keeps its 56-slot plan.
+            self.hf_config = copy.deepcopy(self.hf_config)
+            self.hf_text_config = get_hf_text_config(self.hf_config)
+            text_config = self.hf_text_config
+            text_config.num_hidden_layers = 1
+            text_config.layer_types = ["full_attention"]
+            text_config.full_attention_interval = 1
+            text_config.execution_plan = (0,)
+            text_config.mhc = None
+            text_config.num_nextn_predict_layers = 1
+            self.hf_config.num_nextn_predict_layers = 1
+            self.hf_config.architectures[0] = "Qwen3_5ForCausalLMMTP"
+
         if is_draft_model and self.hf_config.architectures[0] in [
             "Qwen3_5ForConditionalGeneration",
             "Qwen3_5MoeForConditionalGeneration",