[Fix] Fix Qwen3.5 MTP startup with HiCache (#34560)

Co-authored-by: hjzhang <76768149+1e4ves@users.noreply.github.com>
Co-authored-by: YAMY <74099316+YAMY1234@users.noreply.github.com>

python/sglang/srt/configs/model_config.py

@@ -707,6 +707,7 @@ class ModelConfig:
                 )
             self.hf_config.architectures[0] = "Qwen3_5ForCausalLMMTP"
             self.hf_config.num_nextn_predict_layers = 1
+            self.hf_text_config.num_nextn_predict_layers = 1
 
         if is_draft_model and self.hf_config.architectures[0] == "ExaoneMoEForCausalLM":
             self.hf_config.architectures[0] = "ExaoneMoEForCausalLMMTP"

python/sglang/srt/mem_cache/hybrid_cache/hybrid_pool_assembler.py

@@ -930,9 +930,15 @@ def build_full_draft_pools(
     server_args: ServerArgs,
 ) -> tuple[list[SidecarPoolSpec], list[PoolEntry]]:
     """Build draft KV/DSA sidecars whose indices follow target full KV."""
-    from sglang.srt.mem_cache.memory_pool import DSATokenToKVPool
+    from sglang.srt.mem_cache.memory_pool import (
+        DSATokenToKVPool,
+        HybridLinearKVPool,
+    )
 
     pool = draft_kv_pool
+    if isinstance(pool, HybridLinearKVPool):
+        # Hybrid draft runners keep their sole attention layer in this sub-pool.
+        pool = pool.full_kv_pool
     if pool.layer_num == 0:
         return [], []