Support LingV2_5 model (#18598)

Co-authored-by: zhangkaihong.zkh <zhangkaihong.zkh@antgroup.com>
Co-authored-by: 有禾 <zhangdonghao.zdh@antgroup.com>
Co-authored-by: yudian0504 <138860534+yudian0504@users.noreply.github.com>
Co-authored-by: 悠扬 <youyang.zmy@antgroup.com>
Co-authored-by: xinxingyang <xinxing.yangxx@antgroup.com>
Co-authored-by: zmy460290 <zmy460290@antgroup.com>

python/sglang/srt/model_executor/model_runner.py

@@ -31,6 +31,7 @@ import torch.distributed as dist
 from torch import nn
 
 from sglang.srt.configs import (
+    BailingHybridConfig,
     FalconH1Config,
     JetNemotronConfig,
     JetVLMConfig,
@@ -1549,6 +1550,13 @@ class ModelRunner(ModelRunnerKVCacheMixin):
             return config
         return None
 
+    @property
+    def hybrid_lightning_config(self):
+        config = self.model_config.hf_config
+        if isinstance(config, BailingHybridConfig):
+            return config
+        return None
+
     @property
     def hybrid_gdn_config(self):
         config = self.model_config.hf_config.get_text_config()
@@ -1597,7 +1605,12 @@ class ModelRunner(ModelRunnerKVCacheMixin):
 
     @property
     def mambaish_config(self):
-        return self.mamba2_config or self.hybrid_gdn_config or self.kimi_linear_config
+        return (
+            self.mamba2_config
+            or self.hybrid_gdn_config
+            or self.kimi_linear_config
+            or self.hybrid_lightning_config
+        )
 
     def can_run_piecewise_cuda_graph(self):
         if self.is_draft_worker:

python/sglang/srt/model_executor/model_runner_kv_cache_mixin.py

@@ -584,7 +584,13 @@ class ModelRunnerKVCacheMixin:
                     head_dim=self.model_config.head_dim,
                     # if draft worker, we only need 1 attention layer's kv pool
                     full_attention_layer_ids=(
-                        [0] if self.is_draft_worker else config.full_attention_layer_ids
+                        [0]
+                        if self.is_draft_worker
+                        else [
+                            i
+                            for i in config.full_attention_layer_ids
+                            if self.start_layer <= i < self.end_layer
+                        ]
                     ),
                     enable_kvcache_transpose=False,
                     device=self.device,