diff --git a/lightx2v/models/networks/wan/infer/lingbot_fast/transformer_infer.py b/lightx2v/models/networks/wan/infer/lingbot_fast/transformer_infer.py
index f5111186..2d33679d 100755
--- a/lightx2v/models/networks/wan/infer/lingbot_fast/transformer_infer.py
+++ b/lightx2v/models/networks/wan/infer/lingbot_fast/transformer_infer.py
@@ -52,7 +52,11 @@ def _initialize_kv_cache(self, dtype, device):
             return
         kv_cache1 = []
         ws = self._sp_world_size()
-        self.kv_cache_size = self._kv_size // ws
+        if self.local_attn_size != -1:
+            kv_cache_size = self.local_attn_size * self.frame_seq_length // ws
+        else:
+            kv_cache_size = self._kv_size // ws
+        self.kv_cache_size = kv_cache_size
 
         n, d = self.num_heads, self.head_dim
         if self.kv_quant_config is not None: