diff --git a/backends/exllamav2/model.py b/backends/exllamav2/model.py
index ae71b00..6e59dbe 100644
--- a/backends/exllamav2/model.py
+++ b/backends/exllamav2/model.py
@@ -272,6 +272,7 @@ class ExllamaV2Container(BaseModelContainer):
             self.config.max_seq_len = unwrap(
                 user_max_seq_len, min(hf_model.hf_config.max_position_embeddings, 4096)
             )
+            self.cache_size = self.config.max_seq_len
 
         # Set the rope scale
         self.config.scale_pos_emb = unwrap(