diff --git a/cosyvoice/flow/decoder.py b/cosyvoice/flow/decoder.py
index 32b243c..4a89fb1 100644
--- a/cosyvoice/flow/decoder.py
+++ b/cosyvoice/flow/decoder.py
@@ -158,9 +158,12 @@ class CausalAttnProcessor2_0(AttnProcessor2_0):
 
         key_cache = attn.to_k(encoder_hidden_states)
         value_cache = attn.to_v(encoder_hidden_states)
-        # NOTE always concat cache for interface compatibility
-        key = torch.concat([cache[:, :, :, 0], key_cache], dim=1)
-        value = torch.concat([cache[:, :, :, 1], value_cache], dim=1)
+        # NOTE here we judge cache.size(0) instead of cache.size(1), because init_cache has size (2, 0, 512, 2)
+        if cache.size(0) != 0:
+            key = torch.concat([cache[:, :, :, 0], key_cache], dim=1)
+            value = torch.concat([cache[:, :, :, 1], value_cache], dim=1)
+        else:
+            key, value = key_cache, value_cache
         cache = torch.stack([key_cache, value_cache], dim=3)
 
         inner_dim = key.shape[-1]