diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 828a820909..a766ed9366 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -183,8 +183,8 @@ def LlamaAttention_fast_forward_inference(
     pass
 
     # Grouped query attention
+    _, _, cached_len, _ = Knn.shape
     if n_groups != 1:
-        _, _, cached_len, _ = Knn.shape
         Knn = Knn[:, :, None, :, :].expand(bsz, n_kv_heads, n_groups, cached_len, head_dim)
         Vnn = Vnn[:, :, None, :, :].expand(bsz, n_kv_heads, n_groups, cached_len, head_dim)
         Knn = Knn.reshape(bsz, n_heads, cached_len, head_dim)
@@ -195,7 +195,7 @@ def LlamaAttention_fast_forward_inference(
     # pass
 
     # Attention
-    A = torch.matmul(Qn, Knn.transpose(2, 3), out = self.attention[:,:,:,:kv_seq_len])
+    A = torch.matmul(Qn, Knn.transpose(2, 3), out = self.attention[:,:,:,:cached_len])
     A *= self.scalar
     A[:] = torch.nn.functional.softmax(A, dim = -1, dtype = torch.float32)#.to(A.dtype)
     A = torch.matmul(A, Vnn, out = Qn)