diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 8d985aa9d2..27b92282c3 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -342,9 +342,9 @@ def LlamaAttention_fast_forward_inference(
         A = torch_matmul(A, Vnn, out = Qn)
     else:
         if SDPA_HAS_GQA:
-            A = scaled_dot_product_attention(Qn, Knn, Vnn, attn_mask = attention_mask, is_causal = False, enable_gqa = True)
+            A = scaled_dot_product_attention(Qn, Knn, Vnn, attn_mask = attention_mask, is_causal = is_causal, enable_gqa = True)
         else:
-            A = scaled_dot_product_attention(Qn, Knn, Vnn, attn_mask = attention_mask, is_causal = False)
+            A = scaled_dot_product_attention(Qn, Knn, Vnn, attn_mask = attention_mask, is_causal = is_causal)
     pass
     A = A.transpose(1, 2)
     A = A.reshape(bsz, 1, attention_size)