diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index a2ff511992..e2edd9c4da 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -147,9 +147,9 @@ def LlamaAttention_fast_forward_inference(
         RH_Q = torch.empty((bsz, n_heads, 1, head_dim), dtype = dtype, device = "cuda")
     pass
 
-    Qn = fast_linear_forward(self.q_proj, Xn, out = self.temp_QA[0])
-    Kn = fast_linear_forward(self.k_proj, Xn, out = self.temp_KV[0])
-    Vn = fast_linear_forward(self.v_proj, Xn, out = self.temp_KV[1])
+    Qn = fast_linear_forward(self.q_proj, Xn, out = temp_QA[0])
+    Kn = fast_linear_forward(self.k_proj, Xn, out = temp_KV[0])
+    Vn = fast_linear_forward(self.v_proj, Xn, out = temp_KV[1])
     Qn = Qn.view(bsz, 1, n_heads,    head_dim).transpose(1, 2)
     Kn = Kn.view(bsz, 1, n_kv_heads, head_dim).transpose(1, 2)
     Vn = Vn.view(bsz, 1, n_kv_heads, head_dim).transpose(1, 2)
@@ -193,7 +193,7 @@ def LlamaAttention_fast_forward_inference(
     A = torch.matmul(A, Vnn, out = Qn)
     A = A.transpose(1, 2)
     A = A.reshape(bsz, 1, self.hidden_size)
-    A = fast_linear_forward(self.o_proj, A, out = self.temp_QA[1])
+    A = fast_linear_forward(self.o_proj, A, out = temp_QA[1])
     return A, (Kn, Vn)
 pass