diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index d61dbf842b..a51860a14c 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -141,7 +141,7 @@ def LlamaAttention_fast_forward_inference(
         self.paged_attention_V = self.paged_attention[:,1]
         self.attention.resize_((bsz, n_heads, 1, self.attention.shape[-1]+KV_CACHE_INCREMENT))
     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])
@@ -197,7 +197,6 @@ def fast_mlp_inference(self, X):
     # gate = self.gate_proj(X)
     # up   = self.up_proj(X)
     bsz, _, hd = X.shape
-    mlp_size = self.config.intermediate_size
     temp = torch.empty((2, bsz, 1, mlp_size), dtype = X.dtype, device = "cuda")
 
     gate = fast_linear_forward(self.gate_proj, X, out = temp[0])