diff --git a/unsloth/models/mistral.py b/unsloth/models/mistral.py
index bc00e7a982..4f69484dde 100644
--- a/unsloth/models/mistral.py
+++ b/unsloth/models/mistral.py
@@ -132,7 +132,7 @@ def MistralAttention_fast_forward(
         Q = Q.transpose(1, 2)
         K = K.transpose(1, 2)
         V = V.transpose(1, 2)
-        sw = getattr(self.config, "sliding_window", None)
+        sw = isinstance(causal_mask, xformers.attn_bias.BlockDiagonalCausalMask)
         sw = kv_seq_len if (sw is None or sw == "null") else sw
         window = (-1, -1) if (kv_seq_len <= sw) else (sw, sw)
         A = flash_attn_func(Q, K, V, causal = True, window_size = window)
diff --git a/unsloth/save.py b/unsloth/save.py
index 70242eaa7b..c516ed2365 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -87,10 +87,10 @@ def _merge_lora(layer, name, max_vram):
         # First check if memory has been exceeded
         # If yes, move over to CPU to do computation.
         nbytes = W.numel() * 4 # float32 4 bytes
-        if (torch.cuda.memory_allocated() + nbytes) >= max_vram:
-            W = W.to("cpu", non_blocking = True)
-            out_of_memory = True
-        pass
+        # if (torch.cuda.memory_allocated() + nbytes) >= max_vram:
+        #     W = W.to("cpu", non_blocking = True)
+        #     out_of_memory = True
+        # pass
         W = W.to(torch.float32).t()
 
         if A is not None: