diff --git a/unsloth/__init__.py b/unsloth/__init__.py
index 9dce29cd83..7080c92894 100644
--- a/unsloth/__init__.py
+++ b/unsloth/__init__.py
@@ -99,10 +99,14 @@ except:
         cdequantize_blockwise_fp32 = bnb.functional.lib.cdequantize_blockwise_fp32
         libcuda_dirs()
     except:
-        raise ImportError("Unsloth: CUDA is not linked properly.\n"\
-                          "We tried running `ldconfig /usr/lib64-nvidia` ourselves, but it didn't work.\n"\
-                          "You need to run in your terminal `sudo ldconfig /usr/lib64-nvidia` yourself, then import Unsloth.\n"\
-                          "Also try `sudo ldconfig /usr/local/cuda-xx.x` - find the latest cuda version.")
+        warnings.warn(
+            "Unsloth: CUDA is not linked properly.\n"\
+            "Try running `python -m bitsandbytes` then `python -m xformers.info`\n"\
+            "We tried running `ldconfig /usr/lib64-nvidia` ourselves, but it didn't work.\n"\
+            "You need to run in your terminal `sudo ldconfig /usr/lib64-nvidia` yourself, then import Unsloth.\n"\
+            "Also try `sudo ldconfig /usr/local/cuda-xx.x` - find the latest cuda version.\n"\
+            "Unsloth will still run for now, but maybe it might crash - let's hope it works!"
+        )
 pass
 
 from .models import *
diff --git a/unsloth/kernels/rms_layernorm.py b/unsloth/kernels/rms_layernorm.py
index 8af602a857..4db89b7816 100644
--- a/unsloth/kernels/rms_layernorm.py
+++ b/unsloth/kernels/rms_layernorm.py
@@ -119,7 +119,7 @@ def _gemma_rms_layernorm_forward(
     W_row = tl.load(W + col_offsets, mask = mask, other = 0).to(tl.float32)
 
     row_var = tl.sum(X_row * X_row, axis = 0) / n_cols
-    inv_var = 1.0 / tl.sqrt(row_var + eps)
+    inv_var = 1.0 / tl.sqrt(row_var + eps) # Must be 1/sqrt to match Deepmind's impl
     tl.store(r, inv_var)
     normed = X_row * inv_var
     output = normed * (W_row + 1.0)