diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index f5d00eab22..8d0eadb968 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2025.2.8"
+__version__ = "2025.2.9"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index ec6706e515..1eae97ff1c 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -708,7 +708,7 @@ def LlamaModel_fast_forward(
     if attention_mask is None:
         padding_mask = None
     elif self.training:
-    # elif attention_mask is not None and self.training:
+    # elif attention_mask is None:
         attention_mask = None
         padding_mask = None
     else:
@@ -724,7 +724,8 @@ def LlamaModel_fast_forward(
             past_key_values_length,
             sliding_window = getattr(self.config, "sliding_window", None),
         )
-        attention_mask = attention_mask.to(torch.bool)
+        if attention_mask is not None:
+            attention_mask = attention_mask.to(torch.bool)
     pass
 
     hidden_states = inputs_embeds
diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index fc094b0839..3d601b0af1 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -565,8 +565,8 @@ pass
 
 
 def PatchFastRL(algorithm = None, FastLanguageModel = None):
-    return
-    # if FastLanguageModel is not None: PatchRL(FastLanguageModel)
-    # patch_trl_rl_trainers()
-    # if algorithm is not None: PatchRLStatistics(algorithm)
+    if FastLanguageModel is not None: PatchRL(FastLanguageModel)
+    patch_trl_rl_trainers()
+    if type(algorithm) is str and algorithm.islower():
+        PatchRLStatistics(algorithm)
 pass
diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index 4d7a4dbe09..82fd3f8d3c 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -101,23 +101,20 @@ RL_FUNCTIONS["sft_trainer"].append(sft_trainer_prepare_dataset)
 
 # Ignore mean_token_accuracy since it needs logits
 # We override it directly with our version
-def _sft_trainer_compute_loss(self, model, inputs, return_outputs = False, num_items_in_batch = None):
-    (loss, outputs) = super().compute_loss(
-        model,
-        inputs,
-        return_outputs = return_outputs,
-        num_items_in_batch = num_items_in_batch,
-    )
-    return (loss, outputs) if return_outputs else loss
-pass
-
 def sft_trainer_compute_loss(function_name, function):
     if  function_name != "compute_loss": return function
 
-    function = inspect.getsource(_sft_trainer_compute_loss)
-    function = function.replace("def _sft_trainer_compute_loss", "def compute_loss")
-    function = function.split("\n")
-    function = "\n".join(" "*4+x for x in function)
+    def compute_loss(self, model, inputs, return_outputs = False, num_items_in_batch = None):
+        outputs = super().compute_loss(
+            model,
+            inputs,
+            return_outputs = return_outputs,
+            num_items_in_batch = num_items_in_batch,
+        )
+        return outputs
+    pass
+
+    function = inspect.getsource(compute_loss)
     return function
 pass
 RL_FUNCTIONS["sft_trainer"].append(sft_trainer_compute_loss)