diff --git a/unsloth/kernels/fp8.py b/unsloth/kernels/fp8.py
index 435a6e83d2..3093bf61b1 100644
--- a/unsloth/kernels/fp8.py
+++ b/unsloth/kernels/fp8.py
@@ -346,7 +346,7 @@ class FP8BlockQuantLinear(torch.autograd.Function):
         m, n = weight.shape
         p, q = weight_scale.shape
         block_size = getattr(weight, "block_size", None) or getattr(
-            weight_scale, "block_size", None
+            weight_scale, "block_size", [128, 128]
         )
         assert block_size is not None, "block_size is not set"
         if triton.cdiv(m, block_size[0]) != p or triton.cdiv(n, block_size[1]) != q:
diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py
index 672f6257dd..3b5204dfe2 100644
--- a/unsloth/models/rl.py
+++ b/unsloth/models/rl.py
@@ -312,7 +312,7 @@ class Unsloth{RLTrainer_name}(_Unsloth{RLTrainer_name}):
             if getattr(args, "_n_gpu", 1) != 1:
                 args._n_gpu = 1
         if "model" in locals() and hasattr(model, "for_training"):
-            model.for_training()
+            model.for_training(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))
         super().__init__({RLTrainer_call_args}{RLTrainer_kwargs})
         if "model" in locals() and hasattr(model, "for_inference"):
             model.for_inference()
@@ -565,7 +565,7 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"):
     if "model" in call_args:
         training_check = (
             "if model is not None and hasattr(model, 'for_training'):\n"
-            "    model.for_training()\n"
+            "    model.for_training(use_gradient_checkpointing=getattr(args, 'gradient_checkpointing', True))\n"
             "if 'tokenizer' in locals() and hasattr(tokenizer, 'padding_side'): tokenizer.padding_side = 'right'\n"
             "if 'processing_class' in locals():\n"
             "    if hasattr(processing_class, 'padding_side'): processing_class.padding_side = 'right'\n"