diff --git a/arctic_training/model/hf_factory.py b/arctic_training/model/hf_factory.py
index a56ee94a..b785e1ea 100644
--- a/arctic_training/model/hf_factory.py
+++ b/arctic_training/model/hf_factory.py
@@ -83,7 +83,12 @@ def post_create_model_callback(self, model):
             )
 
         if not self.config.disable_activation_checkpoint:
-            model.gradient_checkpointing_enable()
+            gc_kwargs = None
+            if self.trainer.config.activation_checkpoint_cpu_offload:
+                # CPU-offload monkey patch (see arctic_training.monkey_patches)
+                # only intercepts torch.utils.checkpoint's reentrant path.
+                gc_kwargs = {"use_reentrant": True}
+            model.gradient_checkpointing_enable(gradient_checkpointing_kwargs=gc_kwargs)
             model = self.make_model_gradient_checkpointing_compatible(model)
 
         return model