diff --git a/src/openpi/models_pytorch/gemma_pytorch.py b/src/openpi/models_pytorch/gemma_pytorch.py
index 203b36be8a..d9a1c1dda3 100644
--- a/src/openpi/models_pytorch/gemma_pytorch.py
+++ b/src/openpi/models_pytorch/gemma_pytorch.py
@@ -134,13 +134,6 @@ def forward(
                 and self.training
             ) or (hasattr(self, "gradient_checkpointing") and self.gradient_checkpointing and self.training)
 
-            # Force enable gradient checkpointing if we're in training mode and the model supports it
-            if self.training and hasattr(self.gemma_expert.model, "gradient_checkpointing"):
-                if not self.gemma_expert.model.gradient_checkpointing:
-                    print("Forcing gradient checkpointing to be enabled for Gemma expert model")
-                    self.gemma_expert.model.gradient_checkpointing = True
-                use_gradient_checkpointing = True
-
             # Debug gradient checkpointing status
             if hasattr(self, "_debug_gc_printed") and not self._debug_gc_printed:
                 print(f"Gemma expert model gradient checkpointing: {use_gradient_checkpointing}")