diff --git a/src/diffusers/pipelines/controlnet/pipeline_controlnet.py b/src/diffusers/pipelines/controlnet/pipeline_controlnet.py
index 486f9fb764d1..582f51ab480e 100644
--- a/src/diffusers/pipelines/controlnet/pipeline_controlnet.py
+++ b/src/diffusers/pipelines/controlnet/pipeline_controlnet.py
@@ -31,6 +31,7 @@
 from ...utils import (
     USE_PEFT_BACKEND,
     deprecate,
+    is_torch_xla_available,
     logging,
     replace_example_docstring,
     scale_lora_layers,
@@ -42,6 +43,13 @@
 from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
 
 
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -1323,6 +1331,8 @@ def __call__(
                         step_idx = i // getattr(self.scheduler, "order", 1)
                         callback(step_idx, t, latents)
 
+                if XLA_AVAILABLE:
+                    xm.mark_step()
         # If we do sequential model offloading, let's offload unet and controlnet
         # manually for max memory savings
         if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: