diff --git a/nodes_sampler.py b/nodes_sampler.py
index 32b51c6a..dac368fb 100644
--- a/nodes_sampler.py
+++ b/nodes_sampler.py
@@ -1182,7 +1182,7 @@ def predict_with_cfg(z, cfg_scale, positive_embeds, negative_embeds, timestep, i
             with torch.autocast(device_type=mm.get_autocast_device(device), dtype=dtype) if autocast_enabled else nullcontext():
 
                 if use_cfg_zero_star and (idx <= zero_star_steps) and use_zero_init:
-                    return z*0, None
+                    return z*0, None, cache_state
 
                 nonlocal patcher
                 current_step_percentage = idx / len(timesteps)
diff --git a/skyreels/nodes.py b/skyreels/nodes.py
index a188edea..c41622f8 100644
--- a/skyreels/nodes.py
+++ b/skyreels/nodes.py
@@ -457,7 +457,7 @@ def predict_with_cfg(z, cfg_scale, positive_embeds, negative_embeds, timestep, i
             with torch.autocast(device_type=mm.get_autocast_device(device), dtype=dtype, enabled=("fp8" in model["quantization"])):
 
                 if use_cfg_zero_star and (idx <= zero_star_steps) and use_zero_init:
-                    return latent_model_input*0, None
+                    return latent_model_input*0, teacache_state
 
                 nonlocal patcher
                 current_step_percentage = idx / len(init_timesteps)