diff --git a/samplers.py b/samplers.py
index 127174f3..3d9f4a3d 100644
--- a/samplers.py
+++ b/samplers.py
@@ -1348,7 +1348,8 @@ class FunPackLTXAVSceneChainSampler:
                         try:
                             grad = _vf.gradient(cond)
                             d = torch.nn.functional.normalize(grad.float(), dim=-1).to(cond.dtype)
-                        except Exception:
+                        except Exception as _e:
+                            print(f"[FunPackSceneChain] embed_guidance: value function gradient failed ({_e}), using fixed direction")
                             d = _fixed.to(cond.device, cond.dtype).expand_as(cond)
                     else:
                         d = _fixed.to(cond.device, cond.dtype).expand_as(cond)
diff --git a/value_function.py b/value_function.py
index 606cc60a..f53c92cc 100644
--- a/value_function.py
+++ b/value_function.py
@@ -73,11 +73,12 @@ class OnlineValueFunction(nn.Module):
 
     def gradient(self, conditioning):
         """∂reward/∂conditioning — same shape as input."""
+        mlp_device = next(self.parameters()).device
         with torch.inference_mode(False), torch.enable_grad():
-            c_in = conditioning.detach().clone().float().requires_grad_(True)
+            c_in = conditioning.detach().clone().float().to(mlp_device).requires_grad_(True)
             reward = self.forward(self.compress(c_in).unsqueeze(0))
             reward.backward()
-        return c_in.grad.to(conditioning.dtype)
+        return c_in.grad.to(conditioning.device, conditioning.dtype)
 
     def is_ready(self):
         return len(self.buffer_c) >= self.MIN_SAMPLES