Fix FiLM temporal context gradient leak: detach EMA to prevent cross-step backprop

vil_tracker/models/film_temporal.py

@@ -143,10 +143,10 @@ class TemporalModulationManager(nn.Module):
         """
         context = self.context_proj(features.detach())
         if self._temporal_context is None:
-            self._temporal_context = context
+            self._temporal_context = context.detach()
         else:
-            # EMA update
-            self._temporal_context = 0.7 * self._temporal_context + 0.3 * context
+            # EMA update — detach to prevent cross-step gradient leakage
+            self._temporal_context = (0.7 * self._temporal_context + 0.3 * context).detach()
     
     def modulate(
         self,