In the file interactive_causal_inference.py line 37-41, the code explicitly sets cache["k"].zero_(). This operation zeroes out the KV cache, including the sink tokens.

However, during cache updates (see line 260 in causal_model.py), the is_recompute logic is designed to protect the sink region from being overwritten. Because the sink portion of the cache has already been zeroed out, it will never be updated again, resulting in the sink KV cache remaining permanently zero throughout subsequent video generation.

This behavior effectively causes a persistent corruption of the KV cache when global_sink: false.