RAVE's causal convolution mode lowers latency at the cost of quality by removing future context

By default, RAVE uses non-causal convolutions that can see both past and future context within a receptive field window. Adding --config causal switches all convolutions to causal mode (no future context), which lowers the overall latency of the model. This enables lower-latency embedded deployment (e.g., Raspberry Pi) and is required for true realtime performance where future samples are unavailable. The tradeoff is slightly lower reconstruction quality, since the encoder can no longer use future context to improve its estimates. Causal mode is composable: --config discrete --config causal.

Examples

rave train --config discrete --config causal ... to make the model causal and lower its overall latency.

Assessment

A performer needs the minimum possible processing latency in their RAVE nn~ patch. What two changes — one at training time and one at export time — are required?