First of all, thank you for the excellent work on this project and for open-sourcing it to the community. It has been very helpful for my research/experiments.
I noticed that the backward implementation of causal_conv1d relies on the atomicAdd instruction. As a result, the computation results are not guaranteed to be deterministically reproducible across runs.
May I ask if there are any plans to provide a deterministic backward implementation in the future? Such functionality would be very useful for scenarios where reproducibility is critical.
First of all, thank you for the excellent work on this project and for open-sourcing it to the community. It has been very helpful for my research/experiments.
I noticed that the backward implementation of causal_conv1d relies on the atomicAdd instruction. As a result, the computation results are not guaranteed to be deterministically reproducible across runs.
May I ask if there are any plans to provide a deterministic backward implementation in the future? Such functionality would be very useful for scenarios where reproducibility is critical.