← Home

Blog

Compiling updates in Pytorch

Updating each weight inside the backward walk frees each gradient immediately, cutting peak training memory.

Constructing Loss functions with the MDMM method.

Replacing linear loss combinations with constrained optimization makes optimisation more interpretable.

ALE Async send/recv

Adding a torch C++ interface to the ALE.

Multi-ROM support for the ALE

Adding support for multiple ROMs in the ALE.