How did we do it? By loading the model weights into memory once before training begins and inserting them as numpy arrays into the #Ray object store, we can then zero-copy read the weights directly from shared memory into each GPU worker process.

→ View original post on X — @predibase,