To synchronize model weights across globally distributed clusters for Composer 2 training, Cursor and Fireworks developed a compression algorithm that ships only the deltas, which can be up to 20 times smaller than the full 1 terabyte model.

Dima from Fireworks states that Mixture-of-Experts (MoE) models are highly sensitive to numerical mismatches between inference and training, which can cause the model to try and update the weights of an expert that was not actually used during the forward pass.

To improve system efficiency during reinforcement learning, Cursor and Fireworks use a pipelined or asynchronous approach where the trainer and rollout environments operate continuously and in parallel.