Frontier-lab training infrastructure, now as a service

This post details the offering of Fireworks' internal frontier-lab training infrastructure as a managed service, starting with GLM 5.2. It highlights the critical importance of batch invariance and zero KLD (Kullback-Leibler divergence) between training and serving for successful reinforcement learning on large models. The post explains the technical challenges, such as floating-point arithmetic non-associativity and its impact on attention reductions, MoE routing, and cross-GPU reductions, and how Fireworks' infrastructure addresses these to ensure numerical identity between training and serving. It provides validation data showing improved reward stability and reduced token clipping when using the Fireworks stack compared to approaches relying on importance sampling and clipping.