Our earlier companion blog, Primus Projection: Estimate Memory and Performance Before You Train, showed how to answer two questions about a single training configuration before launching a run: “Will it fit?” and “How fast will it be?” But planning a large-scale training job raises a harder question that projection alone doesn’t answer: out of the thousands of legal configurations, which one should I actually run?
This blog covers the Primus Tuning Agent — a tool that treats the projection engine as a fast scoring oracle and uses it to automatically search the joint configuration space (parallelism, micro-batch size, recomputation, pipeline schedule, MoE communication backend, and precision) for a high-throughput, memory-legal recipe. In a Mixtral 8×22B case study, the agent discovered a configuration that delivers +27% measured throughput over AMD’s published 4-node BF16 reference — in under 30 minutes of single-node exploration, with no hand-written Megatron configs and no full-cluster profiling pass.
