Primus Tuning Agent: Closing the Configuration-Search Loop

2026-07-06 · Source: AMD ROCm Blogs · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, long

Summary

The Primus Tuning Agent is a tool designed to automate the search for optimal distributed LLM training configurations. It addresses the challenge of selecting from thousands of possible configurations (parallelism, micro-batch size, recomputation, pipeline schedule, MoE communication backend, and precision) by treating the Primus projection engine as a fast scoring oracle. The agent operates in two phases: a deterministic seed planner that sweeps single axes, and an LLM investigation loop using DSPy to propose cross-axis combinations. This approach significantly reduces the time and cost associated with trial-and-error on real hardware. In a Mixtral 8x22B case study, the agent discovered a configuration delivering +27% measured throughput (4,402 tok/s/GPU) over AMD's published 4-node BF16 reference (3,475 tok/s/GPU) in under 30 minutes of single-node exploration, achieving a projected 4,908 tok/s/GPU. The deterministic phase alone reached 4,036 tok/s/GPU.

Key takeaway

For MLOps Engineers optimizing large-scale LLM training, the Primus Tuning Agent significantly reduces configuration search time and cost. You can automate finding high-throughput, memory-legal recipes for models like Mixtral 8x22B, achieving substantial performance gains (e.g., +27% throughput) in minutes rather than hours of multi-node profiling. Start with "--seed-only" to establish a baseline, then leverage the full agent to explore complex cross-axis optimizations, always validating the final configuration on your target cluster.

Key insights

The Primus Tuning Agent automates LLM configuration search using a projection oracle and a two-phase LLM-driven approach.

Principles

• Combine deterministic sweeps with LLM exploration.
• Utilize fast projection as a scoring oracle.
• Pre-filter OOM candidates at zero GPU cost.

Method

The agent uses a deterministic seed planner for single-axis sweeps, followed by a DSPy-based LLM investigation loop to propose cross-axis configurations. Both phases score candidates via a projection oracle and write to a shared ledger.

In practice

• Install Primus Tuning Agent via pip.
• Configure LLM provider credentials.
• Run agent with workload and cluster YAMLs.

Topics

• Primus Tuning Agent
• LLM Training Optimization
• Distributed Training
• Configuration Search
• GPU Performance
• DSPy

Code references

• ROCm/rocm-libraries

Best for: NLP Engineer, Machine Learning Engineer, AI Engineer, MLOps Engineer

Related on AIssential

• AutoPass: Evidence-Guided LLM Agents for Compiler Performance Tuning — LLMs effectively guide compiler optimization by integrating compiler-internal signals and runtime feedback without offline training.
• Primus Projection: Estimate Memory and Performance Before You Train — Primus projection estimates LLM training memory and performance, reducing costly trial-and-error on AMD Instinct™ GPUs.
• PriFT: Prior-Support Guided Supervised Fine-Tuning — PriFT improves SFT generalization by reweighting tokens based on a frozen pretrained model's "prior support" for stability.
• Self Improving AI actually solves everything — Pioneer Agent automates the entire lifecycle of fine-tuning small, open-source language models, enabling autonomous self-improvement and performance gains.
• AutoPass: Evidence-Guided LLM Agents for Compiler Performance Tuning — AutoPass guides LLM agents with compiler and runtime evidence for superior, training-free compiler performance tuning.
• Production Sub-agents for LLM Post Training — Sub-agent architectures with Claude Code can drastically reduce ML post-training time by parallelizing tasks and mitigating context limitations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AMD ROCm Blogs.