Primus tools
August 12, 2026 · View on GitHub
A quick catalog of the tools that ship with Primus and the sibling projects around it—command-line tools, the tuning agent, ecosystem projects, and auxiliary utilities. Each row gives a short description and a how-to starting point; follow a tool's link for the full reference.
| Tool | Type | What it does | How to use |
|---|---|---|---|
train | CLI | Launch pretraining or post-training on any backend from a YAML configuration. | primus-cli <mode> -- train pretrain --config <yaml> |
benchmark | CLI | GEMM, RCCL, and attention microbenchmarks for hardware and stack validation. | primus-cli direct -- benchmark gemm --M 4096 --N 4096 --K 4096 |
preflight | CLI | Host, GPU, and network health checks before long jobs. | primus-cli slurm srun -N 4 -- preflight --host --gpu --network |
projection | CLI | Estimate per-GPU memory and throughput without occupying a full cluster. | primus-cli direct -- projection both --config <yaml> |
| Tuning agent | Agent | LLM-driven search for a near-optimal training configuration, scored by projection. | python -m primus.agents.tuning_agent --workload <yaml> --target-cluster <yaml> |
| Primus-LM | Ecosystem | The training framework in this repository (multi-backend, unified CLI). | See the Quickstart |
| Primus-Turbo | Ecosystem | High-performance ROCm operators (attention, GEMM, grouped GEMM, DeepEP, FP8/FP4). | Bundled in the rocm/primus image; enabled via configuration flags |
| Primus-SaFE | Ecosystem | Kubernetes-native stability, scheduling, and fault-tolerance platform. | Deployed separately on Kubernetes (Helm) |
| IRLens | Auxiliary | Parse XLA HLO dumps into a communication-vs-compute execution skeleton. | See the README |
| model_stats | Auxiliary | Chart model dimensions from the config registry. | See the README |
| Pipeline visualization | Auxiliary | Render pipeline-parallel schedules in a local web UI. | See the README |
| Auto benchmark | Auxiliary | Interactive Megatron/TorchTitan benchmark menu with metrics collection. | See the README |
Related documentation
- CLI reference—full launcher grammar and subcommand options.
- Tooling—developer-guide index of the
tools/utilities. - Project overview—how the Primus ecosystem layers fit together.