Get up and running with Primus in 5 minutes!
- AMD ROCm drivers (β₯ 7.0)
- Docker (β₯ 24.0) with ROCm support
- AMD Instinct GPUs (MI300X, MI325X, etc.)
# Quick verification
rocm-smi && docker --version# Pull Docker image
docker pull docker.io/rocm/primus:v26.2
# Clone repository
git clone --recurse-submodules https://github.com/AMD-AIG-AIMA/Primus.git
cd Primus# Run a quick benchmark in container
./primus-cli container --image rocm/primus:v26.2 \
-- benchmark gemm -M 4096 -N 4096 -K 4096Expected output:
[BENCH] Markdown saved: ./gemm_report.md (overwrite)
[β] GEMM benchmark finished. Results saved to ./gemm_report.md
β See this output? You're ready to train!
Use the Docker image you just pulled:
# Run training in container (recommended for getting started)
./primus-cli container --image rocm/primus:v26.2 \
-- train pretrain --config examples/megatron/configs/MI300X/llama2_7B-BF16-pretrain.yamlOther modes:
# Direct mode (if running on bare metal with ROCm installed)
./primus-cli direct -- train pretrain \
--config examples/megatron/configs/MI300X/llama2_7B-BF16-pretrain.yaml
# Slurm mode (for multi-node cluster)
./primus-cli slurm srun -N 8 -p gpu -- container --image rocm/primus:v26.2 \
-- train pretrain --config examples/megatron/configs/MI300X/llama2_7B-pretrain.yamlprimus-cli [options] <mode> [mode-args] -- [command]
β β β β
Global Mode Mode-specific Training
options parameters commandCommon options:
--debug- Verbose logging--dry-run- Preview without executing
Learn More:
- CLI User Guide - Complete reference
- CLI Architecture - Design deep dive
- Configuration Guide - YAML configuration
- Examples - Real-world templates
Need Help?
- FAQ - Common questions
- GitHub Issues - Report bugs
That's it! Start training with primus-cli π