Skip to main content

Ctrl+K

ROCm™ AI Ecosystem

Core SDK
AI Ecosystem
GPU Systems and Infrastructure
Toolkits

Blogs
Developer Hub

ROCm AI Ecosystem

AMD Primus 26.9

Getting started

Quickstart
Installation and setup
- Bare-metal installation (Megatron/TorchTitan)
- Bare-metal installation (JAX MaxText)
Release notes
Glossary

User guide

Primus tools
CLI reference
Configuration system
Pretraining workflows
End-to-end training recipes
Post-training workflows
Node-smoke test instruction
Preflight diagnostics
Run preflight without a container
Micro-benchmarking suite
Memory and performance projection
Tuning agent

Configuration reference

Megatron backend
TorchTitan backend
Megatron Bridge backend
MaxText backend
Environment variables

Technical guides

Parallelism strategies
Parallelism configuration
Multi-node networking
AINIC bundle versions
Collective operations (NCCL/RCCL)
Data preparation
Checkpoint management
Turbo cross entropy
MoE training
MegaMoE fused MoE layer
Performance tuning
Profiling and observability
Logging and experiment tracking
Fault tolerance and elastic training
Determinism and reproducibility
Native SFT LoRA
Hybrid recurrent-attention models
Diffusion models
SpecForge

Operations

Deployment
Monitoring and logging
Troubleshooting
Security considerations

Developer guide

Architecture overview
CLI architecture
Contributing
Adding model configurations
Extending backends
Model support matrix
Testing
Tooling
Backend patch notes

Technical blogs

Blog index

About

License

Repository
Open issue

Index

Terms and Conditions
ROCm Licenses and Disclaimers
Privacy
Trademarks
Supply Chain Transparency
Fair and Open Competition
UK Tax Strategy
Cookie Policy
Cookie Settings

© 2026 Advanced Micro Devices, Inc