AMD Primus
Train AI Models at Scale — Without the Complexity
Primus brings training setup, optimized transformers, and production operations into one modular experience. Move from first run to fleet-scale training quickly with reproducible workflows, kernel acceleration, cluster validation, and monitoring baked in.
With Primus, teams can:
Monitor distributed jobs and debug issues faster
Launch reproducible training experiments with a single CLI
Accelerate transformer workloads using optimized operators
Validate clusters before large training runs
Key Components
The Primus Platform
Primus provides modular components that support the entire training lifecycle — from experiment setup to cluster-scale operations.
- Validate & Observe
- Production Pre-Training
- Production Scale RL
Reliability Tools
Validate & Observe
Plan training configurations, validate cluster readiness, and monitor runs at scale.
Finding the Right Training Configuration
Primus Tuning + Primus Projection
- Find training settings for your GPU memory budget.
- Project GPU memory use and training throughput.
- Identify viable configurations before full-cluster runs.
Validate the Cluster before Launch
Primus-Bench + Primus Preflight
- Check cluster hardware and network health.
- Benchmark performance and flag node outliers.
- Test training runs before full-cluster launches.
Spot Bottlenecks in Live Runs
Primus-Lens
- Monitor training jobs and cluster health.
- View distributed runs in dashboards.
- Use alerts to spot performance issues.
Primus Core
Production Pre-Training
Support large-model training with optimized operators and resilient cluster operations.
Launch Large-Model Training Faster
Primus-LM
- Launch large-model training from one CLI.
- Set up experiments with YAML and training recipes.
- Use Megatron-LM, PyTorch TorchTitan, or JAX MaxText.
Boost Workload Performance
Primus-Turbo
- Improve transformer training throughput with optimized operators.
- Overlap compute and communication to improve GPU utilization.
- Integrate with custom training workflows.
Reliable Cluster-Scale Runs
Primus-SaFE
- Schedule jobs with topology-aware placement.
- Check cluster readiness before training.
- Retry failed jobs and resume training from checkpoints.
RL Post Training
Production-Scale RL
Scale RL post-training on AMD Instinct™ GPUs with ready-to-run, validated AMD ROCm™ workflows.
Start RL Post-Training with a Ready-to-Run Stack
veRL on ROCm
- Start from a prebuilt container with verified software dependencies.
- Run PPO, GRPO, and DAPO post-training workflows.
- Customize software versions and GPU targets for your environment.
Scale Rollouts and Training Independently
vLLM and SGLang Rollouts
- Choose AITER-accelerated vLLM or SGLang for rollout.
- Size rollout and training GPU pools independently in fully asynchronous mode.
- Overlap generation and training to reduce waiting on long responses.
Validate Model Quality Across Training Modes
Accuracy and Convergence
- Use validated examples for dense and MoE reasoning models.
- Compare convergence across co-located and fully asynchronous execution.
- Track validation accuracy, training rewards, and rollout–training consistency.
Performance Benchmarks
Primus benchmarks show how customers can turn AMD Instinct™ GPU capacity into fast model development — scaling from a single node to multi-node clusters while maximizing GPU utilization, shortening time-to-train, and improving training efficiency across diverse model architectures.
ROCm 10 vs. ROCm 7
Accelerating Training Improvement
DeepSeek-V3-16B
DeepSeek-V2-Lite
Qwen3-30B-A3B
Release-to-Release: Performance Improvement
2.01x
2.54x
2.76x
On Average
2.4X
faster than previous release across models
Parallelization & Scheduling
Pipeline parallelism, sharded data parallel, gradient all-reduce overlap
Memory Management
Activation checkpointing, FP8 mixed precision, fused gradient accumulation
Optimized Kernels
Fused FlashAttention (fwd + bwd), grouped GEMM • fused optimizer step
Real World Results
Tested at Scale, Proven in Production
Discover how Zyphra built ZAYA1, from scratch on a full-stack AMD platform. This collaboration shows a competitive model, trained on over 1K GPUs, converging in production settings on an E2E Instinct stack.
ZAYA1-base
Qwen3-4B
Gemma3-12B-PT
Llama3.1-8B
OLMoE-1B-7B
MMLU-Pro(5) Accuracy (%)
40.43
41.92
42.25
ZAYA1-base
Qwen3-4B
Gemma3-12B-PT
Llama3.1-8B
OLMoE-1B-7B
MATH-hard(4) Exact Match (%)
54.15
47.05
17.98
ZAYA1-base
Qwen3-4B
Gemma3-12B-PT
Llama3.1-8B
OLMoE-1B-7B
MBPP+(3) Pass@1 (%)
75.40
76.46
75.40
Get Started with Primus
The Journey from Model Selection to Training at Scale
Experience the production tested, reference architecture for building frontier AI models from scratch on Instinct.
Icon
Use Primus LM within the Primus Ecosystem.
Icon
View the latest pre-release features & source code on GitHub.
Resources
