Primus-Turbo

Primus-Turbo is a high-performance acceleration library dedicated to large-scale model training on AMD GPUs. Built and optimized for the AMD ROCm platform, it covers the full training stack — including core compute operators (GEMM, Attention, GroupedGEMM), communication primitives, optimizer modules, low-precision computation (FP8), and compute–communication overlap kernels.

With High Performance, Full-Featured, and Developer-Friendly as its guiding principles, Primus-Turbo is designed to fully unleash the potential of AMD GPUs for large-scale training workloads, offering a robust and complete acceleration foundation for next-generation AI systems.

Note: JAX support is under active development. Optim support is planned but not yet available.

🚀 What's New

• [2025/12/16] 🔥MoE Training Best Practices on AMD GPUs
• [2025/12/01] 🔥Efficient MoE Pre-training at Scale on 1K AMD GPUs with TorchTitan.
• [2025/09/19] Primus-Turbo introduction blog.
• [2025/09/11] Primus-Turbo initial release, version v0.1.0.

🧩 Primus Product Matrix

Module	Role	Key Features
Primus-LM	E2E training framework	- Supports multiple training backends (Megatron, TorchTitan, etc.) - Provides high-performance, scalable distributed training - Deeply integrates with Primus-Turbo and Primus-SaFE
Primus-Turbo	High-performance operators & modules	- Supports core training operators and modules (FlashAttention, GEMM, GroupedGemm, DeepEP etc.) - Integrates multiple high-performance backends (e.g., CK, hipBLASLt, AITER) - High performance and easy to integrate
Primus-SaFE	Stability & platform layer	- Cluster sanity check and benchmarking - Kubernetes scheduling with topology awareness - Fault tolerance - Stability enhancements

📦 Quick Start

Requirements

Software

• ROCm >= 6.4
• Python >= 3.10
• PyTorch >= 2.6.0 (with ROCm support)
• rocSHMEM (optional, required for experimental DeepEP). Please refer to our DeepEP Installation Guide for instructions.

Hardware

Architecture	Supported GPUs
GFX942	✅MI300X, ✅MI325X
GFX950	✅MI350X, ✅MI355X

See AMD GPU Architecture to find the architecture for your GPU.

1. Installation

Docker (Recommended)

Use the pre-built AMD ROCm image from Docker Hub:

# PyTorch Ecosystem
rocm/primus:v26.1

# JAX Ecosystem
rocm/jax-training:maxtext-v25.9

Install from Source

git clone https://github.com/AMD-AGI/Primus-Turbo.git
cd Primus-Turbo

pip3 install -r requirements.txt
pip3 install --no-build-isolation .

# (Optional) Set GPU_ARCHS environment variable to specify target AMD GPU architectures.
GPU_ARCHS="gfx942;gfx950" pip3 install --no-build-isolation .

2. Development

For contributors, use editable mode (-e) so that code changes take effect immediately without reinstalling.

git clone https://github.com/AMD-AGI/Primus-Turbo.git
cd Primus-Turbo

pip3 install -r requirements.txt
pip3 install --no-build-isolation -e . -v

# (Optional) Set GPU_ARCHS environment variable to specify target AMD GPU architectures.
GPU_ARCHS="gfx942;gfx950" pip3 install --no-build-isolation -e . -v

# (Optional) Set PRIMUS_TURBO_FRAMEWORK to compile for a specific framework.
# Supported values: PYTORCH (default), JAX.
# For example, to compile for JAX:
PRIMUS_TURBO_FRAMEWORK="JAX" pip3 install --no-build-isolation -e . -v

3. Testing

Option 1: Single-process mode (slow but simple)

pytest tests/pytorch/    # run all PyTorch tests
pytest tests/jax/        # run all JAX tests

Option 2: Multi-process mode (faster)

# PyTorch tests
## single-GPU tests (parallel)
pytest tests/pytorch/ -n 8
## deterministic tests (parallel)
pytest tests/pytorch/ -n 8 --deterministic-only
## multi-GPU tests
pytest tests/pytorch/ --dist-only

# JAX tests
## single-GPU tests (parallel)
pytest tests/jax/ -n 8
## multi-GPU tests
pytest tests/jax/ --dist-only

4. Packaging

pip3 install -r requirements.txt
python3 -m build --wheel --no-isolation
pip3 install --extra-index-url https://test.pypi.org/simple ./dist/primus_turbo-XXX.whl

5. Minimal Example

import torch
import primus_turbo.pytorch as turbo

dtype = torch.bfloat16
device = "cuda:0"

a = torch.randn((128, 256), dtype=dtype, device=device)
b = torch.randn((256, 512), dtype=dtype, device=device)
c = turbo.ops.gemm(a, b)

print(c)
print(c.shape)

💡 Example

See Examples for usage examples.

📊 Performance

See Benchmarks for detailed performance results and comparisons.

📍 Roadmap

Roadmap: Primus-Turbo Roadmap H1 2026

📜 License

Primus-Turbo is licensed under the MIT License.

© 2025 Advanced Micro Devices, Inc. All rights reserved.