tritonBLAS

A lightweight Triton-based GEMM library that uses an analytical model to predict optimal configurations without autotuning.

Important: This project is intended for research purposes only.

Quick Start

git clone https://github.com/ROCm/tritonBLAS.git
cd tritonBLAS
pip3 install -e .
export PYTHONPATH=$(pwd)/include/:$PYTHONPATH

import torch
import tritonblas

A = torch.randn(4096, 4096, dtype=torch.float16, device='cuda')
B = torch.randn(4096, 4096, dtype=torch.float16, device='cuda')
C = tritonblas.matmul(A, B)

Supported GPUs

GPU Model	Support Status
MI300X	✅ Supported
MI300A	✅ Supported
MI308X	✅ Supported
MI350X	✅ Supported
MI355X	✅ Supported

License

This project is licensed under the MIT License.