triton-ascend-example-softmax

What problem does it solve?

Provides a complete Triton Ascend softmax reduction implementation example, illustrating how to build a three-phase reduction (max → sum(exp) → normalize) with block tiling and precision strategies to achieve high performance on Ascend hardware.

Core Features & Use Cases

• Three-stage reduction: max, sum(exp), normalize to compute softmax efficiently on Triton Ascend.
• Block tiling and scalar accumulator techniques to improve numerical stability and throughput.
• Use Case: design and benchmark reduce operators for neural networks on Ascend devices, with a ready-to-study kernel and reference PyTorch model.

Quick Start

Run the softmax example to study the Triton Ascend reduction kernel and integrate its patterns into your own reduce operators.