triton-ascend-case-reduction-weighted-swiglu

What problem does it solve?

Accelerates the backward fusion of 3D tensor operators by reshaping dimensions to simplify parallelism and memory access on Ascend devices.

Core Features & Use Cases

• Reshape-based dimensionality reduction to merge the first two axes and streamline kernel scheduling.
• Two-stage partitioning of the reduce axis to maximize unit occupancy while avoiding UB.
• Autotune-driven configuration exploration for grids and blocks to find optimal performance on Atlas A2/A3.
• Applicable to 3D tensor workloads with elementwise operations fused with a final reduce.

Quick Start

Run the autotune workflow to benchmark configurations and select the best setup for your 3D SwiGLU backward fusion on Ascend.