skills.rest
xchang1121
xchang1121COMMUNITY
Explore all skills by this author →
GitHub

triton-ascend-case-reduction-weighted-swiglu

Optimize 3D fused operator backward passes on Ascend via reshape-based dimensionality reduction.

6|1|Updated Apr 19, 2026
Search Tags:#triton#fusion#autotune#ascend#kernel-optimization#3d-tensor#backward
One-click install
npx skills add https://github.com/xchang1121/AutoResearch-CC-hook --skill triton-ascend-case-reduction-weighted-swiglu
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: triton-ascend-case-reduction-weighted-swiglu
Source: https://github.com/xchang1121/AutoResearch-CC-hook/tree/main/skills/triton-ascend/cases/triton-ascend-case-reduction-weighted-swiglu
Command: npx skills add https://github.com/xchang1121/AutoResearch-CC-hook --skill triton-ascend-case-reduction-weighted-swiglu

SYSTEM DOCUMENTATION & REQUIREMENTS

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.

Frequently Asked Questions about triton-ascend-case-reduction-weighted-swiglu

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I optimize 3D tensor backward pass on Ascend with Triton?▼

Reshape dimensions to merge the first two axes of your 3D tensor, streamlining kernel scheduling and memory access. This restructuring simplifies parallelism for elementwise operations fused with reduce axes on Ascend devices.

Does triton autotune work for SwiGLU backward fusion on Atlas A2?▼

Yes, autotune-driven configuration exploration benchmarks grid and block setups to find optimal performance for SwiGLU backward fusion on Atlas A2 and A3. Run the autotune workflow to select the best configuration.

What is the best way to handle reduce axis partitioning for Ascend kernels?▼

Use two-stage partitioning of the reduce axis to maximize unit occupancy while avoiding Unified Buffer (UB) limitations. This memory-access aware scheduling strategy ensures efficient parallelism for 3D tensor workloads.

When do I need reshape-based dimensionality reduction for kernel optimization?▼

You need reshape-based dimensionality reduction when accelerating backward fusion of 3D tensor operators. It streamlines kernel scheduling for workloads fusing elementwise operations with a final reduce, like Weight SwiGLU backward paths.

Why does my 3D fused operator backward path experience UB overflow on Ascend?▼

UB overflow occurs when reduce axis partitioning is not optimized. Applying two-stage partitioning during memory-access aware scheduling maximizes unit occupancy while explicitly avoiding Unified Buffer overflow on Ascend hardware.

Related Skills

Explore More Software Engineering Skills→
xchang1121xchang1121
6

triton-ascend-example-relu

Implement a Triton Ascend elementwise ReLU kernel with 1D tiling and boundary masking.

Community
Intermediate
mindspore-aimindspore-ai
258

triton-ascend-case-reduction-weighted-swiglu

Optimize backward fusion of 3D SwiGLU-like operators on Ascend hardware.

Official
Intermediate
xchang1121xchang1121
6

triton-ascend-example-layernorm

Implement a two-stage LayerNorm reduction kernel with Triton Ascend.

Community
Intermediate
pedestrianlovepedestrianlove

Megatron-LM

Manage tensor parallelism, inference optimizations, and mixed-precision support for Megatron-LM.

Community
Advanced