Write a Triton RMSNorm Kernel

What problem does it solve?

This Skill helps implement a numerically stable RMSNorm kernel in Triton, enabling fast and correct normalization for large language model blocks.

Core Features & Use Cases

• One-pass sum-of-squares with fp32 accumulation
• Masked tail handling for non-divisible H
• Per-feature weight broadcast and optional residual fusion
• Forward + backward capability for training or fine-tuning

Quick Start

Implement a forward RMSNorm kernel in Triton that computes y = x * rsqrt(mean(x^2) + eps) * weight with proper masking, and provide an accompanying backward kernel and Python wrapper for autograd.