ROCm Triton Kernels (RMSNorm, RoPE 3D, GEGLU, AdaLN)

What it does

This skill provides in-depth patterns, kernel implementations, and benchmarking guidance for writing optimized Triton kernels targeting AMD ROCm GPUs (MI355X, R9700). It covers core kernels (RMSNorm, RoPE 3D, GEGLU, AdaLN), integration patterns for HuggingFace diffusers (LTX-Video) and transformers, and performance/benchmarking scripts.

When to use it

Use when developing or optimizing GPU kernels for AMD hardware, injecting custom kernels into diffusers pipelines, running micro-benchmarks and end-to-end comparisons, or debugging ROCm/HIP-specific constraints. Ideal for engineers optimizing inference/compute-heavy models on ROCm.

What's included

• Scripts: benchmark_kernels.py, benchmark_e2e.py, ltx_kernel_injection_example.py, and other scripts in the scripts/ folder.
• References: hardware guides (MI355X, R9700), integration guides, and troubleshooting docs in references/.
• Instructions: concrete kernel code samples (triton.jit implementations), mandatory env vars, integration patterns, performance pitfalls, and profiling commands.

Compatible agents

Best used by code-aware agents and developer-facing assistants that can run benchmark scripts and edit code; suitable for agents with access to compute/testing environments.