Dao-AILab/quack

Health v2 · maintenance only

Adoption not part of the score

1135 stars · 150 forks observed · 2026-09-01

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

QuACK is a collection of high-performance GPU kernels (RMSNorm, LayerNorm, softmax, cross-entropy, GEMM with epilogues) written in NVIDIA's CuTe-DSL for Python. It targets modern NVIDIA GPUs (Hopper and Blackwell) and offers optional JAX bindings.

Use cases

• speed up rmsnorm and layernorm in pytorch training
• fast softmax and cross entropy kernels for llm training
• hopper and blackwell gemm with fused epilogue
• write custom cuda kernels in python with cute dsl
• use optimized kernels from jax
• get speed-of-light memory-bound kernels on h100

When to choose

• you need fast fused normalization, softmax, or cross-entropy kernels on H100/B200/RTX 50 GPUs
• you want to prototype GPU kernels in Python using CuTe-DSL
• you want drop-in kernels with optional JAX bindings

When to avoid

• you are on older GPUs like A100 or consumer pre-Ada cards
• you need CUDA versions below 12.9 or Python below 3.12
• you need a broad production-tested kernel library rather than a curated assortment

Facets

library · maturity active

gpu-computing machine-learning deep-learning gpu-computing machine-learning deep-learning developer-tools python cuda cute-dsl kernels gemm normalization softmax hopper blackwell jax linux gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="Dao-AILab/quack")