CODA: GPU Kernels as GEMM-plus-Epilogue Programs

CODA is a GPU kernel abstraction that expresses Transformer operators as GEMM-plus-epilogue programs, fusing normalization, activations, residual updates, and reductions into the GEMM output tile before it is written to global memory, combining framework-level productivity with hardware-level efficiency. CODA is built on CUTLASS CuTeDSL and targets NVIDIA Hopper (H100) GPUs.

Updates

• October 6, 2026. Released v0.3.1, pinning the quack fork to v0.6.5+fork.1.
• September 28, 2026. Released v0.3.0.
• July 19, 2026. Released v0.2.
• June 23, 2026. We are restructuring CODA. For legacy version, please check v0.1 tag.

Installation

Note: the coda-kernels package on PyPI is deprecated and no longer updated. Install from GitHub instead.

pip install git+https://github.com/open-lm-engine/[email protected]

Or from source:

git clone https://github.com/open-lm-engine/coda-kernels.git
cd coda-kernels
pip install -e .

CODA depends on a fork of quack, pinned to the tag v0.6.5+fork.1. Its version differs from PyPI's quack-kernels 0.6.5, so installing CODA replaces a PyPI install.

Functional API

• linear_swiglu
• linear_sigmoid
• linear_cross_entropy
• linear_cross_entropy_forward
• linear_qknorm_rope