Back to Gap Map Infrastructure / Compilers & Model Optimization

cuLA

open source / Overall score: 1.5

cuLA (CUDA Linear Attention) provides hand-tuned CUDA kernels, written in CuTe DSL and CUTLASS C++, for linear attention variants including GLA, KDA, GDN and Lightning Attention, targeting NVIDIA Hopper and Blackwell GPUs. It shares its interface with the flash-linear-attention library as an early-stage standalone precursor to a planned integration.

Openness

5 high confidence

5.0

license: Apache-2.0(OSI)
source: public
core features withheld: no

Apache-2.0 license body confirmed. The repository is public and unarchived and builds the whole product, and the README describes no paid tier, enterprise edition or license-gated build beside it, so source is public and the core ungated.

https://api.github.com/repos/inclusionAI/cuLA recorded 2026-09-02
Repo metadata for cula's canonical repository - private false, archived false, fork false.
https://github.com/inclusionAI/cuLA/blob/main/LICENSE recorded 2026-10-07
LICENSE body carries the verbatim Apache License 2.0 text under an Ant Group / inclusionAI copyright header.
https://github.com/inclusionAI/cuLA/blob/main/README.md recorded 2026-10-07
README describes cuLA as "High-performance CUDA kernels for linear attention variants", providing "hand-tuned CUDA implementations of these linear attention variants, targeting NVIDIA Blackwell (SM10X) and Hopper (SM90) GPUs", and flags itself as early-stage with "significant room for optimization"; it names no paid tier, enterprise edition or license-gated build beside the published source.

Adoption

1 low confidence

1.0

cuLA has no registry package and is built from source, so GitHub stars are the only adoption signal available. A star is not a use.

https://api.github.com/repos/inclusionAI/cuLA recorded 2026-09-02
Repo metadata - stargazers_count = 544 - for inclusionAI/cuLA.

Capability

2 medium confidence

2.0

cuLA provides CUDA kernels for several linear-attention variants on NVIDIA's newest GPUs, and its README calls it early-stage - a narrow kernel set, two steps below TensorRT.

https://github.com/inclusionAI/cuLA/blob/main/README.md recorded 2026-10-07
README describes cuLA as "High-performance CUDA kernels for linear attention variants", providing "hand-tuned CUDA implementations of these linear attention variants, targeting NVIDIA Blackwell (SM10X) and Hopper (SM90) GPUs", and flags itself as early-stage with "significant room for optimization"; it names no paid tier, enterprise edition or license-gated build beside the published source.

Verified 2026-09-27