Kernels

Production kernels from Tilde Research and verified community contributions.

OKLS

v1.4.2

Optimizer - Zero-staleness, KL-optimal Kronecker-factored preconditioner.

Wall Attention

v0.9.1

Attention - 4k → 160k+ extrapolation. Production Triton kernels.

NSA Kernels

v0.7.0

Sparse attention - Native Sparse Attention, hardware-aligned.

Aurora

v0.5.3

Optimizer - Early access - next-generation optimizer from Tilde.

MoMoE

v0.4.1

MoE memory - Memory-optimized Mixture-of-Experts primitives.

Nitrobrew

v0.3.0

Distillation - Fast distillation for production models.