jax-linear-attention

jax-linear-attention provides highly optimized Pallas Mosaic GPU kernels for linear-attention and state-space models (SSM) targeting JAX on NVIDIA Hopper GPUs (H100/H200). It supports architectures like Mamba, Gated DeltaNet, KDA, RWKV-7, and GLA, offering significant speedups for training large language models with exact kernel implementations and an integrated training stack with FP8 support.