January 2026 performance summary for fla-org/flash-linear-attention. Focused on delivering Context Parallel (CP) support for KDA, GDN, and Conv1d, enabling multi-rank parallelism while preserving causal dependencies. Implemented architecture enhancements, updated core functions to accept a CP context, and added comprehensive tests. Resulted in improved throughput, scalability, and reliability for parallel inference and training workloads.

May 2025 (2025-05) performance kernel optimizations across multiple kernels (DyT, GRPO Loss, Block RMS Normalization) delivered to accelerate training/inference and reduce GPU memory footprint. Introduced optimized element-wise DyT kernel (beta modes), a fully Triton-implemented GRPO Loss kernel with higher precision and reduced memory footprint, and a block RMS normalization kernel that delivers 2-4x speedups for large batches with small head dimensions. These kernels collectively reduce computation time and GPU memory usage while maintaining numerical accuracy. The work directly supports higher throughput, larger batch processing, and reduced training costs.