用 NVIDIA CUDA Tile 高性能前向内核加速 Mamba-2
原标题：Accelerating Mamba-2 with High-Performance Forward Pass Kernels in NVIDIA CUDA Tile
AI 摘要
NVIDIA 技术博客介绍其 Python 并行编程 DSL cuTile，Jane Street 的机器学习工程师作为早期采用者用 cuTile 实现了 Mamba-2 序列模型，并与 NVIDIA 合作优化内核。文章详细讲解 Mamba-2 前向传播的三个自定义内核及其 cuTile 优化，展示如何通过分块矩阵乘法利用 Tensor Cores 加速训练与预填充。
正文节选
NVIDIA CUDA Tile (cuTile) is a Python-based DSL and parallel programming model for NVIDIA GPUs that automatically leverages advanced hardware capabilities to deliver peak performance for tile-level GPU programming. By abstracting thread- and warp-level scheduling behind high-level, tile-centric primitives, cuTile enables you to write GPU kernels in Python while compiler optimizations automatically target advanced hardware features like Tensor Cores. As an early adopter of cuTile, machine learnin