For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo package

sm90

Provides the Nvidia Hopper backend implementations for matmuls.

Modules

• ​config: Provides static configuration and heuristic config builders for SM90 (Hopper) GPU warp-specialized matmul kernels.
• ​dispatch: Dispatches rank-2 matmuls to the SM90 (Hopper) warp-specialized kernel.
• ​grouped_matmul: Implements grouped (batched expert) GEMM for SM90 (Hopper) GPUs.
• ​matmul: Provides SM90 warp-specialized GEMM dispatchers with TMA multicasting, split-K reduction, and optional epilogue fusion for Hopper matmuls.
• ​matmul_kernel_persistent: Implements persistent warp-specialized matmul kernels for NVIDIA SM90 (Hopper) GPUs.
• ​matmul_kernels: Implements Hopper SM90 matrix multiplication kernels using TMA and WGMMA.
• ​matmul_output: Provides the output writer for SM90 warp-group matrix multiply-accumulate kernels.
• ​testbed: Provides a testbed for validating SM90 warp-specialized matmul kernels against a vendor reference.
• ​testbed_swapAB: Testbed for comparing swapAB vs normal matmul execution.
• ​tile_loader: TileLoader module for efficient tile loading in GPU matrix multiplication.
• ​tile_writer: TileWriter module for efficient tile writing in GPU matrix multiplication.
• ​tuning_configs: Defines the SM90 (Hopper) matmul tuning configuration record and the pre-tuned BF16 lookup table of kernel hyperparameters for specific (M, N, K) problem shapes.