For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).
Mojo package
sm100
Provides the Nvidia Blackwell backend implementations for matmuls.
Modules
-
block_scaled_dispatch: Provides dispatch logic for SM100 block-scaled (NVFP4, MXFP4, MXFP8) matmul kernels with optional elementwise epilogue. -
block_scaled_matmul: Implements the warp-specialized block-scaled matmul kernel for SM100 (B200) GPUs, supporting MXFP8, MXFP4, and NVFP4 block-scaled formats with TMA-based loads and UMMA tensor core operations. -
block_scaled_matmul_small_bn: Implements the small-BN block-scaled matmul kernel for SM100 (Blackwell B200). -
blockwise_fp8: Implements blockwise scaled FP8 matrix multiplication kernels targeting the NVIDIA SM100 (Blackwell) architecture using 1D A-scales and 2D B-scales. -
config: SM100 matmul configuration - re-exports from sm100_structured. -
matmul: SM100 matmul - TileTensor-native compatibility layer for grouped_matmul. -
pipeline: SM100 pipeline utilities - re-exports from sm100_structured. -
testbed_block_scaled_fp4: Shared test harness for SM100 block-scaled FP4 matmul tests. -
tile_scheduler: Schedules output-tile work for persistent SM100 matmul kernels using Cluster Launch Control (CLC), rasterization order, and block swizzling. -
warp_specialized_blockwise_fp8: Implements a warp-specialized blockwise FP8 matrix multiplication kernel for NVIDIA SM100 (Blackwell) GPUs.
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!