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 module
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.
Structs
-
B200BlockScaledMatmulSmem: Defines the shared memory layout for the B200 block-scaled matmul kernel, including A/B/C tiles, scale factor tiles, and pipeline barriers for TMA-MMA, accumulator, CLC, and TMEM deallocation.
Functions
-
blackwell_block_scaled_matmul_tma_umma_warp_specialized: Launch block-scaled FP8 matmul kernel on SM100. -
blackwell_block_scaled_tma_umma_warp_specialized_kernel: Implements the warp-specialized block-scaled matmul kernel for SM100 GPUs using TMA for global-to-shared loads and UMMA for tensor core MMA. -
consumer_main_loop: TileTensor-based consumer_main_loop for block-scaled MMA. -
load_AB_SFA_SFB: Issues multicast TMA loads for A, B, and their scale factors (SFA, SFB) into a pipeline stage of shared memory.
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!