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
blockwise_fp8
Implements blockwise scaled FP8 matrix multiplication kernels targeting the NVIDIA SM100 (Blackwell) architecture using 1D A-scales and 2D B-scales.
comptime values
logger
comptime logger = Logger(stdout, prefix=String(""), source_location=False)
Functions
-
matmul_sm100_blockwise_scaled_fp8: Enqueues a blockwise scaled FP8 GEMM on SM100 with 1D A-scales and 2D B-scales onto the supplied device context. -
matmul_sm100_blockwise_scaled_fp8_1d2d_kernel: Performs a blockwise scaled FP8 matrix multiplication on SM100 using 1D A-scales and 2D B-scales with TMA-based asynchronous loads and TCgen05 tensor memory accumulation. -
matmul_sm100_blockwise_scaled_fp8_1d2d_wrapper: Wraps the blockwise scaled FP8 SM100 kernel with the LLVM metadata decorators needed for cluster dimensions and TMA grid-constant arguments.
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!