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
warp_specialized_blockwise_fp8
Implements a warp-specialized blockwise FP8 matrix multiplication kernel for NVIDIA SM100 (Blackwell) GPUs.
Functions
-
blackwell_tma_umma_warp_specialized_blockwise_fp8_kernel: Implements the warp-specialized blockwise FP8 GEMM kernel for SM100 using TMA loads, UMMA tensor-core MMA, and a CLC tile scheduler. -
load_AB: Loads A, B, and A-scales tiles into shared memory via multicast TMA for one pipeline stage. -
multi_stage_reg_epilogue: Casts register accumulators to the output type and stores each stage to global memory via TMA. -
promote_accumulators: Promotes FP8 MMA partial products by applying blockwise A and B scales to the tensor-memory accumulators. -
sm100_warp_specialized_blockwise_fp8: Sets up TMA descriptors, shared-memory layouts, and pipeline barriers, then launches the blockwise FP8 GEMM kernel on SM100.
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!