Skip to main content

Mojo module

mla_prefill_utils

Structs

Functions

  • cvt_block_fp8_to_bf16_with_scale: TileTensor overload — standalone implementation using .ptr and comptime static_shape/static_stride directly.
  • split_smem: Split a shared memory tensor into two TileTensors at the boundary of first_size elements.

Was this page helpful?