Skip to main content

Mojo function

col_max_acc

col_max_acc(mut col_accum: TileTensor[DType.float32, address_space=AddressSpace.LOCAL], src: TileTensor[DType.float32, address_space=AddressSpace.LOCAL], src_accum: TileTensor[DType.float32, address_space=AddressSpace.LOCAL])

Running-max for online softmax: col_accum[j] = max(src_accum[j], max(src[*, j, *])).