Skip to main content

Mojo function

mha_splitk_reduce

mha_splitk_reduce[intermediate_type: DType, output_type: DType, depth: Scalar[DType.uint], num_heads: Scalar[DType.uint], num_threads: Scalar[DType.uint], group: Scalar[DType.uint] = 1, use_exp2: Bool = False](intermediate_ptr: UnsafePointer[Scalar[intermediate_type], ImmutAnyOrigin], output_ptr: UnsafePointer[Scalar[output_type], MutAnyOrigin], exp_sum_ptr: UnsafePointer[Scalar[get_accum_type[output_type]()], MutAnyOrigin], qk_max_ptr: UnsafePointer[Scalar[get_accum_type[output_type]()], MutAnyOrigin], batch_size: Int, num_partitions: Int)

Was this page helpful?