submission 774929

import torch
from task import input_t, output_t

def custom_kernel(data: input_t) -> output_t:
    input_tensor, output_tensor = data
    # .to(float64).sum() allocates a full fp64 tensor in HBM then sums it:
    #   read fp32 -> write fp64 (new alloc) -> read fp64 -> scalar
    # .sum(dtype=float64) fuses the cast into the reduction kernel:
    #   read fp32 -> scalar  (no intermediate tensor)
    # Returns a scalar (shape []) matching the reference's return type.
    return input_tensor.sum(dtype=torch.float64).to(torch.float32)