Fix randint uniformity for large ranges on XPU

[Silv3S]

Fixes #3473

XPU random_from_to_kernel was using threshold condition from fbcode path (switch to 64-bit RNG only when range >= 2³² for specific dtypes), instead of range >= 2²⁸. This patch changes it, to use the default threshold that matches upstream CUDA non-fbcode path.

Upstream covers both paths, but fbcode is skipped from ut: pytorch/pytorch#144819

[Copilot]

Pull request overview

Updates XPU random_from_to_kernel to switch to a 64-bit RNG path for large ranges at range >= 2^28, aligning behavior with upstream CUDA (non-fbcode) and addressing non-uniform randint distributions for large high values.

Changes:

• Change RNG-width selection threshold from a dtype-gated range >= 2^32 condition to an unconditional range >= 2^28.
• Route large-range generation through the uint64_t distribution path to reduce bias.

[guangyey]

Nice!

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

Comments suppressed due to low confidence (1)

src/ATen/native/xpu/sycl/DistributionTemplates.h:485

• This change alters the RNG path selection for torch.randint/random_ over large ranges, but there is no regression test in this repo that exercises the large-range uniformity case (e.g. n_max near 0.75 * 2**32 and checking histogram spread). Please add a targeted XPU test to prevent this from regressing again.

        if (range >= 1ULL << 28) {
          distribution_nullary_kernel<
              scalar_t,
              uint64_t,
              rand4_engine_calls / 2>(
              iter,
              gen,
              Rand4ULL2DistFunctor(),
              UniformIntFromToTransformFunctor<scalar_t, uint64_t>(
                  range, base));
        } else {