Fix logsigmoid buffer shape for vmap+JVP

[Silv3S]

Fixes #2240

[Copilot]

Pull Request Overview

This PR fixes the buffer tensor shape allocation in the log_sigmoid_forward_xpu function to properly support vmap and JVP (Jacobian-vector product) operations. Previously, the buffer was initialized with shape {0}, which is incompatible with batched operations.

Key changes:

• Modified buffer allocation to use at::empty_like(input) instead of at::empty({0}, input.options())

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[guangyey]

May I know if there is the same issue existing on CUDA. I see CUDA use empty as well

std::tuple<Tensor, Tensor> log_sigmoid_forward_cuda(const Tensor& input) {
  auto result = at::empty_like(input);
  auto buffer = at::empty({0}, input.options());
  log_sigmoid_forward_out_cuda(input, result, buffer);
  return std::forward_as_tuple(result, buffer);
}

[EikanWang]

Could you please add a regression test?

[Silv3S]

It works on CUDA, as there are extra checks in log_sigmoid_backward_batch_rule() if input tensors are on CUDA device. We can either go with this patch or keep old version and add analogous check for XPU backend in upstream PyTorch.

[guangyey]

Thanks, @Silv3S.
Regarding the logic in
https://github.com/pytorch/pytorch/blob/2f023bf7b962e69c0de01b89af197388d9df27cc/aten/src/ATen/functorch/BatchRulesBinaryOps.cpp#L320-L331,
I think it may be worth changing the logic here.
@EikanWang, what’s your opinion?

[Silv3S]

CUDA custom batching strategy was introduced in pytorch/pytorch#84892