Bug

The PyTorch backend passed q directly to torch.quantile. NumPy permits quantile arrays with arbitrary shape, but PyTorch accepts only a scalar or one-dimensional q tensor. Consequently, valid calls such as a 2 x 2 quantile grid raised:

RuntimeError: quantile() q must be a scalar or 1D tensor

The failure affected both the public PyTorch backend and the raw backend, including tuple-axis reductions. The existing empty-batch compatibility path also failed while validating multidimensional quantiles.

Fix

• flatten multidimensional q only for the native torch.quantile call
• reshape the result back to NumPy's q.shape + reduced_shape contract
• preserve integer-input promotion before converting quantiles
• retain tuple-axis, keepdims, out, and empty-batch behavior
• apply the repair consistently to public and raw PyTorch backend entry points

Regression coverage

• compares 2 x 2 quantile grids against numpy.quantile
• covers integer inputs and tuple-axis reductions with keepdims=True
• verifies multidimensional quantiles preserve empty non-reduced batch dimensions

Validation

• reproduced the current failure with PyTorch 2.10.0
• syntax-compiled the modified source and regression test
• ran an exact wrapper smoke harness covering public/raw behavior, integer inputs, tuple axes, keepdims, and empty batch dimensions
• branch is limited to two files and is based directly on current main

The full repository and backend matrices should run in GitHub Actions.