🐛 Bug

The function torch.cosine_similarity allowed tensor input of differing shapes in torch 1.9.1. In the nightly version for 1.10, this does not seem to be allowed anymore. This breaks backwards compatibility - is this intended here?

To Reproduce

1.) Install torch 1.9.1:

pip3 install torch==1.9.1+cpu torchvision==0.10.1+cpu torchaudio==0.9.1 -f https://download.pytorch.org/whl/torch_stable.html

Run:

import torch
torch.cosine_similarity(torch.ones((10, 2, 3, 4)), torch.ones((2, 3, 4)), dim=-1)

=> works fine

2.) Install torch-nightly for torch 1.10:

pip3 install --pre torch torchvision torchaudio -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html

Run the same function:

import torch
torch.cosine_similarity(torch.ones((10, 2, 3, 4)), torch.ones((2, 3, 4)), dim=-1)

=> This function now returns an error:

RuntimeError: cosine_similarity requires both inputs to have the same number of dimensions, but x1 has 4 and x2 has 3

Expected behavior

Is it intended that backwards compatibility is not kept here?

Environment

• PyTorch Version (e.g., 1.0): 1.9.1
• OS (e.g., Linux): Linux
• How you installed PyTorch (conda, pip, source): pip
• Build command you used (if compiling from source): see above
• Python version: Python 3.9.1
• CUDA/cuDNN version: CPU
• GPU models and configuration: /
• Any other relevant information:

Additional context

cc @ezyang @gchanan @zou3519 @bdhirsh @jbschlosser