Addition of Sparse CSC and Dense Tensors Triggers Internal Assertion Failure

[jdenhof]

🐛 Describe the bug

X = torch.randn(2, 2).to_sparse_csc() 
Y = torch.randn(2, 2)
X + Y

Traceback (most recent call last):
File "/Users/jaggerdenhof/Workspaces/gvsu-workspace/autoencoder/Model/example.py", line 57, in
X + Y
RuntimeError: at::sparse_csr::is_sparse_compressed(other)
INTERNAL ASSERT FAILED at "/Users/runner/work/pytorch/pytorch/pytorch/aten/src/ATen/SparseCsrTensorUtils.h":329,
please report a bug to PyTorch.

Versions

PyTorch version: 2.1.1
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: macOS 13.0.1 (x86_64)
GCC version: Could not collect
Clang version: 14.0.3 (clang-1403.0.22.14.1)
CMake version: version 3.27.4
Libc version: N/A

Python version: 3.9.7 (default, Sep 16 2021, 08:50:36) [Clang 10.0.0 ] (64-bit runtime)
Python platform: macOS-10.16-x86_64-i386-64bit
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Apple M1 Pro

Versions of relevant libraries:
[pip3] numpy==1.24.3
[pip3] torch==2.1.1
[pip3] torchvision==0.15.2
[conda] blas 1.0 mkl
[conda] mkl 2023.1.0 h8e150cf_43559
[conda] mkl-service 2.4.0 py39h6c40b1e_1
[conda] mkl_fft 1.3.6 py39h07fba90_1
[conda] mkl_random 1.2.2 py39h07fba90_1
[conda] numpy 1.24.3 pypi_0 pypi
[conda] numpy-base 1.25.2 py39ha186be2_0
[conda] torch 2.0.0 pypi_0 pypi

cc @alexsamardzic @nikitaved @pearu @cpuhrsch @amjames @bhosmer

[cpuhrsch]

@amjames - Can you take a look? This also fails for .to_sparse_csr

>>> import torch
>>> X = torch.randn(2, 2).to_sparse_csc()
<stdin>:1: UserWarning: Sparse CSC tensor support is in beta state. If you miss a functionality in the sparse tensor support, please submit a feature request to https://github.com/pytorch/pytorch/issues. (Triggered internally at ../aten/src/ATen/SparseCsrTensorImpl.cpp:53.)
>>> Y = torch.randn(2, 2)
>>> X + Y
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
RuntimeError: at::sparse_csr::is_sparse_compressed(other) INTERNAL ASSERT FAILED at "../aten/src/ATen/SparseCsrTensorUtils.h":329, please report a bug to PyTorch.
>>> X = torch.randn(2, 2).to_sparse_csr()
>>> X + Y
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
RuntimeError: at::sparse_csr::is_sparse_compressed(other) INTERNAL ASSERT FAILED at "../aten/src/ATen/SparseCsrTensorUtils.h":329, please report a bug to PyTorch.
>>> torch.version.__version__
'2.2.0.dev20231205+cu121'

[cpuhrsch]

The addition of COO and dense is dense. I think we can follow the same convention for CSR/CSC and dense.

[amjames]

@cpuhrsch @jdenhof
Two bugs involved here.

First, sparse compressed add for dense/sparse requires that the lhs is the dense argument, if that condition is not met it always falls into the sparse-sparse path. #115432 addresses this.

Second, the underlying algorithm for dense + sparse_csr is not generic on the compressed dimension of the sparse argument #115433 fixes this adding support for csc.