Pytorch Sparse Versions Save

What's Changed

• Support sparse element-wise multiplication in SparseTensor via sparse_mat1 * sparse_mat2 (https://github.com/rusty1s/pytorch_sparse/pull/323)
• Fix gradient computation in to_symmetric (https://github.com/rusty1s/pytorch_sparse/pull/327)
• Fix empty edge indices handling (https://github.com/rusty1s/pytorch_sparse/pull/332)
• Add mps Apple silicon GPU Acceleration support (https://github.com/rusty1s/pytorch_sparse/pull/335)
• Fix shape of node_weight in metis computation (https://github.com/rusty1s/pytorch_sparse/pull/342)
• Add PyTorch 2.1.0 support (https://github.com/rusty1s/pytorch_sparse/pull/344)

New Contributors

• @AndreasBergmeister made their first contribution in https://github.com/rusty1s/pytorch_sparse/pull/323
• @ezyang made their first contribution in https://github.com/rusty1s/pytorch_sparse/pull/325
• @jamesmyatt made their first contribution in https://github.com/rusty1s/pytorch_sparse/pull/330
• @AgarwalSaurav made their first contribution in https://github.com/rusty1s/pytorch_sparse/pull/331
• @NripeshN made their first contribution in https://github.com/rusty1s/pytorch_sparse/pull/335

Full Changelog: https://github.com/rusty1s/pytorch_sparse/compare/0.6.17...0.6.18

• PyTorch 2.0 support (#317)
• Integrated faster index_sort in case pyg-lib is installed as well (#306)
• Added a balance_edge option to the METIS graph partitioning algorithm (#309)
• Added SparseTensor.to_torch_sparse_csc_tensor functionality (#319)

• Fix spspmm on newer CUDA versions/GPUs
• PyTorch 1.13 support
• torch.bfloat16 support in spmm
• Use faster hash-map routine in CPP code paths

• Temporal sampling is now correctly performed in disjoint mode (#267)
• Replace std::unordered_map with phmap::flat_hash_map for faster sampling (#266)
• Neighborhood sampling on heterogeneous graphs is now fully-deterministic (#265)

• Internal C++ method for sampling neighbors based on temporal constraints (#202, #225, #226)
• Sampling operators now respect torch.manual_seed (#217)

• SparseTensor: __eq__ functionality
• SparseTensor: add functionality of two sparse matrices (#177)
• SparseTensor: to_torch_csr_tensor and from_torch_csr_tensor functionality
• SparseTensor: Allow indexing via np.array (#194)
• SparseTensor: Skip unnecessary assertions and enable non-blocking data transfers (#195)
• Allow loading of CPU wheels in a PyTorch CUDA installation

PyTorch 1.10 is now required.

• (Internal) heterogeneous neighbor sampling support via torch.ops.torch_sparse.hetero_neighbor_sample
• (Internal) heterogeneous graph transformer sampling support via torch.ops.torch_sparse.hgt_sample (thanks to @chantat)
• Fixed a bug in set_diag in case SparseTensor does not hold any non-zero elements

• Half-precision support (torch.half) for all operators in torch-sparse

This release brings PyTorch 1.9.0 and Python 3.9 support to torch-sparse.

Additional functionality

• Added a check for row.max() < sparse_sizes[0] and col.max() < sparse_sizes[1] when creating a SparseTensor in order to avoid unexpected behavior (thanks to @Adam1679)
• partition now supports the additional optional argument node_weight (thanks to @Spazierganger)

• Reduced the size of shared library files
• CUDA wheels can now also operate on CPU-only devices
• spmm now supports torch.half
• Added parallelization strategies for CPU functionalities
• Fixed a bug in which sample_adj did not return a sparse matrix with sorted indices
• Fixed a bug in spmm in case num_edges < num_nodes