[laifenxiawucha]

Summary

Enables torch.sparse.spsolve on Intel XPU. aten::_spsolve previously had only a SparseCsrCUDA dispatch, so on XPU it raised:

RuntimeError: Could not run 'aten::_spsolve' with arguments from the 'SparseCsrXPU' backend

This adds the SparseCsrXPU dispatch line (one line), mirroring the existing SparseCsrCUDA registration:

- func: _spsolve(Tensor A, Tensor B, *, bool left=True) -> Tensor
  python_module: sparse
  dispatch:
    SparseCsrCUDA: _sparse_csr_linear_solve
    SparseCsrXPU: _sparse_csr_linear_solve_xpu   # added

Coordinated out-of-tree change

The XPU kernel _sparse_csr_linear_solve_xpu is implemented in torch-xpu-ops (intel-sandbox/torch-xpu-ops-exp#1965): since XPU has no native sparse direct solver, it densifies the CSR matrix and defers to the dense at::linalg_solve, mirroring the CUDA cuDSS backend's semantics and validation contract. That PR also carries the identical dispatch line in the torch-xpu-ops codegen yaml (as triangular_solve already does across both yamls).

Verification (Intel Data Center GPU Max 1100, source build torch 2.14.0a0)

With this dispatch line plus the torch-xpu-ops kernel built in:

• Issue reproducer: torch.sparse.spsolve on an identity CSR system returns [1,1,1,1] (was: dispatch RuntimeError).
• Regression suite test_spsolve_sparse_csr.py passes 5/5 (identity, dense reference + residual A@x==b, CPU-dense parity, float32, size-mismatch error).

🤖 Generated with Claude Code