torchmatch
torchmatch - assignment and transport solvers for PyTorch.
- assignmentInteger linear assignment problem (LAP) solvers.
- assignment.opsDirect access to ``torch.ops.assignment.*`` ops.
- benchBenchmark collection CLI.
- transportContinuous optimal-transport (OT) solvers.
- transport.matrixCost-matrix optimal-transport solvers.
- transport.matrix.opsDirect handles for torch.ops.transport.* matrix-face ops.
- transport.samplesSamples-face optimal transport.
- transport.samples.kernelsTriton kernels for streaming optimal transport.
- transport.samples.kernels.apply_fused_sqeuclidFused Schur complement matvec kernel for HVP CG acceleration.
- transport.samples.kernels.apply_rawraw-form apply kernels (mat5 for HVP, deprecated vec/mat wrappers).
- transport.samples.kernels.apply_shiftedshifted-form apply kernels (shifted potentials, s_I cancellation).
- transport.samples.kernels.apply_sqeuclidBackward-compatible re-exports for apply kernels.
- transport.samples.kernels.c_transform_sqeuclidC-Transform (hard argmin) kernel for squared Euclidean cost.
- transport.samples.kernels.cg_denseDense CG solver with cached transport plan matrix.
- transport.samples.kernels.cg_python_batchedPython-level batched CG for HVP acceleration.
- transport.samples.kernels.grad_sqeuclid
- transport.samples.kernels.streaming_sqeuclidSinkhorn OT with streaming softmax and shifted potentials.