Add a lazy Triton CUDA backend for Linux NVIDIA GPUs with compute capability 8.0 or newer and Torch 2.4 or newer.
Add exact SciPy parity, dtype, autocast, non-finite input, non-default stream, and large-batch regression coverage for the Triton solver.
Add reproducible author-run GPU validation and JSONL benchmark reporting with paired SciPy and legacy 0.0.6 comparisons.

Ship a pure-source package by default; normal installation no longer compiles the legacy CUDA extension.
Require Python 3.10 or newer.
Preserve float64 solver precision, promote float16, bfloat16, and integer inputs to float32, and match SciPy's infeasible and invalid-cost behavior.
Fall back to SciPy on unsupported CUDA runtimes and return the assignment to the input device.

Existing batch_linear_assignment and assignment_to_indices calls do not require API changes.
Users who require the compiled CUDA implementation, including execution on T4 (sm_75), should remain on the maintenance-only 0.0.x release line. Release 0.0.6 can be installed with python -m pip install "torch-linear-assignment==0.0.6".
The Triton path requires Linux, an NVIDIA GPU with compute capability 8.0 or newer, Torch 2.4 or newer, and Triton. Other supported environments continue through SciPy.

Exact assignment parity and all GPU tests pass on NVIDIA L4 (sm_89), A100 (sm_80), and RTX PRO 6000 Blackwell (sm_120) author-run environments.
H100 validation and isolated fresh-process cold-start measurements remain pending. Published warm timings are paired with same-machine SciPy and legacy baselines; they are not generalized across hardware.

Authors

Navigation Menu