cudaErrorMemoryAllocation during feasibiliy polishing #42
Description
Description
When running feasibility polishing on the world-shipping instance on H100, the solver reports a cudaErrorMemoryAllocation error. This suggests that the current feasibility-polishing step is using more GPU memory than expected.
We likely need to optimize or reduce memory usage in feasibility polishing for large-scale instances. @Lhongpei
Instance: https://www.oliverhinder.com/large-scale-lp-problems
Error log:
CUDA Error at cuPDLPx/src/solver.cu:1305: cudaErrorMemoryAllocation