Build against cuOpt 26.10 nightly: multi-GPU PDLP, new options, iteration/node reporting - #20

Open

0x17 wants to merge 19 commits into

mainfrom

multigpu-pdlp-nightly

Commits on Oct 2, 2026

Build against latest cuOpt nightly wheels from RAPIDS nightly index
0x17
committed
5b13f78
Support multi-GPU PDLP via num_gpus -1/>1 and add multigpu_pdlp_partitioner option
0x17
committed
9f2afcb
Adapt link and bundles to split cuOpt nightly libraries (libcuopt_mathopt/client, new cuDSS mtlayer)
0x17
committed
56556f9
Drop PDLP reduced-cost workaround, fixed upstream in cuOpt 26.10 (NVIDIA/cuopt#1797)
0x17
committed
b7b2871
Document multi-GPU PDLP usage in README
0x17
committed
ba7bdee
Remove options dropped in cuOpt 26.10 from optcuopt.def
0x17
committed
e5bfb6d
Make multi-GPU PDLP work: create ranged problem and skip initial solutions for multi-GPU PDLP
0x17
committed
990bd1f
Document multi-GPU PDLP limitations (no starting values, different convergence)
0x17
committed
235e7cc
Add primal simplex method and SeDuMi barrier initial point, fix changed default in optcuopt.def
0x17
committed
c7a7712
Add options introduced in cuOpt 26.10 to optcuopt.def
0x17
committed
9a14f86
Report iterations and nodes to GAMS via cuOpt solution attributes
0x17
committed
5db50e0
Pass GAMS nodlim to cuOpt node_limit
0x17
committed
1654974
Update README: system library requirements, manual gamsconfig setup, GPU arch per CUDA version, gamslib test [skip ci]
0x17
committed
5cdd066
Use a harder market split instance in statuses.gms so the time-limit case cannot end early [skip ci]
0x17
committed
cadddb7
Allow model-specific time limits in the gamslib baseline, give poutil 300s [skip ci]
0x17
committed
11e690c
Add mip_hyper_presolve_indicator_strengthening option (cuOpt #1983) [skip ci]
0x17
committed
21c682b

Commits on Oct 4, 2026

build: adapt cuOpt link packaging to modular 26.10 wheels
0x17
committed
8edd916
Revert "build: adapt cuOpt link packaging to modular 26.10 wheels"
0x17
committed
4855ada
ci: pin libcuopt to last pre-modular nightly 26.10.0a223
0x17
committed
ba44887

Navigation Menu

Build against cuOpt 26.10 nightly: multi-GPU PDLP, new options, iteration/node reporting - #20

Build against cuOpt 26.10 nightly: multi-GPU PDLP, new options, iteration/node reporting#20
0x17 wants to merge 19 commits into
mainfrom
multigpu-pdlp-nightly

Commits on Oct 2, 2026

Build against latest cuOpt nightly wheels from RAPIDS nightly index

Support multi-GPU PDLP via num_gpus -1/>1 and add multigpu_pdlp_partitioner option

Adapt link and bundles to split cuOpt nightly libraries (libcuopt_mathopt/client, new cuDSS mtlayer)

Drop PDLP reduced-cost workaround, fixed upstream in cuOpt 26.10 (NVIDIA/cuopt#1797)

Document multi-GPU PDLP usage in README

Remove options dropped in cuOpt 26.10 from optcuopt.def

Make multi-GPU PDLP work: create ranged problem and skip initial solutions for multi-GPU PDLP

Document multi-GPU PDLP limitations (no starting values, different convergence)

Add primal simplex method and SeDuMi barrier initial point, fix changed default in optcuopt.def

Add options introduced in cuOpt 26.10 to optcuopt.def

Report iterations and nodes to GAMS via cuOpt solution attributes

Pass GAMS nodlim to cuOpt node_limit

Update README: system library requirements, manual gamsconfig setup, GPU arch per CUDA version, gamslib test [skip ci]

Use a harder market split instance in statuses.gms so the time-limit case cannot end early [skip ci]

Allow model-specific time limits in the gamslib baseline, give poutil 300s [skip ci]

Add mip_hyper_presolve_indicator_strengthening option (cuOpt #1983) [skip ci]

Commits on Oct 4, 2026

build: adapt cuOpt link packaging to modular 26.10 wheels

Revert "build: adapt cuOpt link packaging to modular 26.10 wheels"

ci: pin libcuopt to last pre-modular nightly 26.10.0a223