NVIDIA/cuopt v26.10.00 on GitHub

What's Changed

🚨 Breaking Changes

build: don't install libcuopt into conda env by default by @ramakrishnap-nv in #1650
Presolve work and iteration limits by @akifcorduk in #1656
refactor: make to_optimization_problem a free function by @ramakrishnap-nv in #1802
Migrate stream APIs from rmm::cuda_stream_view to cuda::stream_ref by @bdice in #1828
build(cmake): split libcuopt into cuopt_client / cuopt_mathopt / cuopt_routing component libs by @ramakrishnap-nv in #1622
build(cmake): give each component its own install component by @ramakrishnap-nv in #1926
modify docker image to run proxy+grpc by default by @tmckayus in #2007
build(wheel): make libcuopt a thin metapackage by @ramakrishnap-nv in #2013

🐛 Bug Fixes

GF2 presolve rank-deficiency fix by @aliceb-nv in #1660
Fix crash when symmetry generators are pruned by @nguidotti in #1689
[BUG] Fixed uninitialized FJ constraint weights when the problem grows by @nguidotti in #1698
Use configured feasibility tolerance in Papilo by @hlinsen in #1688
Fix GF2 incorrectly treating some one-sided rows as equality by @aliceb-nv in #1707
Fix dual residual computation in final solution output by @yuwenchen95 in #1718
Remove split compile by default by @Kh4ster in #1722
Report absent solution values instead of returning success by @ramakrishnap-nv in #1734
fix: avoid uncapturable RMM copies inside routing CUDA graph captures by @ramakrishnap-nv in #1753
fix: keep host sources alive for async route-view copies (#1756) by @ramakrishnap-nv in #1759
Revert changes to PCG by @nguidotti in #1762
Fix QCQP dual solution overrun and constraint count reporting by @yuwenchen95 in #1761
fix: build-images CI failures (deadsnakes PPA flakiness + nightly version mismatch) by @ramakrishnap-nv in #1780
Fix mod2 integrality tolerance by @akifcorduk in #1770
Fix cuopt_assert macro argument expansion by @aliceb-nv in #1784
Fix device_scalar constructors after RMM change by @aliceb-nv in #1795
ci: pass the build datetime through to image builds by @ramakrishnap-nv in #1792
Correctly return reduced costs for PDLP (stable3) by @mlubin in #1797
Add missing TransactionGuard in the SLDA presolver by @aliceb-nv in #1815
Fix cuopt_mvn retry loop being silently skipped under set -e by @ramakrishnap-nv in #1823
Route Java console logging through System.out instead of native stdout by @ramakrishnap-nv in #1825
Fix asynchronous clique table handoff by @hlinsen in #1834
Disable clique/zero-half cuts in determinism mode by @nguidotti in #1841
Forward routing gRPC settings and map node name strings to ints by @tmckayus in #1837
fix(build): pin RAPIDS_BRANCH and shared-workflows to release/26.10 by @ramakrishnap-nv in #1870
change Python grpc client so wait loop happens in Python by @tmckayus in #1857
Concurrent halt fix on barrier by @akifcorduk in #1810
fix(build): repair cuda::stream_ref migration breakage from #1828 by @ramakrishnap-nv in #1894
fix(docker): keep libgomp1 in the Ubuntu runtime image (cuDSS threading layer, #1670) by @vgong-nv in #1888
ci(thirdparty-cvxpy): fix cwd import-shadowing that masked the installed wheel by @ramakrishnap-nv in #1891
Routing: avoid division by zero when vehicle capacity is zero by @vitor1001 in #1853
Guard against nullptr dereference when checking concurrent solver status by @vitor1001 in #1845
Routing: Guard diagonal element lookup against negative offsets in top_k by @vitor1001 in #1855
Routing: Fix vector lifetimes in breaks unit test by @vitor1001 in #1854
docs: update NVIDIA Brev documentation links by @ZedingZhang in #1925
Pin cuDSS's dlopen'd threading-layer plugin to prevent a teardown race by @ramakrishnap-nv in #1883
Fix a few CI issues by @KyleFromNVIDIA in #1927
presolve: refresh the cached linear column count after removing empty… by @yuwenchen95 in #1923
Fix issue in dual simplex where cut off used perturbed objective; use original objective and require dual feasibility by @chris-maes in #1901
grpc server align ping handling with client behavior by @tmckayus in #1937
Fix run mip tolerance by @akifcorduk in #1956
wire MIP initial solutions into gRPC by @tmckayus in #1868
[BUG] Correctly stop sub-MIP when main solve finishes by @nguidotti in #1911
refactor(routing): stop cpu_routing_problem.hpp from leaking raft/rmm by @ramakrishnap-nv in #1969
Fix Java JNI build on Ubuntu 22.04, add a Java smoke test to image CI by @ramakrishnap-nv in #1974
Recompute dual variables and check dual feasibility after basis repair by @chris-maes in #1965
fix(grpc): regenerate cuopt_mcp_schema.json, fail codegen drift earlier in CI by @ramakrishnap-nv in #1980
proxy server make msgpack the default for wildcard ACCEPT by @tmckayus in #1964
Fix false optimality declaration in the determinsitic mode by @aliceb-nv in #1954
fix: escape parser error messages by @neuriv in #1947
bugfix: handle cuda graph errors by @neuriv in #1945
proxy: always include warnings and notes on result-file stubs by @tmckayus in #1990
proxy: DELETE of an unknown request returns 200 with zero counts by @tmckayus in #1991
proxy: say file vs stream in optimization-data 422s by @tmckayus in #1992
fix(ci/docker): manifest --amend duplicate entries + unused gdb-gdbserver by @ramakrishnap-nv in #1997
proxy server match legacy behavior on missing logs by @tmckayus in #1989
track and return solve time for VRP solves in the gRPC server to support proxy server parity with legacy by @tmckayus in #1985
proxy: use a stable job 404 and stop UUID-gating job routes for parity with legacy by @tmckayus in #1987
Restore parameter logging by @nguidotti in #1996
Bug fix: barrier determinism by @yuwenchen95 in #1856
Fix concurrent clique table publication by @derluke in #1775
Fix solution queue memory overflow by @akifcorduk in #2002
fix(routing): declare cusparse, which libcuopt_routing.so links by @ramakrishnap-nv in #2008
fix(cmake): ship constants.h with the component that needs it by @ramakrishnap-nv in #2016
[BUG] Fix delayed termination during root loop by @nguidotti in #2003
[BUG] Fix long interval between progress report by @nguidotti in #2019
Fix ranged constraint conversion and MPS round trips by @rg20 in #2027
Fix erroneous early solve termination by @aliceb-nv in #2037
Fix/mip time limit overruns by @akifcorduk in #2046
grpc pdlp warmstart data was not transmitted and read by the server by @tmckayus in #2041
fix(wheel): declare libraft, whose headers the public API needs by @ramakrishnap-nv in #2014
Reject unsupported mixed-integer quadratic problems before solving by @rg20 in #2032
Bug fix: Insert barrier slack columns in front of the cone block for range rows by @yuwenchen95 in #2029
Fix MPS writing for empty CSR constraint matrices by @mlubin in #2039
fix(ci): pin remaining shared-workflows refs to release/26.10 by @ramakrishnap-nv in #2040
Fix knapsack rational coefficient scaling by @aliceb-nv in #1754
fix(test): update grpc_integration_test to submit_mip's new solver_settings_t signature by @ramakrishnap-nv in #2055
fix(wheel): symlink component VERSION files to the repo-root VERSION by @ramakrishnap-nv in #2061
fix(docker): build the Java JNI library against the split libcuopt wheels by @ramakrishnap-nv in #2062
grpc server add a check for successful cuda init in worker by @tmckayus in #2063
Fix continuous variable elimination tolerance in MIP presolve by @chris-maes in #2066
Add extra termination checks in LP and submip by @aliceb-nv in #2054
fix(docker): install the unsuffixed nvidia-cuda-cccl for every CUDA major by @ramakrishnap-nv in #2075
Dual simplex objective logging fixes by @chris-maes in #2072
grpc and proxy server, return the bound value for incumbents by @tmckayus in #2079
skip job cancellation tests temporarily by @tmckayus in #2081
Fix scale fallback direction in dual simplex by @chris-maes in #2080
Fix FJ score alignment for CUDA atomic loads by @mlubin in #2088
fix(linear_programming): implement neg, pos, and abs for LinearExpression (fixes #2056) by @reginaldalfret in #2068
FIX : respecting project_initial_primal in mPDLP by @Bubullzz in #2050
Fix rationalize coefficient bug and too conservative knapsack separation for fractional capacity by @chris-maes in #2083
grpc switch worker creation from fork to posix spawn with exec by @tmckayus in #2085
Local patch for papilo by @mlubin in #2102
Keep presolve enabled for MIP set-solution callbacks by @akifcorduk in #2098
fix(packaging): make single-component installs consumable from CMake by @ramakrishnap-nv in #2094
build.sh: find libcuopt build tree for the cuopt Python build without --install by @ramakrishnap-nv in #2120
Downgrade PSLP to v0.1.0 by @rg20 in #2116
Fix leak of stale duals in compute_dual_solution_from_basis by @mlubin in #2124
Fix duplicate substitution records emitted in presolve probing by @aliceb-nv in #2126
ci: skip the PuLP nightly step on Python < 3.12 by @ramakrishnap-nv in #2138
fix(mip): ask every early heuristic to stop before waiting for any by @ramakrishnap-nv in #2139

📖 Documentation

docs: enforce heading title case with Vale by @ramakrishnap-nv in #1668
Show mamba commands directly instead of a conda-then-swap note by @ramakrishnap-nv in #1827
docs(grpc): document the VRP gRPC client (RoutingClient) by @ramakrishnap-nv in #1838
docs: make Java bindings discoverable, documented, and installable via the selector by @ramakrishnap-nv in #1999
docs build server spec from proxy server by @tmckayus in #2084
cuopt server documentation update by @tmckayus in #2115
docs: Release Notes 26.10 and missing settings docs by @ramakrishnap-nv in #1972

🚀 New Features

Upgrade cmake version to include the PDLP objective offset fix by @Kh4ster in #1695
Expose solver statistics as scalar solution attributes by @ramakrishnap-nv in #1715
Add a solver setting for controlling adaptive regularization in barrier by @rg20 in #1743
Java bindings for LP, MIP and QP by @nvidiacbrissette in #1524
move server LPData transform into a standalone function by @tmckayus in #1848
Extract LP create_data_model/create_solver into conversion.py by @tmckayus in #1849
extract routing server data creation routines into standalone module by @tmckayus in #1850
move server data compression handling into a standalone module by @tmckayus in #1851
extract result file support logic to be shared by proxy by @tmckayus in #1869
Move deprecated server modules into utils/deprecated. by @tmckayus in #1874
move cuopt server job_queue and worker runtime into deprecated utils by @tmckayus in #1876
routing: add distance breaks by @np96 in #1196
Keep permanent utils from importing the legacy HTTP runtime. by @tmckayus in #1879
add cancel and status for grpc routing client by @tmckayus in #1886
add basic http proxy server that forwards to grpc by @tmckayus in #1892
Add solver caching to support re-solves for barrier QP by @Iroy30 in #1821
proxy server add VRP request and solution over gRPC by @tmckayus in #1917
proxy server make the health check dependent on live grpc server by @tmckayus in #1919
proxy server add pdlp warmstart support to match legacy by @tmckayus in #1918
add cuopt/cuopt sync endpoint to the proxy server for backward compatability by @tmckayus in #1933
Expose distributed (multi-GPU) PDLP settings in the Java API by @ramakrishnap-nv in #1961
Implement Curtis Reid Scaling by @rg20 in #1934
proxy server implement incumbent_set_solutions by @tmckayus in #1971
Fix logging bug when asking for more GPUs than available by @Bubullzz in #1935
change CI to test proxy+grpc instead of legacy http by @tmckayus in #1993
grpc support a map<string, string> for parameters in the server by @tmckayus in #2010
Add C API support for multiGPU PDLP by @Bubullzz in #1958
grpc modify the client to use the new parameter map<string, string> by @tmckayus in #2022
Distributed curtis reid by @Bubullzz in #2004
grpc add a gpu health check before each job or after idle by @tmckayus in #2031
standardize interface and remote solves on solver_settings_t by @tmckayus in #2034
grpc add standard grpc health problems and fix up helm chart by @tmckayus in #2047
grpc for the bundled client with new server handle only new parameter map by @tmckayus in #2057
proxy server add legacy openapi spec endpoint for parity by @tmckayus in #2077
Fix multi-GPU PDLP dispatch in the gRPC server worker by @ramakrishnap-nv in #1962
Implement column reduction for PDLP for large problems by @rg20 in #2067

🛠️ Improvements

ci: sync skill-card.md versions in update-version flow by @ramakrishnap-nv in #1669
fix: add statuses: read permission to request-nvskills-ci workflow by @ramakrishnap-nv in #1681
docs: fix Vale heading case errors in gRPC async client docs by @ramakrishnap-nv in #1682
Symbol visibility controls and exports (remove-internal approach) by @ramakrishnap-nv in #1625
ensure nightly builds always produce new packages by @jameslamb in #1661
ci: update NVSkills CI request template and add require-nvskills-status workflow by @ramakrishnap-nv in #1690
chore: trigger NVSkills CI for cuopt-multi-objective-exploration by @ramakrishnap-nv in #1673
chore: trigger NVSkills CI for cuopt-numerical-optimization-formulation by @ramakrishnap-nv in #1675
chore: trigger NVSkills CI for cuopt-numerical-optimization-api by @ramakrishnap-nv in #1674
Add iterative refinement to the GPU ADAT solve path by @rg20 in #1680
Update to rapids-logger 0.3 by @bdice in #1659
ci: make pr-test-summary and NVSkills CI checks non-blocking by @ramakrishnap-nv in #1696
chore: trigger NVSkills CI for cuopt-routing-api-python by @ramakrishnap-nv in #1676
chore: trigger NVSkills CI for cuopt-user-rules by @ramakrishnap-nv in #1679
Add a hyperparameter that controls bounding of free variables by @yuwenchen95 in #1713
Print total setup time by @yuwenchen95 in #1714
make 'vale' check optional locally by @jameslamb in #1708
fix: exec the binary from console-script wrappers instead of spawning it by @ramakrishnap-nv in #1723
Update cpp/benchmarks codeowners to cuopt-solver-codeowners by @ramakrishnap-nv in #1742
Release pslp presolve memory by @hlinsen in #1735
wheels: avoid vendoring libnvJitLink by @jameslamb in #1741
Build cuopt as an abi3 wheel using the CPython Limited API by @ramakrishnap-nv in #1731
Routing over gRPC: VRP server + compiled C++/Cython client by @ramakrishnap-nv in #1597
test: isolate library wheel smoke test by @bdice in #1624
Build conda python packages independently of the Python version by @ramakrishnap-nv in #1744
Mod-2 cut separator for zero-half cuts by @akifcorduk in #1726
Revert presolve tolerance by @akifcorduk in #1757
Block Bounded Value Elimination presolving pass for binary problems by @aliceb-nv in #1683
Add a binary 2opt move fed from the probing cache into CPUFJ by @aliceb-nv in #1738
Reuse existing cuOptGetProblemIntAttribute to simplify code in c API by @yuwenchen95 in #1769
ci: stop nightly builds for release/26.08 by @ramakrishnap-nv in #1777
fix: give each solver its own seed instead of a process-wide counter by @ramakrishnap-nv in #1717
Root heuristics by @nguidotti in #1700
RENS by @nguidotti in #1719
chore: trigger NVSkills CI for cuopt-install by @ramakrishnap-nv in #1672
Add Kitmaker Portal wheel publishing via Charon Ferry by @ramakrishnap-nv in #1763
ci: report step timeouts explicitly, and skip the test that was hitting them by @ramakrishnap-nv in #1773
Add a dual aggregation presolver for binary variables with a single lock by @aliceb-nv in #901
B&B Halt by @nguidotti in #1760
Resolve NVSkills Tier 1 security findings in cuopt-server-api-python by @ramakrishnap-nv in #1677
Add parameters for tuning barrier primal and dual regularization by @yuwenchen95 in #1807
workaround for signal during grpc test framework teardown by @tmckayus in #1691
Make MIP root relaxation honor LP method by @hlinsen in #1829
Python api performance improvement by @Iroy30 in #1615
Add more strategies of init point computation in barrier by @yuwenchen95 in #1577
Remove unused cusolver/curand deps from libcuopt by @ramakrishnap-nv in #1830
Reduce GPU/host synchronization overhead in barrier termination check by @yuwenchen95 in #1808
refactor(logger): give each component library its own logger by @ramakrishnap-nv in #1778
Fix vendored PSLP writing a stray infeasible-message printf to native stdout by @ramakrishnap-nv in #1836
refactor: split host-only members out of CUDA translation units by @ramakrishnap-nv in #1801
Exhaustive coefficient reduction for binary rows by @aliceb-nv in #1686
Heuristic for arc-flow formulations by @aliceb-nv in #1833
RoutingClient: match Client's (host, port, tls=...) signature by @ramakrishnap-nv in #1840
Tune Maven's own retry/concurrency to reduce Maven Central rate-limiting by @ramakrishnap-nv in #1835
feat(c-api): add cuOptSetLogCallback by @ramakrishnap-nv in #1636
Diving on cut passes by @nguidotti in #1812
Replace process-wide seed_generator with per-worker RNGs in MIP heuristics by @ramakrishnap-nv in #1809
Replaced cusparse wrappers with simple unique_ptrs and more RAII by @Bubullzz in #1342
Set Papilo back to the main branch by @nguidotti in #1694
Adopt CUDA stream compatibility accessors by @bdice in #1858
Bump cuopt-java compile target from Java 11 to Java 17 by @ramakrishnap-nv in #1865
add a smoke test to ensure server images start correctly by @tmckayus in #1646
refactor: keep the GPU warm-start path out of host-only translation units by @ramakrishnap-nv in #1803
update variable type by @Iroy30 in #1783
build: add the CUDA-free cuopt_client library by @ramakrishnap-nv in #1804
Improve duplicate cut detection by @hlinsen in #1764
refactor: build the routing gRPC arm into cuopt_client by @ramakrishnap-nv in #1884
Improvement for qcqp canonicalization and symmetric check by @yuwenchen95 in #1864
Presolve: remove empty columns with a diagonal quadratic term by @rg20 in #1885
simplify test for incumbent solution arrival during wait by @tmckayus in #1898
[CPUFJ PR 1] Improve time-to-first-incumbent, unify solution callback publication, scaffolding changes by @aliceb-nv in #1880
chore(skills): trigger benchmark regeneration for cuopt-server-api-python by @ramakrishnap-nv in #1895
Don't abort the process when cudssSetThreadingLayer fails by @ramakrishnap-nv in #1897
[CPUFJ PR2] Add a fast-path to the CPUFJ engine for binary problems with integer activities by @aliceb-nv in #1881
fix(test): stop DeleteQueuedJobPreventsRun/DeleteRunningJobCancelsWorker from inheriting worker respawn debt by @ramakrishnap-nv in #1877
add a proxy server CPU path for routing conversion to DataModel by @tmckayus in #1910
Speedup barrier second-order cone conversion in general path by @yuwenchen95 in #1914
test: skip incumbent-callback tests until #1909 lands by @ramakrishnap-nv in #1921
Fix compilation with Clang / LLVM by @vitor1001 in #1843
refactor(routing): standardize cost terminology by @Juan-Francisco-Robles in #1299
Describe solver settings and defaults in the gRPC codegen field registry by @ramakrishnap-nv in #1701
build(cython): link each extension module against the component it uses by @ramakrishnap-nv in #1920
Update Doxygen to 1.18.0 by @vuule in #1900
Give cuDSS a threading layer built against cuOpt's own libgomp (wheels) by @ramakrishnap-nv in #1905
Use new verify-dependencies hook from pre-commit-hooks by @KyleFromNVIDIA in #1915
Handle PSLP false-infeasible on badly-scaled LPs by @rg20 in #1922
Clarify factorization issue causes residual error by @yuwenchen95 in #1866
build(conda): split libcuopt into per-component packages by @ramakrishnap-nv in #1928
cuopt_mcp: add an MCP server for LP/MILP solving over gRPC by @ramakrishnap-nv in #1916
Explore self-contained Java classifier JARs with a statically linked libcuopt by @ramakrishnap-nv in #1818
Enable all NVCC warnings as errors by @mlubin in #1948
build(cmake): stop compiling routing/mathopt sources twice by @ramakrishnap-nv in #1960
Scale the variables with variable bounds by @rg20 in #1652
fix(client): hold the GPU variant alternatives behind an opaque pointer by @ramakrishnap-nv in #1940
Fusing Halpern Kernels for performances by @Bubullzz in #1904
Build the Java bindings dynamically against the pip-installed libcuopt in the official Docker images by @ramakrishnap-nv in #1959
codeowners: route gRPC paths under cpp/ to infra team by @ramakrishnap-nv in #1975
build(wheel): split libcuopt into per-component wheels by @ramakrishnap-nv in #1929
Update PSLP to v0.0.12 by @hlinsen in #1943
Add NNZ cutoff for concurrent mode by @hlinsen in #1932
[CPUFJ PR3] Modular CPUFJ engine, simple portfolio by @aliceb-nv in #1950
cuopt-developer skill: sanitize internal context before it becomes public by @ramakrishnap-nv in #1988
Primal and dual simplex improvements by @chris-maes in #1889
Eliminate zero-cost free variables by equality substitution in QP/SOCP barrier presolve by @rg20 in #1912
Fix O(n^2) loop in the flow cover cut generation by @aliceb-nv in #1842
build(wheel): drop the CUDA suffix and dimension from libcuopt-client by @ramakrishnap-nv in #1967
fix(barrier): resolve cuDSS mtlayer path at runtime by @ramakrishnap-nv in #1982
[CPUFJ PR4] Structure-aware starts and repairs for CPUFJ by @aliceb-nv in #1951
Rename CUOPT_DISTRIBUTED_PDLP_PARTITIONER to CUOPT_MULTIGPU_PDLP_PARTITIONER; remove CUOPT_USE_DISTRIBUTED_PDLP by @ramakrishnap-nv in #1984
Include objective gap in QP termination criteria by @rg20 in #1733
[Java] Remove setMethod/setPDLPSolverMode/setNumGpus/setMpdlpPartitioner wrappers by @ramakrishnap-nv in #2000
[CPUFJ PR5] Portfolio diversification, more starts, move batching by @aliceb-nv in #1953
Move Ruiz scaling and building KKT system to the GPU by @yuwenchen95 in #1913
Mutation by @nguidotti in #2001
grpc integration test tighten up warmup job handling by @tmckayus in #2012
Indicator Strengthening by @nguidotti in #1983
fixing mPDLP logs, adding header back by @Bubullzz in #2017
perf(wheel): skip gRPC/abseil build for mathopt and routing wheels by @ramakrishnap-nv in #2035
Replace CUB umbrella header with specific includes by @bdice in #2033
Remove remaining rmm::cuda_stream_view usage by @bdice in #2030
ci: gate wheel publish on the whole libcuopt/cuopt family building by @ramakrishnap-nv in #2051
perf(wheel): skip mathopt/routing/grpc_server when building libcuopt-client by @ramakrishnap-nv in #2043
Add update apis - RHS by @Iroy30 in #1941
ci: remove issue auto-labeling, round-robin assignment, and inactivity reminder by @ramakrishnap-nv in #2053
chore(release): add manual script to publish placeholder release images by @ramakrishnap-nv in #2038
fix(conda): declare libraft-headers, and stop routing inheriting nccl by @ramakrishnap-nv in #2060
ci: use proxy cache for pypi.anaconda.org, remove pip indices from dependencies.yaml by @jameslamb in #2064
fix(conda): drop nccl and libnvjitlink from the client's run exports by @ramakrishnap-nv in #2070
Flag to expose symbols by @nguidotti in #2074
ci: bump C++ gtest step timeout to 90 minutes by @ramakrishnap-nv in #2087
grpc follow up to #2063 by @tmckayus in #2071
Remove Papilo unconditional NDEBUG definition by @aliceb-nv in #2078
Wire cuOpt Java bindings into the shared maven-publish workflow by @ramakrishnap-nv in #1970
Reduce PDLP memory use and avoid warm-start copies by @rg20 in #2065
fix(ci): report crashed gtest binaries in pr-test-summary by @ramakrishnap-nv in #2086
Update PSLP to v0.1.1 by @rg20 in #2092
Fix bug #2042 absence of logs on mPDLP error with C API and not using rhs when present by @Bubullzz in #2091
Propagate primal simplex to APIs. Ensure consistency with dual simplex. by @chris-maes in #2096
[CPUFJ PR6] CDCL heuristic start for CPUFJ by @aliceb-nv in #2049
Persistent LNS ruin-repair search by @akifcorduk in #2058
update code base to replace occurences of "distributed PDLP" with multi-GPU PDLP by @Bubullzz in #2076
Reduce barrier refinement allocations and launch overhead by @rg20 in #2069
Update APIs - SOCP support by @Iroy30 in #1979
Exploit degeneracy in MIP by @chris-maes in #1685
Fix nightly build: barrier_cache_t template argument in SOCP barrier test by @ramakrishnap-nv in #2110
Map LP initial solutions through PSLP presolve for PDLP by @rg20 in #2024
Fix excessive setups and low diversity in local search CPUFJ calls by @aliceb-nv in #2105
Skip duplicate banner print in sequential solve by @yuwenchen95 in #2109
add a cli toggle for PDLP hyper parameter project_initial_primal by @Bubullzz in #2089
ci: use shared NVSkills team-request workflow to support fork PRs by @ramakrishnap-nv in #2100
CPUFJ: no per-move objective rebuild while infeasible and add halt checks in the epigraph projection loop by @hlinsen in #2117
Preserve recoverable barrier iterates with converged objective gaps by @rg20 in #2103
add a C API test to ensure mPDLP works with cuOptCreateProblem from RHS by @Bubullzz in #2108
Limit column reduction log to first activation by @chris-maes in #2125
ci(thirdparty-cvxpy): test the cvxpy release when no C++ compiler is available by @nicoloangileri in #2101

New Contributors

@nvidiacbrissette made their first contribution in #1524
@vgong-nv made their first contribution in #1888
@Juan-Francisco-Robles made their first contribution in #1299
@ZedingZhang made their first contribution in #1925
@vuule made their first contribution in #1900
@neuriv made their first contribution in #1947
@derluke made their first contribution in #1775
@reginaldalfret made their first contribution in #2068
@nicoloangileri made their first contribution in #2101

Full Changelog: v26.10.00a...v26.10.00