[NUMERICAL_ISSUES] Documented 3-variable convex QP returns NumericalError in cuDSS symbolic analysis #1861

Describe the numerical issue

The documented three-variable continuous convex portfolio QP returns NumericalError on a pip/uv-installed cuOpt 26.2.0 environment. The runtime prints a cuDSS CSR creation failure during symbolic analysis and all variable values remain zero.

This is the same model shape shown in the maintained QP example: a PSD quadratic objective, three bounded continuous variables, one budget equality, and one minimum-return inequality.

Steps / minimal reproduction

from cuopt.linear_programming.problem import CONTINUOUS, MINIMIZE, Problem
from cuopt.linear_programming.solver_settings import SolverSettings

problem = Problem("documented_portfolio_qp")
x = problem.addVariable(lb=0, ub=1, vtype=CONTINUOUS, name="stock_a")
y = problem.addVariable(lb=0, ub=1, vtype=CONTINUOUS, name="stock_b")
z = problem.addVariable(lb=0, ub=1, vtype=CONTINUOUS, name="stock_c")

problem.setObjective(
    0.04*x*x + 0.02*y*y + 0.01*z*z
    + 0.02*x*y + 0.01*x*z + 0.016*y*z,
    sense=MINIMIZE,
)
problem.addConstraint(x + y + z == 1, name="budget")
problem.addConstraint(0.12*x + 0.08*y + 0.05*z >= 0.08, name="min_return")

settings = SolverSettings()
settings.set_parameter("time_limit", 60)
problem.solve(settings)
print(problem.Status.name, x.getValue(), y.getValue(), z.getValue())

Observed output:

FAILED: CUDSS call ended unsuccessfully with status = 3, details: "cudssMatrixCreateCsr"
Error in symbolic analysis
Barrier Solve status A numerical error was encountered.
NumericalError 0.0 0.0 0.0

Expected vs actual behavior

Expected: the feasible convex QP returns Optimal or another documented usable primal status with nonzero weights satisfying the constraints.

Actual: NumericalError; cuDSS CSR creation fails during symbolic analysis; values remain all zero. Status gating is essential because the zero vector is not a feasible solution.

Environment details

• Environment location: bare-metal remote Linux GPU host
• Install: cuopt-cu12 wheel installed with uv
• OS: Ubuntu 22.04, Linux 5.15
• Python: 3.12.14
• GPU: NVIDIA A100-PCIE-40GB
• Driver: 580.173.02
• CUDA bindings: 12.9.7
• cuopt-cu12: 26.2.0
• nvidia-cudss-cu12: 0.8.0.10
• NumPy: 2.2.6

Additional context

The same runtime solves LP and MILP examples successfully. This blocks a continuous-QP subproblem in an otherwise supported active-set decomposition. A useful resolution would include a supported cuOpt/cuDSS package matrix and this documented QP as an installation smoke test.