[Bridges] throw GetAttributeNotAllowed in ZerosBridge instead of general error

[joaquimg]

import HiGHS
import MathOptInterface as MOI

# works, of course
# model = MOI.instantiate(HiGHS.Optimizer, with_bridge_type = Float64)

# fails
opt = () -> MOI.instantiate(HiGHS.Optimizer, with_bridge_type = Float64)
model = MOI.instantiate(opt, with_cache_type = Float64)

s = MOI.add_variable(model)

vs = MOI.add_constraint(model, MOI.VectorOfVariables([s]), MOI.Zeros(1))

MOI.optimize!(model)

MOI.get(model, MOI.ConstraintDual(), vs)

ERROR: Unable to query the dual of a variable bound that was reformulated using `ZerosBridge`. This usually arises in conic solvers when a variable is fixed to a value. As a work-around, instead of creating a fixed variable using variable bounds like `p == 1`, add an affine equality constraint like `1 * p == 1` (or `[1 * p - 1,] in Zeros(1)`).
Stacktrace:
  [1] error(s::String)
    @ Base ./error.jl:35
  [2] get(::MathOptInterface.Bridges.LazyBridgeOptimizer{…}, ::MathOptInterface.ConstraintDual, ::MathOptInterface.Bridges.Variable.ZerosBridge{…})
    @ MathOptInterface.Bridges.Variable ~/.julia/dev/MathOptInterface/src/Bridges/Variable/bridges/ZerosBridge.jl:96
  [3] (::MathOptInterface.Bridges.var"#3#4"{…})(bridge::MathOptInterface.Bridges.Variable.ZerosBridge{…})
    @ MathOptInterface.Bridges ~/.julia/dev/MathOptInterface/src/Bridges/bridge_optimizer.jl:324
  [4] (::MathOptInterface.Bridges.Variable.var"#21#22"{MathOptInterface.Bridges.Variable.Map, MathOptInterface.Bridges.var"#3#4"{…}, Int64})()
    @ MathOptInterface.Bridges.Variable ~/.julia/dev/MathOptInterface/src/Bridges/Variable/map.jl:642
  [5] call_in_context(map::MathOptInterface.Bridges.Variable.Map, bridge_index::Int64, f::MathOptInterface.Bridges.Variable.var"#21#22"{…})
    @ MathOptInterface.Bridges.Variable ~/.julia/dev/MathOptInterface/src/Bridges/Variable/map.jl:627
  [6] call_in_context
    @ ~/.julia/dev/MathOptInterface/src/Bridges/Variable/map.jl:642 [inlined]
  [7] call_in_context
    @ ~/.julia/dev/MathOptInterface/src/Bridges/bridge_optimizer.jl:284 [inlined]
  [8] call_in_context(b::MathOptInterface.Bridges.LazyBridgeOptimizer{…}, ci::MathOptInterface.ConstraintIndex{…}, f::MathOptInterface.Bridges.var"#3#4"{…})
    @ MathOptInterface.Bridges ~/.julia/dev/MathOptInterface/src/Bridges/bridge_optimizer.jl:304
  [9] call_in_context
    @ ~/.julia/dev/MathOptInterface/src/Bridges/bridge_optimizer.jl:321 [inlined]
 [10] get(b::MathOptInterface.Bridges.LazyBridgeOptimizer{…}, attr::MathOptInterface.ConstraintDual, ci::MathOptInterface.ConstraintIndex{…})
    @ MathOptInterface.Bridges ~/.julia/dev/MathOptInterface/src/Bridges/bridge_optimizer.jl:1646
 [11] _get_fallback(model::MathOptInterface.Utilities.CachingOptimizer{…}, attr::MathOptInterface.ConstraintDual, index::MathOptInterface.ConstraintIndex{…})
    @ MathOptInterface.Utilities ~/.julia/dev/MathOptInterface/src/Utilities/cachingoptimizer.jl:1004
 [12] get(model::MathOptInterface.Utilities.CachingOptimizer{…}, attr::MathOptInterface.ConstraintDual, index::MathOptInterface.ConstraintIndex{…})
    @ MathOptInterface.Utilities ~/.julia/dev/MathOptInterface/src/Utilities/cachingoptimizer.jl:949
 [13] top-level scope
    @ REPL[361]:1
Some type information was truncated. Use `show(err)` to see complete types

Why would someone do it?

In DiffOpt we need:

1. bridges to be general and accept a wider range of models
2. copy to move the model to a matrix form used in differentiation

However, some bridges do not support geting ConstraintFunction, so we need the cache outside.

[joaquimg]

This came originally from test_conic_linear_VectorOfVariables_2

[odow]

This is expected behaviour.

vs = MOI.add_constraint(model, MOI.VectorOfVariables([s]), MOI.Zeros(1))

Why you adding this constraint?

The issue is that there is a ZerosBridge which substitutes out s with 0.0. So the variable isn't in the model anymore and we don't have a constraint dual to return.

[odow]

I have periodically raised this with @blegat. I would be in favour of removing the ZerosBridge being included by default. It's just annoying.

[odow]

The solution is this, but it's a bit clunky:

julia> import HiGHS

julia> import MathOptInterface as MOI

julia> model = MOI.instantiate(; with_cache_type = Float64) do
           inner = MOI.instantiate(HiGHS.Optimizer, with_bridge_type = Float64)
           MOI.Bridges.remove_bridge(inner, MOI.Bridges.Variable.ZerosBridge{Float64})
           return inner
       end
MOIU.CachingOptimizer
├ state: EMPTY_OPTIMIZER
├ mode: AUTOMATIC
├ model_cache: MOIU.UniversalFallback{MOIU.Model{Float64}}
│ ├ ObjectiveSense: FEASIBILITY_SENSE
│ ├ ObjectiveFunctionType: MOI.ScalarAffineFunction{Float64}
│ ├ NumberOfVariables: 0
│ └ NumberOfConstraints: 0
└ optimizer: MOIB.LazyBridgeOptimizer{HiGHS.Optimizer}
  ├ Variable bridges: none
  ├ Constraint bridges: none
  ├ Objective bridges: none
  └ model: A HiGHS model with 0 columns and 0 rows.

julia> s = MOI.add_variable(model)
MOI.VariableIndex(1)

julia> vs = MOI.add_constraint(model, MOI.VectorOfVariables([s]), MOI.Zeros(1))
MathOptInterface.ConstraintIndex{MathOptInterface.VectorOfVariables, MathOptInterface.Zeros}(1)

julia> MOI.optimize!(model)
Running HiGHS 1.11.0 (git hash: 364c83a51e): Copyright (c) 2025 HiGHS under MIT licence terms
LP   has 0 rows; 1 cols; 0 nonzeros
Coefficient ranges:
  Cost   [0e+00, 0e+00]
  Bound  [0e+00, 0e+00]
Solving LP without presolve, or with basis, or unconstrained
Solving an unconstrained LP with 1 columns
Model status        : Optimal
Objective value     :  0.0000000000e+00
P-D objective error :  0.0000000000e+00
HiGHS run time      :          0.00

julia> MOI.get(model, MOI.ConstraintDual(), vs)
1-element Vector{Float64}:
 0.0

[joaquimg]

Why you adding this constraint?

I am not. test_conic_linear_VectorOfVariables_2 is.

[odow]

Ah. Perhaps what is needed then is to fix the test so that if it errors because of ZeroBridge we don't fail completely?

For now, just skip the test.

[joaquimg]

For now, just skip the test.

yep