Repository navigation
Fix 3679 cupdlp infeas - #3689
Conversation
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## latest #3689 +/- ##
==========================================
+ Coverage 75.96% 75.98% +0.01%
==========================================
Files 467 467
Lines 116460 116543 +83
Branches 18162 18163 +1
==========================================
+ Hits 88471 88555 +84
+ Misses 27749 27748 -1
Partials 240 240 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
In fact, it's targeting (I don't have an option to change it, I think that's something you have to do @Yanyu000) |
Thanks @BenChampion Fixed! Guess I have the necessary admin privilege |
- Remove scaling variables no longer used in PDHG_Compute_Primal/Dual_Infeasibility - Comment why ray residuals are measured in scaled space and label the ifScaled=0 kernel arguments - Remove stray whitespace lines - pdlp-small-row-coefficient: require kIterationLimit rather than just "not kUnboundedOrInfeasible"; add trailing newline
Measuring the ray residual only in unscaled space gave false positives on LPs with tiny coefficients (#3679). Measuring it only in scaled space could introduce false positives the original test would not give. Compute the ray residual in both the scaled and unscaled spaces and return the larger, on both the CPU and GPU paths. The rays accepted are therefore a subset of those accepted before, so no new false positives are possible, while #3679 is still fixed.
|
Merged latest into the branch。 On false positives: I haven’t run a systematic test over other instances. However, I’ve changed the fix so that it cannot create new false positives relative to the original code. The ray residual is now computed in both the scaled and unscaled spaces, and the ray is only accepted if it passes the tolerance in both. So the set of rays accepted is a subset of those accepted before, and any infeasibility/unboundedness declared now would also have been declared by the original code. The cost is that some infeasible/unbounded LPs that were detected before might now only be detected later, or not at all. The existing [pdlp] unit tests (including pdlp-infeasible-lp and pdlp-unbounded-lp) pass, and the new test now checks for kIterationLimit on the #3679 LP and kOptimal on its mirror. |
|
Thanks, further experiments will determine whether this conservative strategy is needed |
Description
Fixes #3679
PDLP (cuPDLP-C) wrongly reported
UnboundedOrInfeasibleon LPs with a verysmall coefficient and presolve off, e.g.
max x s.t. 1e-8 x <= 1, 1e8 x >= 1, x >= 0(optimum 1e8), after one iteration. The mirror imagemin x s.t. 1e-8 x >= 1, x >= 0was wrongly reported as primal infeasible(final status
Unbounded). It was not a step size problem.Cause: in
PDHG_Compute_Dual_InfeasibilityandPDHG_Compute_Primal_Infeasibilitythe residual of the candidate ray wasconverted back to unscaled units (
rowScale/colScale). For these LPs theresidual/objective ratio was 1e-8, equal to
dFeasTol, so the ray wasaccepted.
Fix: measure the ray residuals in scaled space (CPU and GPU kernel
paths). A ray has no rhs to compare against, so the test should not depend on
row/column scaling. This differs from the optimality criteria, which still
use unscaled residuals, so I'd welcome a review of whether that is the right
choice.
Tests:
pdlp-small-row-coefficientcovers both LPs. The first LP needs~8e7 iterations to reach optimality (~23 s on CPU), so the test only checks
that the status is no longer
UnboundedOrInfeasibleafter 1000 iterations.The mirror LP is solved to
kOptimal.Checked:
pdlp-infeasible-lpandpdlp-unbounded-lpbehave as before. Alsorun on a CUDA build (
CUPDLP_GPU=ON): no longer reportsUnboundedOrInfeasible.Checklist
latestbranch