{"project":{"id":"019ce14d-e313-4a5f-88e2-de0ba75d5ad3","name":"CoolPDLP.jl","repository":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl.git","description":"A pure-Julia, hardware-agnostic parallel implementation of Primal-Dual hybrid gradient for Linear Programming (PDLP) and its variants.","default_branch":"main","json_path":"/julialang/coolpdlp-dot-jl.json","path":"/julialang/coolpdlp-dot-jl","settings_path":"/julialang/coolpdlp-dot-jl/settings","builds_path":"https://buildkite.com/julialang/coolpdlp-dot-jl/builds","builds_count":396,"scheduled_builds_path":"/julialang/coolpdlp-dot-jl/builds?state=scheduled","scheduled_builds_count":0,"running_builds_path":"/julialang/coolpdlp-dot-jl/builds?state=running","running_builds_count":0,"permissions":{"edit":{"allowed":false,"reason":"anonymous","message":"You must be signed in to change this pipeline"},"trigger_build":{"allowed":false,"reason":"anonymous","message":"You must be signed in to trigger a build in this pipeline"},"favorite_project":{"allowed":false,"reason":"anonymous","message":"You must be signed in to favorite this pipeline"}}},"pipelineSettings":null,"builds":[{"id":"01a0f851-db45-45c2-a27a-ab535ee508f3","number":380,"message":"perf: split CSR rows across sub-groups, and give Reactant a CSR product it can raise","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/380","commit_id":"726b0a5e402cfa4764323864107fe3789d625c36","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/726b0a5e402cfa4764323864107fe3789d625c36","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:perf/csr-subgroup-spmv","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Aperf%2Fcsr-subgroup-spmv","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-01T16:35:09.267Z","scheduled_at":"2026-10-01T16:35:09.303Z","started_at":"2026-10-01T16:35:11.486Z","canceled_at":null,"finished_at":"2026-10-01T16:51:21.462Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0f8f4-2058-4bab-b9f2-7a90035ca222","number":383,"message":"perf: split CSR rows across sub-groups, and give Reactant a CSR product it can raise","state":"failed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/383","commit_id":"3210d328bba8786a12c34c16ed67e541f34179cc","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/3210d328bba8786a12c34c16ed67e541f34179cc","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:perf/csr-subgroup-spmv","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Aperf%2Fcsr-subgroup-spmv","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-01T19:32:23.774Z","scheduled_at":"2026-10-01T19:32:23.804Z","started_at":"2026-10-01T19:32:24.385Z","canceled_at":null,"finished_at":"2026-10-01T19:45:35.115Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a10242-2a89-48bd-bdcd-228533060fe7","number":396,"message":"chore: merge main into gd/reac2 and trace the objective constant","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/396","commit_id":"d226f530d208315c4fecd680dacdbc5f27b5a633","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/d226f530d208315c4fecd680dacdbc5f27b5a633","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:gd/reac2-merge-main","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Agd%2Freac2-merge-main","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-03T14:54:13.136Z","scheduled_at":"2026-10-03T14:54:13.174Z","started_at":"2026-10-03T14:54:14.424Z","canceled_at":null,"finished_at":"2026-10-03T17:24:51.831Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0f6c4-b243-4949-9ed2-9ccdc5143b0a","number":373,"message":"feat: let a Reactant-compiled solve use the custom sparse formats (#160) * feat: let a Reactant-compiled solve use the custom sparse formats `solve!` could not be Reactant-compiled with `GPUSparseMatrixCSR`, `GPUSparseMatrixELL` or `GPUSparseMatrixCOO`: every combination of format, batching and backend failed before a single operation was emitted. Four independent blockers, each hiding the next. 1. `Ti <: Integer` made the wrappers untraceable: a traced index array has element type `TracedRNumber{Int32}`, a `Number` that is not an `Integer`, so unifying the array parameter with `Ti` landed outside the bound. Drop the bound, as `MILP` already does for `int_var`. 2. The `Int` shape fields: any mode that promotes Julia numbers derives `TracedRNumber{Int64}` for `m` and `n`, which the struct declares as `Int`, and Reactant gives up with a `NoFieldMatchError`. Not avoidable from the call site, since the `@trace while` loop promotes the numbers of everything it carries. Two overloads in the Reactant extension now trace these wrappers with number tracking off: a sparse matrix's shape and sparsity pattern are structure, not data. 3. Reactant overlays `mul!` for every `AbstractMatrix` and lowers the product to a dense `dot_general`, and an overlay beats ordinary dispatch, so the kernel-backed `mul!` was never reached -- the overlay recursed in `get_ancestor_and_indices` until the stack overflowed. The kernels move to `spmul!`, `mul!` forwards to it, and a more specific overlay sends the call back there. 4. `kkt_errors!` passed `-one(T), zero(T)`, which under tracing are traced numbers, so `spmul!` had to branch on `isone`/`iszero` of a `TracedRNumber{Bool}`. Plain `-1, false` keeps the choice a compile-time one and keeps the strong zero. Also fixes a pre-existing bug exposed by the same path: `write_time!` wrote its callback buffer with a scalar `setindex!`, which CUDA.jl refuses on a device array, so the clock callback failed at run time whenever CUDA.jl was loaded. `fill!` works on any backend. A compiled `mul!` and a compiled single-instance `solve!` now agree with the plain loop for all three formats on Reactant's CPU and CUDA backends, and a batched solve agrees on the CPU backend. A batched solve on CUDA still deviates; the bisection is recorded next to the test that leaves it out. Co-Authored-By: Claude Opus 5 (1M context) Claude-Session: https://claude.ai/code/session_01M6hNg5PcDLe4kQs9ANXVbA * fix: run batched sparse products on flat buffers under Reactant A batched solve compiled for Reactant's CUDA backend came out wrong from its second iteration, for all three sparse formats identically, while single-instance solves and the CPU backend were exact. The cause is a Reactant miscompilation: a 2-D array written by a KernelAbstractions kernel inside a `@trace for` and reduced after the loop is silently relayouted. XLA's layout assignment lets the reduction pick a row-major layout for the loop-carried array, and the `enzymexla.kernel_call` pins neither its operand nor its result layouts, so the kernel's column-major bytes are read back transposed. The MRE, the HLO showing the `{0,1}` layout, and the upstream fix (the lowering must forward `result_layouts`, which it currently drops) are recorded on PR #160. A vector has a single layout and is immune, which is why only batched solves were affected. The batched kernels therefore index flattened operands now, at no cost to the plain path, which passes `vec` views; and the Reactant extension runs a batched product through `spmul_batched!`, on `Reactant.Ops.reshape`d operands and a fresh 1-D output. Placing a copy, a broadcast or an `optimization_barrier` between the loop and the reduction does not help, and a fresh 2-D output is elided by XLA in a solve-sized program; the flat form is sound by construction. The `cuda` group now asserts the batched compiled solve for all three formats. Co-Authored-By: Claude Fable 5.1 Claude-Session: https://claude.ai/code/session_01M6hNg5PcDLe4kQs9ANXVbA * refactor: keep flat buffers inside the Reactant extension The kernels in `src` index 2-D arrays again, and `spmul!`/`spmm!` are gone: a format only implements `mul!`. The extension's `mul!` overlays adapt the operands into `FlatArray`s (1-D buffers viewed as arrays), call the format's own `mul!` natively, and relaunch its kernels through Reactant from a `FlatBackend`. Comments and docstrings are shortened. Co-Authored-By: Claude Opus 5 (1M context) Claude-Session: https://claude.ai/code/session_01FG5tVesvgqgRR7bkyofz36 * refactor: flatten only the batched products, per review - `Ti <: Number` on the three formats, as suggested. - The flat buffers are now confined to `flat_mul!`, which only the batched overlays call, and marked temporary until EnzymeAD/Reactant.jl#3269. The single-instance product passes its operands straight through. - `β` is applied to the destination before the call and the kernel gets `β = true`, so the result needs no addition afterwards. Co-Authored-By: Claude Opus 5 (1M context) Claude-Session: https://claude.ai/code/session_01FG5tVesvgqgRR7bkyofz36 * refactor: refuse Reactant SpMM on custom formats instead of flattening The batched product ran on flat buffers to dodge the XLA layout bug of EnzymeAD/Reactant.jl#3269. Drop that workaround: the matrix overlays now throw an error pointing to the issue, and KernelArray loses its separate dims. Single-instance solves are unaffected. Co-Authored-By: Claude Opus 5.5 (1M context) Claude-Session: https://claude.ai/code/session_01HqqFD5rvt9bbFLdca8Md13 --------- Co-authored-by: Claude Opus 5 (1M context) ","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/373","commit_id":"4dd8fee82baf4f601297d47cf6a43be097764ff6","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/4dd8fee82baf4f601297d47cf6a43be097764ff6","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"gd/reac2","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gd%2Freac2","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-01T09:21:20.970Z","scheduled_at":"2026-10-01T09:21:21.006Z","started_at":"2026-10-01T09:21:28.440Z","canceled_at":null,"finished_at":"2026-10-01T10:08:49.362Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a08b0d-8feb-4828-8f9d-1eed54e3bf48","number":357,"message":"feat: add extensible presolve API + default PaPILO implementation","state":"canceled","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/357","commit_id":"15143fe11423aec4fa704ee6c78a66e053112674","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/15143fe11423aec4fa704ee6c78a66e053112674","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:feat/papilo-presolve","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Afeat%2Fpapilo-presolve","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-09-10T11:21:56.980Z","scheduled_at":"2026-09-10T11:21:57.003Z","started_at":"2026-09-10T11:22:14.931Z","canceled_at":"2026-09-10T11:44:14.500Z","finished_at":"2026-09-10T11:44:14.500Z","cancel_reason":"build_skipping","project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0f85e-f49a-4f41-92f2-9a81a4ac8d18","number":381,"message":"feat: add an objective constant to MILP","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/381","commit_id":"e4ce3c1c90380327546249d9b5eaeda720e1a4f0","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/e4ce3c1c90380327546249d9b5eaeda720e1a4f0","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:feat/objective-constant","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Afeat%2Fobjective-constant","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-01T16:49:27.713Z","scheduled_at":"2026-10-01T16:49:27.747Z","started_at":"2026-10-01T16:49:35.341Z","canceled_at":null,"finished_at":"2026-10-01T16:57:22.049Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a03ed8-e227-4c12-803c-9ccc390e0eef","number":310,"message":"refactor!: use MOI.TerminationStatusCode instead of custom TerminationStatus (#141) * refactor: use MOI.TerminationStatusCode instead of custom TerminationStatus enum Drops CoolPDLP's own OPTIMAL/TIME_LIMIT/ITERATION_LIMIT/STILL_RUNNING enum in favor of MOI.TerminationStatusCode (MOI.OPTIMIZE_NOT_CALLED replaces STILL_RUNNING), so ConvergenceStats reports status codes that are already MOI-compatible instead of requiring a translation step in the MOI wrapper. Co-Authored-By: Claude Sonnet 5 * fix: update new termination tests for MOI status codes Rebasing onto upstream/main pulled in tests from #128 that referenced the just-removed bare OPTIMAL name; switch them to MOI.OPTIMAL to match. Co-Authored-By: Claude Sonnet 5 * fix: drop unreachable termination-status branch to restore coverage The else branch handling MOI.OPTIMIZE_NOT_CALLED in MOI.optimize! could never execute: solve!/termination_check! only return once the status has left OPTIMIZE_NOT_CALLED, so the branch was always 0% covered. Replace it with an unconditional assert of the same invariant (executed on every call, so it stays covered) and drop the now-redundant NO_SOLUTION mapping. Co-Authored-By: Claude Sonnet 5 --------- Co-authored-by: Claude Sonnet 5 ","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/310","commit_id":"8c51367c216a648cc933452fbe47ce5219bcaab5","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/8c51367c216a648cc933452fbe47ce5219bcaab5","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"main","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=main","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-08-26T16:13:16.209Z","scheduled_at":"2026-08-26T16:13:16.260Z","started_at":"2026-08-26T16:13:19.180Z","canceled_at":null,"finished_at":"2026-08-26T18:12:36.225Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a03e28-68a4-40e4-85dd-dc66baaf863a","number":305,"message":"fix: avoid infinite step size on constraint-free LPs (#128) * fix: avoid infinite step size on constraint-free LPs `fixed_stepsize` computed `invnorm_scaling * inv(spectral_norm(A, At))` unconditionally. Whenever `A` has no rows (or is all-zero with rows), its spectral norm is exactly 0, so the step size became `Inf`. The first primal step then corrupted every zero-objective-coefficient variable to NaN (`Inf * 0`) and every nonzero-coefficient variable relied on `clamp` to accidentally saturate an infinite step back into range. `solve`'s early exit for `nbcons(milp) == 0` only fired when the whole objective was zero, so nonzero-objective constraint-free problems fell through to the broken general loop. That branch also silently claimed `OPTIMAL` for an infeasible box (`lv > uv`, where `clamp` just returns `uv`), and left `stats.err`/`stats.time_elapsed` at their uninitialized `NaN`/`0.0` defaults instead of reflecting the solution actually returned. Fixes both: - `fixed_stepsize` now falls back to a finite step size when the spectral norm is zero, since `A'y` never contributes to the primal step in that case regardless of `η`. This alone makes the general iterative loop safe for any all-zero/empty `A`, not just the `nbcons == 0` case `solve` special-cases. - `solve`'s early exit now covers any constraint-free LP whose box is feasible and bounded in the direction the objective pushes towards, reading the optimum off `c`/`lv`/`uv` directly; an infeasible or unbounded box falls through to the general loop instead of a wrong `OPTIMAL`, consistent with how this package already has no dedicated status for infeasibility/unboundedness elsewhere. The early exit also now populates `stats.err` and `stats.time_elapsed` correctly. Fixes #96. Co-Authored-By: Claude Sonnet 5 Claude-Session: https://claude.ai/code/session_01UQcf3oERECinXrDV3pqaag * fix: close codecov/patch gap in constraint-free-LP early exit Codecov flagged the multi-line `&&` condition (nbcons==0, box feasibility, and the two unboundedness checks) as only partially branch-covered, even though a passing and an infeasible/unbounded test already exercised it: with inline `@.` broadcasts spread across the `&&` chain's continuation lines, Julia's coverage instrumentation doesn't reliably attribute hits to each sub-condition. Extracted `box_feasible`/`bounded_below`/`bounded_above` into their own named statements so each is unambiguously a single coverable line, and added the missing symmetric test case (`c[1] < 0, uv[1] == Inf`) — the existing tests only ever drove the `bounded_below` (lv == -Inf) side of the unboundedness check false, never `bounded_above`. Co-Authored-By: Claude Sonnet 5 Claude-Session: https://claude.ai/code/session_01UQcf3oERECinXrDV3pqaag --------- Co-authored-by: Claude Sonnet 5 ","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/305","commit_id":"4594f7a83c795e53411eb8b8c59121a03dd04410","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/4594f7a83c795e53411eb8b8c59121a03dd04410","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"main","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=main","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-08-26T13:00:30.767Z","scheduled_at":"2026-08-26T13:00:30.799Z","started_at":"2026-08-26T13:01:09.137Z","canceled_at":null,"finished_at":"2026-08-26T13:17:17.859Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a03f8b-ced8-48b8-9e36-77cd8f836e89","number":314,"message":"build based on fd81727","state":"not_run","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/314","commit_id":"5bf2822e769207c0d72e60c8fcae8b378f3e33b6","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/5bf2822e769207c0d72e60c8fcae8b378f3e33b6","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"gh-pages","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gh-pages","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-08-26T19:28:42.208Z","scheduled_at":"2026-08-26T19:28:42.252Z","started_at":null,"canceled_at":null,"finished_at":"2026-08-26T19:28:42.319Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0fd90-95ce-4db3-b928-24e93ff9cdff","number":395,"message":"build based on d1458aa","state":"not_run","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/395","commit_id":"4c1ca9fb1299e92b75cc8d1ef3c12122ab60af5d","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/4c1ca9fb1299e92b75cc8d1ef3c12122ab60af5d","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"gh-pages","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gh-pages","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-02T17:01:46.324Z","scheduled_at":"2026-10-02T17:01:46.365Z","started_at":null,"canceled_at":null,"finished_at":"2026-10-02T17:01:46.421Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a08adc-5677-46cb-8ea5-d362c25e2668","number":354,"message":"feat: trace the termination status inside a Reactant-compiled solve (#156) `ConvergenceStats.termination_status` was a plain `MOI.TerminationStatusCode`, written from inside three `@trace if` blocks. An enum cannot be traced, so only the trace-time value of those assignments survived: a compiled run that converged still reported `MOI.OPTIMIZE_NOT_CALLED`, and `max_kkt_passes` / `time_limit` decided when to stop without ever saying so. Store the status as an integer code instead. `termination_status_code` is an ordinary number, so `to_rarray` tracks it and the compiled program writes it like any other stat, and `termination_status(stats)` decodes it back to the enum. `solve` still returns the same `MOI.TerminationStatusCode` it always did, one call away. That also removes the `elseif` workaround. With a traceable code, the criteria can be selected with nested `ifelse` calls instead of three separate `@trace if` blocks ordered by increasing priority and relying on last-write-wins (EnzymeAD/Reactant.jl#2563), so the priority between simultaneous criteria is stated in one place. The Reactant coherence tests now compare statuses between the plain and compiled runs, which #153 had to leave out, and the time limit test checks that the compiled solve reports `MOI.TIME_LIMIT` rather than just stopping at the right time. Verified locally on CPU: all three statuses (`OPTIMAL`, `ITERATION_LIMIT`, `TIME_LIMIT`) come back out of a compiled `solve!` for both PDHG and PDLP, and the `Core`, `MOI` and `Reactant` test groups pass. Claude-Session: https://claude.ai/code/session_01JPzdTAQVAXUNUAQsL4Nz71 Co-authored-by: Claude Opus 5 ","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/354","commit_id":"04927b935d1be9d7bf3b060b1ac7c02737ca850d","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/04927b935d1be9d7bf3b060b1ac7c02737ca850d","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"gd/reac2","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gd%2Freac2","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-09-10T10:28:11.012Z","scheduled_at":"2026-09-10T10:28:11.056Z","started_at":"2026-09-10T10:28:21.003Z","canceled_at":null,"finished_at":"2026-09-10T10:38:28.181Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0fcc0-8179-4a55-aba5-19b8c6f516ef","number":386,"message":"perf: give the inner loop of a Reactant-compiled solve a static trip count (#165) With check_every traced, the inner @trace for loop of solve! has a dynamic trip count, and XLA reads its condition back to the host after every KKT pass. Trace GenericParameters without number tracking, like the sparse formats, so that check_every is a compile-time constant. The iteration budget and counters, and the tolerances, stay traced. Co-authored-by: Claude Opus 5.5 ","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/386","commit_id":"21733b62b12c7e6659b7c350041d0b0e31722e4c","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/21733b62b12c7e6659b7c350041d0b0e31722e4c","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"gd/reac2","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gd%2Freac2","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-02T13:14:29.639Z","scheduled_at":"2026-10-02T13:14:29.674Z","started_at":"2026-10-02T13:14:36.312Z","canceled_at":null,"finished_at":"2026-10-02T13:28:24.268Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a03e31-023d-4459-b234-eae91078ac78","number":306,"message":"build based on 4594f7a","state":"not_run","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/306","commit_id":"673a6e24f0056db56812e41c3eb9bfd08e023e68","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/673a6e24f0056db56812e41c3eb9bfd08e023e68","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"gh-pages","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gh-pages","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-08-26T13:09:54.381Z","scheduled_at":"2026-08-26T13:09:54.429Z","started_at":null,"canceled_at":null,"finished_at":"2026-08-26T13:09:54.518Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0f970-7f0c-447c-8d6e-dae8705c5ab4","number":384,"message":"perf: split CSR rows across sub-groups, and give Reactant a CSR product it can raise","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/384","commit_id":"33a8fa0a3475c70ed8e2bac849cd71ac17de2094","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/33a8fa0a3475c70ed8e2bac849cd71ac17de2094","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:perf/csr-subgroup-spmv","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Aperf%2Fcsr-subgroup-spmv","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-01T21:48:14.488Z","scheduled_at":"2026-10-01T21:48:14.523Z","started_at":"2026-10-01T21:48:17.725Z","canceled_at":null,"finished_at":"2026-10-01T22:00:54.014Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0f9ef-cb8a-40d3-a514-fd1e8d2722b8","number":385,"message":"perf: split CSR rows across sub-groups in the GPU SpMV/SpMM kernels","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/385","commit_id":"c7a4bf9fe93e48af7180987dd08ab28c2a6488b6","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/c7a4bf9fe93e48af7180987dd08ab28c2a6488b6","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:perf/csr-subgroup-spmv","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Aperf%2Fcsr-subgroup-spmv","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-02T00:07:17.137Z","scheduled_at":"2026-10-02T00:07:17.175Z","started_at":"2026-10-02T00:07:25.001Z","canceled_at":null,"finished_at":"2026-10-02T00:18:34.643Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a03f88-fff9-4209-912c-3247a0890e74","number":313,"message":"fix: guard spectral_norm's convergence and decouple its iteration budget from problem size (#142) * fix: safer spectral norm estimate * fix: guard spectral_norm's convergence and add non-regression tests The reused fix from gd/spec (safer spectral norm estimate) asserted `λ > zero(λ)`, which rejects the legitimate zero spectral norm of a constraint-free or all-zero `A` and reintroduces the crash #128 fixed; relaxed to `λ >= zero(λ)`. It also left `maxiter` at IterativeSolvers' default `size(B, 2)`, which ties the iteration budget to problem size instead of the (unrelated) convergence rate of the power method — exactly what let the estimate diverge silently on large matrices per #95. Decoupled it with a fixed, generous cap, which also lets ordinary matrices converge to near machine precision instead of the `rtol = 1.0e-1` the old test needed. Co-Authored-By: Claude Sonnet 5 * fix: expose spectral_norm's maxiter as spectral_norm_maxiter Per review: make the power method's iteration budget configurable from StepSizeParameters/Algorithm, the same way spectral_norm_tol already is, instead of hardcoding it inside spectral_norm. Co-Authored-By: Claude Sonnet 5 --------- Co-authored-by: Guillaume Dalle <22795598+gdalle@users.noreply.github.com> Co-authored-by: Claude Sonnet 5 ","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/313","commit_id":"fd81727756df9624dee6d96b7551254c9ca294c7","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/fd81727756df9624dee6d96b7551254c9ca294c7","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"main","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=main","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-08-26T19:25:38.180Z","scheduled_at":"2026-08-26T19:25:38.238Z","started_at":"2026-08-26T19:25:41.746Z","canceled_at":null,"finished_at":"2026-08-26T20:13:14.054Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a08aee-52e7-4242-9d35-a8b16e570565","number":356,"message":"feat: add extensible presolve API + default PaPILO implementation","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/356","commit_id":"991eed2d63b3d842b48f844b85dff686b2345dea","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/991eed2d63b3d842b48f844b85dff686b2345dea","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:feat/papilo-presolve","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Afeat%2Fpapilo-presolve","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-09-10T10:47:49.744Z","scheduled_at":"2026-09-10T10:47:49.767Z","started_at":"2026-09-10T10:47:51.718Z","canceled_at":null,"finished_at":"2026-09-10T10:54:30.003Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0fd7a-c5d8-48d3-ba6f-10ae48617fe7","number":393,"message":"feat: add an objective constant to MILP","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/393","commit_id":"abd6164cae24e933a671b65a6f262d3bb40c1ac4","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/abd6164cae24e933a671b65a6f262d3bb40c1ac4","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"gdalle-bot:feat/objective-constant","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Afeat%2Fobjective-constant","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-02T16:37:56.832Z","scheduled_at":"2026-10-02T16:37:56.866Z","started_at":"2026-10-02T16:38:02.072Z","canceled_at":null,"finished_at":"2026-10-02T17:41:14.419Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0fd8b-bd06-41f4-9acf-7a8408e83009","number":394,"message":"feat: add an objective constant to MILP (#166) * feat: add an objective constant to MILP `MILP` gains a `c0` field, so that the problem it describes is `min cᵀx + c0`. The constant is a single number of the problem's element type, shared by every instance of a batch, and defaults to zero. - `objective_value` includes it. - `kkt_errors!` adds it to the primal and dual objectives: it cancels out in the gap but not in the scale `1 + |primal| + |dual|` the gap is compared to, so a problem with a large constant terminates on a looser absolute gap. - `MILP(::QPSData)` and `mps_to_milp` read it, `milp_to_mps` writes it, and a maximization problem stores its opposite, like it does for `c`. - The JuMP wrapper passes the constant of the objective function to the `MILP` instead of adding it to the reported objective values afterwards. - Preconditioning, permutation, conversion and `relax` carry it over. `random_milp_and_sol` now draws a nonzero constant, so that the existing tests exercise it. Co-Authored-By: Claude Code * test: cover the objective constant of a maximization MPS round trip Co-Authored-By: Claude Opus 5.5 * Apply suggestion from @gdalle --------- Co-authored-by: Claude Code Co-authored-by: Guillaume Dalle <22795598+gdalle@users.noreply.github.com>","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/394","commit_id":"d1458aac0de876b152658c45dbd59a93c6c6116f","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/d1458aac0de876b152658c45dbd59a93c6c6116f","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"main","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=main","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-02T16:56:28.683Z","scheduled_at":"2026-10-02T16:56:28.723Z","started_at":"2026-10-02T16:56:30.865Z","canceled_at":null,"finished_at":"2026-10-02T17:43:52.639Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a08a99-9f55-4aa6-b4d9-001096d6afb4","number":352,"message":"feat: enforce the time limit inside a Reactant-compiled solve (#155) * feat: enforce the time limit inside a Reactant-compiled solve `termination_check!` measured elapsed time with `time()`, an ordinary Julia call that tracing folds into the compiled program as a constant. `time_elapsed` never advanced inside a compiled loop, so `MOI.TIME_LIMIT` could never fire and `max_kkt_passes` was the only real budget. Route the clock through `CoolPDLP.current_time()`, and overlay that method in a new `Reactant` extension with a `Reactant.Ops.julia_callback` host callback, so it is re-evaluated at every iteration. Two details of `julia_callback` shaped the callback. It calls the Julia function as `f(outputs..., inputs...)`, and a `()`-shaped output arrives dereferenced as a plain `Float64` with nothing to write into, so the output is declared with shape `(1,)` and reduced back to a scalar (which compiles to a `stablehlo.reshape`). And `has_side_effect = true` is what keeps the compiler from hoisting the call out of the loop, sharing it across iterations or dropping it. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01YGgoU6rLLT72Qis2Ku8pB6 * ci: collect coverage from the Reactant test group The `Reactant` group was the only place exercising the new Reactant extension, and it ran uninstrumented, so those lines could never be covered. Instrumentation is not as expensive as feared: the group goes from ~6m30s to ~9m00s locally, well inside the job's 45 minute budget, and every line of the extension is then reached. The Buildkite GPU job stays uninstrumented, since it runs the same test files as the GitHub Actions one and would not reach any new line. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01YGgoU6rLLT72Qis2Ku8pB6 * fix: only read the host clock on backends that can run a callback The Buildkite `reactant gpu` jobs hit their 90 minute timeout: on a GPU backend `Reactant.Ops._wrap_buffers` routes the callback's buffers through `CUDA.jl`, which that job deliberately does not load, so the callback raises. Reactant's trampoline catches that, logs it and reports failure on every single call rather than stopping, which surfaces as a hang instead of an error. Emit the callback only where Reactant can service it — the host, or CUDA with the `CUDA.jl` extension loaded — and otherwise fall back to the previous frozen clock, with a warning saying the time limit will not be enforced. The time limit tests are gated on the same check. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01YGgoU6rLLT72Qis2Ku8pB6 * test: cover the frozen-clock fallback The fallback branch was the last uncovered line of the extension: it only runs on a backend that cannot host a callback, which no instrumented CI job uses. Give it a name and a test. That also pins the behaviour it introduced, which was untested until now: a plain `Float64` read once, and a warning that the time limit will not be enforced. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01YGgoU6rLLT72Qis2Ku8pB6 * chore: retrigger CI Buildkite build #350 failed after 2 seconds with a single job, before any Julia ran: an infrastructure blip at the pipeline upload step, not a test failure. The pipeline file parses and has not changed since build #349, which passed. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01YGgoU6rLLT72Qis2Ku8pB6 --------- Co-authored-by: Claude Opus 5 ","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/352","commit_id":"e93b5a96b7892eafe574e2cf2f6f60dd1be4d95b","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/e93b5a96b7892eafe574e2cf2f6f60dd1be4d95b","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"gd/reac2","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gd%2Freac2","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-09-10T09:15:18.749Z","scheduled_at":"2026-09-10T09:15:18.783Z","started_at":"2026-09-10T09:15:25.218Z","canceled_at":null,"finished_at":"2026-09-10T09:43:03.737Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0f773-c824-4f7e-b566-a2f46ba284ce","number":376,"message":"feat: add extensible presolve API + default PaPILO implementation","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/376","commit_id":"2b52e8c72722777652ef8c35a0a0845c8df61ddc","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/2b52e8c72722777652ef8c35a0a0845c8df61ddc","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:feat/papilo-presolve","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Afeat%2Fpapilo-presolve","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-01T12:32:35.373Z","scheduled_at":"2026-10-01T12:32:35.400Z","started_at":"2026-10-01T12:32:50.147Z","canceled_at":null,"finished_at":"2026-10-01T12:50:07.415Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0fcd9-b322-44d6-a4fd-b1b168928d17","number":391,"message":"feat: add an objective constant to MILP","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/391","commit_id":"4b8314e64c7d2a68e5deab09e64517e4063e1600","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/4b8314e64c7d2a68e5deab09e64517e4063e1600","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:feat/objective-constant","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Afeat%2Fobjective-constant","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-02T13:42:00.743Z","scheduled_at":"2026-10-02T13:42:00.767Z","started_at":"2026-10-02T13:42:10.940Z","canceled_at":null,"finished_at":"2026-10-02T14:07:53.434Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0f7c0-a657-49fd-94e9-635063e81ca7","number":378,"message":"perf: give the inner loop of a Reactant-compiled solve a static trip count","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/378","commit_id":"03066c82f29c9e56149f9140faa9a2527e4d2b18","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/03066c82f29c9e56149f9140faa9a2527e4d2b18","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:gd/reac2-static-loop","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Agd%2Freac2-static-loop","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-01T13:56:32.988Z","scheduled_at":"2026-10-01T13:56:33.010Z","started_at":"2026-10-01T13:56:42.966Z","canceled_at":null,"finished_at":"2026-10-01T14:30:26.303Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0fcc3-58d7-41be-a7a9-ad9bbea59438","number":388,"message":"feat: add an objective constant to MILP","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/388","commit_id":"2dea08be566b381b7a492b5c32b58842fcd10448","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/2dea08be566b381b7a492b5c32b58842fcd10448","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:feat/objective-constant","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Afeat%2Fobjective-constant","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-02T13:17:35.839Z","scheduled_at":"2026-10-02T13:17:35.868Z","started_at":"2026-10-02T13:17:43.725Z","canceled_at":null,"finished_at":"2026-10-02T13:26:15.911Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0fcdd-d3bf-412f-9a12-8708fab0b633","number":392,"message":"build based on 11305fa","state":"not_run","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/392","commit_id":"8f88034e946e70d8ed2d90a10bb47a7cf553b1bc","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/8f88034e946e70d8ed2d90a10bb47a7cf553b1bc","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"gh-pages","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gh-pages","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-02T13:46:31.238Z","scheduled_at":"2026-10-02T13:46:31.262Z","started_at":null,"canceled_at":null,"finished_at":"2026-10-02T13:46:31.313Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0f771-9b96-4f3d-92e8-288a802b53c8","number":375,"message":"feat: add constraint names, relaxation and MPS round trip for MILP","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/375","commit_id":"84bee69dbf2b70422ef179ed3212ab34fbc1b483","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/84bee69dbf2b70422ef179ed3212ab34fbc1b483","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:feat/milp-con-names","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Afeat%2Fmilp-con-names","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-01T12:30:12.895Z","scheduled_at":"2026-10-01T12:30:12.939Z","started_at":"2026-10-01T12:30:16.548Z","canceled_at":null,"finished_at":"2026-10-01T12:47:40.759Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a08b67-5d55-4b8b-8d52-a78ecdbdb5fc","number":363,"message":"fix: let a batched solve compile under Reactant (#157) * fix: let a batched solve compile under Reactant A batched `solve!` could not be compiled at all: it aborted with TypeInstabilityError: Instability detected in `termination_check!` ... Inferred to be `Reactant.TracedRNumber`, which is not a concrete type Reactant types every reduction over a `TracedRArray` as `Union{TracedRArray, TracedRNumber}`. For a batch, the two reductions that collapse per-instance quantities into the single decision the `@trace while` loop branches on -- `batched_all` in `set_termination_status!!` and `batched_mean` in `should_restart` -- return exactly that boolean, so the abstract type escaped into the return type of `termination_check!` and `restart_check!`. A single instance keeps those quantities scalar and needs no reduction, which is why only batches were affected. Assert the scalar type of each reduction instead. `batched_mean` returns one element of the batch, and `batched_all` a boolean whose flavour follows the array type, through the new `batched_bool_type` hook that the Reactant extension overloads. The Reactant suite now solves a batch as well as a single instance, on both the CPU and the GPU backend. Its comparison against the plain Julia loop is compiled with `dot_general_precision = HIGHEST`, since a batch multiplies matrices where a single solve multiplies a matrix by a vector, and XLA lowers the former to TF32 tensor cores on NVIDIA: three digits that a `Float32` batch feeds back into its own trajectory until it no longer resembles the plain run. Co-Authored-By: Claude Opus 5 (1M context) Claude-Session: https://claude.ai/code/session_01M6hNg5PcDLe4kQs9ANXVbA * Apply batched suggestions from code review Co-authored-by: Guillaume Dalle <22795598+gdalle@users.noreply.github.com> --------- Co-authored-by: Claude Opus 5 (1M context) Co-authored-by: Guillaume Dalle <22795598+gdalle@users.noreply.github.com>","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/363","commit_id":"9dde973594306de2ab38a6381a88344462a4b053","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/9dde973594306de2ab38a6381a88344462a4b053","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"gd/reac2","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gd%2Freac2","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-09-10T13:00:02.267Z","scheduled_at":"2026-09-10T13:00:02.289Z","started_at":"2026-09-10T13:00:14.224Z","canceled_at":null,"finished_at":"2026-09-10T14:10:23.967Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a03eeb-f771-4c86-8ffe-fd7f02afb8f6","number":312,"message":"build based on 8c51367","state":"not_run","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/312","commit_id":"070c5320f6f9ab11c176cc4681549dfdd30adf6d","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/070c5320f6f9ab11c176cc4681549dfdd30adf6d","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"gh-pages","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gh-pages","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-08-26T16:34:06.839Z","scheduled_at":"2026-08-26T16:34:06.874Z","started_at":null,"canceled_at":null,"finished_at":"2026-08-26T16:34:06.954Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0fcc3-84df-4f03-9ba2-22a2073d697e","number":389,"message":"feat: add extensible presolve API + default PaPILO implementation","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/389","commit_id":"61f9d7e9114e315bf7eab5c44bb1e33d3b9c8fdb","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/61f9d7e9114e315bf7eab5c44bb1e33d3b9c8fdb","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:feat/papilo-presolve","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Afeat%2Fpapilo-presolve","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-02T13:17:47.110Z","scheduled_at":"2026-10-02T13:17:47.135Z","started_at":"2026-10-02T13:17:56.657Z","canceled_at":null,"finished_at":"2026-10-02T13:25:05.398Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0fcd8-644b-4177-9572-481060acb04d","number":390,"message":"feat: add constraint names, relaxation and MPS round trip for MILP (#163) * feat: add constraint names, relaxation and MPS round trip for MILP `MILP` gains a `con_names` field mirroring `var_names`: it defaults to the constraint indices, is read from `qps.connames` by `MILP(::QPSData)`, and is carried through every transformation (permuted with the rows by `sort_rows_columns`). `relax(milp)` returns the continuous relaxation of a `MILP`, sharing every array but `int_var`. `milp_to_mps` writes a `MILP` to an MPS file through MathOptInterface, under names the caller may override, and `mps_to_milp` reads one back through QPSReader. Rows and columns keep their order across the round trip. All three are prerequisites for presolve (#144), split out to keep that pull request reviewable. Co-Authored-By: Claude Code * refactor: keep the MPS helpers internal and document maximization problems Co-Authored-By: Claude Opus 5.5 --------- Co-authored-by: Claude Code ","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/390","commit_id":"11305fa25c2bffe04dc07f62441e555ca404c7b7","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/11305fa25c2bffe04dc07f62441e555ca404c7b7","author_name":"Guillaume Dalle","author_avatar":"https://avatars.githubusercontent.com/u/22795598?v=4","branch_name":"main","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=main","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-02T13:40:35.030Z","scheduled_at":"2026-10-02T13:40:35.057Z","started_at":"2026-10-02T13:40:47.919Z","canceled_at":null,"finished_at":"2026-10-02T14:06:19.876Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0f86b-eba4-437e-a05e-9a4c08c09d16","number":382,"message":"perf: split CSR rows across sub-groups, and give Reactant a CSR product it can raise","state":"failed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/382","commit_id":"0b98510cfb229e4dc508cae3d97ba48f835df6f0","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/0b98510cfb229e4dc508cae3d97ba48f835df6f0","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:perf/csr-subgroup-spmv","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Aperf%2Fcsr-subgroup-spmv","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-01T17:03:37.385Z","scheduled_at":"2026-10-01T17:03:37.422Z","started_at":"2026-10-01T17:03:53.804Z","canceled_at":null,"finished_at":"2026-10-01T17:14:08.437Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0fcc1-db1d-44f6-9571-cfe7cad5282c","number":387,"message":"feat: add constraint names, relaxation and MPS round trip for MILP","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/387","commit_id":"6bfbe7af165b372b5aad172c750a27299d50c519","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/6bfbe7af165b372b5aad172c750a27299d50c519","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:feat/milp-con-names","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Afeat%2Fmilp-con-names","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-02T13:15:58.118Z","scheduled_at":"2026-10-02T13:15:58.150Z","started_at":"2026-10-02T13:16:02.817Z","canceled_at":null,"finished_at":"2026-10-02T13:29:19.894Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a08b21-f7d9-4bd4-8b27-934a55d9b275","number":358,"message":"feat: add extensible presolve API + default PaPILO implementation","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/358","commit_id":"e682afeaae7b05e672f4b79307580413d2c37054","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/e682afeaae7b05e672f4b79307580413d2c37054","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:feat/papilo-presolve","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Afeat%2Fpapilo-presolve","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-09-10T11:44:14.304Z","scheduled_at":"2026-09-10T11:44:14.325Z","started_at":"2026-09-10T11:44:24.107Z","canceled_at":null,"finished_at":"2026-09-10T12:23:30.373Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"},{"id":"01a0f77e-5be6-4e0a-a53b-1e6325120660","number":377,"message":"feat: let a Reactant-compiled solve take cuSPARSE matrices","state":"passed","blocked_state":null,"path":"/julialang/coolpdlp-dot-jl/builds/377","commit_id":"f552bd0af316e12e9533ef392f0760e1dd494b5c","commit_short_length":7,"commit_url":"https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/commit/f552bd0af316e12e9533ef392f0760e1dd494b5c","author_name":"Guillaume Dalle (bot)","author_avatar":"https://avatars.githubusercontent.com/u/185490550?v=4","branch_name":"gdalle-bot:gd/reac2-cusparse","branch_path":"/julialang/coolpdlp-dot-jl/builds?branch=gdalle-bot%3Agd%2Freac2-cusparse","github_workflow_name":null,"github_workflow_path":null,"github_workflow_url":null,"github_workflow_event_description":null,"created_at":"2026-10-01T12:44:08.558Z","scheduled_at":"2026-10-01T12:44:08.603Z","started_at":"2026-10-01T12:44:17.068Z","canceled_at":null,"finished_at":"2026-10-01T13:01:52.873Z","cancel_reason":null,"project":{"provider":{"frontend_icon":"github"}},"merge_queue_base_branch":null,"merge_queue_path":"/julialang/coolpdlp-dot-jl/builds"}]}