[FabianHofmann]

Closes # (if applicable).

Changes proposed in this Pull Request

Note

The following content was generated by AI.

Highs._build_solver_model (direct API, also used by linopy.io.to_highspy) built the HiGHS model in several calls: addVars, changeColsIntegrality, changeColsCost, then addRows.

This PR passes the model in bulk and hands HiGHS the cached CSR matrix as is:

• MILP: one h.passModel(...) call (raw-array overload, MatrixFormat.kRowwise) with a full-length int32 integrality array (0 continuous, 1 binary/integer, 2 semi-continuous). HiGHS rejects an empty array.
• Pure LP: one addCols (cost and bounds, no matrix entries), then one row-wise addRows. passModel is not used here because an all-continuous integrality array makes HiGHS write empty bin/gen/semi sections to LP files, which CBC, GLPK and COPT misread.
• Row bounds compare M.sense as integer codes (view(np.uint32)) instead of unicode strings.
• Names, the Hessian and the objective sense work as before.

An earlier version of this PR converted A to CSC first. That held a second copy of the matrix during the build, and CodSpeed flagged ~20% more memory in test_to_solver[highs-*]. Passing the row-wise matrix directly is about as fast and removes that copy.

Benchmark and test details

Synthetic model, 1.26M variables, 4.9M nonzeros. Build stage only, 3 fresh processes per variant, cloud VM. Peak RSS is the increase during the build.

variant	LP build	LP peak RSS	MILP build	MILP peak RSS
master	386–450 ms	315 MB	383–458 ms	321 MB
previous PR head (CSC)	291–320 ms	348 MB	216–251 ms	235 MB
this PR	305–316 ms	311 MB	245–314 ms	184 MB

Solve time on the same model is unchanged.

Tests: test_highs_direct_known_optimum covers a semi-continuous model and a model without constraints. test_to_highspy_pure_lp_writes_no_integer_sections checks that LPs get no empty integer sections. pytest test/test_io.py test/test_optimization.py: 655 passed, 16 skipped. ruff check, ruff format and mypy . pass.

Checklist

• AI-generated content is marked (see AGENTS.md).
• Code changes are sufficiently documented; i.e. new functions contain docstrings and further explanations may be given in doc.
• Unit tests for new features were added (if applicable).
• A note for the release notes doc/release_notes.rst of the upcoming release is included.
• I consent to the release of this PR's code under the MIT license.

---

Generated by Claude Code

[codspeed]

Merging this PR will improve performance by 29.8%

⚡ 6 improved benchmarks
✅ 175 untouched benchmarks
⏩ 181 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_to_solver[highs-knapsack-n=100]	100.3 KB	70.3 KB	+42.71%
⚡	test_to_solver[highs-milp-n=10]	101.6 KB	75.1 KB	+35.24%
⚡	test_to_solver[highs-piecewise-n=10]	101.1 KB	77 KB	+31.26%
⚡	test_to_solver[highs-piecewise-n=1000]	2.5 MB	2 MB	+26.84%
⚡	test_to_solver[highs-knapsack-n=10000]	1.4 MB	1.1 MB	+24.61%
⚡	test_to_solver[highs-milp-n=50]	512.2 KB	428.7 KB	+19.47%

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing perf/highs-bulk-passmodel (6b8cc1a) with master (22a1882)}

Footnotes

1. 181 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[FBumann]

Saw the same speedup in specsolve.

[FabianHofmann]

Saw the same speedup in specsolve.

nice!

[FBumann]

Just to mention it to you right now:

Gurobi has more to gain: in specsolve, one gurobipy.loadModel call made loading 3–6× faster than addMVar + addMConstr (fluxopt/specsolve#1942). That's worth a separate PR.

[FabianHofmann]

Just to mention it to you right now:

Gurobi has more to gain: in specsolve, one gurobipy.loadModel call made loading 3–6× faster than addMVar + addMConstr (fluxopt/specsolve#1942). That's worth a separate PR.

ah wonderful! let's try that out too!