compile_all: a template for compiling everything a package provides

[sshin23]

Stacked on #311 (base is that branch) so it can land after it, or fold in if you'd rather.

ExaModelsCompiler declares the verb and owns the keyword contract; a package that ships models implements it for itself in an extension, so providing models never costs a dependency on a compiler toolchain — a benchmark set is a modelling package, and juliac is not something you should acquire by loading one.

compile_all(LuksanVlcekBenchmark; path = "@lvb", sizes = 1000)

Callers name the package; providers implement compile_all(::Val{TheirModule}; …). Val is the uniform handle — every provider spells it the same way, and it works for a package that has no marker type of its own (ExaModelsPower has none today). The module method is a one-line wrapper over it.

One library rather than one per model, because sizes are deferred: the models share a runtime, one compile amortises across all of them, and a consumer selects with CNLPModel(lib, :bearing, n).

Verification

Against LuksanVlcekBenchmark on its recipes branch, with a draft extension over its eighteen models: compiled :rosenrock and :broyden_banded into one library, instantiated both at a size never compiled (37), exact against the in-Julia models on objective and gradient. ExaModelsCompiler suite 246/246.

The docstring carries a complete worked extension (weakdeps stanza, ext module, the argument-function spelling for data-defined models) so a provider implements it by copying. @ohm and @nabla are testing it against ExaModelsPower and COPS/LVB respectively — emp is the interesting case, since its models take an argument function rather than sizes.

🤖 Generated with Claude Code

[github-actions]

Your PR requires formatting changes to meet the project's style guidelines.

Please run:

julia --project=@runic -e 'using Pkg; Pkg.add("Runic")'
julia --project=@runic -e "using Runic; exit(Runic.main(ARGS))" -- --fix <files>

(or git runic anvil/rename-compiler if you have the git wrapper installed)

Note: the full diff is omitted because it can exceed GitHub Actions input limits.

[github-actions]

Benchmark Results

Relative timing: current / main  (values < 1.0 are improvements)

================================================================================
  backend-instance-param          |      obj     cons     grad      jac     hess
================================================================================
  CUDA-OPF-case1354               |    1.003    0.990    0.992    0.978    0.982
  CUDA-OPF-case14                 |    1.023    0.976    0.987    0.990    0.992
  CUDA-OPF-case30000              |    1.051    0.978    0.980    0.995    0.986
  CUDA-chain-10                   |    1.019    0.999    0.982    0.985    0.997
  CUDA-chain-100                  |    0.949    0.989    0.979    0.980    0.984
  CUDA-chain-1000                 |    1.011    1.003    0.994    0.981    0.990
  CUDA-elec-10                    |    1.004    0.985    0.987    0.990    0.988
  CUDA-elec-100                   |    0.974    1.104    1.001    0.995    0.994
  CUDA-elec-1000                  |    1.033    0.939    3.138    0.993   51.854
  CUDA-rosenrock-1000             |    0.988    0.990    0.980    1.001    1.007
  CUDA-rosenrock-10000            |    0.990    0.997    0.982    0.986    0.974
  CUDA-rosenrock-100000           |    1.000    1.000    0.988    0.990    1.013
--------------------------------------------------------------------------------
  AMDGPU-OPF-case1354             |    1.015    0.954    1.049    0.974    1.008
  AMDGPU-OPF-case14               |    1.024    1.015    0.998    0.990    1.007
  AMDGPU-OPF-case30000            |    1.022    1.000    1.000    1.029    1.008
  AMDGPU-chain-10                 |    0.983    0.960    0.989    0.985    0.972
  AMDGPU-chain-100                |    1.004    0.988    0.945    1.002    1.045
  AMDGPU-chain-1000               |    1.003    0.979    0.982    1.033    1.025
  AMDGPU-elec-10                  |    1.041    0.934    0.983    0.996    0.967
  AMDGPU-elec-100                 |    1.034    0.978    1.026    0.996    0.999
  AMDGPU-elec-1000                |    1.013    0.987    1.181    1.003    1.169
  AMDGPU-rosenrock-1000           |    1.003    0.955    0.998    0.966    1.062
  AMDGPU-rosenrock-10000          |    0.950    0.951    0.998    0.974    0.985
  AMDGPU-rosenrock-100000         |    1.440    0.965    1.003    0.960    0.735
--------------------------------------------------------------------------------
  oneAPI-OPF-case1354             |    1.027    2.649    1.497    1.981    1.480
  oneAPI-OPF-case14               |    1.133    0.752    1.574    2.987    2.378
  oneAPI-OPF-case30000            |    4.799    3.819    4.011    1.369    1.312
  oneAPI-chain-10                 |    1.428    5.655    6.817    2.522    6.039
  oneAPI-chain-100                |    1.987    1.404    1.639    1.477    1.191
  oneAPI-chain-1000               |    1.031    1.462    3.264    5.383    2.117
  oneAPI-elec-10                  |    0.891    1.281    2.683    2.781    2.297
  oneAPI-elec-100                 |    0.745    1.619    1.063    2.455    1.443
  oneAPI-elec-1000                |    0.925    1.009    0.964    1.424    0.613
  oneAPI-rosenrock-1000           |    1.252    1.736    2.460    9.476    0.326
  oneAPI-rosenrock-10000          |    4.170    1.085    1.254    1.746    2.065
  oneAPI-rosenrock-100000         |    2.338    1.196    0.859    2.631    1.158
--------------------------------------------------------------------------------
  nothing-OPF-case1354            |    1.027    1.006    1.062    1.026    0.642
  nothing-OPF-case14              |    1.062    1.055    0.967    1.086    1.000
  nothing-OPF-case30000           |    1.007    0.999    1.004    0.990    1.013
  nothing-chain-10                |    1.000    1.011    0.858    0.952    2.230
  nothing-chain-100               |    0.998    0.945    0.879    1.016    1.004
  nothing-chain-1000              |    0.948    0.944    0.943    0.998    1.060
  nothing-elec-10                 |    1.067    1.043    1.010    1.076    1.015
  nothing-elec-100                |    1.006    1.014    1.061    1.057    1.013
  nothing-elec-1000               |    0.999    1.000    1.000    0.999    0.987
  nothing-rosenrock-1000          |    1.013    1.002    1.004    0.976    1.005
  nothing-rosenrock-10000         |    1.003    1.001    0.999    0.980    1.894
  nothing-rosenrock-100000        |    1.010    1.001    0.999    0.982    0.337
================================================================================

[sshin23]

Two commits since the last review state:

• a23e698d drops select — a two-line filter does not earn hub API; providers filter their own lists.
• 4d367dc8 makes the emitted library self-describing, answering the two questions a caller holding only a path could not ask: cnlp_nmodels / cnlp_model_name (the catalogue — the only fixed-name symbols, since every per-model entry point needs a prefix and this is where a prefix comes from) and P_argtype (what a model instantiates from, as a typed signature with optional descriptions — "int|size", "string|…", or a structured model's full field list; every model answers, schema or not, and the signature identifies the entry point without symbol probing).

Gate on the tip: 242/242, including signature assertions for all four instantiation surfaces.

Consumer counterparts are on both masters (CNLPModels.jl d4070e5, cnlpmodels-py 45d3f08): available_models, argtype, schema answering nothing instead of throwing for models without one — plus, Julia-side, the two standard-API methods ExaModel had and CNLPModel lacked: cons_nln! and hprod! (assembled COO Hessian-vector product, the piece matrix-free solvers need).

🤖 Generated with Claude Code

[sshin23]

This merge landed on anvil/rename-compiler after #311 had already carried that branch into main, so none of it reached main — and the catalogue/argtype commit was pushed after the merge snapshot. Superseded by #313, which is the same three commits cherry-picked onto current main and re-gated there (242/242).

🤖 Generated with Claude Code