cuopt-model-mapper

作成者: nvidia

解釈された最適化問題を、最小限の確認質問で高速パス用のcuOptネイティブモデルにマッピングします。

npx skills add https://github.com/nvidia/cuopt-examples --skill cuopt-model-mapper

cuOpt Model Mapper

Use this skill in the fast path after the request has already been identified as an optimization problem and the data has been interpreted enough to support model construction.

This skill takes the working interpretation of the problem and maps it directly into cuOpt-native model objects.

Purpose

The fast path should avoid unnecessary architecture.

This skill exists to:

  • build directly in cuOpt
  • avoid a heavyweight intermediate model
  • ask only the smallest number of blocking questions
  • solve and return the result quickly

This skill is not for replayable/auditable artifact design. That belongs to a different path.

Preconditions

Use this skill only when the following are already mostly clear:

  • the request is genuinely an optimization problem
  • the problem family is identified as LP / MILP / QP / routing
  • the uploaded data has been inspected enough to infer likely roles
  • only a small number of unresolved modeling questions remain

If those conditions are not met, first use:

  • optimization-intent-router
  • optimization-mode-router if execution mode may matter for replayability, audit, export, or reuse
  • tabular-optimization-ingestion

Core rule

For the fast path, map directly from the interpreted data into cuOpt structures. Do not introduce a replayable intermediate artifact unless the user asks for replayability, auditability, export, or reuse.

In NemoClaw sandbox: before building the cuOpt model, confirm cuopt-sandbox gates completed (probe → select gRPC capability → smoke test). Do not build a parallel heuristic assigner "first" — cuOpt is the first and only solver for assignments/schedules.

Workflow

1. Confirm the minimum viable formulation

Before building, confirm internally:

  • what the decisions are
  • what the objective is
  • what the hard constraints are
  • whether integrality is required
  • whether the problem is numerical optimization or routing

Use the unresolved blocker list from ingestion as the starting point; do not reopen broad exploratory questioning unless the current interpretation is clearly inconsistent.

If one non-retrievable modeling choice would change the meaning of the solve, ask exactly one concise blocking question.

Scheduling with no stated objective: Feasibility (no double-booking, respect unavailability, assign every item) belongs in hard constraints. Do not deliver a greedy feasible schedule first. Ask one objective question if needed, or state a default secondary objective (e.g. balance slot times, minimize penalty slacks), then solve with cuOpt.

Examples:

  • "Do these production quantities need to be whole numbers?"
  • "Must all demand be met, or can unmet demand be allowed with a penalty?"
  • "Are these time windows mandatory, or just preferred?"

2. Choose the cuOpt path

If LP / MILP

  • hand off through numerical-optimization-formulation for formulation discipline
  • use cuopt-numerical-optimization-api-python (or cuopt-numerical-optimization-api-cli for MPS inputs)
  • preserve direct mappings from source data into variables, coefficients, RHS values, bounds, and objective terms

If QP

  • hand off through numerical-optimization-formulation for formulation discipline
  • use cuopt-numerical-optimization-api-python
  • preserve direct mappings from source data into variables, quadratic objective terms, RHS values, and bounds

If routing

  • hand off through routing-formulation
  • use cuopt-routing-api-python
  • map source data directly into locations, demands, vehicles, travel metrics, time windows, and pickup-delivery relationships

3. Preserve source-to-model traceability lightly

Even in fast mode, keep enough working traceability to avoid confusion during the same interaction.

At minimum, be able to state:

  • which tables/columns drove the objective
  • which tables/columns drove major constraints
  • what each decision variable represents

Do this without building a full replayable artifact.

4. Solve with cuOpt

  • use cuOpt whenever available and appropriate for the identified problem family
  • verify solver status
  • verify that the result is sensible against the modeled constraints

5. Explain results in user language

Return:

  • recommended decisions
  • objective value
  • what the objective means in business terms
  • any important bottlenecks or binding tradeoffs
  • key assumptions that materially affect the answer

Direct mapping guidance

Numerical optimization mapping

Typical direct mappings include:

  • per-row profit/cost columns → objective coefficients
  • per-row resource consumption columns → constraint coefficients
  • capacity table values → RHS limits
  • min/max fields → variable bounds or additional constraints
  • yes/no/open-close decisions → binary variables
  • count decisions → integer variables

Routing mapping

Typical direct mappings include:

  • customer/location rows → stops
  • coordinate or matrix tables → travel metric inputs
  • demand fields → stop demand
  • vehicle table → fleet definition
  • depot fields → start/end nodes
  • time window columns → service time constraints
  • pickup/delivery ids → paired stop relationships

Examples

Example 1: direct LP/MILP mapping

Interpreted problem:

  • one row per product
  • profit column is objective coefficient
  • labor_hours and steel_units are resource coefficients
  • a separate capacity table provides RHS limits

Fast-path mapping:

  • create one decision variable per product
  • maximize total profit
  • add one constraint per capacity/resource
  • if products must be whole units, use integer variables

Example 2: direct routing mapping

Interpreted problem:

  • one row per customer stop
  • vehicle table defines fleet
  • travel-time matrix defines movement cost
  • customer demand and time-window fields are available

Fast-path mapping:

  • map stops, fleet, and travel data directly into cuOpt routing inputs
  • ask only the final blocker question if needed, such as whether time windows are hard constraints
  • solve and explain route recommendations

Example 3: do not overbuild

Interpreted problem:

  • user wants a one-off answer from uploaded data
  • no replay, export, or audit requirement is mentioned

Correct behavior:

  • map directly into cuOpt
  • do not stop to build a replayable structured artifact
  • keep only enough traceability to explain the result clearly in the current interaction

Integrality rule

Do not default to continuous variables when the decisions represent discrete units, counts, assignments, vehicles, workers, items, or yes/no choices.

Ask if unclear, but if the problem statement strongly implies discreteness, model it as MILP rather than LP.

Fast-path guardrails

  • Do not build a heavyweight intermediate representation for one-off solves.
  • Do not skip a critical clarifying question if it changes the model meaning.
  • Do not ask a long questionnaire when one focused question will unblock the solve.
  • Do not use cuOpt for tasks that are really descriptive analytics.
  • Do not force routing when the problem is actually allocation without path construction.
  • Do not force QP when the objective is linear.

Result reporting requirements

Always report at least:

  • solver status
  • objective value
  • brief explanation of what the objective means
  • recommended decisions
  • any assumptions or caveats that materially affect the result

If relevant, also mention which constraint or resource appears to be most limiting.

Handoff guidance

  • For LP / MILP:

    • use numerical-optimization-formulation
    • then use cuopt-numerical-optimization-api-python (or cuopt-numerical-optimization-api-cli for MPS inputs)
    • follow cuopt-user-rules
    • in sandbox contexts, follow cuopt-sandbox (gates + selected gRPC path) before any LP/MILP solve
  • For QP:

    • use numerical-optimization-formulation
    • then use cuopt-numerical-optimization-api-python
    • follow cuopt-user-rules
    • in sandbox contexts, follow cuopt-sandbox (gates + selected gRPC path) before any QP solve
  • For routing:

    • use routing-formulation
    • then use cuopt-routing-api-python
    • follow cuopt-user-rules
    • in sandbox contexts, follow cuopt-sandbox (gates + REST) before any routing solve

Success criterion

This skill succeeds when the agent can go from:

  • interpreted data
  • plus zero or a few blocking clarifications

to:

  • direct cuOpt model construction
  • a validated solve
  • and a clear business-facing answer

without introducing unnecessary replay/audit machinery.

nvidiaのその他のスキル

compileiq-debug
nvidia
何かがおかしいときに使用:Search()がハングする、すべての評価がINVALID_SCOREを返す、スコアが改善しない、すべての設定が同じ数値を返す、ptxasエラー…
create-github-pr
nvidia
gh CLIを使用してGitHubのプルリクエストを作成します。ユーザーが新しいPRを作成したい、コードをレビューに提出したい、またはプルリクエストを開きたい場合に使用します。トリガーキーワード -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
他のオープンなIssueをスキャンし、特定のPRが修正する可能性があるものや、誤って壊す可能性があるものを見つけます。隣接修正の機会や矛盾リスクをfile:line…と共に出力します。
fhir-basics
nvidia
エージェントにFHIR R4 APIの動作方法、利用可能なリソース、検索パラメータを使ったクエリ方法、およびすべてのレスポンス形式を正しく解析する方法を教えます…
compileiq-validate-result
nvidia
検索が完了した後、かつスピードアップの申請やACFの発送の前に使用します。dump_results CSVを読み込み、トップK候補(単一目的)を抽出します…
changelog-audit
nvidia
リリース前にWarp CHANGELOG.mdを監査:失われたエントリを復元、ユーザー影響で並べ替え、エントリの文言を洗練、行折り返し、および(リリースブランチモードで)比較をバンプ…
maintain-dynamic-plugins
nvidia
NeMo Relayの動的プラグインローダー、マニフェスト、RustネイティブSDK、gRPCワーカープロトコル、PythonワーカーSDK、ドキュメント、テスト、およびリリースワークフローのカバレッジを維持する
dgx-diagnose
nvidia
一般的なDGX Station GB300の問題(CUDAクラッシュ、誤ったGPUターゲット、vLLM/SGLangコンテナのバグ、MIG状態の問題、NVLink/Fabric Managerエラーなど)を診断します。