generic-max-supply

作成者: nvidia

多期間サプライチェーン計画モデル:データファイル、BOM構造、max-supplyベースモデルの変数・制約リファレンス。

npx skills add https://github.com/nvidia/cuopt-examples --skill generic-max-supply

Generic Max-Supply Planning

Core

  • Execution order (mandatory): read scripts/model.py first, then decide relevant CSV files from the prompt, then read only those CSVs, then run baseline model.py, then copy to model_whatif.py, apply a targeted edit, and run model_whatif.py.
  • Source data and reference model: scripts/ for Python (data.py, model.py); scripts/data/ for CSV inputs. load_data(data_dir, num_periods) reads from that data directory; optional cost/backlog files are read separately if needed.

Workspace layout

LocationContents
scripts/Python: data.py (load_data), model.py (build/solve). Run from repository root (or your run directory) with python scripts/model.py.
scripts/data/Required and optional CSV files (items, families, processes, supply, demand, etc.). All files from the data file map live here.

Using in the working directory

Do not modify the skill files. To run or change the model: (1) Copy skills/generic-max-supply/scripts/ into your run directory. (2) Edit only the copy. (3) Run from that run directory or repository root. Do not assume a fixed sandbox cwd.

cp -r skills/generic-max-supply/scripts . && python3 scripts/model.py

(From repository root, this copies skills/generic-max-supply/scripts/ to ./scripts/.)

What-If Variants

To run a what-if scenario (e.g. change opening inventory, tighten a supply cap, modify a cost):

  1. Read model.py first (required), identify which model block changes, and infer which CSV files are relevant to the requested change.

  2. Read only relevant CSV files for the prompt. Do not read all CSVs unless explicitly requested.

  3. Copy the scripts to the working directory and fix the data path:

    cp skills/generic-max-supply/scripts/data.py data.py
    cp skills/generic-max-supply/scripts/model.py model.py
    

    Then update DATA_DIR in the copied model.py to match your CSV location.

  4. Run baseline first and capture base objective:

    python model.py
    
  5. Fork the working model:

    cp model.py model_whatif.py
    
  6. Use edit_file to make only the targeted change in model_whatif.py.

  7. Run the what-if model and compare against baseline:

    python model_whatif.py   # ← this is the what-if result
    

Do not rewrite the model from scratch. One targeted edit keeps the change isolated and the comparison meaningful.

Problem overview

Type: MILP (Mixed-Integer Linear Program) Sense: MAXIMIZE Horizon: configurable num_periods (default 10)

The planner decides how many units of each manufacturing process to run on each resource in each period. The goal is to maximise the weighted sum of finished-good inventory at the end of the last period. Constraints capture multi-level BOM structure, lead-time offsets, integer yield truncation, material supply caps, and machine-hour capacity.

Data file map

All files live in the same directory (scripts/data/). load_data(data_dir, num_periods) reads all required CSVs from that directory; the three optional cost/backlog files are read separately if needed.

FileKey columnsWhat it populates
items.csvitem_id, name, family_idall_items, item_family lookup
families.csvfamily_id, name, is_constrainedconstrained_families, unconstrained_families
processes.csvprocess_id, name, lead_time, hours_per_unitall_processes, process_lead_time, process_hours
process_inputs.csvprocess_id, item_id, quantityprocess_input_qty[(p,i)], item_consuming_processes
process_outputs.csvprocess_id, item_id, quantityprocess_output_qty[(p,i)], item_producing_processes, derives produced_items
process_resources.csvprocess_id, resource_idprocess_resource_pairs, process_to_resources, resource_to_processes
resources.csvresource_id, name, period, available_hoursall_resources, resource_capacity[(r,t)]
supply.csvitem_id, period, quantitysupply_qty[(i,t)] — upper bound on procurement per period
demand.csvitem_id, period, quantity, priority_weightfinal_items, demand_weight[i] (only priority_weight is used in the base model)

Optional files — NOT loaded by data.py; read separately if needed:

FileKey columnsPurpose
item_costs.csvitem_id, unit_cost, holding_costPer-unit purchase cot and per-period holding cost
resource_costs.csvresource_id, production_cost_per_hourCost per machine-hour by resource
backlog_params.csvparameter, valueSingle row: backlog_cost_rate

Key sets and derivation

produced_items      = set(process_outputs_df["item_id"])
procured_items      = all_items - produced_items
final_items         = set(demand_df["item_id"])

constrained_families   = {f where is_constrained == True}
unconstrained_families = {f where is_constrained == False}

# A process is constrained if ANY of its output items is in a constrained family.
process_is_constrained[p] = any(
    item_family[i] in constrained_families
    for i in outputs_of(p)
)

In the sample dataset: FAM_FG, FAM_SA, FAM_RM1 are constrained; FAM_RM2 is not. RM3 (family FAM_RM2) is procured but unconstrained — no supply cap applied. All four processes (PROC1PROC4) are constrained because they produce items in constrained families.

Variable reference

VariableRepresentsDomainInitialisation
x[p, r, t]Units of process p executed on resource r starting in period tContinuous ≥ 0addVariable(lb=0, vtype=CONTINUOUS)
produced[i, t]Continuous (possibly fractional) output of item i in period tContinuous ≥ 0addVariable(lb=0, vtype=CONTINUOUS)
used[i, t]Usable (floor-truncated) integer output of item i in period tInteger ≥ 0addVariable(lb=0, vtype=INTEGER)
buy[i, t]Units of procured item i purchased in period tContinuous ≥ 0addVariable(lb=0, vtype=CONTINUOUS)
inventory[i, t]On-hand inventory of item i at end of period tContinuous ≥ 0addVariable(lb=0, vtype=CONTINUOUS)
inventory[i, 0]Opening balance (fixed to zero)Continuous, lb=ub=0addVariable(lb=0, ub=0, vtype=CONTINUOUS)

inventory[i, 0] is a fixed variable (lb = ub = 0). For a non-zero opening balance, set ub = lb = opening_balance.

Constraint map

LabelName patternWhat it enforcesVariables linked
C1prod_def_{i}_{t}produced[i,t] equals the sum of BOM-output quantities from all processes that started at t − lead_timeproduced, x
C2atrunc_upper_{i}_{t}used[i,t] ≤ produced[i,t] (integer cannot exceed real output)used, produced
C2btrunc_lower_{i}_{t}used[i,t] ≥ produced[i,t] − (1 − ε) with ε = 1e-4 (forces integer to be the floor)used, produced
C3abal_proc_{i}_{t}Material balance for procured items: inv[i,t] = inv[i,t−1] + buy[i,t] − consumptioninventory, buy, x
C3bbal_prod_{i}_{t}Material balance for produced items: inv[i,t] = inv[i,t−1] + used[i,t] − consumptioninventory, used, x
C4supply_{i}_{t}buy[i,t] ≤ supply_qty[(i,t)] — only applied to procured items in constrained familiesbuy
C5cap_{r}_{t}Σ hours_per_unit[p] · x[p,r,t] ≤ capacity[(r,t)] — only processes where process_is_constrained[p] is Truex

Lead-time note: C1 uses start_t = t − lead_time[p]. If start_t < 1, the term is skipped (process cannot have started before period 1).

Objective

last_t = max(data.periods)
obj = Σ  demand_weight[i] * inventory[i, last_t]   for i in final_items

Maximise the priority-weighted inventory of finished goods at the end of the planning horizon. demand_weight comes from the priority_weight column in demand.csv (the quantity column in demand.csv is present but not used by the base model).

Solver settings (default)

settings = SolverSettings()
settings.set_parameter("mip_relative_gap", 0.01)
prob.solve(settings)

mip_relative_gap = 0.01 (1% gap). Do not change solver precision (e.g. do not tighten to 0.1% or 0.0); use this default only.

Sanity-checking results

After solving, verify the model is correct:

  1. Check model size changed when you add items/constraints. The solver log prints Solving a problem with N constraints, M variables. If you added a new item or constraint but N and M are unchanged, your data edits did not take effect — check that you edited the CSV files in the working directory.
  2. Check item count in the Loading data... output matches your expectation (e.g., adding RM4 should increase the item count by 1).
  3. Check objective value actually changed meaningfully. Differences smaller than the MIP gap (~1%) may just be solver noise, not real model changes.

Quick links by task

TaskUse
Implement or extend the base modelData file map, key sets, variable reference, constraint map, objective.
Load or modify input dataData file map; optional files read separately.
Tune or run the solverSolver settings (default); data in scripts/data/ by default.
Verify results after changesSanity-checking results.

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。