Compiles the mathematical equations into the SciPy format. Args: objective_fn: The mathematical function to MINI

Executes a grid search across the defined space by passing each parameter combination to the provided evaluation function.

Lowers the workload into a hardware-agnostic computation graph. Default implementation uses ``parameters`` and ``inference_flops``

Fetch a persona by its identifier key (cloud, edge, mobile, tiny). Args: key: Persona identifier. Re

Retrieve cached key-value pairs for attention computation. TODO: Return only the valid cached portion for this layer APPROA

Compute gradients (placeholder — gradient computation is separate). For now, this is a stub that students can ignore.

Find maximum values along specified axis. TODO: Find maximum element(s) along axis (same pattern as sum). APPROACH: 1. Use n

Run the CLI with given arguments.

Sum tensor along specified axis. TODO: Sum all elements or along specific axes. APPROACH: 1. Use np.sum with axis and keepdi

Parse the table. Returns True if successful, False otherwise. Issues are stored in self.issues.

* Get count of records in a store

* Close the database connection

* Read a record by ID

Apply standard MLSys plot aesthetics.

Reshape tensor to new dimensions. TODO: Reshape tensor while preserving total element count. APPROACH: 1. Handle both reshap

Solves the Roofline performance profile for a single hardware node. Parameters ---------- model : Workload

Return all registered items.

Update cache with new key-value pairs for given layer. TODO: Efficiently append new K,V to cache without data copying APPRO

Compute mean of tensor along specified axis. TODO: Calculate average of elements along axis (same pattern as sum). APPROACH:

Escape command examples before rendering in Rich tables.

Run a test and track results.

Assert a value is within tolerance of expected.

Attach provenance to a scalar used in registries or analyses.

Resolve a canonical MLSysIM registry string such as `Hardware.Cloud.H100`.

Attach provenance to a unit-bearing pint Quantity (the Quantity analogue of ``sourced``). The Quantity is returned unchanged for arithmetic/``.m_a

Return list of trainable parameters. Returns: List of Tensor objects (weights and biases)

Execute the command with error handling.

Return list of trainable parameters. TODO: Return all tensors that need gradients APPROACH: 1. Start with weight (a

Perform SGD update step with momentum. TODO: Implement SGD parameter update by composing helpers APPROACH: 1. For e

Return the canonical track profile, accepting legacy category aliases.

Return the variant for a lab and canonical or legacy track ID.

Render MLSysIM APIs, registry refs, equations, and assumptions.

Return total number of elements (PyTorch-compatible).

Format a Pint Quantity (or plain number) for narrative text. Returns a MarkdownStr so Quarto inserts the value verbatim (no escape). The

Set layer to evaluation mode.

Attach a unit if val is a raw number; verify dimensional correctness AND convert to ``expected_unit`` if it is already a Pint Quantity.

Generate a deterministic issue ID like 'vol1-h3-titlecase-0042'.

Run the CLI application.

Ensure a numeric value is >= minimum.

Runs a full multi-level evaluation of the scenario.

Clear gradients from all parameters. TODO: Reset all parameter gradients to None APPROACH: 1. Iterate through all p

Canonical currency formatter — the single, blessed way to render any dollar amount in prose, tables, or callouts. This is the currency m

Format a pint Quantity in ``display_unit`` with a canonical unit suffix. Required OUTPUT path for physical quantities in LEGO cells. The valu

Build a Markdown report and JSON snapshot for a lab submission.

Forward pass through linear layer. TODO: Implement y = xW + b APPROACH: 1. Matrix multiply input with weights: xW

Set layer to training mode.

Transpose tensor dimensions. TODO: Swap tensor dimensions (default: swap last two dimensions). APPROACH: 1. If no dims speci

Alias for reshape, matching PyTorch's Tensor.view() for contiguous tensors.

Pick precision when caller passes precision=None. Auto mode should produce the shortest display that still carries the value's useful magnitu