Matrix multiplication of two tensors. Validates shapes via _validate_matmul_shapes, then computes the product. For 2D matrices, uses

Return local download controls plus an optional visible Markdown fallback.

Render the selected track's cross-lab narrative without registry provenance.

Return version metadata for a lab path such as ``vol1/lab_01_ml_intro.py``.

Render the selected track profile as a student-facing mission.

Return a Marimo radio selector for the canonical student tracks.

Return a Pint Unit from a Pint unit-like object or unit string.

Async load for WASM environments using IndexedDB.

Add repo root to sys.path for editable/local mlsysim imports.

* Clear all records from a store

Apply action and simulate one timestep. Returns: (next_state, reward, done, info)

Creates a Provenance object for peer-reviewed literature or academic whitepapers.

Forward pass: lookup embeddings for given indices. Args: indices: Token indices of shape (batch_size, seq_len) or (seq_l

* Clean up the component

Format a **count** (a dimensionless tally of things), optionally with a magnitude scale glyph. Replaces the ``fmt(value / MILLION, suffi

Format a duration with time-specific checks and defaults. ``fmt_qty`` remains the generic physical-quantity formatter. ``fmt_time`` exists be

Load a YAML file with all hardening rules enforced.

Format a non-physical count throughput such as QPS or tokens/s. Physical rates (``GB/s``, ``TFLOP/s``, ``W``) remain ``fmt_qty`` values. This

Executes the benchmark scenario. Args: query_handler: An asynchronous function that processes a batch of QuerySa

Enable gradient tracking for all Tensor operations. This function enhances the existing Tensor class with autograd capabilities. Call th

Return all track variants for a lab ID.

Format a value as an integer for narrative text. Explicit source-level opt-in for integer display of computed values. Rounds half away f

Returns (train_loader, val_loader) for the given workload. This is the single entry point that auto_trainer.py should use.

A small toggle to reveal the underlying physics.

Creates a Provenance object for widely accepted industry conventions.

True if `pos` falls inside any of the (start, end) spans. Half-open: [start, end). Linear scan; spans are typically <10 per line.

Creates a Provenance object for a hardware datasheet or technical specification.

Ensure a numeric value is non-negative (>= 0).

Ensure tensor has gradient attributes (for tensors created before enable_autograd).

Loads the ledger from the best available persistent storage.

set_book_style() (book Helvetica stack + semantic palette) + margin overrides (~5.5pt, no grid) + the device's native figsize. `svg.fonttype='path

Parse a value as a Quantity and validate its physical dimension. This catches obvious unit errors such as watts in a latency field. The unit-

Creates a Provenance object for expert estimates or rules of thumb.

Advance sequence position after processing current token. Call this after all layers have processed the current token and up

Run a tito command and return (exit_code, stdout, stderr).

* Delete a record

Forward pass through multi-head attention. TODO: Compose the helpers into the complete multi-head attention forward pass AP

Get the module number to folder name mapping (auto-discovered).

Manually register an item.

Return a plain float magnitude for fmt safety checks. All numeric formatters funnel their value extraction through here, so the finite guard

Create a release stage with a stable ID and exact Binder argv.

Assert analytical gradient from backward() matches finite-difference gradient. Both paths go through float32 (Tensor's dtype) so precision i

Load every YAML under vault/questions/. Returns loaded + errors (never raises). v1.1: chains are joined from chains.json after parsing. Question

Measure model inference latency with statistical rigor. TODO: Implement accurate latency measurement APPROACH: 1. R