Educational tensor - the foundation of machine learning computation. This class provides the core data structure for all ML operations: - dat

Linear (fully connected) layer: y = xW + b This is the fundamental building block of neural networks. Applies a linear transformation to

ReLU activation: f(x) = max(0, x) Sets negative values to zero, keeps positive values unchanged. Most popular activation for hidden laye

One scope inside a binder check group. A "scope" is a single check that lives inside a group (e.g. `refs.citations`, `prose.contractions`). S

Stochastic Gradient Descent with momentum. SGD is the foundational optimization algorithm that moves parameters in the direction opposit

Mean Squared Error loss for regression tasks.

2D Convolution layer for spatial feature extraction. Implements convolution with explicit loops to demonstrate computational complexity

Provenance for registry entries (hardware, models, fabrics).

Data loader with batching and shuffling support. Wraps a dataset to provide batched iteration with optional shuffling. Essential for eff

One field inside a BibTeX entry.

Sigmoid activation: σ(x) = 1/(1 + e^(-x)) Maps any real number to (0, 1) range. Perfect for probabilities and binary classification.

A single editorial issue found by the scanner. Fields map 1:1 to the JSON schema in Pass 15 plan section 3. Required fields (scanner must se

Dataset wrapping tensors for supervised learning. Each sample is a tuple of tensors from the same index across all input tensors. All te

2D Max Pooling layer for spatial dimension reduction. Applies maximum operation over spatial windows, preserving the strongest activatio

Professional-grade ML model profiler for performance analysis. Measures parameters, FLOPs, memory usage, and latency with statistical rigor.

The 'Save Game' manager for the MLSys curriculum. Ensures that decisions made in Lab 00 persist through Lab 32.

Multi-head attention mechanism. Runs multiple attention heads in parallel, each learning different relationships. This is the core compo

Named rate (pint quantity) with provenance for economics tables.

Learnable embedding layer that maps token indices to dense vectors. This is the fundamental building block for converting discrete tokens

How we know a numeric value (package audit trail; not BibTeX).

Adam optimizer with adaptive learning rates. Adam computes individual adaptive learning rates for different parameters from estimates of

Result of a single preflight check.

Analyzes the two-phase LLM serving lifecycle: Pre-fill vs. Decoding. LLM inference is not a single mathematical operation; it is a stateful

Complete training orchestrator for neural networks. Handles the full training lifecycle: forward pass, loss computation, backward pass,

Resolves single-node hardware Roofline bounds and feasibility. This model handles the 'Iron Law' of machine learning systems, calculatin

Cross-entropy loss for multi-class classification.

Efficient key-value cache for autoregressive generation. Stores K,V matrices for each transformer layer to avoid recomputation during se

Complete Transformer Block with self-attention, MLP, and residual connections. This is the core building block of GPT and other transformer

A string that ALSO renders as raw Markdown when consumed by Quarto/Jupyter. Quarto's inline ``{python} x`` substitution escapes commas and decima

A single ML Systems Wall. Attributes ---------- number : int Canonical wall number (1–22). name : str Short human-rea

Represents the inter-node network interconnect. Captures the topology, raw bandwidth, latency, and oversubscription ratio of the cl

Tanh activation: f(x) = (e^x - e^(-x))/(e^x + e^(-x)) Maps any real number to (-1, 1) range. Zero-centered alternative to sigmoid.

One line item in the preflight report.

Analyzes model compression trade-offs (Accuracy vs. Efficiency). This model simulates the 'Compression Tax' — the accuracy degradation t

Layer D (Systems/Topology): The complete distributed cluster environment. A Fleet composes a `Node` configuration, a `NetworkFabric`, an

Stateful walker over the lines of a QMD file. Use as an iterator. Each `next()` returns `(line, state, line_num)` where `state` is the LineSt

NBGrader integration command group.

Container that chains layers together sequentially. After you understand explicit layer composition, Sequential provides a convenient wa

Resolves fleet-wide communication, synchronization, and pipelining constraints. This model analyzes the constraints of distributed scale for

Softmax activation: f(x_i) = e^(x_i) / Σ(e^(x_j)) Converts any vector to a probability distribution. Sum of all outputs equals 1.0.

Result of a test phase.

A lint violation against the §5 rules.

Char-level decoder-only transformer used for the NanoGPT workload. Default config (~11M params) is the iter-1/iter-3 small variant. Pass Nano

Complete embedding system combining token and positional embeddings. This is the production-ready component that handles the full embedding

Cosine annealing learning rate schedule. Starts at max_lr, decreases following a cosine curve to min_lr over T epochs. This provides agg

Dropout layer for regularization. During training: randomly zeros elements with probability p, scales survivors by 1/(1-p) During infere

Outcome of a progress-sync attempt. ``ok`` -- the request succeeded AND the server confirmed our progress was persist

Calculates Total Cost of Ownership (TCO) including Capex and Opex. Combines hardware costs, energy consumption, and maintenance into a s

Base class for differentiable operations. Every operation that needs gradients (add, multiply, matmul, etc.) will inherit from this clas

Calculates Datacenter-scale Sustainability metrics. Handles Power Usage Effectiveness (PUE), Carbon Intensity, and Water Usage Effective

2D Average Pooling layer for spatial dimension reduction. Applies average operation over spatial windows, smoothing features while reduc

Professional benchmarking system for ML models and operations. Provides latency, accuracy, and memory benchmarking with statistical rigo

Benchmark report for model performance. Measures and stores: - Model characteristics (parameters, size) - Performance metrics (accur

Learnable positional encoding layer. Adds trainable position-specific vectors to token embeddings, allowing the model to learn positiona

Simple 2-layer MLP for benchmarking demonstration. This is a toy model to demonstrate the benchmarking workflow. Students can later appl

Layer Normalization for transformer blocks. Normalizes across the feature dimension (last axis) for each sample independently, unlike ba

Diagnose PDF page-break whitespace issues.

A single StaffML interview question (schema v1.0). Every classification axis is a YAML field. The filesystem path encodes only `track` for na

Decomposes per-accelerator training memory into teachable components. This model answers a different question than ``SingleNodeModel``. Roof

Comprehensive benchmark suite for ML systems evaluation. Orchestrates multiple benchmark types (latency, accuracy, memory, energy) and g

Raised when the Buttondown API returns an error.

Represents a physical server chassis containing multiple accelerators. Essential for modeling the bandwidth gap between fast intra-node

Concrete executable case: a workload on a system with constraints. ``application`` names the product/domain context, while the scenario itse

Batch Normalization for 2D spatial inputs (images). Normalizes activations across batch and spatial dimensions for each channel, then ap

Complete GPT (Generative Pre-trained Transformer) model. This combines embeddings, positional encoding, multiple transformer blocks, and

A transparent chain of models and solvers for macro-level analysis. Parameters ---------- resolvers : list[BaseResolver] Ordered

Reusable storage tier or service in a system design.

Analyzes queueing delays and P99 tail latency for deployed inference models. Models inference servers as M/M/c queues to determine if the de

Binary cross-entropy loss for binary classification.

Character-level tokenizer that treats each character as a separate token. This is the simplest tokenization approach - every character in th

Analyzes the 'Data Wall' — the throughput bottleneck between storage and compute. This model simulates the data pipeline constraints, compar

A single tier in the Hierarchy of Constraints.

MLPerf-style standardized benchmarking for edge ML systems. MLPerf® is a trademark of MLCommons. Used here purely for educational purposes.

Workload representation of an autoregressive Transformer (e.g., LLMs).

Static visual attached to a question. v0.1.2 hardened: kind is a closed enum (svg only), path must match `^[a-z0-9-]+\\.svg$`, alt ≥10 chars,

Container for benchmark measurements with statistical analysis. TODO: Implement a robust result container that stores measurements and metad

A category of preflight checks.

Analyzes production LLM serving with Continuous Batching and PagedAttention. Traditional static batching suffers from severe memory fragment