Recommended Reading

AI Engineering by Chip Huyen

Choose this broad survey when the next gap spans the full practice of building applications with foundation models.

Build a Large Language Model (From Scratch)

Sebastian Raschka offers a hands-on route to a detailed understanding of large language model architecture.

AI Systems Performance Engineering by Chris Fregly

Choose this performance-focused text to go deeper on model workloads, GPUs, CUDA, and PyTorch.

Baseten blog

Use the book's linked blog as one route to newer implementation notes, while checking dates and applicability to your stack.

Transformer foundations

Read the original attention architecture to connect token interactions, parallel training structure, and the autoregressive serving path.

Input-output-aware attention

Use the first FlashAttention paper to understand why reducing memory traffic can matter more than reducing the arithmetic count.

Diffusion foundations

Study denoising diffusion to see why image and video generation repeat a model over a schedule instead of emitting one token at a time.

Frozen-encoder multimodality

BLIP-2 is a useful bridge for understanding how visual encoders and language models exchange representations without retraining every component together.

Selective state spaces

Mamba offers a contrasting sequence model for asking which transformer costs are architectural necessities and which are design choices.

vLLM

A broad serving baseline for learning continuous batching, paged attention state, and common application programming interfaces.

SGLang

A serving project worth comparing for model support, structured generation, caching, and mixture-of-experts execution.

TensorRT-LLM

NVIDIA's engine documentation exposes architecture-tuned kernels, low precision, and build-versus-runtime trade-offs.

PyTorch Profiler

Use the profiler recipe to attribute host and accelerator time only after a stable workload reproduces the target regression.

CUTLASS

Explore kernel building blocks when framework-level evidence shows that matrix operation shape, tiling, or precision is the remaining constraint.

DeepSeek AI collection

A model family for studying dense and sparse language-model serving choices across reasoning and general workloads.

Alibaba Qwen collection

A broad family spanning language and multimodal checkpoints, useful for comparing sizes and task-specific variants.

Mistral AI collection

A source of open language and sparse expert models for evaluating capability, size, and serving topology together.

Black Forest Labs collection

An image-generation family for examining pipeline components, model variants, memory demand, and license constraints.

OpenAI model collection

Includes Whisper checkpoints that ground automatic speech recognition experiments in a widely implemented reference family.

NVLink and NVSwitch

Use the interconnect overview to understand the topology assumptions behind tensor and expert parallel deployments.

Kubernetes documentation

Study the orchestration primitives that place, restart, expose, and roll out stateful GPU-serving workloads.

Designing Data-Intensive Applications

Use the systems text to strengthen reasoning about replication, partitioning, consistency, streams, and failure beyond model execution.

Site Reliability Engineering

Apply service objectives, error budgets, monitoring, and incident response to inference as a user-facing production dependency.

Paged attention memory management

Read the vLLM systems paper when attention-cache fragmentation or scheduler utilization is the suspected constraint.

Speculative decoding

Study draft-and-verify generation when decode is serial and extra parallel compute may buy lower per-user token latency.

Activation-aware weight quantization

Use AWQ to examine how calibration and sensitive channels affect the quality-performance trade-off of low-bit weights.

FlashInfer

Explore a customizable attention-kernel library when profiles point to serving attention rather than general framework overhead.

A pragmatic disaggregation study

Use the systems analysis to challenge whether separating prefill and decode improves the actual workload after state-transfer costs.

Massive multitask language understanding

A broad knowledge benchmark useful for learning the strengths and limitations of aggregate academic-task scores.

HumanEval

A compact code-generation reference for studying functional correctness, sampling, and benchmark contamination concerns.

Massive text embedding benchmark

Use MTEB to compare representation models across retrieval, similarity, classification, clustering, and other embedding tasks.

SWE-bench

A repository-level software task benchmark that illustrates the gap between isolated code snippets and tool-using workflows.

ARC Prize

A route into evaluations of abstraction and generalization that do not reduce capability to familiar language tasks.

Orientation

Printed page 233 explains why continued learning must follow a fast-changing field.

Architecture

Printed pages 234 to 238 collect foundations for language, vision, audio, attention, diffusion, and sequence models.

Developer tools

Printed pages 239 to 242 collect runtimes, GPU libraries, profilers, and model tooling.

Frontier open models

Printed pages 243 to 245 index open model publishers across several modalities.

GPU infrastructure

Printed pages 246 to 248 link accelerator systems with orchestration and reliability resources.

Inference optimization

Printed pages 249 to 253 survey research on quantization, caching, attention, speculation, parallelism, and disaggregation.

Intelligence evaluation

Printed pages 254 to 255 collect evaluations for general reasoning, code, mathematics, embeddings, and applied tasks.