THE DEV BENCH
⚡

AI & GPU Infrastructure

Run the infrastructure behind AI — NVIDIA cert prep, MLOps, GPU job scheduling with Slurm, and the datacenter operations behind model training and inference.

Start here

AI/ML Vocabulary51 cards — step 1 of 16

Then get hands-on: Slurm lab — 8 scenarios.

AI Infrastructure & Platform Engineer (LLMOps + GPU/Cloud Ops) — learning pathThe combined role in one place — a hands-on GPU/LLMOps lab track (ten runnable runbooks), the concept decks behind the vocabulary, and the cert paths that feed it. Free-tier-first: most labs run on a 16GB card at $0.NVIDIA AI Infrastructure & Operations — learning pathCert roadmap, decks, labs, a skill map, and verified reading — one workspace for the GPU-infra ops skill set.AI Hardware — Rack, Interconnect and Identification — the courseThe physical layer rather than the operations — safety, rack, node, the PCIe/SXM/OAM split, then cabling, testing and bring-up, which are the bulk of the real job. Anchored on NVIDIA's NCP-ARI technician blueprint. Syllabus scaffolded; units not built yet.

Flashcard decks

16 decks · 704 cards · in study order

Vocabulary — start here

1 deck · 51 cards
1AI/ML VocabularyEssential AI/ML terminology -- definitions and practical examples51

LLM foundations

1 deck · 16 cards
2The Numbers: Weights, Parameters, Vectors & Embeddings (the vocabulary everyone assumes you know)Plain-language cards that untangle the two big 'lists of numbers' in LLM work: model weights/parameters (the model's frozen knowledge, loaded into GPU VRAM, shrunk by quantization) versus vectors/embeddings (meaning turned into numbers for semantic search and RAG). Covers tokens, precision, VRAM sizing, vector databases, and cosine similarity, with the one distinction interviewers listen for: weights are the machine, vectors are the data flowing through it.16

The physical stack & what it costs

3 decks · 97 cards
3AI Hardware · Accelerators & SystemsThe physical compute layer: GPU families and what actually separates them, SXM vs PCIe, NVLink domains, DGX vs HGX vs MGX, the AMD and cloud-silicon alternatives, and the consumer-card licensing trap. Decision rules first, specs second. Verified 2026-07-31.384AI Hardware · Fabric, Storage & the Host PlatformEverything around the accelerator that decides whether it's actually busy: InfiniBand vs Spectrum-X vs RoCE, GPUDirect, checkpoint bandwidth and parallel filesystems, PCIe lanes and NUMA, and the power and cooling limits that now gate AI buildouts. Verified 2026-07-31.305AI Infrastructure Economics · Sizing, Cost & TCOThe commercial layer: cloud instance families mapped to real GPUs, why the same card varies 5-8x in price, buy vs rent, full TCO beyond the box, and how to size a cluster and defend the number. Prices verified 2026-07-31 and move monthly — the ratios are more durable than the figures.29

How the models work

1 deck · 30 cards
6Transformers & 'Attention Is All You Need' (self-attention, the architecture, and how modern LLMs evolved)Drillable interview cards on the transformer architecture from the 2017 'Attention Is All You Need' paper up to today's decoder-only LLMs: scaled dot-product and multi-head self-attention, Q/K/V, positional encoding, the encoder-decoder vs decoder-only split, feed-forward blocks, residuals and layer norm, causal masking, why attention replaced RNNs, and modern changes (RoPE, GQA/MQA, FlashAttention, RMSNorm, SwiGLU, pre-norm). Explains the O(n^2) cost that motivates the whole serving-optimization stack.30

Model serving

1 deck · 35 cards
7vLLM Production Serving (PagedAttention, batching, tensor parallelism, quantization)Drillable interview cards on operating self-hosted LLM serving engines - vLLM and SGLang - covering PagedAttention, RadixAttention, KV-cache, continuous batching, tensor/pipeline parallelism, quantization (AWQ/GPTQ/FP8), the serving knobs, and benchmarking throughput and latency under load.35

LLM app orchestration

1 deck · 34 cards
8LLM App Orchestration (LangChain, LangGraph, LlamaIndex, agents, RAG)Drillable interview cards on the LLM application layer that sits above the serving engine: LangChain (LCEL, chains, tools), LangGraph (stateful agent graphs, checkpointing, human-in-the-loop), LlamaIndex (RAG-first indexing and query engines), RAG pipeline design (chunking, retrieval, reranking, evaluation), agent patterns (ReAct, tool calling, multi-agent), MCP, and when to skip the framework and build directly on the API.34

GPU operations

1 deck · 39 cards
9NVIDIA GPU Ops Stack (CUDA, Triton, DCGM, MIG, NIM)Drillable interview cards on the hands-on NVIDIA GPU operations stack - CUDA, NVIDIA Container Toolkit, Triton, Kubernetes GPU Operator, DCGM, MIG, and NIM - focused on deploying, monitoring, partitioning, and serving.39

Cloud AI platforms compared

1 deck · 31 cards
10Cloud AI/ML Platforms Compared — SageMaker vs Vertex AI vs Azure ML (+ Bedrock)The cross-cloud translation layer for AI/ML platform interviews: what SageMaker, Vertex AI and Azure ML each actually are, component by component; how the managed foundation-model APIs (Bedrock, Gemini/Model Garden, Azure OpenAI + AI Foundry) compare; and when a managed service beats self-hosting vLLM on Kubernetes. Built for the question you keep getting asked.31

Legacy cert decks — being merged

4 decks · 313 cards
11NVIDIA NCA-AIIO: AI Infrastructure & Operations (Associate)Cert prep for the NVIDIA-Certified Associate AI Infrastructure & Operations exam (NCA-AIIO): 50 MCQ, 60 min, no labs. Exam-style multiple-choice cards weighted by the official domains (Essential AI Knowledge 38%, AI Infrastructure 40%, AI Operations 22%), each answer with a plain-language jargon decoder.5012AI-300: MLOps & GenAIOps (Microsoft)Comprehensive prep for Microsoft AI-300 (Machine Learning Operations Engineer Associate). Covers all 5 exam domains: MLOps infrastructure, model lifecycle, GenAIOps infrastructure (Microsoft Foundry), generative AI quality assurance, and RAG/fine-tuning optimization. Targets ~100 cards weighted by official domain percentages.11613AWS ML Engineer MLA-C01AWS Certified Machine Learning Engineer Associate -- all 4 domains11114Google Cloud MLGoogle Professional Machine Learning Engineer -- Vertex AI, BigQuery ML, TFX36

Interview prep

2 decks · 58 cards
15AI Infra & Ops — Interview DrillActive-recall drill for a GPU / AI-infrastructure operations interview: the behavioral and technical-conversational questions these screens ask, each with a model answer. Say your answer out loud before flipping.2116AI Ops & Infra — Resume Talk TrackTalk-track drill for a conversational AI-operations/infrastructure screen, weighted to a DevOps/AI-ML resume: containers & Kubernetes, CI/CD & IaC, RAG/MLOps, honest GPU ops, cleared/regulated work, Python building blocks, and STAR stories. Say your answer out loud before flipping — you're rehearsing fluency, not memorizing.37

Slurm lab

live
Job scheduling on a real single-node cluster — the scheduler that runs GPU clusters. Learn mode covers sbatch, resources and the queue; Break-fix mode fixes drained nodes, impossible resource requests and dead daemons.

Curated resources

verified July 2026

The core subject for the AI Infrastructure & Platform Engineer role. Vendor and project docs lead — this stack moves fast enough that third-party tutorials go stale within months. The two papers are worth reading in full; everything else is reference you will return to.

Model serving — the centre of the role

vLLM documentationfree · officialThe reference implementation for high-throughput LLM serving — PagedAttention, continuous batching, tensor parallelism, quantization, and the OpenAI-compatible server you will actually deploy.Efficient Memory Management for LLM Serving with PagedAttention (the vLLM paper)free · paperWhy vLLM exists: KV-cache fragmentation is the bottleneck, and paging it like virtual memory is the fix. Read this once and the serving vocabulary stops being jargon.SGLang documentationfree · officialThe main alternative serving runtime — RadixAttention and structured output. Knowing where it beats vLLM is a real interview differentiator.NVIDIA NIM documentationfree · officialNVIDIA's packaged inference microservices — the enterprise-deployment shape of everything above, and what a defence or regulated customer is most likely to be running.Ray Serve documentationfree · officialMulti-model composition and autoscaling across a cluster — the layer above a single serving engine when you need to route between models.

GPU operations

NVIDIA MIG User Guidefree · officialMulti-Instance GPU end to end — partitioning a single A100/H100 into isolated instances, which is how you get multi-tenancy out of expensive hardware.NVIDIA DCGM documentationfree · officialThe standard for GPU telemetry, health checks and diagnostics — what you wire into Prometheus when someone asks 'is the fleet healthy?'Scheduling GPUs (Kubernetes docs)free · officialHow GPUs are actually requested and allocated in a cluster — device plugins, resource limits, and the constraints that surprise people the first time.

How the models work

Attention Is All You Need (the transformer paper)free · paperThe 2017 paper every model on this page descends from. You do not need the maths to benefit — read it for the architecture and the vocabulary.Hugging Face Transformers documentationfree · officialThe practical library reference — model classes, tokenizers, and the loading/quantization mechanics you hit as soon as you serve something real.

LLM application layer

LangGraph documentationfree · officialStateful agent and workflow orchestration — the current serious answer to 'how do I build something more than a single prompt call'.LlamaIndex documentationfree · officialThe RAG-focused framework — ingestion, indexing, and retrieval patterns, which is where most production LLM work actually lives.

Certification

NVIDIA-Certified Associate: AI Infrastructure and Operations (NCA-AIIO)free · officialThe official exam page — objectives, format and booking for the entry cert that maps most directly to this subject.