InferQuest — verified paths into LLM serving and training
Serve LLMs. Train LLMs. Prove it.
InferQuest is two free roadmaps, built from a close read of what labs currently hire for. One teaches you to serve LLMs fast and cheap in production. The other teaches you to train them as well as possible on cheap hardware. You don’t check off the big milestones yourself; the site verifies them against your real endpoints, kernels, and training runs.
The full curriculum is open to browse — sign in (free) to track progress, take the drills, and unlock the verifiers.
What is InferQuest?
InferQuest is a free, open, non-commercial web application for learning inference engineering and LLM training. It offers two paths — serving large language models fast and cheaply, and training them as good as possible on minimal hardware — organized into quests and tasks. It tracks your progress with XP, levels, and streaks, drills you with graded quizzes and spaced-repetition reviews, and automatically verifies major milestones like deployed endpoints, GPU kernels, training runs, and merged open-source pull requests.
Who it’s for: working software engineers who want to move into ML systems. You should be comfortable programming (Python in particular), and you should have seen a training loop at some point, because terms like training loss and overfitting show up early without much introduction. PyTorch and the matrix math get refreshed along the way, but nothing here teaches programming itself. If you’re starting from nothing, do Karpathy’s Neural Networks: Zero to Hero first and come back.
Signing in (with Google or email) is used only to save that progress to your account — see the privacy policy.
What “verified” means here
Live endpoint probes
Deploy an OpenAI-compatible endpoint (your own engine first, then production vLLM) and InferQuest probes it: streaming framing, usage accounting, max_tokens cutoffs, error shapes, latency targets.
GPU-graded kernels & training runs
A local harness grades your kernels for correctness and measured speed: attention, KV cache, Triton softmax, flash attention, quantizer, ring all-reduce. Training runs get the same treatment — first convergence, a measured ≥1.5× speedup, an adapter fine-tune that keeps its base skills — under fixed token budgets on your own hardware.
Merged-PR checks
The open-source milestones check against the GitHub API that your PRs into vLLM, SGLang, FlashInfer, TRL, torchtitan, nanochat & co. exist, merged, and amount to more than typo fixes.
Graded interview drills
KV-cache sizing math, rooflines, speculative-decoding acceptance, scaling-laws and data-curation calls, parallelism tradeoffs. Graded server-side; the answers never reach your browser.
One trunk, two paths
Everyone starts in Foundations (transformer internals, GPU architecture, kernels), then branches. XP and levels are shared: one ladder from Token to Foundation Model, whichever path you walk. The titles Inference Engineer and Training Engineer are earned separately, as certificates for finishing a path.
Frequently asked questions
What does an inference engineer do?
Inference engineers make large language models fast and cheap to serve in production: writing and tuning GPU kernels, managing KV-cache memory, batching requests, quantizing weights, and operating engines like vLLM, SGLang, and TensorRT-LLM against latency and cost targets. It's one of the fastest-growing specialist roles in AI infrastructure.
What skills do I need to become an inference engineer?
The core inference engineering skills are transformer internals (attention, KV caching, sampling), GPU architecture and CUDA or Triton kernel writing, quantization, continuous batching and paged attention, distributed serving (tensor and pipeline parallelism), and profiling with tools like Nsight. InferQuest's roadmap covers all of these in order, with a verifier gating each major skill.
Can InferQuest teach me to train my own LLM?
Yes — the Model Training path covers exactly that: backprop and optimizers from scratch, data curation with real Common Crawl pipelines, scaling-laws math, the NanoGPT-speedrun efficiency toolkit (Muon, FP8, fused kernels), a GPT-2-class pretraining capstone you can run on one consumer GPU or ~$50 of rented compute, then SFT, LoRA, DPO, and GRPO post-training on a single GPU. It leads to the pretraining, post-training, and RL engineering roles labs are actively hiring for.
Is InferQuest free? Do I get a certificate?
InferQuest is completely free and open. There is no paper certificate — instead, milestones are auto-verified: live probes against your deployed endpoint, GPU-graded kernel submissions, and merged-PR checks against real open-source repos. The result is a portfolio of receipts, which hiring teams weigh far more than a certificate.
How long does the roadmap take?
Both paths together span 183 tasks across 38 quests (21,780 XP), sharing a common trunk of fundamentals. An experienced software engineer studying part-time should expect roughly six months to a year for one path end to end — less if you already know PyTorch and CUDA, since early phases are skimmable.
Was this site written by AI?
Substantially, yes. The curriculum is curated and maintained with heavy AI assistance, and much of the site's text was AI-drafted. Two things are not AI: the sources and the verifiers. Every quest points at the field's primary material — the papers, textbooks, and codebases practitioners actually use — and every major milestone is checked for real: live endpoint probes, GPU-graded kernels, calibrated training runs, merged-PR checks against the GitHub API. Treat the prose as connective tissue. The sources and the verifiers are the curriculum.
Do I need my own GPU?
For the kernel-engineering phases, yes — the grading harness runs on your own hardware, and any modern NVIDIA GPU works. Everything before that (transformer internals, the inference-engine capstone, quizzes and drills) runs on CPU or free cloud notebooks.
Walk into the interview with receipts.
Every verified milestone leaves a receipt: probe results, harness metrics with your GPU’s name on them, merged PRs anyone can look up. By the end you have a portfolio a hiring team can check for themselves.