Exam · Inference engines explained

Take the exam and earn your certificate.

A real, server-scored exam for this course. Score strictly above 80% to pass and get a genuine, verifiable certificate.

Before you start

We ask for your name and email first, the same as a PDF download, so we know you are here and your certificate can carry the right name.

We privately record your name, email and request IP for this exam request. Read the privacy notice and retention policy.

Visit Mickai on LinkedIn

Complete the security check before submitting.

Cloudflare Turnstile helps prevent automated submissions. How this check uses data.

You can retake the exam any time. Each attempt draws a fresh set of questions.

Written practice, ungraded

Write your response in your own notes, then compare it with the sample answer. This practice is separate from the scored exam and does not affect your certificate.

Practice 1

Before using an unfamiliar model download, prepare a short inspection note. List the three main kinds of files or information you expect to find, identify one format-related loading risk from the chapter, and state what you would verify before running it.

Check your answer against these points:

  • Names weights, tokenizer and model configuration as separate parts.
  • Identifies the stated pickle loading risk and a cautious alternative.
  • Includes a source or documentation check before execution.
Compare with a sample answer

I would identify the weight files and their formats, the tokenizer files, and the configuration that describes the architecture and context length. A pickle-based checkpoint can execute code when loaded, so I would prefer a safer format such as safetensors and check the publisher and model documentation before running anything. A file list alone does not establish the model licence or that the files are trustworthy.

Ungraded paper practice. This is a pre-run inspection plan based on the chapter examples, not a security audit of a real download.

Practice 2

Use the chapter’s illustrative assumptions to estimate how many 8,192-token chats fit on a 16 GB card. Show weights, cache and allowance separately, apply the 90% usable-memory margin, then explain one reason the estimate is only a planning example.

Check your answer against these points:

  • Keeps fixed weights separate from per-chat KV cache.
  • Applies the stated allowance and 90% usable-memory margin.
  • Labels the result as an estimate under stated assumptions.
Compare with a sample answer

Under the chapter assumptions, weights use 4.00 GB and each 8,192-token chat adds about 1.07 GB of cache. The worked example fits eight chats: 4.00 + 8.59 + 1.26 = 13.85 GB, below the 14.40 GB usable limit. Nine need about 15.03 GB and exceed it. The allowance is a stated planning assumption and real engine overhead and measured cache requirements can differ.

Ungraded paper practice. Use decimal GB and the chapter’s illustrative cache and allowance rules; a real engine must be measured.

Practice 3

A fictional team can run an 8-billion-parameter model either at 16-bit or 4-bit precision. Compare the chapter’s approximate weight-memory figures, then write a small test plan for deciding whether the lower-precision version is suitable for one real task.

Check your answer against these points:

  • Compares the supplied approximate weight sizes without treating them as total memory.
  • Mentions memory or speed alongside task quality.
  • Proposes representative testing and avoids a universal quality claim.
Compare with a sample answer

The chapter’s simple figures are about 16 GB of weights at 16-bit and about 4 GB at 4-bit, with actual 4-bit storage a little higher due to extra values. I would run both versions on representative examples from the intended task, compare outputs against a stated quality check, and record memory and speed on the target hardware. I would not assume the lower-precision version is acceptable for every task.

Ungraded paper practice. The figures are approximate weight sizes, not complete runtime memory or a reliable quality comparison.

Practice 4

Design a paper measurement sheet for two inference runs: one with a short prompt and one with a longer prompt, requesting the same reply length. Specify what to record for time to first token and for the streaming phase, then predict which phase is more directly affected by the longer prompt and why.

Check your answer against these points:

  • Separates time to first token from the streaming decode interval.
  • Connects longer prompts primarily with additional prefill work.
  • Records conditions and treats measured results as setup-specific.
Compare with a sample answer

For each run I would record the exact prompt length, time until the first output token, and the time from the first token to the final token. The longer prompt is expected to increase prefill work and often the time to first token. Decode generates sequentially and is affected by memory bandwidth and other setup details, so I would compare observed streaming rates rather than infer them from prompt length alone.

Ungraded paper practice. This is a measurement plan and a qualitative prediction, not a performance claim.

Practice 5

A small team handles sensitive prompts, expects occasional bursts of work, and has one machine with limited GPU memory. Compare a local setup with a hosted API for this scenario. State which trade-offs need checking before choosing, and avoid assuming either option is automatically preferable.

Check your answer against these points:

  • Compares data route, hardware or scale, cost shape and upkeep.
  • Connects burst demand and limited GPU memory to capacity and performance questions.
  • Makes the decision conditional on sensitivity and measured needs.
Compare with a sample answer

A local setup can keep prompts on the team’s hardware if configured to stay there, but model size and speed are limited by available hardware and the team handles upkeep and security. A hosted API can offer provider capacity and avoid buying hardware, but prompts go to the provider and the team must review the provider’s terms and costs. I would measure burst volume, required model size, memory and latency, and assess the sensitivity and acceptable data route before deciding.

Ungraded paper practice. The answer should use the course’s comparison dimensions; it does not assess any current provider or make a recommendation for real sensitive data.

Practice 6

Write a short safe-access checklist for a local model server that a colleague may need to use. Cover the bind address, what to add before network access, one API-key limitation to verify, risky built-in tools, prompt logging, and how you would check the final exposure.

Check your answer against these points:

  • Distinguishes localhost binding from exposure on all interfaces.
  • Includes authentication/encryption and verifies API-key route coverage.
  • Addresses tools and prompt logging, then includes an exposure check.
Compare with a sample answer

Keep the server on 127.0.0.1 unless remote access is needed. For remote use, place an authenticating and encrypted reverse proxy in front instead of opening the port directly, and check which routes an API key protects. Do not enable file or shell tools on a reachable server without a justified and reviewed design. Keep prompt-content logs off by default, protect and expire any required logs, update the engine, and inspect listening addresses and logs as part of the check.

Ungraded paper practice. This is a planning checklist from the supplied course, not a security certification or a substitute for testing the actual engine and network.

About this exam

Is this exam graded?

Yes. Every question is scored on our server, never in your browser, and you need a score strictly above 80% to pass. The questions are a real, reviewed bank for this course, drawn fresh for each attempt.

Is it only multiple choice?

Yes. This is a multiple-choice knowledge assessment, with at least one question from each chapter in every attempt. Written exercises are separate practice and are not automatically scored. Passing does not establish practical professional competence or independently verify identity.

Is there a time limit?

Yes, 45 minutes, shown as a countdown once you start. If time runs out, whatever you have answered is submitted automatically.

Do I need to download the PDF first?

No. This course's full text is on its own page, so you can read it online and come straight here when you feel ready.

What do you do with my name and email?

We privately record your name, email, request IP, course and privacy acknowledgement for this request. Newsletter consent is separate and optional. See /privacy.html for storage and retention. A passed exam creates a server-stored named certificate. An email confirmation is not identity assurance.

Can I retake it if I fail?

Yes, straight away. Each attempt draws a fresh set of questions and shuffles the answer order, so it will not be identical to your last attempt, and an attempt can only be submitted once.

Is the certificate real?

It is a genuine, server-issued certificate with its own verification page and a unique ID, not a self-declared one. It records that you passed this course's exam on this date. It is not an accredited qualification, a professional licence or proof of employment.