Frequently asked questions

Interview preparation for the engineers who make models fast and cheap to serve: inference engineers, ML systems and performance engineers, GPU kernel engineers and the platform teams that run LLM fleets. It covers the path below the API, from roofline arithmetic and KV-cache budgets to batching schedulers, attention kernels, quantization, speculative decoding, parallelism, serving engines, fleet design and production debugging.

Engineers interviewing at frontier labs, inference providers, accelerator vendors, big tech and trading firms. Each topic starts from the calculation every candidate is expected to do on a whiteboard and climbs to the questions that separate a hire from a strong hire at senior and staff level. If you already serve models in production, start in the middle of a track; the early questions will be quick.

Premium is one payment of US$25 by card or PayPal, or ₹2,000 by UPI, card or netbanking. It includes every worked answer and derivation in all 12 topics, every concept page, and the full practice-test bank. Access runs for six months, covers anything added in that time, and does not renew on its own.

All question titles are public, and so are the opening answers of each topic and the first concept of every track. A free Google sign-in raises the per-topic window and saves progress and bookmarks. The free answers are full answers, not previews, which is the fairest test of whether the paid ones are worth it.

Hardware figures are the dense peaks of a named part, mostly H100 SXM, H200 and B200, checked against the vendor's published specifications. Engine behavior is pinned to a release, for example vLLM v0.31.0, and was read from that tagged source. When a default changed recently, the answer gives the old value too, because the engine in your interview may be older than the newest one.

No. Almost every question is answerable with arithmetic, and the answers show that arithmetic step by step. The code samples run on a CPU with NumPy or plain Python unless they are explicitly GPU kernels, and those say so.

No. Premium opens digital content the moment it is paid for, so sales are final; the Terms have the full policy. That is the reason the free windows are generous: read the free answers in the topics you care about first. If something has gone wrong with a purchase, write to us anyway. A person reads every message.

Send us your interview date. We can keep your access active until the interview without another payment, so an expiry a few days before your onsite does not interrupt preparation.

The account reverts to free access with your history intact: progress, bookmarks and every free answer. Buying another six months is a separate choice, made by you, whenever you want it.

Yes. Razorpay and PayPal handle the payment, and a card works without a PayPal account. Your card and UPI details never reach us. Access opens as soon as the provider confirms the payment.

Use the form on the Accuracy page: the figure, the page, and the value you get. Every report is reviewed by hand. Each confirmed error is corrected and adds a day of full access, and if we think the original is right you get the calculation behind it.

Yes. Submit it while signed in. If we publish it, it carries your @handle and a day of full access is added to your account automatically.

No. Every answer follows a written standard: numbers come with the chain that produces them, code is executed before it is published, and engine claims cite a version. Content linters reject answers that skip required sections or read like stock generated prose. The free answers are some of the hardest in the bank, so you can check this yourself.

No. A company appears here because it hires for these roles or runs a loop our questions reflect. None of them has reviewed, sponsored or approved this site, and each name and mark remains its owner's property.

A person. Use the contact form or write to support at inferenceinterviews.com about billing, access that did not open, or anything else. Most messages get a reply within two business days.