Medium inference interview questions

Decode reads every weight once per token. The answer turns that sentence into a millisecond floor, then accounts for the three things that keep a real server above it.

Decode and prefill run the same weights through the same kernels, yet one waits on memory and the other on tensor cores. The answer finds the prompt length where they trade places and turns it into a TTFT.

One division tells you whether a headline number is physics or marketing. The answer runs it, then works out which unstated condition would make the claim true and what to ask the vendor for.

The datasheet ratios are close to each other, so the lazy answer is one speedup for everything. The answer splits it into three regimes, and the biggest win turns out to come from capacity, not bandwidth.

MBU is one division, and that is why it is so easy to get wrong. The answer defines both operands precisely on a real server and then shows six ways a dashboard can report a number 30 points off.

The crossover falls out of one division, and the number it gives lines up exactly with something else on the GPU. Spotting that coincidence is what the interviewer is waiting for.

Joules per token is power over throughput, and the power term is the one candidates obsess over. The answer shows why the denominator moves the result by about 50x while the numerator is pinned under a datasheet cap.

The formula takes one line. The answer interviewers score is the memory budget around it: what the engine reserves first, what tensor parallelism does to the heads, and why the count you get is a ceiling.

Weights and KV are the two numbers everyone knows. The answer that gets the offer accounts for the other few gigabytes, says which ones vLLM measures and which it guesses, and knows where an OOM comes from when the sum looks fine.

Doubling the GPUs past the KV head count halves the weights per GPU and leaves the KV per GPU exactly where it was. The answer derives that, prices the duplicate copy, and picks the parallelism that does not pay it.

A prefix cache only hits from the first token forward, so evicting the wrong block can strand gigabytes that stay resident and can never be hit. The answer derives how much, and which eviction order guarantees it never happens.

Count the blocks with and without sharing and the saving looks large. The answer that scores also knows when the copy fires, what a misaligned prompt does, and the workload where the saving all but disappears.

The per-token formula gives a scary number for GQA and a reassuring one for MLA. Put both on real GPUs with tensor parallelism and the reassuring one is the request that does not fit.

The gain is not a constant. It is a statistic of your output-length distribution, and the right token budget is a point on the roofline. The answer derives both, then says where the paper number and the real knee disagree.

Per token, moving KV over PCIe is several times cheaper than recomputing it, yet vLLM recomputes. The answer derives the break-even bandwidth, shows how small copies fall below it, and names what makes recompute nearly free.

One 32K-token prompt in front of ten short ones costs each of them more than a second of TTFT. The answer separates the half of the job length you know from the half you do not, and schedules each differently.

Chunked prefill keeps the decoders running, so the damage looks bounded. The walkthrough shows the inter-token gap climbing from 10 ms to a full second, new arrivals locked out for 8.6 seconds, and a slowdown that outlives the prefill.

The default of 1,024 sequences is never reached by a 2K-context workload on an 8B model, and is past the compute ridge for a 256-token one. The answer finds the three ceilings and picks the lowest.

With no latency target, every scheduler knob points the other way from chat serving. The answer sizes the fleet from the roofline, finds that prefill is three quarters of the bill, and lists the levers in order of dollars saved.

FlashAttention does the same arithmetic as naive attention and is several times faster. The answer is a three-line recurrence and one IO count, and interviewers want both.

The one-query recurrence is a warm-up. A real kernel carries a vector of maxima and a vector of sums for a tile of rows, and the rescale becomes a diagonal matrix. The answer writes it, runs it, and shows the edge case that turns it into NaN.

The textbook IO count says FlashAttention re-reads K and V once per query block, which at 8K is only 2x better than naive. The answer explains why the real HBM number is 65x lower, and which cache makes the difference.

The indirection adds 4 bytes per 64 KiB of KV, so the byte count barely moves. The answer shows where the real cost hides: dependent loads, tiles that span several pages, and copy engines that want one contiguous box.

GQA shrinks the cache by the group size, but a kernel that loops over query heads reads it back group-size times anyway. The answer packs the group into the matrix dimension and shows what that does to intensity and tensor-core utilization.

A 4K window over 32K tokens keeps one eighth of the keys, and the speedup over causal attention is not 8x. The answer counts tiles exactly, classifies which need masking, and explains where the rest of the gap goes.

Packing removes padding but puts different requests' tokens side by side in one tensor, so the kernel's tile boundaries no longer match sequence boundaries. The answer shows the indexing and a bug that a causal test can never catch.

Occupancy is the minimum of four separate budgets, and candidates usually compute one of them. The answer works all four for an H100, finds which one binds, and then asks whether raising occupancy would help at all.

The one-column pad is the textbook fix, and it works for FP32 scalars. The answer derives why, then shows what happens with BF16 and with 16-byte vector stores, where the pad stops working and kernels swizzle instead.

Eight tiles too many turns a one-wave GEMM into two. The answer prices the waste in microseconds for a real prefill shape, then compares the three fixes, including the one that lives in the scheduler, not the kernel.

RMSNorm does almost no arithmetic, so its cost is bytes, and the eager PyTorch version moves nine times more of them than it needs to. The answer writes the one-pass kernel, tests it on awkward hidden sizes, and says where fusion stops paying.

Every kernel in the step can run near peak bandwidth while the step itself runs at two thirds of it. The answer reads the three rows of the timeline that tell you why, and turns the gaps into milliseconds.

A kernel time is a distribution measured under conditions, and a reviewer will ask about both. The answer builds the timing loop, says what each line defends against, and adds the one sanity check that catches most bad numbers.

Finer scales buy accuracy, and the bytes they cost are easy to count. The answer counts them for a 70B model, then shows the cost that matters more: which scales can wait for the epilogue and which ones invade the inner loop.

The two methods ship identical file formats and run on identical kernels, so the difference is entirely in how the rounding was chosen. The answer shows the one design choice that makes one of them fragile when traffic stops looking like the calibration set.

Both formats store the same eight 4-bit values per element. The difference is all in the scale, and the answer shows which of the two scale choices buys more accuracy and what it costs in bits.

Perplexity averages over the tokens a model finds easy and scores it on text it did not write. The answer builds the evaluation that catches what quantization actually breaks, and shows why a one-point drop on a benchmark can be noise or a real regression.

Local-inference formats spend every bit where it buys the most accuracy, and datacenter formats spend them uniformly. The answer works out what a Q4_K_M file actually costs per weight, and explains why the trade that wins on one consumer GPU loses at batch 64.

FP8 halves the weights and still leaves almost no memory for the cache. The answer does the budget line by line, picks the format pair that leaves room for real concurrency, and says which hardware change makes the opposite choice correct.

One option drafts four tokens for six percent of a target step but guesses each position blind; the other conditions every guess and costs nearly half a step. The answer prices both on Llama 3.3 70B and finds the batch size where the winner flips.

A one-layer drafter that beats an 8B model on acceptance sounds wrong until you ask what it is reading. The answer prices the EAGLE head on Llama 3.3 70B, explains the feature trick, and names the three changes in EAGLE-3.

A drafter that costs nothing and needs no training sounds too good, and on open chat it does nothing. The answer traces a code-rename request step by step, counts accepted tokens under three match-length settings, and says when to turn it off.

Compute acceptance from counter snapshots, reject reset-contaminated windows, and handle EOS before inferring request-step counts. Then decide what a matched-load baseline must measure.

Four beams over a 1,000-token prompt cost 600 MiB if each keeps its own cache and as little as 156 MiB if they share blocks. The answer implements the fork, the copy-on-write and the prune, then shows what vLLM actually does today.

A token still has to pass through every layer in order, so splitting the layers across nodes cannot make one token faster. The answer shows what pipelining buys instead, prices the bubble, and finds the all-reduce latency at which it beats cross-node tensor parallelism.

Collectives run at the speed of their slowest link, so one NIC that trained at a quarter of its rate slows every GPU in the instance by the same amount. The answer shows the arithmetic that makes a 4x slower link a 2x slower step, then bisects the cluster to the culprit.

Pulling every replica's copy from object storage turns a 500 GB model into 62.5 TB of reads. The answer finds the three bounds that actually limit a fleet load, shows a pipelined broadcast reaching the tightest one, and lists the startup work that hides behind the bytes.

Five processes, one scheduler with no prefill phase, and a busy loop that should never wait on Python. The answer maps each stage to the component that owns it in the current source, not the 2023 paper.

A graph freezes shapes, so engines capture dozens of them and round every batch up. The answer prices that rounding in the regime where decode lives and finds the real costs somewhere else.

A block's identity is a hash of its tokens and of every block before it. The answer traces one request through hashing, lookup, the last-token rule and eviction, with a runnable model of the lookup.

Most answers still describe building a TensorRT engine per GPU and batch shape. The current release accepts only the PyTorch backend, and its default admission policy can halve your concurrency. The answer covers both, then says when it is still the right engine.

Decode each token on its own and an emoji streams as two replacement characters. The answer builds the offset-window decoder that fixes it, runs it, and adds the stop-string holdback that most streaming bugs come from.

The mask is a 16 KB bitmap per request. Whether it costs nothing or stalls every step depends on which thread computes it and what it is allowed to overlap with.

Same command line, new version, lower throughput. The answer is a method: capture the resolved configuration on both sides, diff the handful of defaults that move every quarter, then pin and bisect.

One engine is built to run a model anywhere with one user; the other is built to share a datacenter GPU across hundreds. The answer turns that into memory-bandwidth arithmetic and finds the cases where the laptop wins.

Sending every request with the same prefix to the same replica halves the fleet, until one customer's prompt gets popular. The answer derives the spill threshold from the prefill a cache hit saves, in milliseconds.

A web-service autoscaler watching GPU utilization will scale this fleet too late and in the wrong direction. The answer turns the four-minute load time into a headroom number, then decides what absorbs the bursts no scaler can catch.

Input and output tokens burn different resources, so they get different prices. The answer prices each phase from first principles, then shows which assumption moves the list price most.

A canary is only as good as the number of requests behind each verdict. The answer sizes the canary from the statistics of the regression it must catch, then stages the rollout around that number.

141 GB of weights on 160 GB of GPUs, and the engine refuses to start. The answer walks the five phases of engine startup, shows which one failed from the log, and does the arithmetic that says the model never fit the way it was configured.

Prompts fit, but their outputs exhaust the pool. Work through a page-level memory ledger, choose the next diagnostic check, and test a concurrency fix against queueing and goodput.

Greedy decoding is deterministic on paper and the customer has two different answers in hand. The answer finds where the bits change, shows a measured example of it, and turns the finding into a promise the platform can keep.

Same model, same image, same instance type, and one replica drags the fleet's p99. The answer first proves the replica is slower at equal load rather than busier, then walks the hardware causes from GPU clocks to a link that trained at the wrong speed.

A dashboard of forty graphs is a dashboard nobody reads at 2 a.m. The answer builds eight panels in the order an incident unfolds, ties each to the action it decides, and shows how two of them check each other with Little's law.

A scheduler is a loop over three queues and a budget. The answer gets the order of those loops right, shows what the budget costs in milliseconds of TPOT, and tests the invariants a grader will check.

RoPE is a rotation per pair of dimensions. The bugs live in which dimensions pair up, which position each row of a ragged batch is at, what left padding does to position ids, and the precision of the angle at position 100,000.

Two nibbles per byte and a scale per 128 weights sounds like bookkeeping. The grader checks the sign extension, the nibble order, what happens when K is odd or not a multiple of the group, and whether your error test has a bound or just a vibe.

The tutorial kernel loads a whole row into one block and breaks the moment the row is a 128,256-entry vocabulary. The answer writes the looped version, keeps it free of NaNs when a whole tile is masked, and tests it the way a grader will.

Two loops and an index formula, and most attempts get the index wrong by one. The answer writes reduce-scatter and all-gather over a simulated ring, proves the result matches a direct sum when p does not divide the length, and turns the byte count into the reason small messages avoid the ring.

A request costs anything from ten tokens to a hundred thousand, and you only learn the real cost when it finishes. The answer reserves at admission, refunds at completion, fails fast on requests that can never fit, and tests the bound the bucket is supposed to guarantee.

The stop string arrives as three tokens, the euro sign arrives as two, and the client has already been sent half of both. The answer holds back exactly the text that could still become a stop, decodes bytes incrementally, and tests the stream against a reference that sees the whole output at once.