NVIDIA Forward Deployed Engineer interview questions

Modern LLMs share one architectural skeleton. Learn the five components and you can read almost any model card cold, plus the single distinction that proves to an interviewer you actually understand it rather than the buzzwords.

Three transformer families, one distinction that decides everything: which tokens are allowed to see which. The attention mask is the whole answer, and it explains why GPT generates and BERT does not.

Rotary embeddings encode position as a rotation of the query and key vectors, so attention scores depend only on the gap between two tokens. That relative property is why RoPE generalizes past its training length where learned absolute tables fall off a cliff.

Where you put the LayerNorm changes whether gradients survive a hundred layers. Pre-norm keeps a clean residual highway and well-scaled gradients so deep models train without warmup heroics; post-norm normalizes the output but fights vanishing gradients at depth.

Attention is bottlenecked by moving the N x N score matrix to and from HBM, not by FLOPs. FlashAttention tiles the computation in fast on-chip SRAM and fuses softmax with an online running statistic, so it computes the exact same output while never materializing the full matrix.

Each chat turn resends the whole conversation, so the model re-prefills the same prefix every time. Server-side KV caching and client-side cache breakpoints kill that redundant compute, but exact-prefix matching and stale entries are where teams silently get it wrong.

Without a KV cache, generating token N re-attends over all N-1 prior tokens from scratch every step, turning decoding quadratic. The cache stores each layer's past keys and values so each new token costs one forward step, and its size is what caps your batch and context.

Brute force is O(N*d) per query and dies at ten million vectors. The answer is approximate nearest neighbor: trade a little recall for orders of magnitude less work, and know exactly which knob buys what.

Ten X-ray studies a second, a 2-second budget, and a missed cancer is unacceptable. The hard parts are recall-first thresholding, a monitor that watches the radiologist, and surviving a new scanner.

A back-of-the-envelope round where the envelope has a purchase order attached. The arithmetic is the easy half. The scoring is in which assumptions you write down, which way you round, and what you say about the number being wrong.

Aggregate metrics hide exactly the customer who is about to churn. Adding a tenant label to everything is the obvious fix and it is also how you get a cardinality explosion that takes down the monitoring before it takes down the platform.

Nobody assigned it, nobody asked for it, and the people you needed had their own roadmaps. The winning answer isn't 'I pushed through the resistance,' it's a small undeniable proof that turned skeptics into a coalition before you ever asked for headcount.

An experience probe with a depth gauge: 'we used Kubernetes' fails, and so does reciting the KServe README. The stack-process-role-warstory structure, the resource and probe details that prove hands-on work, and where KServe earns its complexity.

CPU-based HPA on GPU inference never fires, the trap half of all candidates fall into within a minute. The signals that actually track load, the anatomy of a five-minute cold start, and which mitigations are worth their cost at each layer.

The cluster shows eight free GPUs and your four-GPU pod still won't schedule, the fragmentation puzzle GPU-platform rounds open with. Bin-packing vs. spreading, MIG/MPS/time-slicing for the small-model problem, and the pod-per-model math that breaks at scale.

The GPU-platform capstone: dozens of models, spiky traffic, a fixed monthly GPU budget, and a p99 SLO that scale-to-zero would wreck. Token-based autoscaling, the three layers of LLM cold start, and the residency tiering that makes the budget math work.

The 'works on my GPU box, CUDA error 803 in prod' bug, solved at the source. What the host driver actually pins, why the toolkit version is fungible, and the multi-stage build that ships a lean reproducible image.

The dataset is too big to fit on disk, so you stream it. The components that keep the GPU fed, the throughput math that says whether you'll be I/O-bound, and the failure handling that makes a multi-day run resumable.

The NVIDIA warm-up that exposes whether you've written a kernel or just called .cuda(). The mapping from software hierarchy to silicon explains every GPU performance conversation that follows.

Kernel optimization is mostly the art of feeding 100+ TFLOPs from the small fast end of this hierarchy. Know the rough latencies and bandwidths, interviewers listen for the numbers.

A staple NVIDIA screen with a precise mechanical answer, and a follow-up trap about which branches actually diverge that catches most candidates.

The two access-pattern killers behind most 'my kernel is 10x slower than cuBLAS' mysteries. The matrix transpose walks straight into both, which is why NVIDIA loves asking about it.

Everyone says 'maximize occupancy.' The candidates who get hired at NVIDIA know when 25% occupancy beats 75%, and can explain the resource math that decides it.

The roofline model in one ratio: FLOPs per byte against the hardware's ridge point. Get the H100 arithmetic right and you can classify any kernel, including why LLM decode will never be compute-bound at batch 1.

Sounds like small talk; it's actually a depth probe with follow-ups pre-loaded. A scoping framework that works whether you've run 4 GPUs or 4,000, and the bluff that ends interviews.

The backbone question of every frontier-lab infra loop. The answer that scores is organized around what each strategy communicates and how often, not just what it splits.

The one derivation every GPU-infra loop expects on a whiteboard: scatter-reduce plus all-gather, 2(N−1)K/N per GPU, and why that's provably near-optimal. Plus the latency catch that motivates tree algorithms.

A real infra-loop question that candidates fumble by guessing. The correct map, TP on NVLink, PP and DP across the scale-out fabric, falls out of three traffic profiles, and the reasoning is the answer.

NCCL's topology tricks, GPUDirect RDMA, and the single most misleading metric in distributed training. If you've ever trusted nvidia-smi on a slow run, this question was written for you.

One formula, 2 × layers × KV heads × head_dim × bytes, unlocks every LLM serving capacity question. Worked through for Llama-70B, plus the mitigation stack from GQA to paging.

The OS virtual-memory trick that made vLLM the default serving stack. The answer interviewers want quantifies the fragmentation it killed, and knows what it cost in exchange.

One forward pass, two opposite performance regimes. The arithmetic-intensity argument that explains TTFT vs TPOT, GPU pool design, and why the big serving stacks split the phases onto different hardware.

Free tokens from idle FLOPs, losslessly, which surprises most candidates. The mechanism, the acceptance-rate math, and the high-batch regime where speculation quietly stops paying.

Bits buy bandwidth, bandwidth buys tokens, until quality quietly craters on code and math. A decision framework with the numbers, what GPTQ and AWQ actually do differently, and the eval discipline that keeps you out of trouble.

TP=2, 4, or 8, the answer is a memory budget, a latency target, and one rule: smallest TP that fits, scale with replicas. The worked arithmetic that turns this from vibes into a defensible call.

The worst failure mode in distributed training: no crash, no log, every GPU pinned at 100%, and the job dead. The triage order that finds a collective mismatch or a dead rank before you burn a day of cluster time.

A sparse model with few active params per token, a brutal latency target, and an all-to-all on the critical path. The parallelism layout that hits the SLA, why expert parallelism not tensor parallelism is the lever, and the tail the all-to-all hides.

The fleet was fine on 8k contexts and falls over when a few 200k-token requests arrive. The math that explains why, the admission and pooling fixes that contain it, and why throwing GPUs at it is the wrong first move.

Three serving runtimes that look interchangeable on a benchmark slide and aren't. The axes that actually separate them (scheduler, kernels, operability) and the one-line rule for which to reach for.

Do the weight math first, then choose your poison: shrink the weights, split them across GPUs, or spill to host memory. Each buys you the fit at a different cost, and one of them is almost always wrong for production.

A very large sparse model, hundreds of experts, and users on every continent. The layout that keeps experts close to compute, routes around regional skew, and decides what 'global KV cache' can and cannot mean.

Constrained JSON/grammar decoding usually means a Python callback per token, which serializes the GPU. The staff answer keeps the DFA on-device and turns the mask into a single coalesced gather.

A 7B VLM at 25 FPS on an embedded GPU sounds impossible until you stop processing every frame fully. The answer is token pruning, frame-to-frame KV reuse, and a three-stage pipeline that hides latency.

Going 4K to 1M is not one trick. It is RoPE rescaling chosen by frequency band, a long-doc continued-pretrain mix, needle evals that actually catch failures, and KV math that forces ring attention.

The warm-up that filters out tutorial-level candidates. Interviewers aren't grading the diagram, they're grading whether you know the three places real pipelines fail and how you'd see it happening.

Everyone says '500 tokens with overlap.' The candidates who get hired explain why chunk boundaries are a recall problem, name three strategies, and describe the experiment that picks between them.

Every queue, webhook, and retry loop you'll ever deploy at a customer hides this distinction. Interviewers use it to separate people who've shipped from people who've read about shipping.

'Evals first' is the #1 hiring signal in FDE design rounds. This is the foundational version: what a 50-case golden set looks like, which metrics split retrieval from generation, and the launch gate that convinces interviewers.

A reported FDE design question that punishes one-size-fits-all answers. Each corpus has a different 'semantic unit': get the three designs, the parsing tools to name, and the eval that proves the split was worth it.

The retrieval design question with a precise rubric: candidates who can say what each stage rescues, and what it costs in latency, clear it. Includes the RRF detail and the latency budget interviewers ask for.

A staple at OpenAI and Anthropic in both coding and design rounds. Most candidates can name the algorithms; few can say which one lets 2x your limit through, and that's the part that gets scored.

The signature FDE debugging genre: a moving fault, an opaque client, and a customer who's sure it's your fault. The winning method splits the problem at the boundary, and knows the four classic culprits hiding on their side.