Together AI Infrastructure Engineer interview questions

Count two matrix shapes to derive the H100 ridge, throughput ceilings and time floors. Compare an illustrative measurement with its bound before deciding what to optimize.

Derive the weight-only shortcut, then add KV traffic and feasible eight-GPU placement. A worked latency floor shows why batch size alone cannot establish the decode bottleneck.

Use a profiler summary to form a memory-traffic hypothesis, then check counters, timing and useful work before choosing an optimization.

Derive how reusing weights changes arithmetic intensity, then separate the roofline ceiling from tile-wave effects and measured GEMM throughput.

Compare the cost of serving the same workload at the same latency and quality. A memory-traffic estimate explains what each GPU might improve; measured throughput and your actual rental price determine the purchase.

Compare GPU memory, decode traffic and measured useful throughput. Work out a capacity ceiling without mistaking it for a performance forecast.

Derive the bank mapping before and after padding, trace a fixed-tile kernel, and measure whether the saved conflict outweighs other resource costs.

Follow intermediate tensors through separate and fused stages, calculate the traffic and whole-block bounds, and check spills and numerical behavior.

Trace stable row normalization, handle independent strides and fully masked inputs, then separate fusion traffic from measured speed.

Follow the running maximum and weighted sum across attention tiles, with executable arithmetic and an honest memory-traffic comparison.

Compare the mechanisms in FlashAttention versions 1, 2 and 3: IO savings, work partitioning, asynchronous dependencies and FP8 accuracy.

Start with the operation, supported hardware and performance target. Compare ordinary Triton, lower-level Triton interfaces and CUDA using a measured gap that is large enough to justify maintaining custom code.

Use a timeline to establish launch overhead, then trace what graph replay changes. Stable buffers, page-table updates, capture-safe operations and batch buckets determine whether the serving engine can use it safely.

Follow a logical token through the block table into a physical KV page. Then combine split attention states correctly, including empty splits, and count lookup traffic at the layer and head the kernel actually processes.

Verify a drafted prefix without biasing the target distribution, then roll back rejected KV state. A six-token prefix exposes the off-by-one: nine committed tokens can have only eight valid KV entries.

Separate per-expert submission, small-grid scheduling and partially filled tiles. A grouped kernel can share work across experts, but routing, packing and output combination still determine the layer’s cost.

Longer sequences exercise more state updates, different boundaries and sometimes different kernel paths. Use a two-tile example to check online softmax, then test masking, indexing and precision against a stable reference.

Separate packed base weights, quantization metadata and trainable adapter state. Count the real projection shapes, then leave room for activations and temporary work before calling a one-card configuration feasible.

Follow the weight bytes through a small decode batch and a long prefill. A scoped roofline calculation explains the difference without pretending to predict actual latency.

One request finishes, another keeps writing, and a third waits outside. Follow the same workload through two schedules to see what continuous batching improves, and what it cannot promise.

Trace one cache position through a block table, calculate the memory recovered from oversized reservations, and explain why sharing still needs careful writes and matching.

Work from each GPU's cache budget, reserve room for generated tokens, then check latency separately. A limiting-rank example shows why fitting today's prompts is not enough to admit tomorrow's work.

Trace an accepted prefix, a rejection and the correction token; derive emitted progress and compare it with a measured round cost.

Separate weight storage, matrix arithmetic and attention-cache precision. Choose a supported candidate using task quality, memory and service measurements.

Filter engine candidates by exact compatibility, then compare correctness, goodput and operational cost on the same workload.

Separate checkpoint memory, expert assignments and retained MLA state. The active parameter count alone cannot predict per-rank fit or token latency.

Compare proposal methods by committed output over complete round time, with explicit sampling guarantees and target compatibility.

Separate device throughput bounds from measured service curves, then choose a feasible operating point using latency, quality and goodput.

Translate request demand into prompt work, generated tokens and retained state, then size a measured serving configuration and its failure reserve.

Derive a deployment candidate from the official tensor index, then account for latent-cache replication and the product workload. A worked example shows why enough aggregate HBM can still leave every GPU out of memory.

Find which startup step runs out of memory, count the allocations on that GPU, and change the setting that controls the shortage.

Turn public metadata into a deployment candidate with explicit uncertainty. Inspect architecture, stored formats and engine support, then work a per-rank memory example that explains why an aggregate fit can fail.

Find whether expert placement, small batches or communication makes the deployment slower. Compare equivalent workloads and inspect the worker that holds up each step.

Measure reachable prefix reuse, separate token savings from latency, and compare both engines on the same agent contract and resource budget.

Use the actual tensor index to size Kimi K3, then separate a published hardware starting point from production capacity. Account for mixed precision, hybrid recurrent state, parser behavior, and the workload behind a throughput claim.

Check the bound’s assumptions, locate the critical path, and quantify which optimization can actually improve one user’s decode latency.

Derive a prompt-plus-output contract, calculate per-rank state capacity, and measure the concurrency that meets latency and overload targets.

The answers sound fluent, but the scores dropped. Compare inputs, generation settings, raw output, parsing and scoring to find where the new deployment differs.

Define a reproducible workload and service contract, account for every offered request, and choose a measured operating region with evidence and margin.

Compare actual encoded tensors, per-rank runtime fit, supported kernels and held-out task quality. Four-bit weight-only kernels can accelerate wider arithmetic, and smaller weights need not reduce the GPU count.

A repeated-prompt test measures one reuse pattern. Production gains depend on token-weighted reuse, routing, residency, alignment and the prefill share of latency; a falling hit rate does not uniquely identify eviction.

Long prefills coincide with decode stalls. Compare a tuned combined pool with separate phases, derive transfer and pool capacity, and test whether the improvement survives connector failures.

The nodes joined the cluster, but the first request hangs. Trace how far it got, check worker assignments, and reproduce the blocked operation on the same communication path.

The output forecast is a starting point. Match the prompt and state workload, derive replica counts from an SLO-tested rate, check simultaneous loss scenarios and keep the cost denominator consistent with the forecast.

Compare committed output per complete draft-and-verify cycle with ordinary decoding. Derive the acceptance-length model, separate it from engine counters, and test latency, goodput and memory across the actual workload.

Confirm that requests are waiting for usable state capacity. Separate allocated memory, live blocks and evictable prefixes, then compare admission, lower-precision state, prefix sharing and hardware changes with explicit assumptions.

A fourfold ratio is a symptom, not a diagnosis. Trace individual requests, separate waiting from execution and delivery, and test workload, scheduling, cache and capacity hypotheses against the service objective.

Build a week around release gates and dependencies. Budget artifact transfer, validate the exact runtime path, measure the intended workload, and size the canary evidence before promising full rollout.

Price the same workload and service guarantees, including input tokens, idle capacity, failures and operations. Then check whether the algebraic crossover fits the fleet before treating it as a buying decision.

An apparent outlier can reflect workload mix, queueing, runtime state or a degraded execution path. Normalize the comparison, run a matched replay and use critical-path arithmetic to test possible causes.

A context-window limit is not a latency or concurrency guarantee. Budget retained state, prefill and shared service time; distinguish cache reuse from eliminating work; choose isolation only when measured interference justifies its cost.

A minor revision can change weights, encoding or response behavior. Compare pinned deployment bundles, evaluate task slices, monitor a representative canary and keep enough capacity to restore the old service.

Design a shared catalogue with explicit service tiers, reproducible deployments and bounded cold starts. Residency depends on burst behavior and resource reuse; placement depends on GPU groups and interference as well as memory.

Apply the project’s offline tuning guidance in context, derive the memory tradeoff, and use workload-specific acceptance criteria instead of universal queue and headroom targets.

Verify the throughput definition and service outcomes, compare resolved configurations and execution traces, then isolate supported changes. Quantify capacity replacement correctly and choose rollback from the full release risk.

Budget growing attention state, bounded recurrent state and retained prefix snapshots separately. Work the short-context crossover, then test whether the cache manager resumes every layer from the same prefix boundary.

Separate pre-routing from a sequential cascade, charge every attempted generation, and measure checker mistakes. A lower escalation rate is not proof of higher quality, and the largest model is not a correctness oracle.

Define whether half the footprint means checkpoint bytes or runtime memory. Choose a supported recipe, measure which tensors it changes, separate calibration from held-out evaluation, and validate quality and serving cost.

Define cold and reused-context objectives, verify the checkpoint's real state layout, and size per-rank memory and transfer demand. Caching and phase separation help only when their compatibility, retention and capacity assumptions hold.

Follow media through fetching, preprocessing, encoding and language-model execution. Derive visual positions from the actual processor output, account for encoder placement and separate the caches and latency stages that text-only monitoring misses.

A partial drop does not rule out the parser. Define success and its denominator, trace failures through encoding, generation, parsing and execution, then replay the relevant boundary with controlled versions.

Test the complete serving bundle with contract fixtures, representative tasks and load-aware outcomes. Preserve paired results, calibrate the scorer and distinguish a relative improvement from meeting an absolute product requirement.

Constrained decoding restricts valid continuations; it does not guarantee completion or correct values. Define success and error contracts, validate supported schemas, and measure compilation, masking and queueing under the actual workload.

Attribute queueing, token work and live state before choosing a cap. Bound new admission, verify recovery, and distinguish fair scheduling, reserved service and physical isolation without promising instant or absolute protection.

Screen exact checkpoints for hard requirements, then compare deployable candidates on held-out product tasks, service limits and total cost. Smaller artifacts, leaderboard scores and token rates each answer only part of the decision.

Measure exposed launch and dispatch delay, actual graph coverage, capture memory and warmup. Batch size influences the result, but does not alone determine whether graphs improve the service.

Group compatible adapters by pinned base, derive their size from target matrices and ranks, and benchmark mixed batches. Catalogue size, CPU residency, GPU slots and active adapters are separate limits; base sharing does not make fifty variants operationally free.

A mutable branch or local model directory can change what a restarted replica loads. Establish the actual artifact identity, separate disk evidence from loaded state, and make rollout and restart behavior reproducible.

Validate the exact model, accelerator generation and serving features. Separate packed storage from kernel execution, derive the real memory fallback, and compare qualified throughput and total migration cost.

Map a checkpoint format to compute capability, then say which part of the speedup survives on each generation. The same file can load on two GPUs and take two different execution paths.

Turn a benchmark number into a bill that survives real utilization, then write the serving contract the gateway enforces and the evaluation the tuned configuration still owes.

A measured run shows where the time goes in a small batch job, then why every result parsed but none was deliverable. The answer turns on startup cost, co-tenant latency and a per-item contract.

Walk a single-GPU vLLM Deployment from pod scheduling to the first token, then show the three settings that turn a working manifest into restart loops and stuck rollouts.

Check model placement first, then derive a TP8 traffic floor and add context. Separate a single stream’s pace from an ideal bandwidth ratio.

Compute cost from a matched bill and token count. Show when an activity factor belongs in the denominator and when it would double-count idle time.

Translate active streams into replicas using growth reservations and matched latency tests. Show peak and failure headroom separately, with no invented benchmark.

Calculate FP8 savings by separating weights, KV cache and other step costs. Compare the same batch first, then check whether larger batches meet latency and quality targets.

Derive the batch-295 shortcut, count the tensors it omits, and test whether the resulting kernel estimate is even inside the replica’s memory budget.

Use bandwidth per dollar as a first comparison, then check feasible memory, supported execution and complete service cost before ranking accelerators.

Define what a throughput claim counts, then compare its required arithmetic and memory traffic with a feasible deployment and matched latency targets.

Compare tensor, pipeline and expert traffic with per-request KV transfer. Work through declared payload and latency assumptions before choosing a serving network.

Design fractional GPU admission around explicit shapes and enforceable guarantees, then reconcile reservations with the runtime.

Count feasible placements for the waiting job, including its per-node and topology requirements. Then compare waiting, reservations and checkpoint/restart consolidation using their actual disruption costs.

Choose the tenant boundary before planning base-model reuse. Account for frozen weights, adapter state and activations separately, then schedule by compatibility, fairness and the real cost of restarting training.

Compare equivalent requests, inspect the replica’s software and device state, and reduce new traffic only after checking where that load will go.

Compare the output of a speculative round with drafting and verification cost. Work the break-even condition, then design a measured rollout with memory and correctness checks.

Trace usage from execution to one settled charge. Distinguish requests, attempts and events, then handle cancellation, late records, replay and reconciliation.

Implement size- and timer-triggered batching, then test cancellation, handler errors and the latency added before dispatch.

Compare three GPU placement heuristics with executable traces. Show how admission and free-space shape change, then add workload and topology constraints.

Build a reference-counted KV block allocator and test that forked sequences keep their data independent when either branch appends.

Derive and simulate speculative output counts, then compare them with explicitly assumed or measured round costs.

Connect a performance change to its measured mechanism. Compare fair baselines and calculate how a local speedup affects the whole workload.

Show a cost reduction with comparable service evidence, realized savings and a feasible recovery plan. Distinguish a quiet trial from proof of unchanged reliability.

Follow a token into retained attention state, derive bytes per position for a 70B model, and distinguish present cache usage from the room needed to finish a response.

Compare raw weight sizes at 16, 8 and 4 bits, then budget KV cache and runtime memory for Llama 3.1 70B. Follow the worked per-GPU example.

Separate quota, borrowing, reclaim latency and fairness, with a demand-conserving example of when pooling improves utilization.

Work from token timestamps to first-token wait, average pace and individual stalls. Then count requests meeting both latency limits, keeping failures and measurement boundaries explicit.

Draw query-head sharing, count actual stored state and compare payloads with explicit model geometry and binary units.

Work from training FLOPs to fleet throughput, days and GPU-hours. Then see how utilization and downtime change the estimate.

Follow a four-GPU request from advertised capacity to concrete devices, and see why free GPUs across nodes cannot satisfy one pod.

DRA gives the scheduler device claims and attributes to work with. Compare that with extended-resource counts, then check the deployed driver and controller support before promising sharing, partitioning or gang placement.