Software Development and Engineering
Free NVIDIA-Certified Associate: Generative AI Multimodal practice — 6 questions on Software Development and Engineering, with explanations. No sign-up. Full 12-question mixed test →
A company deploys a multimodal chatbot that must respond in under 100ms per request even during traffic bursts, and separately runs a nightly offline job captioning 2 million images where total throughput matters most. Which deployment configuration BEST meets BOTH requirements using Triton Inference Server's dynamic batching?
Latency-sensitive interactive workloads need a short max_queue_delay so requests aren't held waiting for a batch to fill, while throughput-oriented offline jobs benefit from larger batches and longer queue delays to maximize GPU efficiency — Triton allows per-model instance configuration to serve both patterns simultaneously.
Which Triton Inference Server config.pbtxt parameter directly controls how long the server waits to accumulate additional inference requests before executing a dynamically formed batch, trading latency for throughput?
max_queue_delay_microseconds sets the maximum time the dynamic batcher waits for more requests to arrive before dispatching the current batch, making it the direct latency/throughput trade-off knob.
A team must guarantee that a PyTorch-based generative model behaves identically across dev, staging, and production environments, including matching CUDA and cuDNN versions. Which approach BEST ensures this using NVIDIA's ecosystem?
NGC containers bundle a fixed, tested combination of CUDA, cuDNN, and framework versions; pinning an exact tag guarantees byte-for-byte identical runtime stacks across environments, which pip alone cannot guarantee since it doesn't control system-level CUDA/cuDNN.
You want Triton Inference Server to load two concurrent instances of the same model, placing one instance on GPU 0 and one on GPU 1 to increase serving concurrency. Which instance_group configuration in config.pbtxt achieves this?
Specifying count: 2 with kind: KIND_GPU and gpus: [0, 1] tells Triton to place one instance on GPU 0 and one on GPU 1, achieving the desired multi-GPU concurrency.
A data science team building a multimodal pipeline combining a CLIP encoder and a diffusion model reports inconsistent inference outputs across teammates' laptops despite everyone using the same Python version. Which practice would MOST reliably resolve this?
Inconsistent results despite matching Python versions typically stem from differing CUDA/cuDNN/driver stacks or transitive dependency versions; a pinned Dockerfile built from an NGC base image locks the entire runtime environment, guaranteeing reproducibility across machines.
When integrating a multimodal model into a latency-sensitive application via Triton Inference Server's client APIs, which protocol choice is generally preferred for lower overhead when sending large tensor payloads such as images?
gRPC uses Protocol Buffers and binary serialization, which is more efficient for large tensor payloads than JSON text encoding, reducing serialization overhead and network payload size for latency-sensitive multimodal inference.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$99.99 $34.99 with code FREETEST33 — valid through September 9.
Get my $34.99 deal →