Cuestionario vinculado a una guía · Medio Nivel

Contenedores GPU con cuestionario del kit de herramientas de contenedores NVIDIA

Separe los controladores de GPU del host de las bibliotecas CUDA del contenedor y rastree los pasos de tiempo de ejecución necesarios para exponer y validar los dispositivos NVIDIA.

Rutas de guía relacionadasContenedores de GPU Nvidia Kit de herramientas de contenedores
Pregunta 1 de 8

¿Qué componente normalmente gestiona la GPU física NVIDIA a nivel de kernel?

Vista previa de cada pregunta8 preguntas

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

  1. ¿Qué componente normalmente gestiona la GPU física NVIDIA a nivel de kernel?
    • The host driver
    • The CUDA container image alone
    • The model checkpoint
    • The Python package lockfile
  2. ¿Qué función proporciona NVIDIA Container Toolkit?
    • A replacement GPU kernel driver stored in each image
    • Runtime integration that exposes devices and driver interfaces to containers
    • A framework-neutral model accuracy benchmark
    • A compiler that converts all Python into CUDA
  3. ¿Por qué un contenedor con bibliotecas CUDA aún no puede inicializar la GPU?
    • CUDA libraries always replace the host driver
    • The container necessarily needs its own physical GPU
    • The host driver or runtime integration may be incompatible or missing
    • The model file does not contain a device node
  4. ¿Qué prueba ver una GPU en un contenedor?
    • That memory allocation can never fail
    • That the model's accuracy is acceptable
    • That CUDA and driver versions are identical
    • Device visibility, but not that every application kernel will run correctly
  5. ¿Por qué es diferente actualizar los paquetes CUDA dentro de una imagen que actualizar el controlador del host?
    • Both update the same kernel module automatically
    • They are different software layers with different responsibilities
    • The container cannot contain any runtime libraries
    • Driver versions are determined by the model weights
  6. Which test is stronger than checking device visibility alone?
    • Inspect the model filename
    • Read the image tag
    • Run a small representative framework operation on the GPU
    • Confirm the container starts without a GPU request
  7. What should an incident record include for reproducibility?
    • Only the framework version and model architecture
    • Only the host driver and GPU model
    • Image tag and Python version without runtime configuration
    • Image digest, host driver, toolkit and runtime configuration
  8. What does GPU containerization not provide by itself?
    • A fully independent hardware driver environment
    • Packaged user-space dependencies
    • A way to request devices at launch
    • An image filesystem