Multimodal & vision

Multimodal & computer vision

Image/video understanding and multimodal models — vision encoders plus generation.

PyTorchTensorRTTritonNVIDIA DALIHugging Face

Architecture

Preprocessing (decode/resize) feeds a vision encoder; multimodal models fuse image and text. Video adds decode throughput and memory pressure.

Sizing

Memory-bound on large images/video; batch aggressively; hardware decode (NVDEC) offloads the CPU.

Networking

North-south; large media benefits from proximity to object storage.

Storage

Object storage for media; local cache; high read throughput for video pipelines.

Scaling

Scale on queue depth; use hardware decode and batching to raise utilization.

Monitoring

Frames/sec, decode utilization, GPU memory, latency per item.

Sample manifest
starting point
apiVersion: apps/v1
kind: Deployment
metadata:
  name: vision-infer
spec:
  replicas: 2
  template:
    spec:
      containers:
        - name: triton
          image: nvcr.io/nvidia/tritonserver:latest
          resources: { limits: { nvidia.com/gpu: 1 } }
Recommended GPUs