VECTOR FEED

Deployment

Docker

Global (HuggingFace models)

git clone https://github.com/opendatalab/VECTOR-FEED.git
cd VECTOR-FEED
docker compose --profile global up

China (ModelScope mirrors)

docker compose --profile china up

Multi-GPU

VECTOR FEED supports distributed inference across multiple GPUs. The Pipeline backend splits batch processing across available devices:

# Auto-detect available GPUs
vector-feed -p ./batch/ --backend pipeline

# Manual device specification
CUDA_VISIBLE_DEVICES=0,1,2,3 vector-feed -p ./batch/ --backend pipeline

Hardware Acceleration

Platform	Support	Notes
NVIDIA CUDA	Full	Volta, Ampere, Hopper architectures
Apple Silicon (MPS)	Full	Native Metal Performance Shaders
CPU-only	Full	Slower, suitable for Office backend

REST API Server

Start Server

vector-feed-api --port 8000

Client Usage

import requests

response = requests.post(
    "http://localhost:8000/v1/parse",
    files={"file": open("document.pdf", "rb")},
    params={"backend": "pipeline"}
)
task_id = response.json()["task_id"]

# Poll for results
result = requests.get(f"http://localhost:8000/v1/task/{task_id}")