对象存储

Vultr 对象存储 – 面向 AI 工作负载的 S3 兼容存储

可扩展的 S3 兼容对象存储，与 Vultr GPU 实例集成。以实惠的价格存储训练数据集、模型检查点和生产工件。

Object Storage Features

🔌

S3-Compatible API

Drop-in replacement for AWS S3. Use boto3, rclone, s3cmd, MinIO client, or the AWS CLI with a single endpoint change — no code rewrite needed.

💰

Predictable Per-GB Pricing

Pay only for what you store. No per-request charges for standard GET/PUT operations. Ideal for large-scale ML dataset storage with high I/O frequency.

🌐

Global CDN Edge

Vultr's anycast CDN delivers static assets from 32+ global PoPs. Serve model inference responses, static ML artifacts, and API results at the edge.

🔒

Access Control & Encryption

Fine-grained S3 ACLs, bucket policies, and server-side AES-256 encryption. CORS configuration for web-facing API integrations.

⚡

GPU Instance Integration

Mount buckets via s3fs-fuse or stream data via boto3 directly from Vultr GPU instances. Co-located storage and compute minimize egress latency.

♾️

Unlimited Scalability

No object count limits. Store petabytes of training data, model weights, and checkpoints without pre-provisioning storage capacity.

Object Storage Pricing vs AWS S3 & GCS

Feature	Vultr	AWS S3	Google Cloud
Storage (per GB/mo)	~$0.020	~$0.023 (S3 Standard)	~$0.020 (Standard)
GET requests (10K)	Free	$0.004	$0.004
PUT requests (1K)	Free	$0.005	$0.005
Egress (per GB)	~$0.01 (CDN)	$0.09 (Internet)	$0.08 (Internet)
S3 API Compatible	Yes	Native	Via XML API
GPU Co-location	Yes (same DC)	Partial	Partial

AI & ML Storage Use Cases

🧠

ML Training Datasets

Store ImageNet, Common Crawl, or custom datasets. Stream multi-TB datasets directly to GPU training nodes using tf.data or PyTorch DataLoader with S3 connectors.

⚖️

Model Weight Storage

Centralize GGUF, safetensors, ONNX, and checkpoint files. Version model weights with bucket versioning and restore previous checkpoints instantly.

🎬

Generative AI Assets

Store raw video/audio for fine-tuning multimodal models. Serve generated images and videos via Vultr CDN to end users without bandwidth spikes.

💾

Database Backups

Automated PostgreSQL, MongoDB, and Redis backups to object storage. Lifecycle rules automatically archive old backups to cold storage tiers.

Quick Integration Examples

boto3 (Python)python

import boto3

s3 = boto3.client(
    "s3",
    endpoint_url="https://ewr1.vultrobjects.com",
    aws_access_key_id="YOUR_ACCESS_KEY",
    aws_secret_access_key="YOUR_SECRET_KEY",
)

# Upload dataset
s3.upload_file("dataset.tar.gz", "my-ml-bucket", "datasets/dataset.tar.gz")

# Stream to GPU instance
obj = s3.get_object(Bucket="my-ml-bucket", Key="models/llama3-70b.gguf")
data = obj["Body"].read()

rclone syncbash

# Configure rclone
rclone config create vultr s3 \
  provider=Other \
  endpoint=ewr1.vultrobjects.com \
  access_key_id=YOUR_KEY \
  secret_access_key=YOUR_SECRET

# Sync datasets to GPU instance
rclone sync vultr:my-ml-bucket/datasets ./datasets/ \
  --transfers=32 --checkers=16 --progress