所有应用程序
人工智能与模型

vLLM

NVIDIA GPU required · Self-hosted only

在自托管 NVIDIA GPU 上服务的 OpenAI 兼容模型。预览待处理的真实 GPU 安装测试。

OpenAI 兼容 API

连接现有的推理客户端。

NVIDIA 推理

请求一个具有私有容器共享内存的 GPU。

持久模型缓存

重用下载和编译缓存。

NVIDIA inference preview

This template uses vLLM's official CUDA image, one NVIDIA GPU and a generated API key for its OpenAI-compatible API. It is listed as a preview and cannot be installed yet: real NVIDIA hardware validation is still required. It is self-hosted only and does not support managed Cloud or CPU-only servers.

The server needs a compatible NVIDIA driver and NVIDIA Container Toolkit. The selected model must fit the GPU's VRAM; ordinary RAM guidance does not measure VRAM or reserve GPU capacity. Openship does not provision GPUs or install drivers.

Model and startup

The initial model is Qwen/Qwen3-0.6B. Choose a different Hugging Face model and, when required, provide a token and accept its license on Hugging Face. The first start downloads and loads the model. Check Logs for progress; a starting health check is not proof of a failed deployment.

The template persists model downloads and compilation caches. It gives the container a private 2 GiB /dev/shm, without sharing the host IPC namespace or enabling privileged mode. Context length, model choice and concurrency determine memory demand.

Use the generated API key as a bearer token for /v1 requests. Keep that key private. The pinned image's CUDA version must be compatible with the host driver; this preview is not a guarantee for every NVIDIA card or model.

Official Docker deployment · GPU requirements

包含在此模板中

由目录配置的服务作为一个项目一起部署。

vllm

vllm/vllm-openai:v0.31.0