GPU server best practices

Deploy a distilled DeepSeek-R1 model on a GPU-accelerated instance

This practice is suitable for scenarios requiring lightweight, high-performance inference. This tutorial describes how to deploy an inference service for a distilled DeepSeek-R1 model on a GPU-accelerated instance.

Deploy a full DeepSeek model in a distributed two-machine setup on GPU-accelerated instances

Expand computing power using a two-machine distributed architecture for very large model training or high-concurrency inference. This tutorial describes how to use vLLM as the inference framework for the DeepSeek model to build an inference service for DeepSeek-V3/R1 on two ebmgn8v instances.

Deploy a full DeepSeek model on a single GPU-accelerated instance

This practice is intended for full model inference or training in single-machine, multi-GPU scenarios. This tutorial describes how to use SGLang as the inference framework for the DeepSeek model to build an inference service for DeepSeek-V3/R1 on one ebmgn8v instance.

Build a model inference environment using a DeepGPU-LLM image

Use the DeepGPU-LLM container image from Alibaba Cloud to quickly build an inference environment for a large language model (LLM). This approach is primarily suited for natural language processing scenarios, such as intelligent chatbots, text analytics, and coding assistants. This tutorial describes how to use a DeepGPU-LLM container image on a GPU-accelerated instance to build an LLM inference service.

Quickly build a model inference environment using a vLLM image

Deploy the vLLM open source inference framework to quickly build an inference environment for an LLM. This approach is primarily suited for natural language processing scenarios, such as intelligent chatbots and text classification or analytics. This tutorial describes how to use a vLLM container image on a GPU-accelerated instance to quickly build an LLM inference service.

Accelerate text-to-image generation using an SD-WebUI container image

For scenarios requiring text-to-image generation, deploy an SD-WebUI container image on a GPU-accelerated instance to achieve faster computing and higher inference performance. This tutorial describes how to deploy an SD-WebUI container image on a GPU-accelerated instance to quickly generate images from text.

Deploy an NGC environment to build a deep learning development environment

This tutorial uses the TensorFlow deep learning framework as an example. It describes how to deploy an NGC environment on a GPU-accelerated instance to pre-install a deep learning development environment.

Use FastGPU to one-click deploy and train applications

Alibaba Cloud provides tutorials for FastGPU training scenarios in the Developer Lab. You can use the real environment provided in the tutorials to learn and complete the experiments.

Quickly configure eRDMA using an eRDMA image

Add the elastic Remote Direct Memory Access (eRDMA) feature to a container environment, such as Docker, to enable faster data transmission and communication. This is useful for containerized applications that require large-scale data transmission and high-performance network communication. This tutorial describes how to use an eRDMA image to quickly configure eRDMA on a GPU-accelerated instance.

Build a model inference environment using TensorRT-LLM

Deploy the TensorRT-LLM open source inference acceleration library to quickly build an LLM inference environment. This approach is primarily suited for natural language processing scenarios, such as intelligent chatbots and text analytics. This tutorial describes how to install and use TensorRT-LLM on a GPU-accelerated instance.

Build a secure large language model inference environment on heterogeneous confidential computing instances

Building a secure inference service for large models to ensure data security during inference is a key concern for businesses. This tutorial describes how to quickly build a secure LLM inference environment using Alibaba Cloud heterogeneous confidential computing instances and the Confidential AI solution.