Docs

GPU Compute Service

Document

GPU Compute Service

What's New

New feature release notes

Image release notes

Overview

What is a GPU instance?

Getting started

NVIDIA driver installation guide

Preparations

Purchasing Heterogeneous Compute-optimized instances

Purchasing HPC Heterogeneous instances

User Guide

Instance types

GPU instance naming and type recommendations

GPU Accelerated

GPU Rendering

HPC-optimized

Instance types about to be discontinued

Instance user guide

Installing NVIDIA drivers

Installing NVIDIA Tesla drivers

Automatically installing Tesla driver during instance creation

Quickly installing Tesla drivers in created instances

Installing GPU drivers manually

Installing CUDA toolkits

Installing NVIDIA-Fabric Manager packages

Installing NVIDIA GRID drivers

Automatically loading GRID drivers during instance creation

Installing GRID drivers manually

Uninstalling NVIDIA drivers

Uninstalling NVIDIA Tesla drivers

Uninstalling NVIDIA-Fabric Manager software packages

Uninstalling NVIDIA GRID drivers

Managing HPC clusters

HPC cluster overview

Creating HPC clusters

Expanding HPC clusters

Deleting HPC clusters

Managing cluster events

Cluster diagnosis

Cluster diagnosis overview

Creating diagnosis tasks

Viewing historical diagnosis reports

Diagnosis item overview

Monitoring

Cloud Monitor

Monitoring overview

RDMA network monitoring (hpcpni2)

GPU public image and driver maintenance cycle

Best Practices

Building models

Overview

DeepSeek series models

GPU - Deploying DeepSeek-V4.1-Flash model

GPU - Deploying DeepSeek-V4-Pro model

GPU - Deploying DeepSeek-V4-Flash model

GPU - Deploying DeepSeek-R1/V3 model

GPU - Deploying DeepSeek-V3.2 model

GPU - Deploying DeepSeek-R1/V3 model with small-scale Expert Parallelism

GPU - Deploying DeepSeek-R1/V3.1 model with large-scale Expert Parallelism

GPU - Deploying DeepSeek-R1/V3 model based on xLLM

HPC-Single Node Deployment Plan of DeepSeek-R1 using W4A8

Kimi series models

GPU - Deploying Kimi K3 model

GPU - Deploying Kimi K2.6 model

GPU - Deploying Kimi K2.5 model

Qwen series models

GPU - Deploying Qwen3.8 27B model

GPU - Deploying Qwen 3.8 model

GPU - Deploying Qwen3.8-Flash-Next model

GPU - Deploying Qwen3.5 model

GPU - Deploying Qwen3 model

GLM series models

GPU - Deploying GLM-5.3 model

GPU - Deploying GLM-5.3-Flash model

GPU - Deploying GLM-5.2 model

GPU - Deploying GLM-5.1 model

GPU - Deploying GLM-5 model

MiniMax series models

GPU - Deploying MiniMax-M3 model

GPU - Deploying MiniMax-M2.7 model

GPU - Deploying MiniMax-M2.5 model

HPC – Compatibility between Elastic Scheduled Instances (Segmented) and veRL

HPC-VeRL Code Sandbox Best Practices

Building applications

GPU - Deploying NGC environment

GPU - Deploying PyTorch applications

GPU - Building AIGC capabilities (Linux)

GPU - Building AIGC capabilities (Windows)

GPU-Setting up Isaac Sim 5.1 simulation environment

HPC - Configuring NCCL

HPC - Setting up Slurm compute cluster

Performance testing/tuning

HPC-Single-node/multi-node NCCL test (Easy_NCCL)

HPC-Multi-node NCCL test optimization scheme: affinity-based instance sorting

Assessing GPU performance through single-node/multi-node GPT training (Easy_GPT)

Optimize NCCL communication performance using the NCCL Tuner Plugin

Using NCCL Softbond Plugin to ensure training traffic recovery in the event of a single network interface failure

Proprietary high-performance collective communication library veCCL

GPU Performance Calculator

HPC - Configuring multiple IPs on a single network interface

GPU - Creating RDMA-enabled container image

Installing RDMA Agent in HPC-optimized instances

Using second-level monitoring feature

Troubleshooting

How to troubleshoot an abnormally interrupted A100 GPU-involved training process?

Cause analysis and solutions for DeepEP errors in Ubuntu 24.04 + Nvidia drive 570 + doca 2.9.1 environment

Common NVIDIA GPU XIDs and solutions

FAQ

GPU instance hardware-related FAQ

Terms

Inter-container sharing mGPU

Overview

What is container sharing mGPU

Billing

User Guide

Installing and using mGPU service through Docker Cloud

Use mGPU through VKE

Usage method

Configuration of mGPU compute allocation policy

Documentation

GPU Compute Service

Best Practices

Building models

Kimi series models

GPU - Deploying Kimi K3 model

Copy page

Download PDF

Kimi series models

GPU - Deploying Kimi K3 model

Copy page

Download PDF

GPU - Deploying Kimi K3 model

Last updated: 2026.08.18 16:44:17

Public security network record in Beijing: No. 11010802032137