User Roles
👩💻 ML Engineer
👨💼 Business Developer
🔧 System Admin
🤖 Agent / 3rd-party App
💡 Data Annotator
↓
🌐
Access Layer — Developer & Open Platform
🔑
Platform Business API
Full RESTful coverage for asset & user management; Swagger/OpenAPI standard docs;
Token auth + rate limiting
REST API
OpenAPI
Rate Limiting
⚡
Inference API Gateway
Fully OpenAI-compatible inference endpoint; Serverless API billed per token; Commercial
API Key management + budget limits
OpenAI Compatible
Serverless
API Key
🔌
MCP Gateway
Centralized registration and management of remote MCP Servers; unified tool-call entry
point; supports SSE / HTTP connection types
MCP Protocol
Tools API
SSE
🐍 Python SDK
Lightweight client with direct access to all core
platform assets: upload/download models & datasets, call inference APIs, manage Spaces, and
more.
pypi install csghub-sdk
dataset / model / inference
↓
📦
Core Asset Management — Full-Spectrum AI Asset Hub
🤖
Model Hub
Model Cards, model tree, cross-asset association, multi-framework metadata
Model CardTree
🗂️
Datasets
Online preview of Parquet/CSV/JSON, Dataset Card, version management
ViewerGit LFS
💻
Code Repos
Standard Git workflow, PR/Issue collaboration, branch management
GitPR
🚀
App Spaces
Host Gradio · Streamlit · Docker · Nginx · MCP Server apps
GradioDocker
💬
Prompt Library
Visual Prompt management with categorization and version iteration
PromptVersioning
🛠️
Skills / MCP
SkillHub components + MCP repository management, one-click deploy
SkillHubMCP Repo
Unified Lifecycle Git version tracking +
cross-asset reproducibility
·
XNet Large-File Acceleration Proprietary
high-speed download
·
Open-Ecosystem Mirror HuggingFace sync
·
Collections Multi-asset aggregation
·
Multi-Source Sync OpenCSG community /
enterprise sources
↓
⚙️
Compute Services — Full Model Lifecycle Development
🌐
Model Inference
Serverless public inference billed per token
Dedicated Instances exclusive GPU always-on service
vLLMTGITritonNIM
🎓
Model Fine-tuning
Visual parameter configuration, multi-framework support, auto-export of fine-tuned
results back to asset library
LLaMA-FactoryMS-SWIFT
📊
Model Evaluation
Standard benchmarks + custom dataset evaluation, multi-dimensional scoring reports
OpenCompassEvalScopelm-eval
📓
Notebook Sandbox
Ready-to-use deep learning dev environment, dynamic lifecycle management
Jupyter LabPyTorchTF
↓
🔬
Data Toolchain — DataFlow Processing Foundation
🕸️
Multi-Source Ingestion
One-click document extraction, database connectors, web crawling — multi-channel
aggregation
Web CrawlDB Connector
🔄
Format Conversion
Multi-format document inter-conversion; unified processing for structured and
unstructured data
PDF/CSV/JSONParquet
🧹
Visual Cleaning Pipeline
100+ professional operators via drag-and-drop; filter / dedup / mask / tag; Celery
distributed scheduling
100+ OperatorsCeleryDAG
🏷️
Data Annotation
Integrated Label Studio; supports text/image/audio multimodal annotation task
distribution
Label StudioMultimodal
↓
🖥️
Compute Management Center
Scheduling & Resource Orchestration
Cluster onboarding, node partitioning, K8s + Volcano scheduling engine
GPU Virtualization & Partitioning
HAMi soft-cut vGPU + MIG hardware partitioning, resource isolation
Compute Observability
Global health dashboard, load monitoring, node alerting system
Instance & Job Management
Always-on services + offline jobs — global hosting and anomaly intervention
Compute Metering & Usage Analytics
GPU-Hours consumption traceability, multi-dimensional statistics, report
export
🛡️
Platform Foundation & Access Control
Users & Organizations (RBAC)
Three-tier permission model, multi-tenant isolation, fine-grained resource
control
Unified SSO Authentication
Pluggable integration with Casdoor, Paraview, custom IdP
Security Audit & Compliance
Sensitive-word filtering, operation audit logs, compliance risk controls
Metering & Billing System
SKU pricing, top-up / bills / invoicing, GPU-Hours / Token multi-dimensional
metering
Admin Console
Asset management, ops configuration, system monitoring dashboard, License
management
↓
💰
Compute Commercialization Platform
🪙
Serverless On-Demand Billing
Pay per actual token consumed, auto-scaling, zero management overhead,
multi-dimensional metering (LLM/Audio/Image)
Token BillingAuto-Scale
📅
Dedicated Instance Time Billing
Minute-level billing for Notebook/inference/fine-tuning/evaluation; monthly/annual
resource reservation with discount model
ReservedDedicated GPU
📈
Commercial Operations
Order management, revenue-sharing model, self-service invoicing
Revenue ShareSpotInvoice
↓
🏗️
Infrastructure Layer
☸️
Kubernetes
Cluster orchestration foundation, Volcano scheduler, multi-cluster management
🎮
GPU Clusters
Support for mainstream GPUs globally and domestic accelerators (Huawei Ascend, etc.)
🗄️
Distributed Storage
Large-file Git LFS + object storage + XNet accelerated download
📨
Messaging & Scheduling
Celery distributed task queue, async task orchestration and monitoring
🔭
Observability
Prometheus + Grafana metrics monitoring, operation audit logs
Edition Matrix
🪶
CSGHub Lite
Desktop single-node edition, offline install, zero
dependencies — ideal for individual development and intranet rapid validation
Local DeployCLIGUI
🌱
CSGHub CE Community
Open-source community edition, full core features,
community resource sync
Open SourceK8sOpenCSG Sync
🏢
CSGHub EE Enterprise
Enterprise edition with compute management, commercial
billing, SSO, multi-source sync quotas, and HA deployment
Multi-Source SyncRBACBillingHA