DOCSEARCHbyAlgolia
Submit Docs

Qubrid AI Documentation

website·Indexed 33 days ago

https://docs.platform.qubrid.com

Platform for deploying and managing Large Language Models (LLMs) and GPU infrastructure. Features include serverless inference with an OpenAI-compatible API, on-demand GPU virtual machines, bare metal servers, GPU clusters, and the Qubrid AI Controller for on-premise server management. Includes tools for RAG, model fine-tuning, Hugging Face model deployment, and integrations with AI coding agents like Cursor, VS Code, Cline, and Claude Code.

Tokens
19.7K
Snippets
63
Records
151
Agent score
58%

What's inside Qubrid

  1. Overview of Qubrid AI Controller capabilities

    The Qubrid AI Controller is an enterprise software solution for on-premise servers that provides IT administrators with a centralized management interface for GPU infrastructure. Key capabilities include:

    • Controlling GPU resource and AI model allocation for users.
    • Monitoring usage across nodes and workloads.
    • Updating drivers and maintaining bare-metal AI environments.
    • Centrally managing AI package installation, model deployment, and Docker workloads.
    • Providing end-users with a personal console to access assigned GPUs, models, and programming environments.

  2. Overview of Qubrid AI Model Studio

    Qubrid AI Model Studio is a unified environment designed for developers, researchers, and enterprises to build, deploy, manage, and scale AI/ML workloads. It provides a code-optional interface for custom model creation, tools for benchmarking and comparing leading AI models, and capabilities for fine-tuning pre-trained models using high-performance compute resources.

  3. Forecast expenses with the API Cost Estimator

    The API Cost Estimator allows users to plan model usage expenses by simulating costs based on estimated daily input and output token usage. It provides real-time daily and monthly cost projections, displays base pricing (input/output token rates) for selected models, and suggests credit top-ups based on usage patterns.

  4. Use Open WebUI for AI model access

    Open WebUI is a self-hosted, offline AI interface that supports Ollama and OpenAI API. It opens automatically to provide a user-friendly UI for accessing models without needing SSH or Jupyter configuration.

  5. Compare serverless models side-by-side

    The Model Comparison tool enables users to evaluate 2 to 4 Qubrid serverless models simultaneously. Users can run the same prompt across selected models to compare streaming responses, latency, and token metrics to determine the best model for production.

  6. Common workflows for Claude Code

    Claude Code is an agentic coding CLI that can be used for several developer tasks: summarizing architecture and tracing features, fixing bugs by searching the repository, writing and running tests, performing multi-file refactors with diff reviews, and automating chores like lint fixes and dependency updates.
    claude "write tests for the auth module, run them, and fix failures"

  7. Understand Qubrid Serverless Inference pricing and billing

    Qubrid serverless inference provides access to hosted models via a standard API using shared capacity. Billing is usage-based, typically calculated per 1 million tokens for text models, meaning costs scale with actual consumption rather than reserved capacity. There are no upfront commitments, idle fees, or cluster management overhead. Serverless endpoints are subject to fair-use rate limits and are recommended for development, benchmarks, and unpredictable workloads. For predictable latency or private hardware, users should use On-Demand GPUs or GPU Clusters instead.

  8. Integrate Claude Code with Qubrid OSS models

    Claude Code is an agentic coding CLI that requires the Anthropic Messages API. Qubrid provides OpenAI-compatible endpoints. To use Qubrid models (like GPT OSS 120B or Qwen3 Coder) within Claude Code, you must deploy a translation proxy that converts Anthropic-format requests to OpenAI-format chat completions and vice versa.

    Prerequisites:

    • Python 3.10+ (for the proxy)
    • Node.js or the native Claude Code installer
    • A Qubrid account and API key
    • Sufficient Qubrid credits

  9. Understand the LLM inferencing process in Qubrid AI Controller

    Inferencing with LLMs on the Qubrid GPU infrastructure involves four primary stages:

    1. Model Deployment: Placing the trained LLM on production environments or inference servers.
    2. Input Processing: Preprocessing data via tokenization or encoding to match the model's input format.
    3. Model Inference: Passing processed input through the LLM to generate predictions or text.
    4. Output Post-processing: Refining the model's output for better readability or application usability.

  10. Test models using the Model Playground

    The Model Playground is a unified interface for experimenting with models and APIs. It allows users to:

    • Evaluate performance with custom prompts and view metrics like TTFT (Time to First Token), TPS (Tokens Per Second), and total tokens.
    • Adjust model parameters and configure system prompts with few-shot examples.
    • Toggle streaming on/off via the UI.
    • Generate API keys and copy API requests directly from the interface.

  11. Set up the claude-code-proxy (Recommended)

    The claude-code-proxy is a lightweight tool that maps Claude Code model tiers (Opus, Sonnet, Haiku) to specific Qubrid OSS models. It listens on port 8082 by default.
    # Clone and install
git clone https://github.com/fuergaosi233/claude-code-proxy.git
cd claude-code-proxy
pip install -r requirements.txt

# Configure .env
cp .env.example .env
# Edit .env with these values:
# OPENAI_API_KEY="your-qubrid-api-key"
# OPENAI_BASE_URL="https://qubrid.com/v1"
# BIG_MODEL="openai/gpt-oss-120b"
# MIDDLE_MODEL="openai/gpt-oss-120b"
# SMALL_MODEL="Qwen/Qwen3-Coder-Next"
# HOST="127.0.0.1"
# PORT="8082"

# Start the proxy
python start_proxy.py

  12. Create a Qubrid account

    To get started with Qubrid, create a new account at qubrid.com. You can sign up using Google or GitHub for faster activation, or use the manual account creation process by providing your name, email, mobile number, organization, and a password.