Quick Start

Choose your backend

Local GPU
Tinker SDK
Modal Cloud (Coming Soon)

Runs SDPO training and vLLM inference on your own GPU. Requires >= 24 GB VRAM.

# Install dependencies
uv sync --extra local

# Docker setup (recommended)
cd docker
cp .env.local.example .env
# Edit .env -- set TELEGRAM_BOT_TOKEN
docker compose --profile local up --build

The first run downloads Qwen3-8B (~16 GB). Expect the vLLM health check to take 10–20 minutes on first start.

Full Local backend reference

Requirements, all config variables, services, and manual setup

Uses Tinker’s hosted inference and training. No GPU required.

# Install dependencies
uv sync --extra tinker

# Docker setup
cd docker
cp .env.tinker.example .env.tinker
# Edit .env.tinker -- set TELEGRAM_BOT_TOKEN + TINKER_API_KEY
docker compose -f docker-compose.tinker.yml --env-file .env.tinker up --build

Full Tinker backend reference

Tinker config, model naming gotchas, and services

Runs distillation remotely on Modal (L40S GPUs). No local GPU needed.

# Install dependencies
uv sync --extra local

# Authenticate with Modal
uv run modal token new

# Deploy
export HF_TOKEN=...
export CLAAS_BASE_MODEL_ID=Qwen/Qwen3-8B
uv run modal deploy -m claas.deploy

The deployed app exposes the same API at https://your-app--claas-distill-fastapi-app.modal.run.

Full Modal backend reference

Deployment, config, and LoRA storage on Modal Volumes

# Check inference endpoint curl http://localhost:8000/v1/models -H "Authorization: Bearer sk-local" # Check CLaaS API curl http://localhost:8080/ # List LoRA adapters curl http://localhost:8080/v1/lora

Getting Started

Training Backends

Evaluation

Reference

Quick Start

Prerequisites

Choose your backend

Full Local backend reference

Full Tinker backend reference

Verify your setup

​Quick Start

​Prerequisites

​Choose your backend

Full Local backend reference

Full Tinker backend reference

Full Modal backend reference

​Verify your setup

Quick Start

Prerequisites

Choose your backend

Verify your setup