Twinkle DOCUMENTATION

Usage Guide

• Quick Start
• Twinkle Installation
• Server and Client
• NPU (Ascend) Quick Start Guide
• Twinkle Training Service on ModelScope
• Qwen3.5 Training Best Practices

Components

• Dataset
  • Basic Dataset Components
  • Lazy Loading Dataset
  • Fixed-Length Packing Dataset
  • Streaming Dataset
  • Streaming Fixed-Length Packing Dataset
• Data Format
  • Message
  • Trajectory
  • Model Input
  • Model Output
  • Sampling Output
  • Model Output
• Template
  • Template
  • Template mapping
• Preprocessor and Filter
  • Preprocessor
  • Built-in Preprocessors
  • Filter
• Data Loading
  • DataLoader
• Task Processor
  • InputProcessor
  • GRPOLossProcessor
• Model
  • TwinkleModel
  • TransformersModel
  • MultiLoraTransformersModel
  • MegatronModel
  • MultiLoraMegatronModel
• Sampler
  • Sampler
  • vLLMSampler
  • TorchSampler
• Reward
  • Reward
  • GSM8K Reward
  • MultiModal Reward
  • OlympiadBench Reward
• Advantage
  • Advantage
  • GRPOAdvantage
  • RLOOAdvantage
• Gym
  • Gym
• Hub
  • Hub
• Checkpoint Engine
  • CheckpointEngine
  • NCCLCheckpointEngine
  • HCCLCheckpointEngine
• Metrics
  • TrainMetric
  • LossMetric
  • Accuracy
  • CompletionRewardMetric
  • DPOMetric
  • Building Metrics
• Loss
  • Cross Entropy
  • Chunked Cross Entropy
  • DPO Loss
  • GKD Loss
  • GRPO Loss
  • MSE Loss
  • Building New Loss
• Loss Scale
  • Loss Scale
• LRScheduler
  • CosineWarmupScheduler
  • LinearWarmupScheduler
• Patch
  • Patch
• Plugin
  • Plugin
  • Service Security
• Kernel
  • Twinkle Kernel Module
• Training Middleware
  • DeviceMesh/DeviceGroup
  • RemoteClass

Indices and tables

• Index

• Module Index

• Search Page