Twinkle Twinkle
  • Home
  • Quick Start
  • Documentation
  • Blog
  • Cookbook
  • Community
EN 中文
ESC

Searching...

No results found

↑↓ Navigate ↵ Select
Powered by Hugo Blox
  • Blog
    • Sequence Parallel & Ring Attention: Training with Ultra-Long Contexts
    • Ascend NPU Support: Fused Operators and Flash Linear Attention
    • Two Execution Modes: torchrun (Local) vs Ray (Distributed)
    • Multi-LoRA: Concurrent Multi-Tenant Training on Shared GPUs
    • TUI & Auto-Research: An AI Agent for Training Control
    • OpenEnv Integration: Connecting External Environments to RL Training
    • Free LLM Training on ModelScope: Twinkle Training-as-a-Service
  • Cookbook
    • Shell Launch
    • SFT (FSDP2)
    • Megatron TP
    • NPU (Ascend)
    • EP + MoE
    • GRPO
    • GKD Distill
    • Embedding
    • Multi-Turn RL
    • DPO
    • Multimodal
  • Documentation
    • Quick Start
    • Concepts & Architecture
      • Components
      • Runtime Modes
      • Multi-Tenancy
      • Server & Client
      • Training as a Service
      • Cookbook
      • NPU Support
      • Architecture
    • API Reference
    • Usage Guide
      • Training Guide
      • Twinkle Installation
      • Server and Client
        • Overview
        • Server
        • Observability
        • Twinkle Client
        • Tinker Client
      • NPU (Ascend) Quick Start Guide
      • Twinkle Training Service on ModelScope
      • Agentic RL Deployment and Training
      • Qwen3.5 Training Best Practices
      • Embedding Training
    • 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
        • DeepSeek-V4 Template
        • Tool Call Parsers
        • Template
      • Preprocessor and Filter
        • Preprocessor
        • Built-in Preprocessors
        • Filter
      • Data Loading
        • DataLoader
      • Task Processor
        • InputProcessor
      • Model
        • Supported Models
        • TwinkleModel
        • TransformersModel
        • MultiLoraTransformersModel
        • MegatronModel
        • MultiLoraMegatronModel
      • Sampler
        • Sampler
        • vLLMSampler
        • TorchSampler
      • Reward
        • Reward
        • GSM8K Reward
        • MultiModal Reward
        • OlympiadBench Reward
      • Advantage
        • Advantage
        • GRPOAdvantage
        • RLOOAdvantage
      • Hub
        • Hub
      • Checkpoint Engine
        • CheckpointEngine
        • NCCLCheckpointEngine
        • HCCLCheckpointEngine
      • Metrics
        • TrainMetric
        • LossMetric
        • Accuracy
        • CompletionRewardMetric
        • DPOMetric
        • GRPOMetric
        • EmbeddingMetric
        • Building Metrics
      • Loss
        • InfoNCE 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
      • Kernel
        • Twinkle Kernel
      • Training Middleware
        • DeviceMesh/DeviceGroup
        • Expert Parallel (EP)
        • Sequence Parallel (SP)
        • Padding-Free Training
        • RemoteClass
        • TwinkleClient
      • CLI
        • CLI
      • Notifier
        • Notifier
      • Agentic
        • Agentic Preprocessor
        • Protocol
        • Multi-Turn Rollout
        • Tools & ToolManager
        • Environments (Envs)
        • Multi-Turn Tool Usage Guide
      • Auto
        • Auto-Research
        • SkillProvider
  • Community
  • Need help?
  • Community
  • GitHub ↗
  • Quick Start
  • Concepts & Architecture
    • Components
    • Runtime Modes
    • Multi-Tenancy
    • Server & Client
    • Training as a Service
    • Cookbook
    • NPU Support
    • Architecture
  • API Reference
  • Usage Guide
    • Training Guide
    • Twinkle Installation
    • Server and Client
      • Overview
      • Server
      • Observability
      • Twinkle Client
      • Tinker Client
    • NPU (Ascend) Quick Start Guide
    • Twinkle Training Service on ModelScope
    • Agentic RL Deployment and Training
    • Qwen3.5 Training Best Practices
    • Embedding Training
  • 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
      • DeepSeek-V4 Template
      • Tool Call Parsers
      • Template
    • Preprocessor and Filter
      • Preprocessor
      • Built-in Preprocessors
      • Filter
    • Data Loading
      • DataLoader
    • Task Processor
      • InputProcessor
    • Model
      • Supported Models
      • TwinkleModel
      • TransformersModel
      • MultiLoraTransformersModel
      • MegatronModel
      • MultiLoraMegatronModel
    • Sampler
      • Sampler
      • vLLMSampler
      • TorchSampler
    • Reward
      • Reward
      • GSM8K Reward
      • MultiModal Reward
      • OlympiadBench Reward
    • Advantage
      • Advantage
      • GRPOAdvantage
      • RLOOAdvantage
    • Hub
      • Hub
    • Checkpoint Engine
      • CheckpointEngine
      • NCCLCheckpointEngine
      • HCCLCheckpointEngine
    • Metrics
      • TrainMetric
      • LossMetric
      • Accuracy
      • CompletionRewardMetric
      • DPOMetric
      • GRPOMetric
      • EmbeddingMetric
      • Building Metrics
    • Loss
      • InfoNCE 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
    • Kernel
      • Twinkle Kernel
    • Training Middleware
      • DeviceMesh/DeviceGroup
      • Expert Parallel (EP)
      • Sequence Parallel (SP)
      • Padding-Free Training
      • RemoteClass
      • TwinkleClient
    • CLI
      • CLI
    • Notifier
      • Notifier
    • Agentic
      • Agentic Preprocessor
      • Protocol
      • Multi-Turn Rollout
      • Tools & ToolManager
      • Environments (Envs)
      • Multi-Turn Tool Usage Guide
    • Auto
      • Auto-Research
      • SkillProvider
  • Need help?
  • Community
  • GitHub ↗
Edit this page
Documentation
Usage Guide

Usage Guide

Training Guide Twinkle Installation Server and Client NPU (Ascend) Quick Start Guide Twinkle Training Service on ModelScope Agentic RL Deployment and Training Qwen3.5 Training Best Practices Embedding Training
Training Guide Twinkle Installation Server and Client NPU (Ascend) Quick Start Guide Twinkle Training Service on ModelScope Agentic RL Deployment and Training Qwen3.5 Training Best Practices Embedding Training
docs

© 2026 ModelScope. Licensed under Apache License 2.0.

Made with Hugo Blox. Create yours →