- 支持transformers模型
- 支持megatron模型
- 支持vLLM采样器
- 支持dataset、dataloader、reward、advantage、权重同步等基本组件
- 支持数据集packing、padding_free、流式数据集
- 支持纯文本模型的PT/SFT
- 支持纯文本模型的GRPO
- 支持kernels
- 兼容NPU生态
- 支持多LoRA租户
- 支持twinkle client训练
- 支持tinker API的兼容性
- 支持租户资源控制、水位控制
- 支持checkpoint的保存上传、下载
- 支持魔搭免费训练集群
- Support for Transformers models
- Support for Megatron models
- Support for vLLM sampler
- Support for basic components including dataset, dataloader, reward, advantage, and weight synchronization
- Support for dataset packing, padding-free, and streaming datasets
- Support for PT/SFT of text-only models
- Support for GRPO of text-only models
- Support for kernels
- Compatibility with NPU ecosystem
- Support for multi-LoRA tenants
- Support for Twinkle client training
- Support for Tinker API compatibility
- Support for tenant resource control and watermark control
- Support for checkpoint saving, uploading, and downloading
- Support for ModelScope free training cluster
- 支持多模态模型
- 支持megatron VPP
- 支持liger kernel
- 支持transformers模型的ulysses/ring-attention
- 兼容transformers v5的tp、pp
- 支持多轮RL
- 支持gym训练
- 支持GAPO、GSPO算法
- 支持GKD、on-policy-distill等蒸馏算法
- 支持DPO对齐训练
- 支持colocate RL训练
- Preprocess支持batched
- 对多replica的支持和粘滞路由
- Support for multimodal models
- Support for Megatron VPP
- Support for Liger kernel
- Support for Ulysses/Ring-Attention for Transformers models
- Compatibility with Transformers v5 TP and PP
- Support for multi-turn RL
- Support for Gym training
- Support for GAPO and GSPO algorithms
- Support for distillation algorithms such as GKD and on-policy distillation
- Support for DPO alignment training
- Support for colocate RL training
- Support for batched preprocessing
- Support for multiple replicas and sticky routing