Training and Tinker service support
Hi! We are the ModelScope community, and we offer some choices for developers to train this model👋
ms-swift provides supervised fine-tuning (SFT) and reinforcement learning (RL) support for Qwen3.8-28B. Parameter update schemes include LoRA and full-parameter tuning, together with Megatron-based parallel training; on the RL side, the GRPO family of algorithms is implemented alongside DPO, PPO and GKD, and multi-turn agentic rollouts may be connected to an OpenEnv environment service.
pip install ms-swift -U
# pip install mcore-bridge -U
swift sft --model Qwen/Qwen3.8-28B --tuner_type lora --dataset <your-data>
# megatron sft --model Qwen/Qwen3.8-28B --tuner_type lora --dataset <your-data>
Where a Tinker-style client-server training paradigm is preferred, Twinkle✨ may be used. Its interface is a superset of the Tinker API: the client is responsible for data construction, loss definition and orchestration of the training loop, while forward-backward computation and optimizer updates are executed on the server over HTTP. ModelScope hosts a Qwen3.8-28B server, so no local GPU is required on the client side; the server may equally be deployed privately, in which case only the client's base_url parameter changes, and a single base-model deployment can host multiple LoRA tenants concurrently. SFT, DPO, GRPO, PPO, GKD and OpenEnv and AgentEnv(Kimi)-based multi-turn rollouts are likewise available in this mode.
pip install twinkle-kit -U
client = init_twinkle_client(base_url='https://www.modelscope.cn/twinkle', api_key=...)
...
Related links: ms-swift docs · Twinkle docs · Twinkle client cookbook