KTransformers

Local Fine-Tuning

KTransformers fine-tuning extends the same local ownership idea from inference to adaptation. If a workstation can run a MoE model through KTransformers, the project direction is that the same class of machine should be able to train LoRA adapters for it.

System Direction

ComponentRole
GPUAttention, shared paths, residual LoRA capacity, and distributed training control.
CPU expert backendLarge MoE expert weights through AMX BF16/INT8/INT4 SFT backends.
LLaMA-FactoryUser-facing training workflow and dataset/config management.
KT integrationuse_kt: true, kt_config, and backend-aware expert execution.

The current public path is LoRA SFT with LLaMA-Factory. This technical page explains the direction; the Fine-Tuning section contains the runnable user docs.