KTransformers 0.7.2 supports LoRA fine-tuning on DeepSeek V4 Flash's native MXFP4 routed experts, without converting those expert weights to BF16 first.
We completed a 20-step, two-RTX-5090 training smoke, saved both standard and expert LoRA weights, and loaded the adapter in a separate inference process.
See the DeepSeek V4 Flash LoRA quickstart for installation, example YAML, and adapter loading.