AMD Ryzen AI Software 1.9 Adds Dynamic LoRA Switching for Selected Text Models

AMD's Ryzen AI Software 1.9 adds dynamic LoRA switching for selected text models, allowing applications to swap task adapters on the NPU while keeping the quantized base model loaded.

AMD announced the update Nov. 2. The supported model list includes Qwen2.5-3B, Qwen3-4B and Qwen3.5-4B.

LoRA, or low-rank adaptation, specializes a model by adding a small set of learned weights while leaving the base model's weights fixed. In the earlier static deployment method, developers merged each adapter into the base model before compiling it. Preparing versions for several tasks could therefore duplicate the base weights. The new approach keeps adapters separate, so an application can select one by name without loading the base model again.

AMD's setup begins with a base model quantized through AMD Quark using INT4 AWQ. Developers train each adapter against that frozen, quantized model and save it separately. They compile the base and adapters as distinct artifacts, then load the base once and switch adapters at runtime. AMD cautioned that an adapter trained against a full-precision model may deliver less benefit when applied to the quantized base.

The demonstration used a Qwen3-4B INT4 base with one adapter for the Spider text-to-SQL benchmark and another for medical multiple-choice questions in MedMCQA. In the demonstration, Spider execution accuracy was 63.0%, compared with 47.5% before adaptation. MedMCQA exact match was 53.0%, up from 42.0%. Each result used 200 examples.

The release also adds long-context support and the ability to import INT4 models produced through quantization-aware training. It includes support for additional models.