AMD adds dynamic LoRA switching to Ryzen AI Software 1.9

AMD says its Ryzen AI Software now lets developers compile a base model once and switch LoRA adapters dynamically at inference time on the NPU, a change the company says can cut duplication when deploying multiple specialist models on an AI PC. In an AMD technical article published Nov. 2, 2026, the company said the feature is available in Ryzen AI Software release 1.9 and is supported for selected text models including Qwen2.5-3B, Qwen3-4B and Qwen3.5-4B.

Evidence and context

LoRA, or Low-Rank Adaptation, fine-tunes a large language model by training small adapter weights while keeping the base model frozen. AMD said the previously supported deployment method on Ryzen AI was static merge, where LoRA weights are folded into the base model before compilation. The company said that approach works for a single adapter but becomes inefficient when multiple specialist adapters are needed, because each merged model duplicates the same base weights.

With dynamic LoRA, AMD said one quantized base model remains resident on the NPU while small adapters are compiled as separate artifacts and activated by name. The article said this lets an application switch from one specialist to another, such as a SQL-focused model to a medical one, without reloading the base model.

AMD also warned that adapters must be trained against the already-quantized base model. In the article, the company said the NPU base is quantized to INT4 in its example, and that training a LoRA adapter against a full-precision model before applying it to the quantized base can reduce the adapter's benefit on device.

The company described a four-step flow: quantize the base model once with AMD Quark using INT4 AWQ; fine-tune each LoRA adapter against that frozen quantized base and save it unmerged; compile the base and each adapter for the NPU with a single model_generate call per adapter; and serve by loading the base once and switching adapters at runtime.

AMD said its example used a Qwen3-4B INT4 base with two adapters: a text-to-SQL adapter trained on Spider and a medical multiple-choice adapter trained on MedMCQA. On the Ryzen AI NPU, AMD reported that the Spider execution accuracy rose from 47.5% to 63.0%, a gain of 15.5 percentage points, while MedMCQA exact match increased from 42.0% to 53.0%, a gain of 11.0 points. The article said both results were based on n=200 examples.

AMD said a runnable example is available in its Ryzen AI Software examples repository and that Ryzen AI Software release 1.9 also introduces long-context support, importing INT4-quantized models produced through quantization-aware training, and support for additional models. The article did not disclose a broader model list beyond the three named text models or provide independent benchmark comparisons outside the cited example.