# VitisAI EP Model Support

The VitisAI EP (Execution Provider) within Windows ML supports input models in the following formats.

## Model Support Table

| Model Type                         | Support                                                                                                                             |
|------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------|
| **CNN Models**                     | - Original float (FP32) model with automatic BF16 conversion during compilation<br/>- Quantized QDQ model using A8W8 configuration  |
| **Transformer Models**             | - Original float (FP32) model with automatic BF16 conversion during compilation<br/>- Quantized QDQ model using A16W8 configuration |
| **LLM Models (via Foundry Local)** | - Quantized and pre-compiled LLM models<br/>- Support for custom models through Olive recipe                                        |

### Note

- For CNN and Transformer models, you can use either the original float model (with automatic BF16 conversion) or a quantized QDQ model. Quantization can reduce model size and improve inference performance.
- For LLMs, Foundry Local provides pre-built models that auto-detect the NPU. Custom LLM deployment may require model preparation using the Olive recipe or Ryzen AI OGA workflow. See [Windows ML LLM examples](https://github.com/amd/RyzenAI-SW/tree/main/WinML/LLM) for details.
- For model conversion and quantization options, see [Model Conversion and Quantization (Foundry Toolkit)](https://ryzenai.docs.amd.com/projects/WinML/en/latest/model_conversion.html).