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 |
|
Transformer Models |
|
LLM Models (via Foundry Local) |
|
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 for details.
For model conversion and quantization options, see Model Conversion and Quantization (AI Toolkit).