Minimal repros for an incompatibility between ONNX Runtime and the AMD NPU (VitisAI) EP, hit when running a model on the NPU via Foundry Local.

Three ways to reproduce it, from easiest to most involved. See Bisection results below for the summary of what causes the bug and where it's fixed.

• Windows on an AMD Ryzen AI (XDNA NPU) machine
• AMD Vitis AI EP installed (bug reproduces on 1.8.63; fixed in 1.8.68+)
• Python 3.11+ (for repros 2 and 3)

No Python required. Download and install the latest Foundry Local CLI (0.10.1) from the Foundry-Local releases page, then run any model with an AMD NPU variant:

foundry model run qwen2.5-0.5b

The CLI ships with Core.WinML 1.2.3 (ORT 1.26.0), which triggers the bug on Vitis EP 1.8.63.

Uses foundry_amd_npu.py to download and run an NPU model via the Foundry Local Python SDK. Same underlying Core.WinML as the CLI, so hits the same bug.

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

python foundry_amd_npu.py

# Or pick a different model alias (default: qwen2.5-0.5b):
python foundry_amd_npu.py --model phi-3.5-mini

Pass any model alias in the catalog that has an AMD NPU variant via --model (check with foundry model list --variants — look for NPU / VitisAIExecutionProvider).

To confirm the diagnosis, downgrade the SDK to 1.1.0 (Core.WinML built against ORT 1.23.2.3) — it runs successfully against Vitis EP 1.8.63:

pip install --force-reinstall "foundry-local-sdk-winml==1.1.0"
python foundry_amd_npu.py

genai_amd_npu.py bypasses Foundry Local and runs the model directly with onnxruntime-genai-winml. This lets you swap the ONNX Runtime version independently of the Foundry Local Core, so you can bisect ORT / ORT-GenAI versions against the AMD NPU EP.

It accepts the same --model alias as foundry_amd_npu.py (default: qwen2.5-0.5b) and looks up the cached model folder in the Foundry Local model cache. Run repro 1 or 2 first so the AMD NPU variant is downloaded.

python -m venv .venv-genai
.\.venv-genai\Scripts\Activate.ps1
pip install -r requirements-genai.txt

# Install the ONNX Runtime version you want to test against
# (windowsml does NOT install onnxruntime unless you use its [with-ort] extra,
# so we install it explicitly to control the version we're bisecting.)
pip install "onnxruntime==1.26.0"

python genai_amd_npu.py

# Or with a different model alias:
python genai_amd_npu.py --model phi-3.5-mini

# To bisect, uninstall and reinstall with a different version:
pip install --force-reinstall "onnxruntime==1.23.2"
pip install --force-reinstall "onnxruntime-genai-winml==0.13.2"

Root cause: The AMD Vitis AI EP 1.8.63 introduced the incompatibility with ONNX Runtime versions 1.25.0 or later. EP 1.8.62 works; 1.8.63 exhibits the bug. The bug is fixed in Vitis EP version 1.8.68.

Rows with blank CLI/SDK were tested via genai_amd_npu.py (direct ORT / ORT-GenAI, no Foundry Local Core).

Cliff: The WinML SDK jumped from ORT 1.23.2.3 (1.1.0) straight to 1.26.0 (1.2.0) — no WinML SDK build was ever shipped against ORT 1.24 or 1.25, which is why the regression only surfaced at SDK 1.2.0.