genai_config.json missing for embeddinggemma-300m_npu_rai_1.8.0_npu_4K. com.ryzenai:CastAvx` Error

[venureddyj]

Bug Title

Ryzen AI 1.8.0 Catalog Mismatch: embeddinggemma-300m_npu_rai_1.8.0_npu_4K fails with com.ryzenai:CastAvx Error

Environment Layout

• Processor: AMD Ryzen AI Strix Point (STX)
• Software Suite: Ryzen AI Software Suite v1.8.0 (Clean Installation)
• Active Kernel Ecosystem: Python 3.11 / Miniforge3
• Core Runtime Packages:
  • onnxruntime v1.19.x / v1.20.x
  • onnxruntime-genai-directml-ryzenai v0.14.0

Description of the Issue

When attempting to load the official 1.8.0 NPU 4K embedding asset (amd/embeddinggemma-300m_npu_rai_1.8.0_npu_4K) via standard onnxruntime.InferenceSession targeting the VitisAIExecutionProvider, initialization fails immediately.

The underlying model binary contains a compiled reference to com.ryzenai:CastAvx. However, this operator has been completely deprecated/removed from the stable 1.8.0 production C++ driver library (onnxruntime_vitis_ai_custom_ops.dll), causing an unrecoverable validation crash on session boot.

Additionally, the repository lacks the standard genai_config.json map despite being hosted under the OGA-optimized release profile layers.

Steps to Reproduce

1. Download repository amd/embeddinggemma-300m_npu_rai_1.8.0_npu_4K.
2. Initialize an Inference Session under Ryzen AI 1.8.0:

import onnxruntime as ort
session = ort.InferenceSession("model.onnx", providers=["VitisAIExecutionProvider"])

Expected Behavior

The pre-compiled 1.8.0 model should load cleanly onto the physical NPU tiles via the Vitis AI execution lanes without referencing deleted 1.7-preview CPU-fallback operator signatures.

Actual Terminal Exception Trace

onnxruntime.capi.onnxruntime_inference_collection.Fail: [ONNXRuntimeError] : 1 : FAIL : Load model failed:Fatal error: com.ryzenai:CastAvx(-1) is not a registered function/op