Fakeaudio Onnx Vaiml Repro

--- tags: - onnx - audio-classification - ryzen-ai - vitis-ai library_name: onnxruntime --- # FakeAudio ONNX used for the VAIML reproduction This repository contains the exact FakeAudio ONNX artifact tested in [`odobias/amd-vaiml-window-partition-repro`](https://github.com/odobias/amd-vaiml-window-partition-repro). It is published separately so AMD and other investigators can reproduce the VAIML compilation result without access to the private benchmark asset store. ## Artifact - File: `model.onnx` - Size: 127,844,774 bytes - SHA-256: `d4b2dde8f4f95862ecb9f5e90821d853adea3774e7c48857f60ee6084d5e85c6` - ONNX opset: 17 - Producer recorded in ONNX: PyTorch 2.1.2 The ONNX file contains no additional provenance metadata. ## Download ```powershell hf download odobias/fakeaudio-onnx-vaiml-repro model.onnx --local-dir . ``` Verify the downloaded artifact before reproducing the result: ```powershell (Get-FileHash .\model.onnx -Algorithm SHA256).Hash.ToLowerInvariant() ``` ## Observed VAIML result With Ryzen AI SDK 1.8.0-beta on an AMD Ryzen AI 9 HX 370, AMD's BF16 VAIML configuration fails flexible L1 buffer placement at layer 224. ONNX Runtime then executes the entire optimized graph on CPU: 0 NPU nodes, 512 CPU nodes, and 100% CPU offload.

Source context: 0 downloads · 0 likes · Pipeline audio-classification · Library onnxruntime · Repo odobias/fakeaudio-onnx-vaiml-repro