FakeAudio ONNX used for the VAIML reproduction
This repository contains the exact FakeAudio ONNX artifact tested in
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
hf download odobias/fakeaudio-onnx-vaiml-repro model.onnx --local-dir .
Verify the downloaded artifact before reproducing the result:
(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.