bge-m3-onnx

ONNX export of BAAI/bge-m3 targeting the AMD Ryzen AI NPU via the VitisAI Execution Provider in ONNXRuntime.

Files

File	Description
bge-m3.onnx	ONNX graph (weights stored externally)
bge-m3.onnx.data	Weight data (~2.17 GB, referenced by the graph)

Export details

The model was exported from the original PyTorch weights using torch.onnx.export with a static input shape of (1, 512) — text is always padded or truncated to exactly 512 tokens. Opset 17 was used.

Because the weight tensor exceeds the ONNX protobuf 2 GB limit, weights are stored in the companion .onnx.data file using the single-file external data format required by VitisAI.

After export, the graph nodes are topologically sorted using Kahn's algorithm. torch.onnx.export emits nodes in Python execution order which is not guaranteed to be topological, and ONNXRuntime rejects such graphs at load time.

License

MIT — inherited from BAAI/bge-m3.