Code to reproduct ORT Gather noder discrepancy.

test.py

import numpy as np
from onnxruntime import InferenceSession
import torch
import torch.nn as nn

# Create simple model with one Gather node
model = nn.Embedding(num_embeddings=10, embedding_dim=5)
x = torch.randint(0, 10, (3,))
torch.onnx.export(
    model,
    args=(x,),
    f="model.onnx",
    input_names=["input_ids"],
    output_names=["output"],
    opset_version=14,
    dynamic_axes={"input_ids": {0: "batch_size"}, "output": {0: "batch_size"}}
)

for provider in ["CPUExecutionProvider", "CUDAExecutionProvider"]:
    print(f"testing {provider} ...")
    session = InferenceSession("model.onnx", providers=[provider])
    io_binding = session.io_binding()
    io_binding.bind_cpu_input("input_ids", np.ones(3, dtype=np.int64) * 10)
    io_binding.bind_output("output")
    try:
        session.run_with_iobinding(io_binding)
        print(f"{provider} did not raise an error")
    except Exception as e:
        print(f"{provider} raised the following error:\n{e}")
    print()