Triton ONNX runtime backend slower than onnxruntime python client on CPU

[Mitix-EPI]

Description
When deploying an ONNX model using the Triton Inference Server's ONNX runtime backend, the inference performance on the CPU is noticeably slower compared to running the same model using the ONNXRuntime Python client directly. This performance discrepancy is observed under identical conditions, including the same hardware, model, and input data.

Triton Information
TRITON_VERSION=2.46.0

To Reproduce

model used:

wget -O model.onnx https://contentmamluswest001.blob.core.windows.net/content/14b2744cf8d6418c87ffddc3f3127242/9502630827244d60a1214f250e3bbca7/08aed7327d694b8dbaee2c97b8d0fcba/densenet121-1.2.onnx

Triton server (ONNX runtime)

config.pbtxt

name: "test_densenet" 
platform: "onnxruntime_onnx"

Python clients

Triton client

import numpy as np
import tritonclient.grpc as grpcclient
import tritonclient.grpc._infer_input as infer_input

grpcclient = grpcclient.InferenceServerClient(url='localhost:9178')

i = infer_input.InferInput('data_0', [1, 3, 224, 224], 'FP32')
i.set_data_from_numpy(np.zeros((1, 3, 224, 224), dtype=np.float32))

%%timeit
res = grpcclient.infer(model_name="test_densenet", inputs=[i])

results: 473 ms ± 87.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

ONNX Runtime

import onnxruntime as ort

ort_sess = ort.InferenceSession('model.onnx')
test_inputs = {"data_0": np.zeros((1, 3, 224, 224), dtype=np.float32)}

%%timeit
ort_sess.run(["fc6_1"], test_inputs)

results: 159 ms ± 23.9 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)