我做cam++模型export的时候，int8量化之后onnx模型15Mb,未量化后的onnx模型35Mb，
推理时内存占用情况：未量化模型占用81Mb，量化后模型占用67Mb
推理速度情况（多次反复测试）：未量化模型占用108ms，量化后模型占用170ms，量化后速度反而更慢了！
请问modelscope-funasr这个可能是哪里出了问题啊？from funasr import AutoModel
import torch
import onnx
from onnxruntime.quantization import QuantType, quantize_dynamic
import os
model = AutoModel(model="damo/speech_campplus_sv_zh-cn_16k-common",
model_revision="v2.0.2",
device="cpu"
)

dummy_input = torch.rand(1, 553,80)
quant=True
model=model.model
model_script = model #torch.jit.trace(model)
model_path = "camplus.onnx"
input_names=["input"]
output_names=["output"]
verbose=True
torch.onnx.export(
model_script,
dummy_input,
model_path,
verbose=verbose,
opset_version=14,
input_names=input_names,
output_names=output_names,
dynamic_axes={'input': [0, 1]}
)

quant

if quant:
quant_model_path = "camplus_quant.onnx"
if not os.path.exists(quant_model_path):
onnx_model = onnx.load(model_path)
quantize_dynamic(
model_input=model_path,
model_output=quant_model_path,
weight_type=QuantType.QUInt8,
)
print("export Done!") 这是我的转换代码