fxmarty/distilbert-base-uncased-sst2-onnx-int8-for-tensorrt

This model is a fork of distilbert-base-uncased-finetuned-sst-2-english quantized with Optimum library 🤗 using static quantization.

This model can be used as follow:

import onnxruntime
from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForSequenceClassification

session_options = onnxruntime.SessionOptions()
session_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_DISABLE_ALL

tokenizer = AutoTokenizer.from_pretrained("fxmarty/distilbert-base-uncased-sst2-onnx-int8-for-tensorrt")
ort_model = ORTModelForSequenceClassification.from_pretrained(
    "fxmarty/distilbert-base-uncased-sst2-onnx-int8-for-tensorrt",
    provider="TensorrtExecutionProvider",
    session_options=session_options,
    provider_options={"trt_int8_enable": True},
)

inp = tokenizer("TensorRT is a bit painful to use, but at the end of day it runs smoothly and blazingly fast!", return_tensors="np")

res = ort_model(**inp)

print(res)
print(ort_model.config.id2label[res.logits[0].argmax()])
# SequenceClassifierOutput(loss=None, logits=array([[-0.545066 ,  0.5609764]], dtype=float32), hidden_states=None, attentions=None)
# POSITIVE

Inspecting the graph (for example here with netron), we see that it contains Quantize and Dequantize nodes, that will be interpreted by TensorRT to run in INT8: