Trying ONNX models on CPU.

app.py

@@ -2,12 +2,12 @@ import spaces
 import torch
 import gradio
 import json
+import onnxruntime
 import time
 from datetime import datetime
-from transformers import AutoTokenizer, pipeline
+from transformers import pipeline
 from fastapi import FastAPI
 from fastapi.middleware.cors import CORSMiddleware
-from optimum.onnxruntime import ORTModelForSequenceClassification
 
 # CORS Config - This isn't actually working; instead, I am taking a gross approach to origin whitelisting within the service.
 app = FastAPI()
@@ -25,15 +25,10 @@ print(f"CUDA device: {torch.cuda.get_device_name(torch.cuda.current_device())}")
 
 # "xenova/mobilebert-uncased-mnli" "typeform/mobilebert-uncased-mnli" Fast but small--same as bundled in Statosphere
 
-model_name = "MoritzLaurer/deberta-v3-base-zeroshot-v2.0"
-tokenizer_name = "MoritzLaurer/deberta-v3-base-zeroshot-v2.0"
+model_name = "MoritzLaurer/roberta-large-zeroshot-v2.0-c"
+tokenizer_name = "MoritzLaurer/roberta-large-zeroshot-v2.0-c"
 
-model_name_cpu = "MoritzLaurer/roberta-large-zeroshot-v2.0-c"
-
-model_cpu = ORTModelForSequenceClassification.from_pretrained(model_id=model_name_cpu, subfolder="onnx", file_name="model.onnx")
-tokenizer_cpu = AutoTokenizer.from_pretrained(model_name_cpu)
-
-classifier_cpu = pipeline(task="zero-shot-classification", model=model_cpu, tokenizer=tokenizer_cpu)
+classifier_cpu = pipeline(task="zero-shot-classification", model=model_name, tokenizer=tokenizer_name)
 classifier_gpu = pipeline(task="zero-shot-classification", model=model_name, tokenizer=tokenizer_name, device="cuda:0")
 
 def classify(data_string, request: gradio.Request):