Getting different sentence embeddings when using model on CPU and GPU

Beginners
1

I am trying to get sentence embedding from pretrained Roberta. When I use the model on CPU and GPU, I get two different sentence embeddings. Following is the code snippet:

from transformers import RobertaConfig, RobertaModel, RobertaTokenizer
import torch
import numpy as np
device = ("cuda" if torch.cuda.is_available() else "cpu")
# Initializing tokenizer
tokenizer = RobertaTokenizer.from_pretrained("roberta-base")

# Initializing a RoBERTa configuration
configuration = RobertaConfig()
configuration.vocab_size = tokenizer.vocab_size

# Initializing a model from the configuration
model = RobertaModel(configuration)
model = model.to(device)
model = model.eval()

with torch.no_grad():
    tokenized_task = tokenizer('random_sentence_check_v000', return_tensors="pt")
    outputs = model(**tokenized_task.to(device))
    embedding = outputs.pooler_output.squeeze(0).cpu().numpy().tolist()

I use colab for my experiments. Shouldn’t the embedding be the same as the weights are pretrained?