i am trying to convert my text into its embeddings using a bert model , when i apply this to my my dataset it works fine for some of my inputs then stops and gives that error

i have set TORCH_USE_CUDA_DSA and CUDA_LAUNCH_BLOCKING to 1 and the inputs are not excedding the number of token limit also and this is my embedding code i have tried to free the memory of my gpu and still it didnt work.

import torch
import numpy as np

def embeddings(text):
    if len(text) > 4000:
        flag = text.split(".")
        t1 = flag[:len(flag) // 2]
        t2 = flag[len(flag) // 2:]
        t1 = ".".join(t1)
        t2 = ".".join(t2)
        emb_avg = np.mean([embeddings(t1), embeddings(t2)], axis=0)
        return emb_avg
    else:
        with torch.no_grad(): 
            encoded_input = tokenizer(text, return_tensors='pt')
            encoded_input.to(device)

            output = model(**encoded_input)
            emb = output.encoder_last_hidden_state
            emb_np = emb.cpu().numpy()  
           
            del emb
            del output
            del encoded_input
            gc.collect()
            
            torch.cuda.empty_cache()  
        emb_avg = np.mean(emb_np, axis=1)
        emb_avg = emb_avg.flatten()
        torch.cuda.empty_cache()
        return emb_avg

and im applying to my data set

from tqdm.auto import tqdm
tqdm.pandas()  
df['emb'] = df['abstract'].progress_apply(embeddings)