use new st caching

Unfortunately no more hash_funcs :(

utils.py

@@ -1,25 +1,16 @@
 from typing import Tuple
 
-import streamlit as st
-
-import torch
-from torch.quantization import quantize_dynamic
-from torch import nn, qint8, Tensor
-from torch.nn import Parameter
-from transformers import PreTrainedModel, PreTrainedTokenizer
 from optimum.bettertransformer import BetterTransformer
 from mbart_amr.constraints.constraints import AMRLogitsProcessor
 from mbart_amr.data.tokenization import AMRMBartTokenizer
+import streamlit as st
+import torch
+from torch.quantization import quantize_dynamic
+from torch import nn, qint8
 from transformers import MBartForConditionalGeneration
 
 
-st_hash_funcs = {PreTrainedModel: lambda model: model.name_or_path,
-                 PreTrainedTokenizer: lambda tokenizer: tokenizer.name_or_path,
-                 Parameter: lambda parameter: parameter.data,
-                 Tensor: lambda tensor: tensor.cpu()}
-
-
-@st.cache(show_spinner=False, hash_funcs=st_hash_funcs, allow_output_mutation=True)
+@st.cache_resource(show_spinner=False)
 def get_resources(multilingual: bool, quantize: bool = True, no_cuda: bool = False) -> Tuple[MBartForConditionalGeneration, AMRMBartTokenizer, AMRLogitsProcessor]:
     """Get the relevant model, tokenizer and logits_processor. The loaded model depends on whether the multilingual
     model is requested, or not. If not, an English-only model is loaded. The model can be optionally quantized
@@ -51,7 +42,6 @@ def get_resources(multilingual: bool, quantize: bool = True, no_cuda: bool = Fal
     return model, tokenizer, logits_processor
 
 
-@st.cache(show_spinner=False, hash_funcs=st_hash_funcs)
 def translate(text: str, src_lang: str, model: MBartForConditionalGeneration, tokenizer: AMRMBartTokenizer, **gen_kwargs) -> str:
     """Translates a given text of a given source language with a given model and tokenizer. The generation is guided by
     potential keyword-arguments, which can include arguments such as max length, logits processors, etc.