good

app.py

@@ -12,7 +12,7 @@ model_name = st.selectbox(
 )
 auth_token = os.environ.get('TOKEN') or True
 
-@st.cache(hash_funcs={tokenizers.Tokenizer: lambda tokenizer: hash(tokenizer.to_str())})
+@st.cache(hash_funcs={tokenizers.Tokenizer: lambda tokenizer: hash(tokenizer.to_str())}, allow_output_mutation=True)
 def load_model(model_name: str):
     with st.spinner('Loading models...'):
         tokenizer = transformers.AutoTokenizer.from_pretrained(model_name, use_auth_token=auth_token)
@@ -93,16 +93,15 @@ def merge_dialog_data(
 @st.cache
 def inference(model_name: str, sample: dict):
     tokenizer, model = load_model(model_name)
-    with st.spinner('Running inference...'):
-        tokenized_dialog = tokenize_dialog_data(tokenizer, sample, MAX_SEQ_LENGTH, sorted_dialog_columns)
-        tokens = merge_dialog_data(tokenizer, tokenized_dialog)
-        with torch.inference_mode():
-            logits = model(**tokens).logits
-            probas = torch.sigmoid(logits)[0].cpu().detach().numpy()
-        return probas
+    tokenized_dialog = tokenize_dialog_data(tokenizer, sample, MAX_SEQ_LENGTH, sorted_dialog_columns)
+    tokens = merge_dialog_data(tokenizer, tokenized_dialog)
+    with torch.inference_mode():
+        logits = model(**tokens).logits
+        probas = torch.sigmoid(logits)[0].cpu().detach().numpy()
+    return probas
 
-
-probas = inference(model_name, sample)
+with st.spinner('Running inference...'):
+    probas = inference(model_name, sample)
 st.metric(
     label='Вероятность того, что последний ответ диалогового агента релевантный',
     value=probas[0]