threading error onnx fix

app.py

@@ -87,9 +87,12 @@ hide_streamlit_style = """
             """
 st.markdown(hide_streamlit_style, unsafe_allow_html=True)
 
-options = ort.SessionOptions()
-options.intra_op_num_threads=1
-options.inter_op_num_threads=1
+#onnx runtime inference threading changes -- session options must be passed in session run
+# os.environ["OMP_NUM_THREADS"] = "1"
+session_options_ort = ort.SessionOptions()
+session_options_ort.intra_op_num_threads=1
+session_options_ort.inter_op_num_threads=1
+# session_options_ort.execution_mode = session_options_ort.ExecutionMode.ORT_SEQUENTIAL
 
 
 @st.cache(allow_output_mutation=True, suppress_st_warning=True, max_entries=None, ttl=None)
@@ -139,7 +142,7 @@ def sentiment_task_selected(task,
     # create_onnx_model_sentiment(_model=model_sentiment, _tokenizer=tokenizer_sentiment)
 
     #create inference session
-    sentiment_session = ort.InferenceSession(f"{sent_onnx_mdl_dir}/{sent_onnx_mdl_name}")
+    sentiment_session = ort.InferenceSession(f"{sent_onnx_mdl_dir}/{sent_onnx_mdl_name}",sess_options=session_options_ort)
     # sentiment_session_quant = ort.InferenceSession(f"{sent_onnx_mdl_dir}/{sent_onnx_quant_mdl_name}")
 
     return tokenizer_sentiment,sentiment_session
@@ -170,7 +173,7 @@ def zs_task_selected(task,
     # create_onnx_model_zs()
 
     #create inference session from onnx model
-    zs_session = ort.InferenceSession(f"{zs_onnx_mdl_dir}/{zs_onnx_mdl_name}")
+    zs_session = ort.InferenceSession(f"{zs_onnx_mdl_dir}/{zs_onnx_mdl_name}",sess_options=session_options_ort)
     # zs_session_quant = ort.InferenceSession(f"{zs_onnx_mdl_dir}/{zs_onnx_quant_mdl_name}")
 
     return tokenizer_zs,zs_session