Set use_cache in ipex model tests (huggingface#649)

Author: jiqing-feng

optimum/intel/ipex/modeling_base.py

@@ -89,6 +89,10 @@ def ipex_jit_trace(model, task, use_cache):
 
     model.config.return_dict = False
 
+    if "past_key_values" in sample_inputs and use_cache:
+        # Make sure the model will output past_key_values in generation tasks
+        model.config.use_cache = True
+
     model = ipex.optimize(model.eval(), dtype=model.dtype, inplace=True)
     # Disable repack while jit tracing to reduce the memory
     ipex._C.disable_jit_linear_repack()

tests/ipex/test_modeling.py

@@ -253,7 +253,7 @@ def test_compare_to_transformers(self, model_arch):
     def test_pipeline(self, model_arch):
         model_id = MODEL_NAMES[model_arch]
         tokenizer = AutoTokenizer.from_pretrained(model_id)
-        model = IPEXModelForCausalLM.from_pretrained(model_id, export=True, use_cache=False)
+        model = IPEXModelForCausalLM.from_pretrained(model_id, export=True)
         model.config.encoder_no_repeat_ngram_size = 0
         model.to("cpu")
         pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)