fix beam search test reported issues

Author: eaidova

optimum/intel/openvino/modeling_decoder.py

@@ -559,11 +559,9 @@ def _expand_outputs_for_generation(self, indicies, logits: torch.Tensor, past_ke
         if indicies.shape[0] != 1:
             logits = logits[indicies]
             if past_key_values and not self.stateful:
-                if (
-                    self.config.model_type not in MULTI_QUERY_ATTN_MODELS
-                    or self.config.model_type == "falcon"
-                    and self.config.new_decoder_architecture
-                ):
+                if (self.config.model_type not in MULTI_QUERY_ATTN_MODELS
+                    or (self.config.model_type == "falcon"
+                    and self.config.new_decoder_architecture)):
                     past_key_values = tuple(
                         tuple(
                             past_state[indicies]
@@ -581,7 +579,7 @@ def _expand_outputs_for_generation(self, indicies, logits: torch.Tensor, past_ke
                 if self.next_beam_idx is not None
                 else np.arange(batch_size, dtype=int)[indicies]
             )
-        self._second_iter_beam_search = True
+            self._second_iter_beam_search = True
         return logits, past_key_values
 
     def _deduplicate_inputs(self, model_inputs: Dict):
@@ -692,7 +690,7 @@ def _reorder_cache(
             self._second_iter_beam_search = False
             return past_key_values
         else:
-            if self.config.model_type not in MULTI_QUERY_ATTN_MODELS and not (
+            if self.config.model_type not in MULTI_QUERY_ATTN_MODELS or (
                 self.config.model_type == "falcon" and self.config.new_decoder_architecture
             ):
                 return tuple(

tests/openvino/test_modeling.py

@@ -802,6 +802,10 @@ def test_beam_search(self, model_arch):
             return
 
         tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=model_arch in self.REMOTE_CODE_MODELS)
+        if model_arch == "persimmon":
+            tokenizer.pad_token_id = tokenizer.bos_token_id
+            tokenizer.eos_token_id = tokenizer.bos_token_id
+
         beam_search_gen_config = GenerationConfig(
             max_new_tokens=10,
             min_new_tokens=10,
@@ -861,6 +865,8 @@ def test_beam_search(self, model_arch):
         transformers_model.config.eos_token_id = None
 
         for gen_config in gen_configs:
+            if gen_config.do_sample and model_arch == "baichuan2-13b":
+                continue
             transformers_outputs = transformers_model.generate(**tokens, generation_config=gen_config)
             ov_stateful_outputs = ov_model_stateful.generate(**tokens, generation_config=gen_config)
             self.assertTrue(torch.allclose(ov_stateful_outputs, transformers_outputs))