Merge branch 'main' into bump-release

Author: echarlaix

optimum/intel/openvino/modeling_decoder.py

@@ -28,10 +28,10 @@
 from transformers import AutoModelForCausalLM, PretrainedConfig
 from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
 from transformers.generation import GenerationMixin
-from transformers.generation.configuration_utils import GenerationConfig, GenerationMode
+from transformers.generation.configuration_utils import GenerationConfig
 from transformers.generation.logits_process import LogitsProcessorList
 from transformers.generation.stopping_criteria import StoppingCriteriaList
-from transformers.generation.utils import GenerateOutput
+from transformers.generation.utils import GenerateOutput, GenerationMode
 from transformers.modeling_outputs import CausalLMOutputWithPast
 
 from optimum.utils.normalized_config import NormalizedConfigManager
@@ -386,10 +386,8 @@ def prepare_inputs(
         inputs = {}
         if not self.stateful:
             if past_key_values is not None:
-                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
                 ):
                     if self._pkv_precision == Type.bf16:
                         # numpy does not support bf16, pretending f16, should change to bf16
@@ -499,10 +497,8 @@ def forward(
             if self.use_cache:
                 # Tuple of length equal to : number of layer * number of past_key_value per decoder layer (2 corresponds to the self-attention layer)
                 past_key_values = tuple(self.request.get_tensor(key).data for key in self.key_value_output_names)
-                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
                 ):
                     # Tuple of tuple of length `n_layers`, with each tuple of length equal to 2 (k/v of self-attention)
                     past_key_values = tuple(
@@ -559,10 +555,8 @@ 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(
@@ -581,7 +575,7 @@ 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 +686,7 @@ 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

@@ -812,6 +812,10 @@ 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,
@@ -872,6 +876,8 @@ 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 in ["baichuan2-13b", "olmo"]:
+                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))