[NPU] Fix hf generate with save/load generation config for Python (cpp backend) (#12509)

* Fix hf generate with save/load generation config

* Small fix

* Fix based on comments

python/llm/src/ipex_llm/transformers/npu_model.py

@@ -440,6 +440,15 @@ class _BaseAutoModelClass:
                 model.kv_len = config_dict['kv_len']
                 model.vocab_size = config_dict['vocab_size']
                 model.logits_buffer = torch.empty(1, 1, model.vocab_size, dtype=torch.float32)
+                if model.can_generate():
+                    try:
+                        model.generation_config = GenerationConfig.from_pretrained(
+                            pretrained_model_name_or_path,
+                            subfolder=subfolder,
+                            **kwargs,
+                        )
+                    except (OSError, TypeError):
+                        pass
             except:
                 invalidInputError(False,
                                   "Fail to InitLLMPipeline.")

python/llm/src/ipex_llm/transformers/npu_pipeline_model/convert_pipeline.py

@@ -466,6 +466,8 @@ def convert_llm_for_deploy(model: torch.nn.Module,
                        "lm_head_low_bit": lm_head_low_bit}
         model.config.update(update_dict)
         model.config.save_pretrained(save_directory)
+        if model.can_generate():
+            model.generation_config.save_pretrained(save_directory)
 
         from .qwen import convert_qwen_layer, convert_fused_qwen_layer
         from .qwen import convert_lm_head_and_embedding