Fix bloom generation (#736)

* Fix bloom generation

* remove unused variable

* add style

* add message error

* update model id

Author: echarlaix

optimum/intel/openvino/modeling_decoder.py

@@ -587,11 +587,11 @@ def _deduplicate_inputs(self, model_inputs: Dict):
         )
         for input_name, input_tensor in model_inputs.items():
             if input_name not in ["input_ids", "beam_idx"]:
-                if not isinstance(input_tensor, Tensor):
+                if input_name not in self.key_value_input_names:
                     upd_model_inputs[input_name] = input_tensor[indicies]
                 else:
-                    shape = input_tensor.shape
-                    dtype = input_tensor.element_type
+                    shape = input_tensor.shape if isinstance(input_tensor, Tensor) else list(input_tensor.shape)
+                    dtype = input_tensor.element_type if isinstance(input_tensor, Tensor) else Type(input_tensor.dtype)
                     upd_batch_size = indicies.shape[0]
                     if self.config.model_type == "bloom":
                         upd_batch_size *= self.config.num_attention_heads

tests/openvino/test_modeling.py

@@ -879,14 +879,14 @@ def test_beam_search(self, model_arch):
         ov_model_stateless.config.eos_token_id = None
         transformers_model.config.eos_token_id = None
 
-        for gen_config in gen_configs:
+        for idx, gen_config in enumerate(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))
+            self.assertTrue(torch.allclose(ov_stateful_outputs, transformers_outputs), f"generation config : {idx}")
             ov_stateless_outputs = ov_model_stateless.generate(**tokens, generation_config=gen_config)
-            self.assertTrue(torch.allclose(ov_stateless_outputs, transformers_outputs))
+            self.assertTrue(torch.allclose(ov_stateless_outputs, transformers_outputs), f"generation config : {idx}")
 
 
 class OVModelForMaskedLMIntegrationTest(unittest.TestCase):

tests/openvino/utils_tests.py

@@ -65,7 +65,7 @@
     "internlm2": "katuni4ka/tiny-random-internlm2",
     "levit": "hf-internal-testing/tiny-random-LevitModel",
     "longt5": "hf-internal-testing/tiny-random-longt5",
-    "llama": "fxmarty/tiny-llama-fast-tokenizer",
+    "llama": "HuggingFaceM4/tiny-random-LlamaForCausalLM",
     "llama_awq": "HuggingFaceH4/tiny-random-LlamaForCausalLM",
     "llama_gptq": "hf-internal-testing/TinyLlama-1.1B-Chat-v0.3-GPTQ",
     "m2m_100": "hf-internal-testing/tiny-random-m2m_100",