Add openvino export for InternLM2 and Orion architectures (#628)

* support more models in export

* add orion

* update tests

optimum/exporters/openvino/__main__.py

@@ -202,7 +202,6 @@ def main_export(
         quantization_config = getattr(config, "quantization_config", None)
         do_gptq_patching = quantization_config and quantization_config["quant_method"] == "gptq"
         model_type = config.model_type.replace("_", "-")
-
         if model_type not in TasksManager._SUPPORTED_MODEL_TYPE:
             custom_architecture = True
         elif task not in TasksManager.get_supported_tasks_for_model_type(
@@ -220,6 +219,20 @@ def main_export(
             )
         if is_transformers_version(">=", "4.36") and model_type in SDPA_ARCHS_ONNX_EXPORT_NOT_SUPPORTED:
             loading_kwargs["attn_implementation"] = "eager"
+        # there are some difference between remote and in library representation of past key values for some models,
+        # for avoiding confusion we disable remote code for them
+        if (
+            trust_remote_code
+            and model_type in {"falcon", "mpt", "phi"}
+            and ("with-past" in task or original_task == "auto")
+            and not custom_export_configs
+        ):
+            logger.warning(
+                f"Model type `{model_type}` export for task `{task}` is not supported for loading with `trust_remote_code=True`"
+                "using default export configuration, `trust_remote_code` will be disabled. "
+                "Please provide custom export config if you want load model with remote code."
+            )
+            trust_remote_code = False
 
     # Patch the modules to export of GPTQ models w/o GPU
     if do_gptq_patching:

optimum/exporters/openvino/convert.py

@@ -345,7 +345,7 @@ def ts_patched_forward(*args, **kwargs):
                     input_dict = dict(zip(keys, tuple_input))
                     kwargs[input_name] = input_dict
                 outputs = patched_forward(*args, **kwargs)
-                return tuple(outputs.values())
+                return tuple([value if not isinstance(value, list) else tuple(value) for value in outputs.values()])
 
             patcher.patched_forward = ts_patched_forward
 

optimum/exporters/openvino/model_configs.py

@@ -74,7 +74,7 @@ def init_model_configs():
 
 
 @register_in_tasks_manager("baichuan", *["text-generation", "text-generation-with-past"], library_name="transformers")
-class BaichaunOpenVINOConfig(TextDecoderOnnxConfig):
+class BaichaunOpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
     DEFAULT_ONNX_OPSET = 13
     NORMALIZED_CONFIG_CLASS = NormalizedTextConfig.with_args(
         num_layers="num_hidden_layers", num_attention_heads="num_attention_heads", hidden_size="hidden_size"
@@ -400,3 +400,21 @@ class Starcoder2OpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
     DUMMY_INPUT_GENERATOR_CLASSES = (DummyTextInputGenerator, MistralDummyPastKeyValuesGenerator)
     DUMMY_PKV_GENERATOR_CLASS = MistralDummyPastKeyValuesGenerator
     NORMALIZED_CONFIG_CLASS = NormalizedTextConfig
+
+
+@register_in_tasks_manager("internlm2", *["text-generation", "text-generation-with-past"], library_name="transformers")
+class InternLM2OpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
+    DEFAULT_ONNX_OPSET = 14
+
+    DUMMY_INPUT_GENERATOR_CLASSES = (DummyTextInputGenerator, MistralDummyPastKeyValuesGenerator)
+    DUMMY_PKV_GENERATOR_CLASS = MistralDummyPastKeyValuesGenerator
+    NORMALIZED_CONFIG_CLASS = NormalizedTextConfig
+
+
+@register_in_tasks_manager("orion", *["text-generation", "text-generation-with-past"], library_name="transformers")
+class OrionOpenVINOConfig(TextDecoderWithPositionIdsOnnxConfig):
+    DEFAULT_ONNX_OPSET = 14
+
+    DUMMY_INPUT_GENERATOR_CLASSES = (DummyTextInputGenerator, MistralDummyPastKeyValuesGenerator)
+    DUMMY_PKV_GENERATOR_CLASS = MistralDummyPastKeyValuesGenerator
+    NORMALIZED_CONFIG_CLASS = NormalizedTextConfig

optimum/exporters/openvino/model_patcher.py

@@ -513,5 +513,5 @@ def __init__(
     ):
         super().__init__(config, model, model_kwargs)
         # model has first inference buffers initialization
-        if self._model.lm_head.first_flag:
+        if hasattr(self._model.lm_head, "first_flag"):
             self._model(torch.ones((1, 10), dtype=torch.int64), torch.ones((1, 10), dtype=torch.int64))

tests/openvino/test_modeling.py

@@ -524,10 +524,12 @@ class OVModelForCausalLMIntegrationTest(unittest.TestCase):
         "stablelm",
         "starcoder2",
         "phi",
+        "internlm2",
+        "orion",
     )
     GENERATION_LENGTH = 100
     IS_SUPPORT_STATEFUL = is_openvino_version(">=", "2023.3")
-    REMOTE_CODE_MODELS = ("chatglm", "minicpm", "baichuan2", "jais", "qwen")
+    REMOTE_CODE_MODELS = ("chatglm", "minicpm", "baichuan2", "jais", "qwen", "internlm2", "olmo", "orion")
 
     @parameterized.expand(SUPPORTED_ARCHITECTURES)
     def test_compare_to_transformers(self, model_arch):

tests/openvino/utils_tests.py

@@ -50,6 +50,7 @@
     "gptj": "hf-internal-testing/tiny-random-GPTJModel",
     "hubert": "hf-internal-testing/tiny-random-HubertModel",
     "ibert": "hf-internal-testing/tiny-random-ibert",
+    "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",
@@ -69,6 +70,8 @@
     "mpt": "hf-internal-testing/tiny-random-MptForCausalLM",
     "mt5": "stas/mt5-tiny-random",
     "nystromformer": "hf-internal-testing/tiny-random-NystromformerModel",
+    "olmo": "katuni4ka/tiny-random-olmo",
+    "orion": "katuni4ka/tiny-random-orion",
     "pegasus": "hf-internal-testing/tiny-random-pegasus",
     "pix2struct": "fxmarty/pix2struct-tiny-random",
     "phi": "echarlaix/tiny-random-PhiForCausalLM",