Many tests of the Python module fail: Unsupported model IR version: 14, max supported IR version: 13 #33429

________ TestLlamaRotaryEmbedding.test_hf_token_rotary_batched_pos_ids _________

self = <test_parity_rotary_embedding.TestLlamaRotaryEmbedding testMethod=test_hf_token_rotary_batched_pos_ids>

    def test_hf_token_rotary_batched_pos_ids(self):
        x_bnsh = torch.randn(self.config.batch_size, self.config.num_heads, 1, self.config.head_size)
        cos_hf, sin_hf = self.llama_hf.get_cos_sin_cache(self.config.sequence_length)
        pos_ids = torch.stack([torch.tensor([2]) for _ in range(self.config.batch_size)])
        output_hf = self.llama_hf(x_bnsh, cos_hf, sin_hf, pos_ids)  # output is BxNxSxH
    
        x_bsnh = x_bnsh.transpose(1, 2)
        x_bsd = x_bsnh.reshape(self.config.batch_size, 1, self.config.hidden_size)
        cos_ms, sin_ms = self.llama_ms.get_cos_sin_cache()
        onnx_graph = self.create_onnx_graph(x_bsd.shape, pos_ids.shape, cos_ms, sin_ms, interleaved=False)
        inputs_ort = {
            "input": x_bsd.detach().cpu().numpy(),
            "position_ids": pos_ids.detach().cpu().numpy(),
        }
    
        # Compare outputs as BxSxNxH
>       self.run_ort_ep_tests(onnx_graph, inputs_ort, output_hf.transpose(1, 2).detach().cpu().numpy())

transformers/test_parity_rotary_embedding.py:405: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
transformers/test_parity_rotary_embedding.py:267: in run_ort_ep_tests
    sess = ort.InferenceSession(onnx_graph, providers=[ep])
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
../../../../stage/usr/local/lib/python3.12/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:683: in __init__
    self._create_inference_session(providers, provider_options, disabled_optimizers)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <onnxruntime.capi.onnxruntime_inference_collection.InferenceSession object at 0x120f1a20aa50>
providers = ['CPUExecutionProvider'], provider_options = [{}]
disabled_optimizers = None

    def _create_inference_session(self, providers, provider_options, disabled_optimizers=None):
        available_providers = C.get_available_providers()
    
        # Validate that TensorrtExecutionProvider and NvTensorRTRTXExecutionProvider are not both specified
        if providers:
            has_tensorrt = any(
                provider == "TensorrtExecutionProvider"
                or (isinstance(provider, tuple) and provider[0] == "TensorrtExecutionProvider")
                for provider in providers
            )
            has_tensorrt_rtx = any(
                provider == "NvTensorRTRTXExecutionProvider"
                or (isinstance(provider, tuple) and provider[0] == "NvTensorRTRTXExecutionProvider")
                for provider in providers
            )
            if has_tensorrt and has_tensorrt_rtx:
                raise ValueError(
                    "Cannot enable both 'TensorrtExecutionProvider' and 'NvTensorRTRTXExecutionProvider' "
                    "in the same session."
                )
        # Tensorrt and TensorRT RTX can fall back to CUDA if it's explicitly assigned. All others fall back to CPU.
        if "NvTensorRTRTXExecutionProvider" in available_providers:
            if (
                providers
                and any(
                    provider == "CUDAExecutionProvider"
                    or (isinstance(provider, tuple) and provider[0] == "CUDAExecutionProvider")
                    for provider in providers
                )
                and any(
                    provider == "NvTensorRTRTXExecutionProvider"
                    or (isinstance(provider, tuple) and provider[0] == "NvTensorRTRTXExecutionProvider")
                    for provider in providers
                )
            ):
                self._fallback_providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
            else:
                self._fallback_providers = ["CPUExecutionProvider"]
        elif "TensorrtExecutionProvider" in available_providers:
            if (
                providers
                and any(
                    provider == "CUDAExecutionProvider"
                    or (isinstance(provider, tuple) and provider[0] == "CUDAExecutionProvider")
                    for provider in providers
                )
                and any(
                    provider == "TensorrtExecutionProvider"
                    or (isinstance(provider, tuple) and provider[0] == "TensorrtExecutionProvider")
                    for provider in providers
                )
            ):
                self._fallback_providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
            else:
                self._fallback_providers = ["CPUExecutionProvider"]
        else:
            self._fallback_providers = ["CPUExecutionProvider"]
    
        # validate providers and provider_options before other initialization
        providers, provider_options = check_and_normalize_provider_args(
            providers, provider_options, available_providers
        )
    
        # Print a warning if user passed providers to InferenceSession() but the SessionOptions instance
        # already has provider information (e.g., via add_provider_for_devices()). The providers specified
        # here will take precedence.
        if self._sess_options is not None and (providers or provider_options) and self._sess_options.has_providers():
            warnings.warn(
                "Specified 'providers'/'provider_options' when creating InferenceSession but SessionOptions has "
                "already been configured with providers. InferenceSession will only use the providers "
                "passed to InferenceSession()."
            )
    
        session_options = self._sess_options if self._sess_options else C.get_default_session_options()
    
        self._register_ep_custom_ops(session_options, providers, provider_options, available_providers)
    
        if self._model_path:
            sess = C.InferenceSession(session_options, self._model_path, True, self._read_config_from_model)
        else:
>           sess = C.InferenceSession(session_options, self._model_bytes, False, self._read_config_from_model)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
E           onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : /usr/ports/misc/py-onnxruntime/work-py312/onnxruntime-1.31.0/onnxruntime/core/graph/model.cc:257 onnxruntime::Model::Model(ModelProto &&, const PathString &, const IOnnxRuntimeOpSchemaRegistryList *, const logging::Logger &, const ModelOptions &) Unsupported model IR version: 14, max supported IR version: 13