Update on "[ONNX] Add huggingface models into CI tests"

1. Add a list of HF models to CI tests. The PR intends to build them from Config, but some of them are not supported with Config. NOTE: Loaded from pre-trained model could potentially hit [uint8/bool conflict](huggingface/transformers#21013) when a newer version of transformers is used.
    - Dolly has torch.fx.Node in OnnxFunction attribute, which is currently not supported.
    - Falcon and MPT has unsupported user coding to Dynamo.
2. Only update GPT2 exporting with real tensor to Config, as FakeMode rises unequal input errors between PyTorch and ORT. The reason is that [non-persistent buffer is not supported](#107211)

[ghstack-poisoned]

.ci/docker/common/install_onnx.sh

@@ -32,3 +32,18 @@ pip_install onnx-weekly==1.15.0.dev20230717
 
 # TODO: change this when onnx-script is on testPypi
 pip_install onnxscript-preview==0.1.0.dev20230809 --no-deps
+
+# Cache the transformers model to be used later by ONNX tests. We need to run the transformers
+# package to download the model. By default, the model is cached at ~/.cache/huggingface/hub/
+IMPORT_SCRIPT_FILENAME="/tmp/onnx_import_script.py"
+as_jenkins echo 'import transformers; transformers.AutoModel.from_pretrained("sshleifer/tiny-gpt2"); transformers.AutoTokenizer.from_pretrained("sshleifer/tiny-gpt2");' > "${IMPORT_SCRIPT_FILENAME}"
+
+# Need a PyTorch version for transformers to work
+pip_install --pre torch --index-url https://download.pytorch.org/whl/nightly/cpu
+# Very weird quoting behavior here https://github.com/conda/conda/issues/10972,
+# so echo the command to a file and run the file instead
+conda_run python "${IMPORT_SCRIPT_FILENAME}"
+
+# Cleaning up
+conda_run pip uninstall -y torch
+rm "${IMPORT_SCRIPT_FILENAME}" || true