Add Llama2 Onnx Model E2E test

[mszhanyi]

Description

1. Leverage hugging face API to verify the converted onnx model.
2. config.json , tokenization.model , etc are generated by Option 3. The test target is the model converted by this repo (option 1), so I only copied the config files.

Motivation and Context

llama related code hasn't a pipeline to run E2E test like https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/python/tools/transformers/models/stable_diffusion/pipeline_stable_diffusion.py

[kunal-vaishnavi]

Couple of questions:

• Can you store the extra files in the same location that the saved PyTorch model is in before running convert_to_onnx? In other words, can you store the extra files in /meta-llama2 instead? This will put the model and its related files in the same location.

onnxruntime/tools/ci_build/github/azure-pipelines/bigmodels-ci-pipeline.yml

Line 315 in 8245dd6

python3 -m models.llama.convert_to_onnx -m meta-llama/Llama-2-7b-hf --output llama2-7b-fp16 --precision fp16 --execution_provider cuda --input /meta-llama2 --small_gpu ;\

• There is already an E2E test using Hugging Face's Optimum at the end of the E2E notebook. Can we use/modify that instead since it shows batch inference? This will allow us to ensure the notebook is up-to-date as well.

[mszhanyi]

Couple of questions:

• Can you store the extra files in the same location that the saved PyTorch model is in before running convert_to_onnx? In other words, can you store the extra files in /meta-llama2 instead? This will put the model and its related files in the same location.

onnxruntime/tools/ci_build/github/azure-pipelines/bigmodels-ci-pipeline.yml

Line 315 in 8245dd6

python3 -m models.llama.convert_to_onnx -m meta-llama/Llama-2-7b-hf --output llama2-7b-fp16 --precision fp16 --execution_provider cuda --input /meta-llama2 --small_gpu ;\

• There is already an E2E test using Hugging Face's Optimum at the end of the E2E notebook. Can we use/modify that instead since it shows batch inference? This will allow us to ensure the notebook is up-to-date as well.

Thank you, I'll leverage your existing example as possible.
By the way, is there a way to run the E2E inference with onnxruntime only .

Maybe we could hold this PR and your example, like tokerizer generation, would be updated by Gen API.
It'd best to cover out own code rather than Optimum in CI. @snnn

import onnxruntime_genai as og
model=og.Model(f'models/microsoft/phi-2', device_type)
tokenizer = model.CreateTokenizer()
....

https://github.com/microsoft/onnxruntime-genai?tab=readme-ov-file#sample-code-for-phi-2-in-python:~:text=import%20onnxruntime_genai%20as%20og%0A%0Amodel%3Dog.Model(f%27models/microsoft/phi%2D2%27%2C%20device_type)%0A%0Atokenizer%20%3D%20model.CreateTokenizer()