Fix pip install path issue for certain packages

[xuechendi]

Fix relative path import.

Error Log before fixing:

After adding complete path, I am able to run serving.

1. install steps

conda create -n llm-on-ray python=3.9
conda activate llm-on-ray
pip install .[cpu,ui] -f https://developer.intel.com/ipex-whl-stable-cpu -f https://download.pytorch.org/whl/torch_stable.html
source $(python -c "import oneccl_bindings_for_pytorch as torch_ccl;print(torch_ccl.cwd)") /env/setvars.sh

# Ray Started

2. test step:

python inference/serve.py --config_file inference/models/llama-2-7b-chat-hf.yaml

[xuechendi]

@carsonwang @KepingYan , please help to take a review.

[xwu-intel]

@KepingYan

From https://docs.python.org/3/library/sys.html#sys.path

python script.py command line: prepend the script’s directory. If it’s a symbolic link, resolve symbolic links.

We should test and make sure it works in the following both cases and should allow the script to be run from any path:

1. user installed the whole package and run serve.py.
2. user only installed the dependencies and run serve.py directly.

[xuechendi]

@KepingYan

From https://docs.python.org/3/library/sys.html#sys.path

python script.py command line: prepend the script’s directory. If it’s a symbolic link, resolve symbolic links.

We should test and make sure it works in the following both cases and should allow the script to be run from any path:

1. user installed the whole package and run serve.py.
2. user only installed the dependencies and run serve.py directly.

Yeah, after all, the readme should provide User a smooth experience to Run a 'hello world' demo.
If adding my cloned llm-on-ray to python path is a must, may need to add it to README as well.
Otherwise, provide full module path during import will help to avoid python package finding issues.

[xwu-intel]

We also need to consider the distributed nature of Ray that all workers should find the package. (may be the cause of this issue)

Following the industrial practices, I would suggest move inference/finetune into a package llm_on_ray namespace and use absolute module path to import module, in this way, ui/inference/finetune can cross import as well.

For users, they can launch serve as python -m llm_on_ray.inference.serve ...
For developers, they can use pip install -e to install package in-place and debug.

@xuechendi We will discuss locally and let you know.

[xwu-intel]

@xuechendi We decided to use package path (move inference/finetune/... into a package llm_on_ray namespace) for import to avoid name conflict such as
from llm_on_ray.inference.inference_config import InferenceConfig

And use module when running
python -m llm_on_ray.inference.serve

Also need to install serve.py and finetune.py as bin commands (llm_on_ray-serve & llm_on_ray-finetune) that user can run it directly from command line.

@KepingYan could you fix this in #79 as discussed?

[xuechendi]

@xwu99 , That sounds great! This is also something I felt quite unusual that after installing llm-on-ray, the package name is not "llm-on-ray" but "inference". Great to know that you are going to fix the package name. :)

Please feel free to close this PR once you have merge the other one. I would be willing to test the new installation by then.

[xwu-intel]

fixed by #106