Adding QEFFAutoModel i.e. model loader for loading any type of model.

[ochougul]

• Added QEfficient.QEFFAutoModel that acts as loader for any kind of GENAI model
• Added QEFFBaseModel that acts as parent class for any GENAI model loaded via QEFFAutoModel.from_pretrained
• Added common interface for export and transform functions and made them compatible with QEFFBaseModel
• Cleaned __init__ files
• cleaned infer API file

[vbaddi]

Also, in general can we start tapping necessary info. in the logger.info() the log file can sit in the same qeff_models root directory.

[anujgupt-github]

@ochougul - it would help if you push readme updates also part of this PR, so that it's easy to visualize how you want users to use the interface, and review can be staged along with the intent.

[ochougul]

Also, in general can we start tapping necessary info. in the logger.info() the log file can sit in the same qeff_models root directory.

Will do this in separate PR.

[ochougul]

@ochougul - it would help if you push readme updates also part of this PR, so that it's easy to visualize how you want users to use the interface, and review can be staged along with the intent.

Updated README and notebooks

[vbaddi]

Also, as discussed, I feel it does make sense to wrap transform() inside the pre_trained() API, since it makes equivalent to transformers repo. ref: https://github.com/huggingface/optimum-nvidia?tab=readme-ov-file#generate

[ochougul]

This is the hierarchy of classes being followed.
QEFFCommonModelLoader recognizes the type of the model using config of the model and tries to initialize the correct type of the model class i.e. QEFFAutoModelForCausalLM etc.
This is used when user only passes name of the model to infer API.

Whereas, if user wants to use low level APIs. user can directly initialize QEFFAutoModelForCausalLM or QEFFAutoModel in the same way user initializes transformers.AutoModelForCausalLM.

In this PR, I am adding support only for from_pretrained method to these classes, and in subsequenst PRs. following method will be enabled for all these classes i.e.

1. transform
2. export
3. execute (based on runtime)
4. compile
   etc.

[vbaddi]

LGTM +1