[etLLM] New config system to export_llama

[iseeyuan]

🚀 The feature, motivation and pitch

Current status

The configuration system in export_llama is limited:

• Only via cli command. There's no python API. We cannot export an arbitrary nn.Module
• The cli args are complicated. It's hard for users to get what can be configured
• It's a bad practice to have flat configs and pass it deeply into various functions

Proposal

• Similar to the torchtune config system, use Hydra and omegaconf, so that the configs can be organized in yaml file, and overrides by cli.
• Use structured config, like config.model, config.quantization, etc. Only pass sub configs to each component.
• Keep config and sub configs at high level and directly pass field values
• An example of config definition in https://gist.github.com/iseeyuan/df9cd283b46dc574d1d506324258f59b

Steps

Since export_llama is used in many places, we do incremental changes.

1. convert args to DictConfig with structure.
2. Refactor functions for composability and unblock internal use
3. Add new entry point using Hydra
4. Deprecate the legacy argparser systme.

Alternatives

No response

Additional context

No response

RFC (Optional)

No response

cc @larryliu0820 @mergennachin @cccclai @helunwencser @jackzhxng

[kimishpatel]

I +1 to moving to config based thing rather than "add yet another arg to cli"

[larryliu0820]

Yes please! I can add some food for thoughts:

• Testability - this config system should be very easy to test
• Composability - allow users to fine tune the yaml based config easily
• Separation of concern - ideally we should have model configs, export configs and optrace, etrecord configs, different configs for different backends should be nicely grouped together (coreml, qnn etc)