Use llama-1b for faster and more accessible examples

[kylesayrs]

Purpose

• Make examples more accessible to users without large GPU resources. Since these are just examples, they should be accessible as possible to encourage easy use and adoption of llm-compressor.

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

[dsikka]

I think we actually want to keep the larger models as most of our questions on vllm are around compressing larger models to run them there. So I dont think we should be making this change as these are an easy reference and are our most common case.

These examples are also used directly in our testing which help identify cases that would otherwise be ignored when dealing with larger memory requirements.

I wouldn't be opposed to adding a smaller model in addition to the lager models.

[horheynm]

I think if the model structure is the same we can use a smaller model.
for example if the difference from the large and small model is N attention heads only then it should be ok, bc the vllm code execution path pathway will be the same.

If the larger model has different architecture, then we should keep the larger model, the execution path will be different

[kylesayrs]

I agree, it makes sense to keep the 8b models. We already have an example of quantizing a small tiny llama model in the readme.