Replace readme paths with urls

[kylesayrs]

Purpose

• Files with the .md extension are not listed in the MANIFEST.in, meaning that they will not be included in the LLM Compressor pypi package. This means that references to these files are left dangling for users who have installed from the pypi package. Rather than including .md in the package and having to also ship all the large images files associated with them, this PR moves the references to urls hosted by github
  • While the github url paths may change between versions, this solution works in lieu of a dedicated readthedoc build for each version
  • This solution also aligns with the practice of other libraries which point to hosted urls rather than file paths
• Note that this does not apply to files which are themselves .md files, as these files will not be included in the pypi distribution
  • src/llmcompressor/transformers/finetune/README.md
  • src/llmcompressor/pipelines/sequential/README.md

Changes

• Replace readme file paths with urls
• Small change to DisableQuantization to better catch cases where exceptions such as tracing exceptions are triggered

Testing

• N/A

[github-actions]

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

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.

[horheynm]

model.apply(enable_quantization)
will always runs regardless of try being success or fail. Is this intended.
Also if it fails I would add a log

[kylesayrs]

model.apply(enable_quantization) will always runs regardless of try being success or fail. Is this intended.

Yes, that is exactly the intention of this change

Also if it fails I would add a log

If an exception is raised, the user will see the raised exception

[horheynm]

sorry why do we want to run model.apply(enable_quantization) if it raises exception? Wouldn't it just exit?

And as long as the exception contains info that we can get from this function, then it should be ok. Ex. have input args useful to log causing the error.

[kylesayrs]

sorry why do we want to run model.apply(enable_quantization) if it raises exception? Wouldn't it just exit?

For example, if the sequential pipeline fails, an exception will be thrown and then caught, in order to fall back to the layer-wise sequential pipeline. This is example of where an exception is raised but caught, and I do not want do have to think about the weird interactions that would result from this context not being properly exited.

This "try finally" pattern is, in general, good for contexts. It allows the user to not have to worry about if the context was exited properly, reducing mental load.

[rahul-tuli]

Could we move this code change to a separate PR, ideally I would like the scope of this PR to be limited to updating readme paths