thanks @Aydin-ab - see also cursor/gemini comments when it comes to code. As long as you are able to run them successfully, should be free of serious bugs now.

Code Review

The pull request introduces new LLM batch inference examples for both text and vision data, along with their corresponding CI configurations and helper scripts. The changes effectively split the existing template into two distinct, independently readable content pieces, which is a good refactoring step. The new examples demonstrate how to use Ray Data LLM APIs for batch inference, including data preparation, processor configuration, and scaling considerations. The addition of CI scripts ensures these examples remain functional.

However, there are a few areas that could be improved for robustness and clarity:

• The nb2py.py scripts use specific string matching to modify dataset limits for CI. This approach is brittle and could break if the exact string in the notebook changes.
• Some comments in the Jupyter notebooks are slightly misleading regarding dataset size limits.
• The standalone Python scripts contain hardcoded configuration values that would ideally be configurable for real-world use cases.

thanks! minor comments but overall LGTM

• datasets version in vision template (4.4.2) shows diff from job.yaml (4.4.1)
• imports at top (not inline - style) for python files
• image mode handling potential issue with batch_inference_vision.py and batch_inference_vision_scaled.py

when you open images with PIL, there can be issues with modes - something like

  image = Image.open(BytesIO(image))                                                                                                  
  if image.mode != 'RGB':                                                                                                             
      image = image.convert('RGB')  

could improve robustness

• partition counts are hardcoded (64/128/256) but explanation says "2-4x the worker (GPU) count" - so with concurrency=4 and 128 partitions, that's 32x the gpu count

overall great - extensive use of the apis and clear/helpful distinction between batch/online inference and well written scaling guidance + bits on model paralellism

approved with comments