Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

LLM offsets logic consolidate w/ checks and test case fix #1422

Closed
wants to merge 9 commits into from

Commits on Oct 26, 2024

  1. Fix for automatic GitHub workflow reruns (pytorch#1419)

    Summary:
    
    Slight bug in the retry GitHub workflow causes failing workflows to rerun indefinitely. This _should_ fix it
    
    Differential Revision: D65008676
    craymichael authored and facebook-github-bot committed Oct 26, 2024
    Configuration menu
    Copy the full SHA
    fd688bb View commit details
    Browse the repository at this point in the history
  2. Add __call__ to TokenizerLike (pytorch#1418)

    Summary:
    
    Add __call__ to TokenizerLike for transformers compatibility
    
    Differential Revision: D64998805
    craymichael authored and facebook-github-bot committed Oct 26, 2024
    Configuration menu
    Copy the full SHA
    eb3eca2 View commit details
    Browse the repository at this point in the history
  3. Improve tokenizer pretty-pretty logic + __call__ method (pytorch#1417)

    Summary:
    
    Use the __call__ method of tokenizers that returns a BatchEncoding with offsets. This allows us to grab text from the fully decoded string and not make assumptions about how many tokens correspond to a single string.
    
    Differential Revision: D64998804
    craymichael authored and facebook-github-bot committed Oct 26, 2024
    Configuration menu
    Copy the full SHA
    b4fe485 View commit details
    Browse the repository at this point in the history
  4. Fix mypy issue in visualization.py (pytorch#1416)

    Summary:
    
    visualize_image_attr_multiple can return a List[Axes], adds proper annotations to satisfy mypy
    
    
    Why casting is necessary: numpy/numpy#24738
    
    https://github.com/matplotlib/matplotlib/blob/v3.9.2/lib/matplotlib/pyplot.py#L1583C41-L1584C1
    
    Differential Revision: D64998799
    craymichael authored and facebook-github-bot committed Oct 26, 2024
    Configuration menu
    Copy the full SHA
    aea894e View commit details
    Browse the repository at this point in the history
  5. Remove mypy note from infidelity.py (pytorch#1415)

    Summary:
    
    Adds enough typing to get rid of
    `captum/metrics/_core/infidelity.py:498: note: By default the bodies of untyped functions are not checked, consider using --check-untyped-defs  [annotation-unchecked]`
    
    Differential Revision: D64998800
    craymichael authored and facebook-github-bot committed Oct 26, 2024
    Configuration menu
    Copy the full SHA
    7b944f9 View commit details
    Browse the repository at this point in the history
  6. Fix remaining pyre errors in infidelity.py (pytorch#1414)

    Summary:
    
    Fix pyre/mypy errors in infidelity.py. Introduce new BaselineTupleType
    
    Differential Revision: D64998803
    craymichael authored and facebook-github-bot committed Oct 26, 2024
    Configuration menu
    Copy the full SHA
    59b7e81 View commit details
    Browse the repository at this point in the history
  7. Fix kwargs annotation in feature_ablation.py (pytorch#1421)

    Summary:
    
    Fix incorrect **kwargs annotation
    
    Differential Revision: D65001879
    craymichael authored and facebook-github-bot committed Oct 26, 2024
    Configuration menu
    Copy the full SHA
    a92c1b8 View commit details
    Browse the repository at this point in the history
  8. Consolidate LLM attr logic (pytorch#1420)

    Summary:
    
    Add base class BaseLLMAttribution to consolidate repeat logic between perturbation/gradient-based LLM attr classes
    
    Differential Revision: D65008854
    craymichael authored and facebook-github-bot committed Oct 26, 2024
    Configuration menu
    Copy the full SHA
    5b60545 View commit details
    Browse the repository at this point in the history
  9. LLM offsets logic consolidate w/ checks and test case fix (pytorch#1422)

    Summary:
    
    Consolidate offsets logic with extra checks to one function. May be used to later group data in gradient LLM attribution. Test case fixed as a result of checks.
    
    Differential Revision: D65010820
    craymichael authored and facebook-github-bot committed Oct 26, 2024
    Configuration menu
    Copy the full SHA
    e232eb1 View commit details
    Browse the repository at this point in the history