Added skip_bos Argument to LLM Attribution To Enable Wider Model Support for Attributing Against a Single Token

[SulRash]

Some models do not generate a bos, so when trying to attribute a single token the code crashes. I added a skip_bos argument (that defaults to True to not change anything in the system) and some minor error handling to tell the user when to use the argument.

Here's my error handling so far:

if self.include_per_token_attr:
    try:
        target_log_probs = torch.stack(
            [total_log_prob, *log_prob_list], dim=0
        ).unsqueeze(0)
    except TypeError:
        print("It seems like you got an empty list of target tokens. If you are attributing only one target token (a single character / word) try using the skip_bos argument in the attribute function.")
        exit()
else:
    target_log_probs = total_log_prob

This allowed me to attribute multiple choice question answers (a single token) with models (like microsoft/Phi-3-mini-4k-instruct) that don't generate a bos token when the target is predefined.

The error that would pop up without this change is the following:

Traceback (most recent call last):
  File "/Users/sultan/Documents/Github/idk-what-this-is-yet/test.py", line 86, in <module>
    attr_res = llm_attr.attribute(inp, target=target)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/sultan/miniconda3/envs/generalai/lib/python3.11/site-packages/captum/attr/_core/llm_attr.py", line 360, in attribute
    cur_attr = self.attr_method.attribute(
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/sultan/miniconda3/envs/generalai/lib/python3.11/site-packages/captum/log/__init__.py", line 42, in wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/Users/sultan/miniconda3/envs/generalai/lib/python3.11/site-packages/captum/attr/_core/feature_ablation.py", line 289, in attribute
    initial_eval = self._strict_run_forward(
                   ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/sultan/miniconda3/envs/generalai/lib/python3.11/site-packages/captum/attr/_core/feature_ablation.py", line 599, in _strict_run_forward
    forward_output = _run_forward(*args, **kwargs)
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/sultan/miniconda3/envs/generalai/lib/python3.11/site-packages/captum/_utils/common.py", line 531, in _run_forward
    output = forward_func(
             ^^^^^^^^^^^^^
  File "/Users/sultan/miniconda3/envs/generalai/lib/python3.11/site-packages/captum/attr/_core/llm_attr.py", line 259, in _forward_func
    target_log_probs = torch.stack(
                       ^^^^^^^^^^^^
TypeError: expected Tensor as element 0 in argument 0, but got int

[facebook-github-bot]

Hi @SulRash!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks!

[facebook-github-bot]

Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks!

[SulRash]

@yucu @vivekmig @aobo-y could you check this pull request out, it's only a few lines worth of change :) let me know if I should change anything

[SulRash]

Thanks for adding this @SulRash , this looks great! From the CI results, it looks like the lint / ufmt checks are failing, would you be able to resolve that?

Also, ideally we can add the same argument to LLMGradientAttribution for consistency, but we can add that in a separate PR as well.

Thanks again for the contribution!

Hey @vivekmig! Oh yeah, that's actually such a good idea, I'll submit a separate PR request for that class!

[SulRash]

Just fixed the linting issue, should be fine now :)