From d90cc6dc5f243bc6b9e3de9380c7f90f9a36f28e Mon Sep 17 00:00:00 2001 From: Gabriele Sarti Date: Wed, 19 Jun 2024 14:19:40 +0200 Subject: [PATCH 1/3] Fix skip special tokens if pre-specified in prompt --- inseq/models/attribution_model.py | 4 +++- inseq/utils/misc.py | 7 +++++++ 2 files changed, 10 insertions(+), 1 deletion(-) diff --git a/inseq/models/attribution_model.py b/inseq/models/attribution_model.py index 42d72b3b..08e645de 100644 --- a/inseq/models/attribution_model.py +++ b/inseq/models/attribution_model.py @@ -411,7 +411,9 @@ def attribute( "Step scores are not supported for final step methods since they do not iterate over the full" " sequence. Please remove the step scores and compute them separatly passing method='dummy'." ) - input_texts, generated_texts = format_input_texts(input_texts, generated_texts) + input_texts, generated_texts = format_input_texts( + input_texts, generated_texts, skip_special_tokens, self.special_tokens + ) has_generated_texts = generated_texts is not None if not self.is_encoder_decoder: for i in range(len(input_texts)): diff --git a/inseq/utils/misc.py b/inseq/utils/misc.py index 628995bc..2f0c178e 100644 --- a/inseq/utils/misc.py +++ b/inseq/utils/misc.py @@ -202,6 +202,8 @@ def isnotebook(): def format_input_texts( texts: TextInput, ref_texts: Optional[TextInput] = None, + skip_special_tokens: bool = False, + special_tokens: list[str] = [], ) -> tuple[list[str], list[str]]: texts = [texts] if isinstance(texts, str) else texts reference_texts = [ref_texts] if isinstance(ref_texts, str) else ref_texts @@ -211,6 +213,11 @@ def format_input_texts( len(texts), len(reference_texts) ) ) + if skip_special_tokens: + for special_token in special_tokens: + texts = [text.replace(special_token, "") for text in texts] + if reference_texts is not None: + reference_texts = [text.replace(special_token, "") for text in reference_texts] return texts, reference_texts From 067351a58e16a0e0972c588fa4e109626b4be7c7 Mon Sep 17 00:00:00 2001 From: Gabriele Sarti Date: Wed, 19 Jun 2024 15:26:19 +0200 Subject: [PATCH 2/3] Move attentions to cpu preemptively --- inseq/models/huggingface_model.py | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/inseq/models/huggingface_model.py b/inseq/models/huggingface_model.py index eb374fbe..fe8a7bca 100644 --- a/inseq/models/huggingface_model.py +++ b/inseq/models/huggingface_model.py @@ -456,6 +456,12 @@ def get_attentions_dict( ) -> dict[str, MultiLayerMultiUnitScoreTensor]: if output.encoder_attentions is None or output.decoder_attentions is None: raise ValueError("Model does not support attribution relying on attention outputs.") + if output.encoder_attentions is not None: + output.encoder_attentions = tuple(att.to("cpu") for att in output.encoder_attentions) + if output.decoder_attentions is not None: + output.decoder_attentions = tuple(att.to("cpu") for att in output.decoder_attentions) + if output.cross_attentions is not None: + output.cross_attentions = tuple(att.to("cpu") for att in output.cross_attentions) return { "encoder_self_attentions": torch.stack(output.encoder_attentions, dim=1), "decoder_self_attentions": torch.stack(output.decoder_attentions, dim=1), @@ -506,6 +512,8 @@ def configure_embeddings_scale(self): def get_attentions_dict(output: CausalLMOutput) -> dict[str, MultiLayerMultiUnitScoreTensor]: if output.attentions is None: raise ValueError("Model does not support attribution relying on attention outputs.") + else: + output.attentions = tuple(att.to("cpu") for att in output.attentions) return { "decoder_self_attentions": torch.stack(output.attentions, dim=1), } From 0e318ceab94a52bc10276f72cdd43d4c3b7df7f4 Mon Sep 17 00:00:00 2001 From: Gabriele Sarti Date: Thu, 27 Jun 2024 16:38:40 +0200 Subject: [PATCH 3/3] Add rescale and readme changes --- CHANGELOG.md | 8 ++++++-- README.md | 7 +++++-- inseq/commands/attribute/attribute.py | 3 +++ inseq/commands/attribute/attribute_args.py | 8 ++++++++ inseq/commands/attribute_context/attribute_context.py | 1 + 5 files changed, 23 insertions(+), 4 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index b4a36ee2..9ef5672e 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -9,15 +9,19 @@ ## 🔧 Fixes and Refactoring - Fix the issue in the attention implementation from [#268](https://github.com/inseq-team/inseq/issues/268) where non-terminal position in the tensor were set to nan if they were 0s ([#269](https://github.com/inseq-team/inseq/pull/269)). - + - Fix the pad token in cases where it is not specified by default in the loaded model (e.g. for Qwen models) ([#269](https://github.com/inseq-team/inseq/pull/269)). - Fix bug reported in [#266](https://github.com/inseq-team/inseq/issues/266) making `value_zeroing` unusable for SDPA attention. This enables using the method on models using SDPA attention as default (e.g. `GemmaForCausalLM`) without passing `model_kwargs={'attn_implementation': 'eager'}` ([#267](https://github.com/inseq-team/inseq/pull/267)). +- Fix multi-device support and duplicate BOS for chat template models ([#280](https://github.com/inseq-team/inseq/pull/280)). + +- Add `rescale_attributions` to Inseq CLI commands for `rescale=True` ([#280](https://github.com/inseq-team/inseq/pull/280)). + ## 📝 Documentation and Tutorials *No changes* ## 💥 Breaking Changes -*No changes* \ No newline at end of file +*No changes* diff --git a/README.md b/README.md index 6f20a99e..7757aa7d 100644 --- a/README.md +++ b/README.md @@ -280,7 +280,7 @@ Our vision for Inseq is to create a centralized, comprehensive and robust set of ## Citing Inseq -If you use Inseq in your research we suggest to include a mention to the specific release (e.g. v0.4.0) and we kindly ask you to cite our reference paper as: +If you use Inseq in your research we suggest to include a mention to the specific release (e.g. v0.6.0) and we kindly ask you to cite our reference paper as: ```bibtex @inproceedings{sarti-etal-2023-inseq, @@ -308,7 +308,7 @@ If you use Inseq in your research we suggest to include a mention to the specifi Inseq has been used in various research projects. A list of known publications that use Inseq to conduct interpretability analyses of generative models is shown below. > [!TIP] -> Last update: May 2024. Please open a pull request to add your publication to the list. +> Last update: June 2024. Please open a pull request to add your publication to the list.
2023 @@ -331,6 +331,9 @@ Inseq has been used in various research projects. A list of known publications t
  • ReAGent: A Model-agnostic Feature Attribution Method for Generative Language Models (Zhao et al., 2024)
  • Revisiting subword tokenization: A case study on affixal negation in large language models (Truong et al., 2024)
  • Exploring NMT Explainability for Translators Using NMT Visualising Tools (Gonzalez-Saez et al., 2024)
    • +
  • DETAIL: Task DEmonsTration Attribution for Interpretable In-context Learning (Zhou et al., 2024)
    • +
  • Should We Fine-Tune or RAG? Evaluating Different Techniques to Adapt LLMs for Dialogue (Alghisi et al., 2024)
    • +
  • Model Internals-based Answer Attribution for Trustworthy Retrieval-Augmented Generation (Qi, Sarti et al., 2024)
    • diff --git a/inseq/commands/attribute/attribute.py b/inseq/commands/attribute/attribute.py index 4869b846..95c4e5d7 100644 --- a/inseq/commands/attribute/attribute.py +++ b/inseq/commands/attribute/attribute.py @@ -11,12 +11,14 @@ def aggregate_attribution_scores( selectors: Optional[list[int]] = None, aggregators: Optional[list[str]] = None, normalize_attributions: bool = False, + rescale_attributions: bool = False, ) -> FeatureAttributionOutput: if selectors is not None and aggregators is not None: for select_idx, aggregator_fn in zip(selectors, aggregators): out = out.aggregate( aggregator=aggregator_fn, normalize=normalize_attributions, + rescale=rescale_attributions, select_idx=select_idx, do_post_aggregation_checks=False, ) @@ -79,6 +81,7 @@ def attribute(input_texts, generated_texts, args: AttributeExtendedArgs): selectors=args.attribution_selectors, aggregators=args.attribution_aggregators, normalize_attributions=args.normalize_attributions, + rescale_attributions=args.rescale_attributions, ) print(f"Saving {'aggregated ' if args.aggregate_output else ''}attributions to {args.save_path}") out.save(args.save_path, overwrite=True) diff --git a/inseq/commands/attribute/attribute_args.py b/inseq/commands/attribute/attribute_args.py index ee707b5f..dfca76dd 100644 --- a/inseq/commands/attribute/attribute_args.py +++ b/inseq/commands/attribute/attribute_args.py @@ -61,6 +61,14 @@ class AttributeBaseArgs: "for each context are normalized to sum up to 1, providing a relative notion of input salience." ), ) + rescale_attributions: bool = cli_arg( + default=False, + help=( + "Whether to rescale the attribution scores for each context. If ``True``, the attribution scores " + "for each context are rescaled to sum up to the number of tokens in the input, providing an absolute" + " notion of input salience." + ), + ) model_kwargs: dict = cli_arg( default_factory=dict, help="Additional keyword arguments passed to the model constructor in JSON format.", diff --git a/inseq/commands/attribute_context/attribute_context.py b/inseq/commands/attribute_context/attribute_context.py index 1eb72126..e2bafe7e 100644 --- a/inseq/commands/attribute_context/attribute_context.py +++ b/inseq/commands/attribute_context/attribute_context.py @@ -211,6 +211,7 @@ def attribute_context_with_model(args: AttributeContextArgs, model: HuggingfaceM selectors=args.attribution_selectors, aggregators=args.attribution_aggregators, normalize_attributions=args.normalize_attributions, + rescale_attributions=args.rescale_attributions, )[0] if args.show_intermediate_outputs: cci_attrib_out.show(do_aggregation=False)