ColPali rebase#228
Open
NohTow wants to merge 66 commits into
Open
Conversation
Update dependency requirements for the new sentence-transformers API, add optional `image` dependency group for multimodal support, and update ruff target-version to py310. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adapt Dense to inherit from the new sentence_transformers.base.modules.Dense (Module base class). Key changes: - Add config_keys list including use_residual - Use module_input_name/module_output_name instead of hardcoded keys - Update save() to use save_config/save_torch_weights - Update load() with new hub-compatible signature and cls.load_config/load_torch_weights - Handle string activation functions in from_sentence_transformers() - Use fullname() for activation function serialization Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Major changes to the ColBERT model: - Update imports to new ST paths (base.modules, util.quantization) - Constructor: keyword-only args, processor_kwargs replaces tokenizer_kwargs, use_auth_token deprecation handling, add backend parameter - Replace _load_auto_model/_load_sbert_model with _load_default_modules/ _load_config_modules/_load_converted_modules - Add _get_model_config()/_parse_model_config() for config management - Rename tokenize() to preprocess() with backward-compat wrapper - Add encode_query()/encode_document() convenience methods - Add multimodal support: _is_text_input(), _preprocess_multimodal(), handle non-text inputs in encode() with skiplist mask fallback - Update save() to overlay PyLate fields on existing config Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Rename the primary parameter from tokenize_fn to preprocess_fn with backward compatibility. Add _is_text_column() helper for multimodal detection and skip prompt prepending for non-text columns. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- colbert_triplet: sentence_transformers.sentence_transformer.evaluation - colbert_distillation: base.evaluation, sentence_transformer.model - nano_beir: sentence_transformer.evaluation.nano_beir - IR evaluator: sentence_transformer.evaluation - model_card: sentence_transformer.model_card, base.evaluation, sentence_transformer.trainer, get_embedding_dimension() Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…mport Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Example script demonstrating text-query to image-document retrieval training with a VLM-based ColBERT model using CachedContrastive loss. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Replace model.tokenize with model.preprocess in code examples - Replace tokenize_fn with preprocess_fn in collator usage - Update Transformer import path to sentence_transformers.base.modules Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The training_args module moved from sentence_transformers.training_args to sentence_transformers.sentence_transformer.training_args in v5.4. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- get_word_embedding_dimension() -> get_embedding_dimension() - truncate_sentence_embeddings() -> truncate_embeddings() Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
_text_length was removed in ST v5.4. The inherited _input_length (BaseModel) handles strings, dicts, lists, tensors, and PIL images. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replaces 16 lines of manual prompt/prompt_name resolution logic with the single-line _resolve_prompt() inherited from BaseModel. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replaces ~40 lines of reimplemented HPU power-of-2 padding with the inherited _pad_features_for_hpu() from SentenceTransformer. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Multimodal inputs (images, etc.) may not have input_ids. Guard all direct input_ids access with .get() or 'in' checks, falling back to attention_mask where appropriate. - contrastive.py: extract_skiplist_mask() guards input_ids for query and document features - cached_contrastive.py: use next(iter(values)) instead of input_ids for batch size - colbert.py: guard input_ids in query expansion mask creation Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Aligns with ST v5.4 API convention. Uses @deprecated_kwargs decorator so that callers using sentences= still work with a deprecation warning. Also renames encode_query/encode_document parameter accordingly. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove dead hasattr(first_module, "tokenize") check in preprocess(): ST v5.4 Transformer only has preprocess() - Remove redundant config JSON re-write in save(): _get_model_config() already includes all PyLate fields, super().save() writes them - Remove unused _embedding_size/_bias attributes: set but never read Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Both _load_config_modules and _load_converted_modules had identical ~15-line filtering logic. Extract to a shared static method. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
_preprocess_multimodal() now accepts is_query parameter, allowing downstream use for different max lengths or metadata. Also removes dead tokenize() fallback (ST v5.4 Transformer only has preprocess). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
ColBERTCollator now extends ST v5.4's BaseDataCollator, inheriting _resolve_router_mapping(), _resolve_prompts(), _get_prompt_for_column(), and maybe_warn_about_column_order(). Overrides __call__() for ColBERT-specific behavior: is_query detection, list flattening, _id column skipping, pad=True, and prompt length tracking. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Load ColPali-family checkpoints (ColQwen2, ColQwen2_5, ColQwen3, etc.) as their base VLM architecture, extracting the projection layer into a separate Dense module. Removes the colpali_engine runtime dependency. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…sion ColPali models use no [Q]/[D] prefixes, empty skiplist, right-padding with the native pad token, and attend_to_expansion_tokens=True by default. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Append N expansion tokens after the full query text instead of pad-to-fixed-length, matching colpali_engine's process_queries behavior. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Disable unpadding when query expansion is active and raise an error if attend_to_expansion_tokens=False is used with Flash Attention, which silently degrades retrieval scores. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
NohTow
force-pushed
the
snapshot-colplai-rebase
branch
from
c3bffac to
a3ce020
Compare
added 2 commits
June 16, 2026 09:24
NohTow
force-pushed
the
snapshot-colplai-rebase
branch
from
2749b49 to
74d55ec
Compare
added 15 commits
June 16, 2026 09:32
…transformers version do not ship the corresponding processor
…y/except for non image users
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.