Improve tokenizer and model handling for Transformers v5 compatibility

[RaykKretzschmar]

No description provided.

[Copilot]

Pull request overview

This PR updates Lightning IR’s model/config/tokenizer integration to stay compatible with Transformers v5, including improved backbone type inference, tokenizer/model class-factory behavior, and test/CI adjustments for external dependencies.

Changes:

• Bump Python/Transformers requirements (Python >= 3.10, Transformers >= 5.0.0) and add pytest markers for model/dataset-downloading tests.
• Improve backbone model type handling across configs/models (e.g., get_backbone_model_type(), dropping None overrides) and add caching/registration in class factories.
• Patch/adjust several models and tests for Transformers v5 behavior (tied weights keys, state-dict loading, CPU device usage, attention mask handling).

Reviewed changes

Copilot reviewed 23 out of 23 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
tests/test_models/test_splade.py	Forces CPU device in test to stabilize behavior across environments.
tests/test_models/test_col.py	Monkeypatches ColBERT internals for Transformers v5 tied-weights/key-ignore behavior.
tests/test_models/test_coil.py	Adds post_init() and custom state-dict loading to match expected checkpoints.
tests/test_lsr_benchmark_integration.py	Marks dataset-downloading tests with a pytest marker.
tests/test_config.py	Extends config serialization assertions for new/internal config keys.
tests/test_callbacks.py	Skips Seismic-dependent callback tests when the package isn’t installed.
tests/conftest.py	Limits BLAS/OpenMP threads; updates typing; bumps test embedding dim defaults.
pyproject.toml	Raises Python minimum to 3.10; bumps Transformers to v5; pins sentence-transformers; adds pytest markers; updates Ruff target.
lightning_ir/retrieve/plaid/plaid_searcher.py	Adjusts FastPlaid initialization (removes preload_index=True).
lightning_ir/models/cross_encoders/set_encoder.py	Updates backbone type handling and attention path for Transformers v5.
lightning_ir/models/cross_encoders/mono.py	Uses get_backbone_model_type() for T5 detection.
lightning_ir/models/bi_encoders/splade.py	Updates tied-weights handling and backbone type inference for mapping MLM weights.
lightning_ir/modeling_utils/batching.py	Moves Callable import to collections.abc (Py3.10+).
lightning_ir/data/dataset.py	Ensures targets tensors are float32 for consistency in training/loss usage.
lightning_ir/cross_encoder/cross_encoder_tokenizer.py	Adjusts preprocessing for T5 template post-processing behavior.
lightning_ir/callbacks/callbacks.py	Moves Callable import to collections.abc (Py3.10+).
lightning_ir/bi_encoder/bi_encoder_model.py	Places mask-scoring ID tensors on CPU (moved to input device at use time).
lightning_ir/bi_encoder/bi_encoder_config.py	Allows explicit embedding_dim override and persists it appropriately in diffs.
lightning_ir/base/tokenizer.py	Improves tokenizer class resolution when TOKENIZER_MAPPING is incomplete in newer Transformers.
lightning_ir/base/model.py	Drops None backbone type overrides when merging config diffs during from_pretrained.
lightning_ir/base/external_model_hub.py	Moves Callable import to collections.abc (Py3.10+).
lightning_ir/base/config.py	Adds backbone type inference and safer serialization/deserialization across derived configs.
lightning_ir/base/class_factory.py	Adds caching and AutoModel registration; improves tokenizer mapping handling for Transformers v5.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Implemented in 84c6285. I removed the duplicate attention_mask extension from forward() and now bind get_extended_attention_mask / attention forward from class methods each call to avoid stacked partials across repeated forwards.

[RaykKretzschmar]

I will clean this up and make a new PR