feat(deps): upgrade transformers to 5.x and sentence-transformers to 5.2+

docs/optimizer_search_space_config.schema.json

@@ -1302,6 +1302,8 @@
                 },
                 "warmup_ratio": {
                     "default": 0.1,
+                    "exclusiveMaximum": 1,
+                    "minimum": 0,
                     "title": "Warmup Ratio",
                     "type": "number"
                 },

pyproject.toml

@@ -51,8 +51,8 @@ dependencies = [
 [project.optional-dependencies]
 catboost = ["catboost (>=1.2.8,<2.0.0)"]
 peft = ["peft (>= 0.10.0, !=0.15.0, !=0.15.1, <1.0.0)"]
-transformers = ["transformers[torch] (>=4.49.0,<5.0.0)"]
-sentence-transformers = ["sentence-transformers (>=3,<4)"]
+transformers = ["transformers[torch] (>=5.0.0,<6.0.0)"]
+sentence-transformers = ["sentence-transformers (>=5.4.0,<6.0.0)"]
 dspy = [
     "dspy (>=2.6.5,<3.0.0)",
 ]

src/autointent/_dump_tools/unit_dumpers.py

@@ -292,7 +292,7 @@ def load(path: Path, **kwargs: Any) -> PreTrainedTokenizer | PreTrainedTokenizer
         require("transformers")
         import transformers
 
-        return transformers.AutoTokenizer.from_pretrained(path)  # type: ignore[no-any-return,no-untyped-call]
+        return transformers.AutoTokenizer.from_pretrained(path)
 
     @classmethod
     def check_isinstance(cls, obj: Any) -> bool:  # noqa: ANN401

src/autointent/_wrappers/embedder/sentence_transformers.py

@@ -299,9 +299,8 @@ def train(self, utterances: list[str], labels: ListOfLabels, config: EmbedderFin
         from sentence_transformers import (
             SentenceTransformerTrainer,
             SentenceTransformerTrainingArguments,
-            losses,
-            training_args,
         )
+        from sentence_transformers.sentence_transformer import losses, training_args
         from transformers import EarlyStoppingCallback
 
         x_train, x_val, y_train, y_val = train_test_split(
@@ -324,7 +323,8 @@ def train(self, utterances: list[str], labels: ListOfLabels, config: EmbedderFin
                 per_device_train_batch_size=config.batch_size,
                 per_device_eval_batch_size=config.batch_size,
                 learning_rate=config.learning_rate,
-                warmup_ratio=config.warmup_ratio,
+                # warmup_steps accepts a float < 1 as a fraction of total steps.
+                warmup_steps=config.warmup_ratio,
                 fp16=config.fp16,
                 bf16=config.bf16,
                 seed=config.seed,

src/autointent/_wrappers/ranker.py

@@ -12,7 +12,7 @@
 import logging
 from pathlib import Path
 from random import shuffle
-from typing import TYPE_CHECKING, Any, Literal, TypedDict
+from typing import TYPE_CHECKING, Any, Literal, TypedDict, cast
 
 import joblib
 import numpy as np
@@ -127,7 +127,7 @@ def __init__(
             revision=self.config.revision,
             trust_remote_code=self.config.trust_remote_code,
             device=self.config.device,
-            max_length=self.config.tokenizer_config.max_length,  # type: ignore[arg-type]
+            max_length=self.config.tokenizer_config.max_length,
         )
         self._train_head = False
         self._clf = classifier_head
@@ -136,7 +136,10 @@ def __init__(
         if classifier_head is not None or self.config.train_head:
             self._train_head = True
             self._activations_list: list[npt.NDArray[Any]] = []
-            self._hook_handler = self.cross_encoder.model.classifier.register_forward_hook(self._classifier_hook)
+            # CrossEncoder is a nn.Sequential of modules; [0] is the Transformer
+            # wrapping the HF model exposed as .auto_model.
+            transformer = cast("Any", self.cross_encoder[0])
+            self._hook_handler = transformer.auto_model.classifier.register_forward_hook(self._classifier_hook)
 
     def _classifier_hook(self, _module, input_tensor, _output_tensor) -> None:  # type: ignore[no-untyped-def] # noqa: ANN001
         """Hook to capture classifier activations.
@@ -161,13 +164,13 @@ def _get_features_or_predictions(self, pairs: list[tuple[str, str]]) -> npt.NDAr
         if not self._train_head:
             return np.array(
                 self.cross_encoder.predict(
-                    pairs,
+                    pairs,  # type: ignore[arg-type]
                     batch_size=self.config.batch_size,
-                    activation_fct=nn.Sigmoid() if self.output_range == "sigmoid" else nn.Tanh(),
+                    activation_fn=nn.Sigmoid() if self.output_range == "sigmoid" else nn.Tanh(),
                 )
             )
 
-        self.cross_encoder.predict(pairs, batch_size=self.config.batch_size)
+        self.cross_encoder.predict(pairs, batch_size=self.config.batch_size)  # type: ignore[arg-type]
         res = np.concatenate(self._activations_list, axis=0)
         self._activations_list.clear()
         return res  # type: ignore[no-any-return]
@@ -311,7 +314,7 @@ def load(cls, path: Path, override_config: CrossEncoderConfig | None = None) ->
 
     def clear_ram(self) -> None:
         """Clear model from RAM and GPU memory."""
-        self.cross_encoder.model.cpu()
+        self.cross_encoder.cpu()
         del self.cross_encoder
         gc.collect()
         torch.cuda.empty_cache()

src/autointent/configs/_transformers.py

@@ -28,7 +28,10 @@ class EmbedderFineTuningConfig(BaseModel):
     batch_size: int
     margin: float = Field(default=0.5)
     learning_rate: float = Field(default=2e-5)
-    warmup_ratio: float = Field(default=0.1)
+    # Fed to TrainingArguments.warmup_steps in v5, which interprets float<1 as
+    # a fraction of total steps and float>=1 as a raw step count. Cap below 1
+    # so warmup_ratio=1.0 doesn't silently become a single step.
+    warmup_ratio: float = Field(default=0.1, ge=0, lt=1)
     early_stopping_patience: int = Field(default=1)
     early_stopping_threshold: float = Field(default=0.0)
     val_fraction: float = Field(default=0.2)

src/autointent/modules/scoring/_bert.py

@@ -130,17 +130,27 @@ def get_implicit_initialization_params(self) -> dict[str, Any]:
     def _initialize_model(self) -> Any:  # noqa: ANN401
         from transformers import AutoModelForSequenceClassification
 
-        label2id = {i: i for i in range(self._n_classes)}
-        id2label = {i: i for i in range(self._n_classes)}
-
+        # huggingface_hub v1 StrictDataclass requires label2id keys to be str
+        # (and id2label values to be str); int-keyed dicts raise
+        # StrictDataclassFieldValidationError on from_pretrained in v5.
+        label2id = {str(i): i for i in range(self._n_classes)}
+        id2label = {i: str(i) for i in range(self._n_classes)}
+
+        # transformers v5 + PEFT triggers find_adapter_config_file on every
+        # from_pretrained; it propagates _commit_hash for the cache lookup but
+        # NOT the outer `revision` to the fall-through hf_hub_download
+        # (auto_factory.py:308 only forwards adapter_kwargs). Set revision
+        # explicitly via adapter_kwargs so the adapter probe stays pinned.
+        revision = self.classification_model_config.revision
         return AutoModelForSequenceClassification.from_pretrained(
             self.classification_model_config.model_name,
             trust_remote_code=self.classification_model_config.trust_remote_code,
-            revision=self.classification_model_config.revision,
+            revision=revision,
             num_labels=self._n_classes,
             label2id=label2id,
             id2label=id2label,
             problem_type="multi_label_classification" if self._multilabel else "single_label_classification",
+            adapter_kwargs={"revision": revision} if revision is not None else None,
         )
 
     def fit(
@@ -152,7 +162,7 @@ def fit(
 
         self._validate_task(labels)
 
-        self._tokenizer = AutoTokenizer.from_pretrained(  # type: ignore[no-untyped-call]
+        self._tokenizer = AutoTokenizer.from_pretrained(
             self.classification_model_config.model_name, revision=self.classification_model_config.revision
         )
         self._model = self._initialize_model()

src/autointent/modules/scoring/_lora/lora.py

@@ -120,7 +120,15 @@ def _initialize_model(self) -> Any:  # noqa: ANN401
         model = super()._initialize_model()
         from peft import get_peft_model
 
-        return get_peft_model(model, self._lora_config)
+        peft_model = get_peft_model(model, self._lora_config)
+        # PEFT's save_pretrained vocab-check (save_and_load.py:380-384) calls
+        # AutoConfig.from_pretrained(base_model_name_or_path) with no revision
+        # during every Trainer checkpoint. On a cold cache this falls through
+        # to an unpinned hf_hub_download. Clearing base_model_name_or_path
+        # short-circuits the check; our dumper saves the base model
+        # separately (HFModelDumper), so the adapter doesn't need to remember it.
+        peft_model.peft_config["default"].base_model_name_or_path = ""
+        return peft_model
 
     def dump(self, path: str) -> None:
         from peft import LoraConfig

src/autointent/modules/scoring/_ptuning/ptuning.py

@@ -154,7 +154,15 @@ def _initialize_model(self) -> Any:  # noqa: ANN401
         model = super()._initialize_model()
         from peft import get_peft_model
 
-        return get_peft_model(model, self._ptuning_config)
+        peft_model = get_peft_model(model, self._ptuning_config)
+        # PEFT's save_pretrained vocab-check (save_and_load.py:380-384) calls
+        # AutoConfig.from_pretrained(base_model_name_or_path) with no revision
+        # during every Trainer checkpoint. On a cold cache this falls through
+        # to an unpinned hf_hub_download. Clearing base_model_name_or_path
+        # short-circuits the check; our dumper saves the base model
+        # separately (HFModelDumper), so the adapter doesn't need to remember it.
+        peft_model.peft_config["default"].base_model_name_or_path = ""
+        return peft_model
 
     def dump(self, path: str) -> None:
         from peft import PromptEncoderConfig

tests/conftest.py

@@ -25,39 +25,6 @@
     from autointent.nodes import NodeOptimizer
 
 
-def _disable_transformers_mistral_regex_patch() -> None:
-    # transformers.PreTrainedTokenizerBase._patch_mistral_regex calls
-    # huggingface_hub.model_info() for every tokenizer load with vocab > 100k
-    # (e.g. XLM-RoBERTa-based models like intfloat/multilingual-e5-*). On CI
-    # that uncacheable API call hammers the HF rate limit (429s). Tests never
-    # load mistralai tokenizers, so the correction is pure overhead — replace
-    # it with a no-op for the whole test session.
-    #
-    # Upstream bug & fix (merged for transformers 5.0.0+, NOT backported to 4.x):
-    #   https://github.com/huggingface/transformers/issues/44843
-    #   https://github.com/huggingface/transformers/pull/45444
-    # Drop this workaround when we upgrade to transformers>=5.0:
-    #   https://github.com/deeppavlov/AutoIntent/issues/295
-    try:
-        from transformers import tokenization_utils_base
-    except ImportError:
-        return
-
-    base = getattr(tokenization_utils_base, "PreTrainedTokenizerBase", None)
-    if base is None or not hasattr(base, "_patch_mistral_regex"):
-        return
-
-    def _noop_patch_mistral_regex(  # type: ignore[no-untyped-def]  # reason: monkey-patched into transformers internal classmethod; transformers is in ignore_missing_imports so signature types are unavailable
-        cls, tokenizer, *args, **kwargs
-    ):
-        return tokenizer
-
-    base._patch_mistral_regex = classmethod(_noop_patch_mistral_regex)
-
-
-_disable_transformers_mistral_regex_patch()
-
-
 def get_dataset_path() -> Path:
     return cast("Path", ires.files("tests.assets.data").joinpath("clinc_subset.json"))
 
@@ -192,7 +159,7 @@ def tiny_sentence_transformer() -> SentenceTransformer:
 
     from autointent.configs._pinned_revisions import DEFAULT_REVISIONS
 
-    return SentenceTransformer(TINY_SENTENCE_TRANSFORMER, revision=DEFAULT_REVISIONS[TINY_SENTENCE_TRANSFORMER])
+    return SentenceTransformer(TINY_SENTENCE_TRANSFORMER, revision=DEFAULT_REVISIONS[TINY_SENTENCE_TRANSFORMER])  # type: ignore[no-any-return]
 
 
 def apply_test_models(pipeline: Pipeline) -> None:

tests/embedder/test_sentence_transformers_backend.py

@@ -43,7 +43,7 @@ def test_model_lazy_loading(self, st_backend: SentenceTransformerEmbeddingBacken
         # cannot see the mutation inside `.embed()`. The post-call assert is
         # the whole point of this test (lazy load: None -> non-None).
         assert st_backend._model is not None
-        assert embeddings.shape == (1, st_backend._model.get_sentence_embedding_dimension())  # type: ignore[unreachable]
+        assert embeddings.shape == (1, st_backend._model.get_embedding_dimension())  # type: ignore[unreachable]
 
     def test_clear_ram(self, st_backend: SentenceTransformerEmbeddingBackend) -> None:
         """Test clearing model from RAM."""

tests/modules/test_dumper.py

@@ -41,8 +41,7 @@ class TestTransformers:
     def init_attributes(self) -> None:
         from transformers import AutoModelForSequenceClassification, AutoTokenizer
 
-        # reason: transformers AutoTokenizer.from_pretrained is untyped in stubs
-        self.tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")  # type: ignore[no-untyped-call]
+        self.tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
         self._tokenizer_predictions = np.array(self.tokenizer(["hello", "world"]).input_ids)
         self.transformer = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased")
 

tests/test_deps.py

@@ -104,7 +104,7 @@ def test_resolve_recurses_into_nested_extra(monkeypatch: pytest.MonkeyPatch) ->
     _patch_metadata(
         monkeypatch,
         {
-            "autointent": ["transformers[torch]>=4.49.0,<5.0.0 ; extra == 'transformers'"],
+            "autointent": ["transformers[torch]>=5.0.0,<6.0.0 ; extra == 'transformers'"],
             "transformers": [
                 "torch>=2.2 ; extra == 'torch'",
                 "accelerate>=0.26.0 ; extra == 'torch'",
@@ -180,13 +180,13 @@ def test_require_detects_missing_nested_accelerate(monkeypatch: pytest.MonkeyPat
     _patch_metadata(
         monkeypatch,
         {
-            "autointent": ["transformers[torch]>=4.49.0,<5.0.0 ; extra == 'transformers'"],
+            "autointent": ["transformers[torch]>=5.0.0,<6.0.0 ; extra == 'transformers'"],
             "transformers": [
                 "torch>=2.2 ; extra == 'torch'",
                 "accelerate>=0.26.0 ; extra == 'torch'",
             ],
         },
-        {"transformers": "4.49.0", "torch": "2.2.0"},  # accelerate absent
+        {"transformers": "5.0.0", "torch": "2.2.0"},  # accelerate absent
     )
     with pytest.raises(ImportError) as exc:
         deps.require("transformers")
@@ -200,7 +200,7 @@ def test_require_reports_extra_package_entirely_missing(monkeypatch: pytest.Monk
     # `transformers[torch]` requirement is flagged as missing with the install hint.
     _patch_metadata(
         monkeypatch,
-        {"autointent": ["transformers[torch]>=4.49.0,<5.0.0 ; extra == 'transformers'"]},
+        {"autointent": ["transformers[torch]>=5.0.0,<6.0.0 ; extra == 'transformers'"]},
         {},  # transformers (and everything else) absent
     )
     with pytest.raises(ImportError) as exc: