# Transformers release review

Reviewed 2026-10-01 against Transformers **5.18.0**. Release metadata and tagged implementation take precedence when an unversioned documentation paragraph still describes v4. The versioned `/v5.18.0/en/main_classes/...` docs routes were not accessible during this review, so the released GitHub source and installed wheel were compared directly.

## Dependency boundaries

- Transformers 5.18.0 requires Hub >=1.31,<3; its Torch extra requires Torch >=2.5 and Accelerate >=1.1. Datasets 5.0.1 requires Hub <2. The tested training stack therefore uses Hub 1.33.0, not the newer Hub 2.1.1 observed on PyPI.
- Native checks: Python 3.11.11, Torch 2.14.1, Datasets 5.0.1, Accelerate 1.15.0, PEFT 0.21.2, Pillow 12.3.0, bitsandbytes 0.50.2. The core test environment excludes optional trackers, ONNX, timm/torchvision, and audio-file/decoder dependencies.
- Current librosa 1.0.0 needs Python >=3.12; use a separate compatible audio environment. EvolutionaryScale's `esm` SDK has a Transformers <5 constraint and must also remain separate.
- Bitsandbytes 0.50.2 publishes macOS 14 ARM64, Linux x86_64/aarch64, and Windows x86_64/ARM64 wheels. CPU NF4 quantize/dequantize runs on this macOS host. This is not a test of model-wide quantized loading, 8-bit optimizers, or every accelerator/backend.

Primary metadata: [Transformers](https://pypi.org/pypi/transformers/5.18.0/json), [Datasets](https://pypi.org/pypi/datasets/5.0.1/json), [Hub](https://pypi.org/pypi/huggingface-hub/2.1.1/json), [librosa](https://pypi.org/pypi/librosa/1.0.0/json), [bitsandbytes](https://pypi.org/pypi/bitsandbytes/0.50.2/json), and the [bitsandbytes installation matrix](https://huggingface.co/docs/bitsandbytes/v0.50.2/installation).

## Verified contracts

| Surface | Current behavior and evidence |
| --- | --- |
| Model loading | Auto classes, task heads, `from_pretrained`, dtype/attention/dispatch kwargs, config construction, parameter inspection, local save/reload, safetensors, and embedding resize checked against [released model source](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/modeling_utils.py) and [Auto mappings](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/models/auto/modeling_auto.py). Legacy `low_cpu_mem_usage` is ignored; `AutoModelForVision2Seq` is absent. |
| Pipelines | All named tasks, aliases, model-head mappings, input/output conventions, and removed tasks checked against the [released registry](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/pipelines/__init__.py). Generative QA/captioning and extractive/seq2seq tasks retain distinct semantics. |
| Audio / image score meaning | The AudioSet AST checkpoint uses independent labels; the [authors' inference source](https://github.com/YuanGongND/ast/blob/master/src/traintest.py) applies sigmoid, while the [audio pipeline](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/pipelines/audio_classification.py) defaults to softmax. The example now selects sigmoid. [DPT pipeline](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/pipelines/depth_estimation.py) visualizes per-image normalized depth; CLIP scores depend on the supplied candidate set. |
| Tokenization | Encoding, tensor/offset/word mappings, special tokens, local serialization, chat dictionary output, and ignored `use_fast` checked against [base tokenizer source](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/tokenization_utils_base.py), [AutoTokenizer](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/models/auto/tokenization_auto.py), and [chat documentation](https://huggingface.co/docs/transformers/en/chat_templating). |
| Generation | Greedy, sampling, beam and beam sampling, token budgets, filtering/repetition controls, config serialization, streaming and cache interfaces checked against [generation source](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/generation/utils.py), [configuration source](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/generation/configuration_utils.py), and [cache source](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/cache_utils.py). Contrastive/constrained strategies route to Hub custom code. |
| Training | `processing_class`, `eval_strategy`, fractional `warmup_steps`, removed `logging_dir`/`warmup_ratio`, optimizer default, custom loss signature, checkpoints and Hub destination semantics checked against [Trainer](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/trainer.py) and [TrainingArguments](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/training_args.py). Dataset map/split reviewed against [Datasets API](https://huggingface.co/docs/datasets/en/package_reference/main_classes). |
| Collators / adapters | Padding, MLM, seq2seq labels, and EOS-as-PAD masking checked against [collator source](https://github.com/huggingface/transformers/blob/v5.18.0/src/transformers/data/data_collator.py). BERT LoRA targets and trainable/frozen parameters checked against [PEFT LoRA](https://huggingface.co/docs/peft/en/package_reference/lora) and native PEFT. |
| Optional workflows | [Accelerate inference dispatch](https://huggingface.co/docs/accelerate/en/concept_guides/big_model_inference), [quantization](https://huggingface.co/docs/transformers/en/quantization/bitsandbytes), [DeepSpeed](https://huggingface.co/docs/transformers/en/deepspeed), [Optimum ONNX export](https://huggingface.co/docs/optimum-onnx/onnx/usage_guides/export_a_model), and [Trainer hyperparameter search](https://huggingface.co/docs/transformers/en/hpo_train) are documentation/source-verified, not full execution claims. |
| Authentication / Hub | [Environment variables](https://huggingface.co/docs/huggingface_hub/en/package_reference/environment_variables), [authentication](https://huggingface.co/docs/huggingface_hub/en/package_reference/authentication), and installed `hf auth login --help` checked. Public metadata and immutable-revision config reads succeeded for 19 model repositories and the Yelp dataset metadata record; no authentication, weight/data download, or upload was performed. |

The [v5 migration guide](https://github.com/huggingface/transformers/blob/v5.18.0/MIGRATION_GUIDE_V5.md) supplies the removal history, but its `warmup_step` spelling is a documentation typo: the released API is `warmup_steps`.

## Executed checks and limits

`tests/transformers/test_workflows.py` executes 21 native CPU cases with tiny random BERT, GPT-2, Llama, T5, ViT, and Wav2Vec2 configurations and a local vocabulary. Coverage includes batched classification/NER/masking, NLI score structure, chat generation, four decoding modes, synchronous streaming, static/dynamic cache equivalence, direct extractive span validity, safetensors roundtrip, resized embeddings, padding/MLM/seq2seq labels, actual optimizer updates, evaluation/best checkpoint/resume, the literal documented weighted Trainer, LoRA gradients/frozen parameters, PIL image and waveform pipelines, eager/SDPA agreement, gradient checkpointing, and CPU NF4.

Twenty-five released source files match the installed wheel byte-for-byte. All 169 Python fences parse; 57 public API objects/signatures were resolved. These are separate checks: parsing/signatures do not execute a placeholder example or prove model/backend compatibility. The tests use synthetic data and random weights, and training runs only a few steps. They do not establish convergence, calibration, factuality, scientific validity, or pretrained performance.

The skill has no custom REST inference service. Its external operations are the Hub SDK's model/dataset reads and optional uploads; those remain illustrative. Public single-record metadata/config probes do not test gated access, rate limits, Xet/LFS weight transfer, dataset streaming, or pagination. CUDA/ROCm/XPU/Gaudi/MPS, OCR, real speech files, VLM/detection/segmentation pretrained inference, multi-device dispatch, custom Hub generation, ONNX export, DeepSpeed, Optuna search, and external trackers were not executed. SHAP wrapper compatibility remains outside this suite and must be checked against actual model outputs.