By the numbers

Data collected on 2026-04-30 from the main branch at commit a752ba7af8.

Size

The repository ships ~149K lines of library Python code under src/transformers/, plus ~412K lines of test code under tests/. The 462 architecture-specific directories under src/transformers/models/ account for the majority of both.

xychart-beta horizontal
    title "Lines of code by area"
    x-axis ["src/transformers (lib)", "tests/", "examples/", "docs/", "utils/", "benchmark*/"]
    y-axis "Lines of code (×1000)" 0 --> 450
    bar [149, 412, 60, 110, 35, 12]

(Values are approximate; computed by wc -l over *.py and *.md files in each tree.)

Area	Files	Notes
Model directories	462	One per architecture under src/transformers/models/.
modeling_*.py	445	Standalone PyTorch modules.
modular_*.py	230	Modular shards expanded by make fix-repo.
configuration_*.py	444	One per architecture (some shared).
tokenization_*.py	99	Includes both slow and fast variants.
image_processing_*.py	194	Vision and multimodal preprocessors.
processing_*.py	131	Multimodal processor classes.
Pipeline tasks	28	Files under src/transformers/pipelines/.
Quantizers	23	Files matching quantizers/quantizer_*.py.
Integrations	40	Files under src/transformers/integrations/.

Top-level files by size (library code)

File	Lines	Purpose
src/transformers/modeling_utils.py	5,042	PreTrainedModel, weight init, mixins, attention dispatch
src/transformers/trainer.py	4,418	Full training loop
src/transformers/generation/utils.py	3,887	generate(), decoding strategies
src/transformers/tokenization_utils_base.py	3,580	Tokenizer base class
src/transformers/training_args.py	2,868	TrainingArguments dataclass
src/transformers/cache_utils.py	1,574	KV cache hierarchy
src/transformers/configuration_utils.py	1,362	PretrainedConfig base class
src/transformers/pipelines/base.py	1,370	Pipeline base class

Repository tooling files (largest)

File	Lines	Purpose
utils/modular_model_converter.py	≈3,400	Expands modular_*.py into modeling_*.py
utils/check_repo.py	≈1,900	Repo-consistency rules
utils/check_docstrings.py	≈2,500	Docstring validation
utils/create_dummy_models.py	≈2,400	Tiny model fixtures for fast tests
utils/tests_fetcher.py	≈1,500	Picks tests covering a diff

Activity

Commits per month over the last ~20 months (truncated; full list in git log):

xychart-beta horizontal
    title "Commits per month (recent)"
    x-axis ["2024-09", "2024-10", "2024-11", "2024-12", "2025-01", "2025-02", "2025-03", "2025-04", "2025-05", "2025-06", "2025-07", "2025-08", "2025-09", "2025-10", "2025-11", "2025-12", "2026-01", "2026-02", "2026-03", "2026-04"]
    y-axis "Commits" 0 --> 450
    bar [241, 352, 173, 181, 259, 229, 289, 398, 314, 326, 389, 394, 374, 369, 285, 283, 284, 303, 330, 266]

Total commits across the project's lifetime: 22,758 (since the initial commit on 2018-10-29).

Bot-attributed commits

Of all 22,758 commits in the history, 230 commits carry a Co-authored-by trailer mentioning a bot account such as dependabot[bot], github-actions[bot], factory-droid[bot], or copilot[bot]. That is roughly 1.0% of the history.

This number is a strict lower bound on AI-assisted work: inline tools such as Copilot or Cursor leave no trace in git log, and many AI-assisted PRs land without a Co-authored-by trailer. The current contributing guidelines explicitly discourage agent-only PRs (see CONTRIBUTING.md).

Test surface

Test grouping	Files	Notes
tests/models/	one dir per architecture	Per-model unit tests
tests/test_modeling_common.py	291,137 LOC	Shared ModelTesterMixin reused by every model
tests/test_processing_common.py	95,360 LOC	Multimodal processor mixin
tests/test_pipeline_mixin.py	37,715 LOC	Per-task pipeline test mixin
tests/test_tokenization_common.py	132,908 LOC	Tokenizer mixin
tests/quantization/	25 subdirs	One per quantizer
tests/pipelines/	per-task tests	Companion to per-model tests
tests/generation/, tests/trainer/, tests/tensor_parallel/, tests/sagemaker/	shared infra

Complexity

The library has a long-tail distribution of file sizes. A handful of central files dominate (see "Top-level files by size" above); most model files are 200-2,000 lines. The ruff config in pyproject.toml allows McCabe complexity up to 75 — the team has historically prioritized readability over short functions.

Dependencies

Mandatory runtime dependencies are declared in setup.py and the consolidated table at src/transformers/dependency_versions_table.py. Highlights:

Dependency	Pinned version	Use
torch	2.4+	Modeling backend
huggingface-hub	1.5+ to <2.0	Hub I/O
tokenizers	from huggingface-hub	Fast tokenizers
safetensors	runtime	Default checkpoint format
accelerate	1.1+	Distributed/mixed precision dispatch
kernels	0.12	Custom CUDA kernels via kernels-community
numpy	1.17+	Tensors and CPU paths
Pillow	10.0.1 to 15.0	Image I/O
tqdm	4.27+	Progress bars
regex	runtime	Tokenizer fast paths
requests / httpx	runtime	HTTP
pyyaml	runtime	Configs
packaging	runtime	Version checks
filelock	runtime	Cache concurrency

Optional extras ([torch], [dev], [quality], [testing], [audio], [vision], [integrations], [deepspeed], [sklearn], …) pull in roughly 200 additional packages spanning DeepSpeed, FSDP wrappers, peft, datasets, evaluate, sentencepiece, librosa, av, kenlm, scipy, Pillow, vlm libraries, and quantization backends (bitsandbytes, auto-gptq, autoawq, optimum-quanto, etc.).

See Dependencies for the full breakdown.