Dependencies

The dependency surface declared by setup.py and the auto-generated mirror at src/transformers/dependency_versions_table.py.

Hard install requirements

Listed in setup.py as install_requires. Installing pip install transformers (no extras) pulls these:

Package	Version pin	Why
huggingface-hub	>=1.5.0,<2.0	Hub I/O
numpy	>=1.17	Tensors and CPU paths
packaging	>=20.0	Version checks
pyyaml	>=5.1	Model card metadata
regex	>=2025.10.22	Tokenizer fast paths
tokenizers	>=0.22.0,<=0.23.0	Fast tokenizers
typer	(latest)	CLI
safetensors	>=0.4.3	Default checkpoint format
tqdm	>=4.27	Progress bars

filelock is bundled transitively. The fallback HTTP stack uses requests (also transitive).

torch is in an extra, not the install requirements

pip install transformers does not install torch. The recommendation is pip install "transformers[torch]", which adds:

torch>=2.4
accelerate>=1.1.0

The library raises clear errors if you try to use any modeling feature without torch. Some pipelines, tokenizers, and processors work without torch.

Major optional extras

Defined in setup.py:

Extra	Adds	Purpose
torch	torch, accelerate	Required for modeling
vision	torchvision, Pillow	Image processors
audio	torchaudio, librosa, pyctcdecode, phonemizer, kenlm (Py<3.13)	Audio processors / ASR
video	av	Video decode
timm	timm	Vision backbones
kernels	kernels	Hub kernels
sentencepiece	sentencepiece, protobuf	Slow tokenizer backends
tiktoken	tiktoken, blobfile	OpenAI BPE
mistral-common	mistral-common[image]	Mistral tokenizer backend
chat_template	jinja2, jmespath	Chat-template engine
sklearn	scikit-learn	Metrics
accelerate	accelerate	Distributed dispatch (also pulled by torch)
retrieval	faiss-cpu, datasets	RAG
sagemaker	sagemaker	AWS SageMaker
deepspeed	deepspeed, accelerate	DeepSpeed
optuna	optuna	Hyperparameter search
ray	ray[tune]	Hyperparameter search (Py<3.14)
integrations	kernels, optuna, codecarbon, ray (Py<3.14)	Aggregate of common integrations
codecarbon	codecarbon	Carbon emissions tracking
serving	openai, pydantic, uvicorn, fastapi, starlette, rich + torch	transformers serve
num2words	num2words	Some preprocessors
benchmark	optimum-benchmark	make benchmark
ja	fugashi, ipadic, unidic_lite, unidic, rhoknp, sudachipy/dict (Py<3.14)	Japanese tokenization
open-telemetry	opentelemetry-api/exporter-otlp/sdk	Metrics in continuous batching
quality	datasets, ruff, GitPython, urllib3, libcst, rich, ty, tomli, transformers-mlinter	Repo quality tooling
docs	hf-doc-builder	Docs build
testing	pytest + asyncio/random-order/rich/xdist/order/rerunfailures/timeout/env, parameterized, psutil, dill, evaluate, rouge-score, nltk, sacremoses, rjieba, beautifulsoup4, tensorboard, sacrebleu, filelock + extras docs, quality, retrieval, sentencepiece, serving, mistral-common	All testing deps
deepspeed-testing	deepspeed + testing + optuna + sentencepiece	DeepSpeed CI
all	aggregate of inference-time extras	"Most users want this"
dev	all + testing + ja + sklearn	Heaviest install

The pip install -e ".[dev]" install pulls roughly 200 packages.

Auto-generated mirror

src/transformers/dependency_versions_table.py is the runtime version of setup.py's _deps list, regenerated by python utils/checkers.py deps_table (a step of make fix-repo). src/transformers/dependency_versions_check.py enforces minimum versions at import time.

Tooling versions

These are pinned for reproducible CI:

Package	Pin	Reason
ruff	==0.14.10	Lint output deterministic
transformers-mlinter	==0.1.1	Custom rules in utils/rules.toml
ty	==0.0.20	Type checker
pandas	<2.3.0	Compatibility (datasets)
pytest	>=7.2.0,<9.0.0	Test framework

Heavy optional dependencies

Quantization backends each declare their own dependency outside this list (consumed via pip install bitsandbytes, pip install autoawq, pip install hqq, etc.) and are checked at runtime by is_<backend>_available helpers in src/transformers/utils/import_utils.py.

See also

• Tooling — make fix-repo regenerates dependency_versions_table.py.
• Quantization — backend dependencies and is_<backend>_available.
• Integrations — third-party adapters live in src/transformers/integrations/.