transformers

Public API of @huggingface/transformers. Everything re-exported from this file is considered stable — other imports are internal and may change.

Start here

• pipeline() — the one-call entry point for every task.
• Environment — env fields and LogLevel enum.

Model loading

• Pipelines — task-specific pipeline classes (TextGenerationPipeline, etc.).
• Models — AutoModel* classes (one per task).
• Tokenizers — AutoTokenizer, chat templates, Message.
• Processors — AutoProcessor, AutoImageProcessor, AutoFeatureExtractor.
• Configs — AutoConfig / PretrainedConfig.

Generation

• Generation config — sampling and beam-search parameters.
• Generation parameters — full shape of generate() arguments.
• Logits processors — modify next-token probabilities.
• Stopping criteria — control when generation halts.
• Streamers — receive tokens as they’re produced.

Data types and I/O

• Tensors — Tensor, shape ops, math, I/O.
• Images — RawImage, load_image().
• Audio — RawAudio, load_audio().
• Video — RawVideo, RawVideoFrame, load_video() (experimental).

Utilities

• Hub options — shared from_pretrained() option shapes.
• Maths — softmax, cos_sim, typed-array helpers.
• Model registry — inspect or clear the model cache.
• Random — seedable MT19937 PRNG matching Python’s random.

Type Definitions

DeviceType

Type: keyof typeof DEVICE_TYPES

DataType

Type: keyof typeof DATA_TYPES