Features

Cross-cutting capabilities that span more than one subsystem. Each page in this section covers a feature whose surface touches modeling, generation, configuration, the Hub, and often the trainer or the CLI.

Page	Spans	One-liner
Auto classes	Configuration, modeling, tokenization, processing	Late-binding factories that pick the right concrete class from model_type.
from_pretrained	Modeling, configuration, hub, quantization, tensor-parallel	The universal load path and its V5 weight-loading API.
Modular models	Modeling, repo tooling	The modular_*.py shard mechanism and make fix-repo.
Continuous batching	Generation, cache, attention, CLI	Paged-KV scheduler that powers transformers serve.
Tensor parallelism	Modeling, integrations, trainer	tp_plan and the integrations TP engine.
Chat templates	Tokenization, generation	Jinja2-based chat / tool-call rendering.
Serving	CLI, generation, continuous batching, pipelines	OpenAI-compatible HTTP serving.