huggingface /

Transformers

Live map
View repository
Snapshotmain · a61d5f9

How one model ID becomes configuration, preprocessing, weights, generation, or a training run.

Click a building or route
Choose a path

Select a pretrained model, preprocess input, run it, and decode generated IDs.

System map7 components · 8 connections
130%
Hugging Face Transformers architectureThe Auto factories and Hub-backed loader form the center; inference runs upward through pipelines and generation, while training runs downward through Trainer.model ID, config, taskconfiguration and tens…encoded tensor batchtask model inputslogits and model cachetoken IDsTokenizerand processorTask pipelineAuto factoriesPretrainedmodelGenerationloopPretrainedloaderTrainer
Connections
BuildingsEntryServiceComputeStorageRuntimeToolingExternal

Terms, in plain English

Auto classA factory that chooses a concrete implementation from pretrained configuration metadata.

How to study this repo

  1. 1
    Resolve one model ID

    Follow AutoConfig and AutoModel mappings into from_pretrained and the chosen model class.

  2. 2
    Compare inference and training

    Trace the same model through Pipeline.generate and Trainer.training_step to see where their control paths diverge.