Why it matters
Pipelines lower the barrier for using models. Understanding them accelerates experimentation.
Advertisement
The architecture
Create: pipeline(task, model, tokenizer). Task-specific defaults for popular tasks.
Invoke: call with input; returns dict of results.
Advertisement
How it works end to end
Supported tasks: text generation, classification, QA, NER, translation, summarization, zero-shot classification, image classification, ASR, TTS.
Batch inference: pass list; pipeline batches automatically.
Device: pipeline(device=0) uses GPU 0.
Custom: PipelineRegistry lets you add tasks.