🤗 Transformers

Hugging Face Transformers is the model-definition framework for state-of-the-art text, vision, audio, video and multimodal models — for both inference and training.

Adapted for this playbook from the 🤗 Transformers documentation by Hugging Face. Official page: https://huggingface.co/docs/transformers/index. Images © Hugging Face (documentation-images) unless noted. This is not a substitute for the upstream docs — verify against the current version.

Source: Hugging Face Transformers documentation (transformers_as_a_model_definition.png).

Executive summary

Use this section when you need a structured, enterprise-friendly walkthrough of 🤗 Transformers: how to install it, load models, run pipelines and generate, fine-tune with Trainer/PEFT, quantise, and plug into engines such as vLLM — always with links back to the official Hugging Face docs.

When to use this page

• Adopting or standardising on Hugging Face Transformers in a client or product stack.
• Briefing engineers on Pipeline vs Trainer vs Generate vs Serve CLI.
• Choosing quantization, serving engines or task recipes.
• Pair with the model landscape, Kimi K3 and Generative AI fundamentals.

Why Transformers matters

Hugging Face positions transformers as the pivot across the ecosystem: if a model definition is supported here, it is compatible with most training stacks (Axolotl, Unsloth, DeepSpeed, FSDP, …), inference engines (vLLM, SGLang, TGI, …) and adjacent libraries (llama.cpp, MLX, …).

There are 1M+ Transformers model checkpoints on the Hub.

Source: Hugging Face Transformers documentation.

Design principles (summary)

1. Fast and easy to use — each model is configuration + model + preprocessor; use Pipeline or Trainer quickly.
2. Pretrained models — reuse checkpoints instead of training from scratch; reduce carbon, cost and time.
3. Three core classes — configuration, model, preprocessor (from_pretrained / save_pretrained / push_to_hub).

See Philosophy for the full tenets.

How this playbook maps to Hugging Face docs

Playbook section	Hugging Face docs area	Start here
Get started	Installation, Quickstart, Philosophy	Installation
Base classes	Models, Preprocessors	Loading models
Inference	Pipeline, Generate, Optimisation, Chat, Serving	Pipeline
Training	Trainer, PEFT, distributed, hardware	Trainer
Quantization	Quantization guides	Concepts
Ecosystem	Community integrations	Inference engines
Tasks	Task how-tos	NLP tasks
Resources	Glossary, timeline, troubleshooting	Glossary
Flagship models	Selected model_doc deep dives	BERT
API map	Main classes + full model API on HF	API map

Credit policy

Every page in this section:

• Credits Hugging Face and links to the official Transformers doc URL
• Hosts selected diagrams from Hugging Face documentation-images with figure captions
• Rewrites content for playbook clarity (British spelling, enterprise notes) rather than mirroring API reference dumps
• Does not replace the upstream docs — always verify against huggingface.co/docs/transformers

Source: Hugging Face brand assets.

Learning path

1. Installation → Quickstart → Philosophy
2. Pipeline for fast inference; Loading models when you need control
3. Text generation and Chat for LLMs
4. Fine-tuning + PEFT for adaptation
5. Quantization and Serving for production

Key takeaways

• Transformers is the shared model-definition layer for Hub checkpoints and most of the open ML stack.
• Prefer Pipeline for tasks, generate for LLMs/VLMs, Trainer/PEFT for training.
• This playbook section is a credited learning guide — Hugging Face remains the source of truth.

What to do next

Start with Installation or jump to Pipeline.