Kenhuang Academy
  • English

      • Introduction to Hugging Face

      What is Hugging Face?

      Hugging Face is both a community and an open-source stack. Two pieces matter most:

      • Hub: a collaboration platform for models, datasets, and Spaces (hosted demos), with 1M+ public checkpoints.
      • Transformers: the model-definition framework for text, vision, audio, video, and multimodal models—inference and training. One definition can plug into PyTorch loops, Accelerate, and engines such as vLLM / TGI.

      This course centers on Transformers + Hub, and introduces datasets, evaluate, accelerate, and peft.

      Ecosystem

      ┌─────────────────────────────────────────────────────────────┐
│                   Hugging Face ecosystem                     │
└─────────────────────────────────────────────────────────────┘
   Hub (models / datasets / Spaces / model cards)
           │  from_pretrained / push_to_hub
   ┌───────┴────────┐
   │  Transformers  │  definitions + Pipeline + Trainer + generate()
   └───────┬────────┘
     datasets   evaluate   accelerate   peft   huggingface_hub
      ComponentRoleTypical use
      TransformersModel classes, tokenizers, Pipeline, TrainerLoad checkpoints, infer, fine-tune
      Hub“App store” for weights and dataSearch, download, publish, Spaces
      datasetsCached / streamed datasetsload_dataset, map
      evaluateMetricsaccuracy, F1
      accelerateDevices and distributed runsdevice_map="auto", multi-GPU
      peftParameter-efficient fine-tuningLoRA / QLoRA
      huggingface_hubHub client and hf CLIhf auth login, uploads

      Three core classes

      Officially a checkpoint is three objects, assembled with AutoClass + from_pretrained:

      ClassJobOn disk
      PreTrainedConfigDepth, hidden size, vocab, …config.json
      PreTrainedModelWeights and forward passmodel.safetensors
      PreprocessorTokenizer / ImageProcessor / Processortokenizer.json, …
      from transformers import AutoConfig, AutoTokenizer, AutoModelForCausalLM

name = "distilbert/distilgpt2"
config = AutoConfig.from_pretrained(name)
tokenizer = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(
    name, device_map="auto", dtype="auto"
)

      device_map="auto" shards layers across GPU/CPU; dtype="auto" picks the checkpoint’s float type. You do not need to write attention from scratch.

      Three official paths

      GoalAPI
      Fast inferencepipeline
      Training / fine-tuningTrainer (or a custom loop + Accelerate)
      LLM continuation / chatmodel.generate()

      vs training from scratch and vs Ollama

      This site’s PyTorch course teaches tensors, nn.Module, and handwritten loops—building networks from scratch. Transformers is load a pretrained checkpoint, then infer or fine-tune. They complement each other: if loss.backward() is unclear, go back to PyTorch; if you need BERT or Qwen from the Hub, stay here.

      TransformersOllamaScratch PyTorch
      On-rampThree-line PipelineOne command local chatYou write the loop
      StrengthResearch, fine-tune, multimodal, custom headsLocal Chat UI / APIFully custom architectures
      CostYou manage deps and VRAMGGUF (and similar) formatsYou own data and compute

      Pick: chat only → Ollama; fine-tune, eval, publish, vision/audio → Transformers.

      Coding agents (Cursor, Claude Code, …) can use Hugging Face Skills to run hf or TRL jobs. That is optional, not the main path of this tutorial.

      Good fit / caveats

      Good fit: text classification / NER, generation and chat, ASR, image classification, reproducing Hub paper checkpoints, LoRA on domain data.

      Watch out: full-finetuning a 7B on a laptop GPU; shipping weights without reading the card; gated Llama-style repos you never accepted; committing tokens.

      Readers who need strong Chinese support can prefer Qwen and DeepSeek (often ungated). Generation examples use distilgpt2 and Qwen/Qwen2.5-0.5B-Instruct.

      Next steps

      评论