Hugging Face
- Category
- AI and LLMs
- Docs outline
- 14 sections
Official docs outline
huggingface.co- Philosophy and core classes
Configuration, model and preprocessor classes, Hub checkpoints, design principles of Transformers.
- Pipeline inference
pipeline() for tasks, task identifiers, device and device_map, batching, large datasets, half precision.
- Models and configuration
PreTrainedModel, from_pretrained/save_pretrained, AutoModel classes, configs, dtype and device_map loading.
- Tokenizers
Tokenization, vocab, special tokens, padding and truncation, fast tokenizers, attention masks.
- Text generation
generate(), decoding strategies (greedy, sampling, beam search), temperature, top-k/top-p, max_new_tokens, streaming.
- Chat templates
apply_chat_template, roles, generation prompts, tool use formatting for chat models.
- KV cache
Key-value caching during generation, cache types, memory and speed tradeoffs.
- Attention implementations
SDPA, FlashAttention and attention backends, attn_implementation setting.
- Trainer
Trainer and TrainingArguments, training loop, evaluation, callbacks, mixed precision, distributed training.
- Efficient training on GPU
Batch size, gradient accumulation and checkpointing, mixed precision, memory optimization.
- PEFT and LoRA
Parameter-efficient fine-tuning with adapters, LoRA, loading and merging adapters.
- Quantization
bitsandbytes, GPTQ, AWQ quantization methods, loading low-bit models, memory reduction.
- Serving models
transformers serve and OpenAI-compatible serving, inference deployment options.
- Sharing models on the Hub
push_to_hub, model cards, repositories, versioning and uploading checkpoints.
Each section links to the official page. Titles and summaries by Extra Credit; the docs belong to their authors.
