Python / Python Modern Generative AI and Agents Interview Questions

What is LoRA and how does the Hugging Face PEFT library simplify fine-tuning large models?

Fine-tuning all parameters of a 7B model requires enormous compute and memory. LoRA (Low-Rank Adaptation) sidesteps this by keeping the original pretrained weights frozen and injecting small trainable rank decomposition matrices into each layer. For a weight matrix W ∈ ℝ^{d×k}, LoRA adds ΔW = BA where B ∈ ℝ^{d×r} and A ∈ ℝ^{r×k} with rank r ≪ min(d,k). Only A and B are trained, reducing trainable parameters by 100–10,000×.

The Hugging Face PEFT (Parameter-Efficient Fine-Tuning) library wraps any transformers model with LoRA (or other methods like Prefix Tuning, IA3) and integrates with the Trainer API for a complete fine-tuning workflow. QLoRA combines 4-bit quantisation with LoRA, enabling fine-tuning a 7B model on a single 24 GB GPU.

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import LoraConfig, get_peft_model, TaskType
import torch

model_id  = 'mistralai/Mistral-7B-v0.1'
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Load in 4-bit for QLoRA
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type='nf4',
)
model = AutoModelForCausalLM.from_pretrained(
    model_id, quantization_config=bnb_config, device_map='auto'
)

# Prepare for k-bit training
from peft import prepare_model_for_kbit_training
model = prepare_model_for_kbit_training(model)

# LoRA configuration
lora_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=16,                    # rank: lower = fewer params = faster, less expressive
    lora_alpha=32,           # scaling factor (typically 2*r)
    lora_dropout=0.05,
    target_modules=[         # which weight matrices to add LoRA to
        'q_proj', 'k_proj', 'v_proj', 'o_proj',
        'gate_proj', 'up_proj', 'down_proj',
    ],
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# trainable params: 83,886,080 || all params: 7,325,491,200 || trainable%: 1.1%

# Save LoRA adapter only (not the full model)
model.save_pretrained('./lora-adapter')

# Load and merge for inference
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16)
merged = PeftModel.from_pretrained(base, './lora-adapter').merge_and_unload()

Take quiz

What does LoRA inject into a model's weight matrices, and what remains frozen?LoRA replaces the original weights with smaller matrices; nothing is frozen

✗ Try again.

LoRA adds two small trainable rank-r matrices (B and A) whose product ΔW=BA is added to the frozen original weight — only these small matrices are updated during training, leaving the vast majority of parameters unchanged

✓ Correct! Well done.

LoRA freezes the small matrices and trains the original full weight matrix

✗ Try again.

LoRA trains a separate model and averages its weights with the original

✗ Try again.

What does merge_and_unload() do after fine-tuning with LoRA?It removes the LoRA adapters to return to the original base model

✗ Try again.

It adds the LoRA weight deltas (BA) directly into the frozen base weights and removes the adapter structure — producing a standard model with the fine-tuned knowledge baked in, which runs without PEFT overhead at inference

✓ Correct! Well done.

It saves the model in a compressed format for faster loading

✗ Try again.

It evaluates the model on a test set and returns accuracy metrics

✗ Try again.

More Related questions...

What are Large Language Models (LLMs) and how do they generate text? What is the Hugging Face Transformers pipeline API and how do you use it for common NLP and vision tasks? How does tokenisation work in Hugging Face and what are the key tokenizer concepts? What is the Auto-class pattern in Hugging Face and how do you run inference with a raw model? What is prompt engineering and what are the most effective techniques for getting better outputs from LLMs? What is Retrieval-Augmented Generation (RAG) and why is it preferred over full fine-tuning for knowledge-intensive tasks? What are vector databases and how do they enable semantic search in RAG pipelines? How do you build a complete RAG pipeline using LangChain? What are the most important text splitting strategies in RAG, and how do chunk size and overlap affect retrieval quality? What are LangChain's core abstractions — Chains, Runnables, and the LangChain Expression Language? How do you add conversation memory to an LLM application with LangChain? What is an AI agent and how does function calling / tool use work in LLM-based agents? What is the ReAct agent pattern and how does LangChain implement it? How do you efficiently load large Hugging Face models for inference, including quantization and device placement? How do you use Hugging Face's text-generation pipeline with open-source chat models like Mistral or Llama? How do you use the Hugging Face Inference API and the InferenceClient for production deployments? What is LoRA and how does the Hugging Face PEFT library simplify fine-tuning large models? How do you use the Hugging Face Datasets library for training and evaluation? How do you fine-tune a model using the Hugging Face Trainer API? How do you evaluate LLM outputs for quality, factual accuracy, and hallucination? How do you stream LLM responses token by token for a better user experience? How do you use multimodal models (vision-language) with Hugging Face for image understanding tasks? How do you reliably get structured JSON output from LLMs, and what tools does LangChain provide? How do you compute semantic similarity between texts using Hugging Face and OpenAI embeddings? What document loaders does LangChain provide, and how do you handle different file types in a RAG pipeline? What is the OpenAI Assistants API and how does it differ from the Chat Completions API? What is the Parent Document Retriever pattern and when does it improve RAG performance? How do you manage, version, and reuse prompts in production LLM applications? How do you generate and manipulate images using Hugging Face's Diffusers library? How do you handle documents or conversations that exceed an LLM's context window? What is LangGraph and how does it differ from LangChain's AgentExecutor for building agents? What embedding models should you use for production RAG systems, and how do you choose between OpenAI and open-source options? How do you add safety guardrails and input/output validation to LLM applications? How do you manage LLM API costs and implement caching to reduce redundant calls? What is LlamaIndex and how does it compare to LangChain for RAG use cases? What is the Hugging Face Hub and how do you push a trained model to share it? How do you build a demo web interface for an LLM application using Gradio? How do you monitor and debug LLM applications in production using LangSmith?

Show more question and Answers...

FastAPI Basics Interview Questions

	Interviews Questions Java Spring Hibernate DevOps Testing API BigData Web Erlang AI Database Integration Cloud Python Scala Tools Golang SAP	About Javapedia.net Javapedia.net is for Java and J2EE developers, technologist and college students who prepare of interview. Also this site includes many practical examples. This site is developed using J2EE technologies by Steve Antony, a senior Developer/lead at one of the logistics based company. Follow Javapedia updates
	contact: javatutorials2016[at]gmail[dot]com
Kindly consider donating for maintaining this website. Thanks.
	Copyright © 2026, javapedia.net, all rights reserved. Privacy Policy · Terms of Service.

Python / Python Modern Generative AI and Agents Interview Questions

What is LoRA and how does the Hugging Face PEFT library simplify fine-tuning large models?

Comments & Discussions

Recently added questions