LLM Fine-Tuning Bootcamp
From fine-tuning fundamentals to distributed training in practice — systematically master core techniques including LoRA / QLoRA / RLHF / DeepSpeed, with hands-on guidance through the full LLM training and deployment pipeline.
Course Overview
Difficult Fine-Tuning Method Selection
With so many fine-tuning approaches available (Adapter, Prefix Tuning, LoRA, QLoRA, etc.), there is a lack of systematic comparison and selection guidance, making it hard to choose the optimal method for real-world scenarios.
Massive Training Resource Consumption
Full-parameter fine-tuning can easily require hundreds of GBs of GPU memory, far beyond what a single GPU can handle. Leveraging quantization and distributed training to lower the resource barrier is a core challenge.
Lack of Production Deployment Experience
There is a huge gap between lab-stage models and production-grade services, involving a series of engineering challenges such as data engineering, model serving, and containerized deployment.
Urgent Need for Domestic Hardware Adaptation
Running LLM training and inference on domestic computing platforms (such as Huawei Ascend) has become essential, yet practical experience and documentation remain scarce.
Course Projects
All course lab code is open source, with 1,000+ Stars on GitHub — one of the most popular LLM fine-tuning starter projects in China.
Core Technology Stack
Hugging Face Transformers
The industry-standard LLM development framework covering core capabilities such as model loading, Tokenizer, and Pipeline, with support for rapid access and fine-tuning of thousands of pre-trained models.
PEFT / LoRA / QLoRA
Hugging Face's parameter-efficient fine-tuning library that dramatically reduces GPU memory requirements through low-rank adaptation (LoRA) and quantized fine-tuning (QLoRA), enabling LLM fine-tuning on consumer-grade GPUs.
Microsoft DeepSpeed
Microsoft's open-source distributed training framework providing ZeRO memory optimization, Hybrid Engine, and an end-to-end RLHF training pipeline, capable of supporting training for models with hundreds of billions of parameters.
Huawei Ascend 910
A domestic AI training processor platform. The course provides in-depth, hands-on instruction on the full workflow of LLM deployment, inference, and fine-tuning on the Ascend 910.
Core Value
Systematic Mastery of Fine-Tuning Techniques
From Adapter Tuning to QLoRA, from RLHF to MoE — systematically map out the LLM fine-tuning technology landscape and build a comprehensive technical knowledge framework.
End-to-End Hands-On Skills
Covers the entire pipeline from data preparation, model training, evaluation, to service deployment, with complete code labs for every stage — ready to apply to real projects upon completion.
Production-Grade Deployment Skills
From personalized ChatBots to RAG-enhanced retrieval, from Docker containerization to distributed inference — master the key engineering skills needed to bring models into production.
Domestic Platform Hands-On Practice
Gain deep hands-on experience with the Huawei Ascend platform, mastering domestic computing adaptation skills and building a solid foundation for LLM application development in localized environments.
Learning Outcomes
- Systematically master the principles and applicable scenarios of mainstream fine-tuning methods including Prompt Tuning, Prefix Tuning, LoRA, and QLoRA
- Proficiently use Hugging Face Transformers and the PEFT library to complete the full LLM fine-tuning workflow
- Fine-tune and privately deploy large models such as DeepSeek v4 on consumer-grade GPUs using QLoRA
- Master the DeepSpeed distributed training framework with the engineering capability to train models with hundreds of billions of parameters
- Gain hands-on experience in model training and inference on domestic computing platforms such as Huawei Ascend
Detailed Syllabus
01 Warm-Up: AI LLM Technology Overview & Fine-Tuning Fundamentals
Four-Stage AI LLM Technology Overview
- In-depth look at the four waves of AI evolution: Narrow AI, Machine Learning, Deep Learning, and Large Language Models
- Four-stage LLM technology: Prompt Engineering, AI Agents, LLM Fine-Tuning, and Pre-Training
- Seizing the wave: LLMs as a tremendous opportunity for building super-individual capabilities
LLM Fine-Tuning Techniques Demystified
- Evolution of LLM technology: from statistical machine learning to Transformer-based large models
- Introduction to PEFT efficient fine-tuning techniques: Adapter Tuning, Prefix Tuning, Prompt Tuning, P-Tuning v1/v2
- Deep dive into the LoRA family: LoRA, QLoRA, AdaLoRA — principles and comparison
- Exploring the unified fine-tuning framework UniPELT and the few-shot method IA3
ChatGPT LLM Training Techniques Explained
- Reinforcement Learning from Human Feedback (RLHF): SFT, Reward Model, and RL-based fine-tuning
- Mixture of Experts (MoE) architecture: Switch Transformer, Expert Choice, and GLaM
- Understanding the critical role of RLHF + MoE in ChatGPT-class models
02 Tooling: Hugging Face Transformers & PEFT in Practice
Hugging Face Transformers Core Features
- Core concepts of the Transformers library: the three pillars — Pipeline, Model, and Tokenizer
- Quickly implement NLP tasks such as sentiment analysis and text summarization using Pipeline
- Detailed walkthrough of pre-trained model loading, saving, and Tokenizer strategies
Hugging Face PEFT: Parameter-Efficient Fine-Tuning Library
- Core concepts of the PEFT library: PeftConfig parameter configuration and PeftModel management
- Simplifying model loading with Auto Classes
- Complete workflow for LoRA model fine-tuning using the PEFT library
03 Hands-On: QLoRA Fine-Tuning & Personalized Application Development
Fine-Tuning DeepSeek v4 with QLoRA
- Preparing fine-tuning datasets: obtaining data from Hugging Face Datasets, data cleaning, and encoding
- Building a PEFT QLoRA fine-tuning pipeline: model definition, training, validation, and saving
- Using GPT-4 to generate high-quality private training data
- Running a private DeepSeek v4 model service end-to-end
Building a Personalized ChatBot with Fine-Tuned LLMs
- Setting up the development environment: MiniConda, Jupyter Lab, LangChain, Chroma vector database
- Implementing core ChatBot logic with LangChain PromptTemplate and Agents
- Production deployment: Docker containerized build and ChatBot service deployment
RAG-Enhanced ChatBot with Retrieval-Augmented Generation
- Embedding technology fundamentals and hands-on practice with OpenAI Embedding models
- Building a ChatBot knowledge base with the Chroma vector database
- Implementing a RAG-enhanced intelligent Q&A system with LangChain
04 Advanced: DeepSpeed Distributed Training & RLHF
DeepSpeed Distributed Training Framework
- DeepSpeed core modules: Training, Inference, Compression, and 4Science
- Distributed architecture: distributed training, inference, and communication strategies
- ZeRO redundancy optimization: memory consumption analysis, ZeRO-Offload, and ZeRO-Stage3
- End-to-end RLHF training system DeepSpeed-Chat and Hybrid Engine
Implementing RLHF Model Fine-Tuning with DeepSpeed Chat
- Introduction to the Meta OPT model family and dataset preparation
- Three-stage RLHF training: Actor Model, Reward Model, and one-click reinforcement learning training
- Deploying and running model inference services with DeepSpeed Chat
05 Extended: Domestic Hardware Adaptation & LLM Pre-Training
Domestic Adaptation: Fine-Tuning on Huawei Ascend 910
- Introduction to Huawei's full-stack AI platform and Ascend 310/910 processors
- Deploying DeepSeek v4 on Ascend 910: installation, inference, and training end-to-end
- Key insights and best practices for domestic computing platform adaptation
Qwen3.6 Pre-Training Techniques Revealed
- Introduction to the Qwen3.6 LLM family and technical architecture analysis
- Using and deploying Qwen3.6 models on Hugging Face
- Detailed analysis of Qwen3.6 strengths, weaknesses, and applicable scenarios
Hands-On: Pre-Training the Qwen3.6 LLM
- Training data preparation: open-source dataset selection, Chinese dataset optimization, and Token encoding efficiency improvements
- Pre-training Qwen3.6 with QLoRA + DeepSpeed: GPU selection and quantized pre-training script walkthrough
- Qwen3.6 pre-training end-to-end hands-on practice
Instructor
Jingtian Peng
Founder / CEOFounding member of Huawei 2012 Lab deep learning team. UC visiting scholar. Former technical partner at Caicloud (acquired by ByteDance in 2020), co-founder & CTO of Pinlan Data (raised ~¥200M). Kubeflow maintainer, TensorFlow contributor, Linux CNCF program committee member.
