LLMs-Lab

# LLMs-Lab  

Full-stack LLM Engineering Lab. Features: Autonomous Agents (ReAct/AutoGPT) | Fine-Tuning Llama/Mistral (SFT/DPO) | Large Model Deployment (DeepSeek 671B / 2.5-bit) | Advanced RAG (Hybrid Search) | Function Calling (Stream/Text-to-SQL/External APIs) | Frameworks (LangChain, Semantic Kernel, OpenAI) | Daily SOTA Paper Tracking. From theory to 0-to-1.
  
## 9. 📂 [DeepSeek](DeepSeek/)
Focused on **Inference Optimization** and **Low-Bit Quantization** strategies for massive-scale MoE models (600B+ parameters).

- **DeepSeek-R1 (671B) 2.51-bit Extreme Quantization Deployment**:
  - Deployed the **2.51-bit quantized version** (via Unsloth) of the 671B MoE model, achieving an **~80% reduction** in memory footprint (from 720GB to ~212GB).
  - Analyzed official and community benchmarks for **1.58-bit vs 2.51-bit** configurations, ultimately selecting the 2.51-bit build (Q2_K_XL) to ensure superior reasoning stability on the H20 GPU cluster.
  - 📄 **[View Hands-on Deployment Log & Benchmarks (PDF)](DeepSeek/Hands-on%20deployment%20of%20671B%20model%20inference%20(2.51-bit%20quantization).pdf)**
  
## 8. 📂 [Fine-Tuning](Fine-Tuning/)

This directory bridges the gap between theoretical architecture analysis and practical, memory-efficient fine-tuning of state-of-the-art open-source models. It covers the full lifecycle from pre-training understanding to post-training alignment.

### Key Modules

- **[Transformer Source Code Analysis](Fine-Tuning/AnnotatedTransformer.ipynb)**: A deep dive into the vanilla Transformer architecture, focusing on a line-by-line implementation analysis of **Self-Attention mechanisms**, Multi-Head Attention, and Layer Normalization to understand the foundational building blocks.

- **[Llama Series: QLoRA & Quantization](Fine-Tuning/Llama%20%20Fine-Tuning%20&%20Quantization%20&%20Running%20&%20Instroducntion%20&%20Deployment.pdf)**: Implementation of **[QLoRA](Fine-Tuning/Lora%20code.ipynb)** (Quantized Low-Rank Adapters) to fine-tune L