TIGER-GGUF

🚀 SKT-OM (TIGER-OM) - Agentic RAG System

Advanced 13B Agentic RAG with Think Mode + Dynamic Plugins + LangGraph

Built for AMD Developer Hackathon 2026 on AMD Developer Cloud.

---

🌟 Project Overview

SKT-OM (also known as TIGER-OM) is a powerful 13B parameter fully agentic Retrieval-Augmented Generation (RAG) system. It goes far beyond traditional RAG by integrating:

• Think Mode — Advanced multi-step reasoning engine
• Dynamic Plugin Architecture — Intelligent tool selection & execution
• LangGraph Multi-Agent Workflow — Stateful agent collaboration
• SKT RAG — High-performance retrieval pipeline

The system takes natural language queries and returns intelligent, reasoned, and accurate responses with tool usage and verification.

---

📊 Model Details

• Model Name: TIGER-OM (SKT-OM)
• Parameters: 13 Billion
• Base Model: Custom trained on AMD hardware
• Quantization: Q4_K_M (Excellent balance between quality and size)
• GGUF Format: Optimized for CPU + GPU inference
• Training Hardware: AMD Developer Cloud GPUs ($100 credits)
• Inference: ROCm 7.0 + vLLM (Full FP16) + GGUF (Q4_K_M)

Q4_K_M Version provides near FP16 level reasoning quality while being much more memory efficient and faster on consumer/pro hardware.

---

✨ Key Features

• Think Mode Engine: Chain-of-Thought, Self-Reflection, Verification Loops, and Self-Critique
• Plugin Ecosystem: Code Runner, Math Solver, Web Search, Data Analyzer, Document Parser + Custom Plugins
• Advanced RAG: SKT RAG with query rewriting, multi-hop retrieval, reranking & contextual compression
• Multi-Agent System: LangGraph powered stateful workflow
• Memory: Persistent conversation state
• Tool Use: Dynamic plugin routing based on query intent

---

🔗 Important Links

• Live Demo: https://huggingface.co/spaces/lablab-ai-amd-developer-hackathon/SKT-OM
• Main Model Repo: Shrijanagain/TIGER-OM
• GGUF Quantized Models (Q4_K_M): Shrijanagain/TIGER-GGUF
• GitHub Repository (RAG + ADK): https://github.com/SHRIJANAGAIN/SKT-AMD-FILES

---

How It Works

graph TD
    A[User Query] --> B[Think Mode]
    B --> C[Decomposition & Planning]
    C --> D[Plugin Router]
    C --> E[SKT RAG Retrieval]
    D --> F[Execute Plugins]
    E --> G[Context Processing]
    F & G --> H[Verification Loop]
    H --> I[LangGraph Synthesis]
    I --> J[Final Response]

---

🛠️ Technologies Used

• LLM: 13B TIGER-OM (Q4_K_M GGUF)
• RAG Framework: SKT RAG + ADK Kit
• Agent Framework: LangGraph
• GPU Stack: ROCm 7.0 + AMD ADK Kit
• Inference: vLLM (FP16) + llama.cpp (GGUF Q4_K_M)
• Hardware: AMD MI300X
• Cloud: AMD Developer Cloud

---

🚀 Quick Start - GGUF Q4_K_M

# Using llama.cpp
./llama-cli \
  -m tiger-om-q4_k_m.gguf \
  -p "Your complex query here..." \
  -n 1024 \
  -t 8 \
  --temp 0.7

Python Example (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(
    model_path="tiger-om-q4_k_m.gguf",
    n_gpu_layers=-1,      # Use all GPU layers
    n_ctx=8192,
    verbose=False
)

response = llm.create_chat_completion(
    messages=[{"role": "user", "content": "Explain..."}],
    temperature=0.7,
    max_tokens=1024
)

print(response['choices'][0]['message']['content'])

---

📁 Repository Structure

• /skt_ai_labs — Core ADK + RAG integration
• /plugins — Plugin system
• /agents — LangGraph workflows
• /examples — Ready-to-use examples
• /docs — Architecture & guides

---

🏆 Hackathon Information

• Event: AMD Developer Hackathon 2026
• Trained on: AMD Developer Cloud ($100 credits)
• Built in Public: Regular technical updates shared
• Goal: Showcasing powerful agentic AI on AMD ROCm ecosystem

---

📄 License

MIT