Status: [LIVE]
Endpoints found:
https://www.kaggle.com/code/YOUR_USERNAME/shadowgpu-llm
Source: https://github.com/Shreyash-Gaur/ShadowGPU
# 🌑 ShadowGPU: Headless Kaggle GPU Tunneling
ShadowGPU is a lightweight, zero-cost architecture that lets you run massive Large Language Models (like the 32B Qwen 3.6) on Kaggle's free dual Tesla T4 GPUs and reach them directly from your local development environment — no browser, no GUI, no cloud bill.
It provides a 100% headless, browser-free experience with a real-time streaming kill switch to preserve your Kaggle compute quotas.
> **Research & Experimentation Use:** This project is designed for personal AI research and local experimentation. Treat your Kaggle GPU quota as the finite shared resource it is — spin kernels up when you need them, kill them promptly when you don't.
---
## 🏗️ Architecture
```mermaid
graph LR
A[Local Script
localhost:11434] --> B[local_proxy.py
Request Router]
B --> C{Auth Headers
Injected}
C -->|LLM requests| D1[Ngrok Tunnel A]
D1 --> E1[Kaggle Kernel A
Dual Tesla T4]
E1 --> F1[Ollama Server A
Qwen 3.6 32B
deploy_secure_llm.py]
C -->|Embedding requests
dual-node only| D2[Ngrok Tunnel B]
D2 --> E2[Kaggle Kernel B]
E2 --> F2[Ollama Server B
Qwen3-Embedding
deploy_secure_emb.py]
style A fill:#1a1a2e,color:#e0e0e0
style B fill:#16213e,color:#e0e0e0
style D1 fill:#0f3460,color:#e0e0e0
style D2 fill:#0f3460,color:#e0e0e0
style E1 fill:#533483,color:#e0e0e0
style E2 fill:#533483,color:#e0e0e0
style F1 fill:#e94560,color:#fff
style F2 fill:#e94560,color:#fff
```
**Data path:** Local script → Proxy (injects auth headers) → Ngrok HTTPS tunnel → Kaggle kernel → Ollama (VRAM-locked)
---
## ✨ Key Features
- **Massive Compute:** Run 24GB+ LLMs on Dual 16GB Tesla T4 GPUs for free.
- **Ngrok Tunneling:** Exposes the remote Ollama server securely to the public internet — OS-agnostic (works across WSL, Windows, Mac).
- **Zero Cold-Starts:** Forces VRAM locking (`KEEP_ALIVE="-1"`) to guarantee 50+ tokens/second infere