Skip to content

🧪 JupyterLab PyTorch LLM Training Lab

Deploy an enterprise JupyterLab Data Science & AI Workspace pre-configured with PyTorch, Hugging Face Transformers, TRL (SFTTrainer), and PEFT (LoRA/QLoRA).

🏛 Architecture Overview

graph TD
    User["👨‍💻 AI Engineer / Data Scientist"] -->|http://jupyter-llm.gbnt.local| Caddy["🔒 Caddy Ingress (:80 / :443)"]
    Caddy -->|Reverse Proxy :8888| Jupyter["🧪 JupyterLab Container"]
    Jupyter -->|Notebooks & Weights| Granaries["📁 Shared Storage Pool (/var/contenedores/jupyter-llm)"]
    Jupyter -->|Loss & Hardware Telemetry| Observability["📊 Gubernator SRE (Loki + Prometheus)"]

🚀 Compose Blueprint

version: "3.8"

services:
  jupyter-llm:
    image: quay.io/jupyter/pytorch-notebook:latest
    container_name: jupyter_llm_lab
    restart: unless-stopped
    ports:
      - "127.0.0.1::8888"
    environment:
      - JUPYTER_TOKEN=gubernator-secret
      - JUPYTER_ENABLE_LAB=yes
    volumes:
      - /var/contenedores/jupyter-llm/work:/home/jovyan/work
      - /var/contenedores/jupyter-llm/hf_cache:/home/jovyan/.cache/huggingface
    deploy:
      resources:
        limits:
          memory: 8G
        reservations:
          memory: 2G
      placement:
        constraints:
          - stack.name == jupyter-llm-stack
          - ingress.host == jupyter-llm.gbnt.local
          - gbnt.caddy.port == 8888
          - node.labels.gbnt.node.role == worker

🔬 Interactive Fine-Tuning Steps

  1. Connect to http://jupyter-llm.gbnt.local with token gubernator-secret.
  2. Open llm_lora_finetuning.ipynb.
  3. Follow the 7-step pipeline:
  4. Step 1: Hardware environment check (PyTorch, CUDA/CPU).
  5. Step 2: Load ChatML domain instruction dataset.
  6. Step 3: Initialize base model (SmolLM-135M or Qwen2.5-0.5B).
  7. Step 4: Attach LoRA adapter config (r=8, alpha=16).
  8. Step 5: Train with SFTTrainer with live loss reporting.
  9. Step 6: Perform comparative inference tests.
  10. Step 7: Save weights to /home/jovyan/work/output/.