Fix Colab: use Colab's native port output instead of ngrok

fix_colab_no_ngrok.py

@@ -0,0 +1,122 @@
+#!/usr/bin/env python3
+"""Fix Colab to not use ngrok — use google.colab.output instead."""
+import subprocess, os, json
+
+TOKEN = "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"
+subprocess.run(["git", "clone", f"https://{TOKEN}@github.com/ticketguy/littlefig.git", "/app/littlefig"], check=True)
+os.chdir("/app/littlefig")
+subprocess.run(["git", "config", "user.name", "0xticketguy"], check=True)
+subprocess.run(["git", "config", "user.email", "0xticketguy@harboria.dev"], check=True)
+
+colab = {
+    "nbformat": 4,
+    "nbformat_minor": 0,
+    "metadata": {
+        "colab": {"provenance": [], "gpuType": "T4"},
+        "kernelspec": {"name": "python3", "display_name": "Python 3"},
+        "accelerator": "GPU"
+    },
+    "cells": [
+        {"cell_type": "markdown", "metadata": {}, "source": [
+            "# 🍐 Little Fig Studio\n",
+            "\n",
+            "**Train any LLM on any hardware.** Select your model, configure, launch.\n",
+            "\n",
+            "| What | How |\n",
+            "|---|---|\n",
+            "| Quantization | FigQuant INT4 (beats NF4 on 156/156 layers) |\n",
+            "| Speed | 7× faster than BnB NF4 on GPU |\n",
+            "| Memory | Train 1.1B models in <4GB VRAM |\n",
+            "| Optimizer | FigMeZO (−18.6% loss, original research) |\n",
+            "\n",
+            "**Run cells below ↓**"
+        ]},
+        {"cell_type": "code", "metadata": {}, "source": [
+            "# Install\n",
+            "!pip install -q torch\n",
+            "!git clone https://github.com/ticketguy/littlefig.git /content/littlefig --quiet 2>/dev/null || true\n",
+            "!cd /content/littlefig && pip install -q -e \".[train]\"\n",
+            "!pip install -q uvicorn fastapi python-multipart\n",
+            "\n",
+            "import torch, sys\n",
+            "sys.path.insert(0, '/content/littlefig/src')\n",
+            "print(f'\\n✅ Ready | PyTorch {torch.__version__} | GPU: {torch.cuda.get_device_name() if torch.cuda.is_available() else \"CPU\"}')"
+        ], "execution_count": None, "outputs": []},
+        {"cell_type": "code", "metadata": {}, "source": [
+            "# Launch Little Fig Studio UI\n",
+            "import subprocess, time, threading\n",
+            "from google.colab import output\n",
+            "\n",
+            "# Start server in background\n",
+            "proc = subprocess.Popen(\n",
+            "    ['python', '-m', 'uvicorn', 'little_fig.web.server:app',\n",
+            "     '--host', '0.0.0.0', '--port', '8888'],\n",
+            "    cwd='/content/littlefig/src',\n",
+            "    stdout=subprocess.PIPE, stderr=subprocess.PIPE\n",
+            ")\n",
+            "time.sleep(3)\n",
+            "\n",
+            "# Serve via Colab's built-in proxy\n",
+            "output.serve_kernel_port_as_window(8888)\n",
+            "print('\\n🍐 Little Fig Studio launched!')\n",
+            "print('   A new tab/window should have opened with the UI.')\n",
+            "print('   If not, click the link above.')\n",
+            "print('\\n   Keep this cell running to keep the server alive.')"
+        ], "execution_count": None, "outputs": []},
+        {"cell_type": "markdown", "metadata": {}, "source": [
+            "---\n",
+            "## Or use Python directly (no UI)\n",
+            "\n",
+            "Change `MODEL` to any HuggingFace model you want to train."
+        ]},
+        {"cell_type": "code", "metadata": {}, "source": [
+            "from little_fig.engine import FigModel, FigTrainer, FigTrainingConfig\n",
+            "\n",
+            "# === CHANGE THIS TO YOUR MODEL ===\n",
+            "MODEL = 'TinyLlama/TinyLlama-1.1B-Chat-v1.0'\n",
+            "# Other options:\n",
+            "# MODEL = 'google/gemma-3-4b-it'\n",
+            "# MODEL = 'Qwen/Qwen2.5-1.5B'\n",
+            "# MODEL = 'microsoft/phi-2'\n",
+            "\n",
+            "model = FigModel.from_pretrained(MODEL, lora_r=16, lora_alpha=32, shared_codebook=True)\n",
+            "print(f'\\n✅ Loaded {MODEL}')\n",
+            "print(f'   Trainable: {sum(p.numel() for p in model.parameters() if p.requires_grad):,} params')"
+        ], "execution_count": None, "outputs": []},
+        {"cell_type": "code", "metadata": {}, "source": [
+            "# Train\n",
+            "config = FigTrainingConfig(\n",
+            "    num_epochs=1,\n",
+            "    learning_rate=2e-4,\n",
+            "    max_seq_length=256,\n",
+            "    batch_size=2,\n",
+            "    gradient_accumulation_steps=4,\n",
+            "    logging_steps=5,\n",
+            "    use_packing=True,\n",
+            ")\n",
+            "\n",
+            "trainer = FigTrainer(model, config)\n",
+            "trainer.load_dataset('tatsu-lab/alpaca', max_samples=200)\n",
+            "trainer.train()\n",
+            "\n",
+            "model.save_adapter('./my_adapter')\n",
+            "print('\\n✅ Done! Adapter saved to ./my_adapter')"
+        ], "execution_count": None, "outputs": []},
+        {"cell_type": "markdown", "metadata": {}, "source": [
+            "---\n",
+            "*0xticketguy / Harboria Labs | AGPL-3.0*"
+        ]}
+    ]
+}
+
+with open("Little_Fig_Colab.ipynb", "w") as f:
+    json.dump(colab, f, indent=2)
+
+subprocess.run(["git", "add", "-A"], check=True)
+subprocess.run(["git", "commit", "-m",
+    "Fix Colab: use google.colab.output for port serving (no ngrok needed)\n\n"
+    "Removed ngrok dependency (requires auth now).\n"
+    "Uses Colab's built-in output.serve_kernel_port_as_window() instead.\n"
+    "No signup, no tokens, just works."], check=True)
+subprocess.run(["git", "push", "origin", "main"], check=True)
+print("✅ Colab fixed — no ngrok needed")