{
 "nbformat": 4,
 "nbformat_minor": 5,
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "name": "python",
   "version": "3.11.0"
  }
 },
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/duoan/TorchCode/blob/master/templates/08_rmsnorm.ipynb)\n",
    "\n",
    "# 🟡 Medium: Implement RMSNorm\n",
    "\n",
    "Implement **Root Mean Square Layer Normalization** — the normalization used in LLaMA, Gemma, etc.\n",
    "\n",
    "$$\\text{RMSNorm}(x) = \\frac{x}{\\text{RMS}(x)} \\cdot w, \\quad \\text{RMS}(x) = \\sqrt{\\frac{1}{d}\\sum x_i^2 + \\epsilon}$$\n",
    "\n",
    "### Signature\n",
    "```python\n",
    "def rms_norm(x: torch.Tensor, weight: torch.Tensor, eps: float = 1e-6) -> torch.Tensor:\n",
    "    # Normalize over the last dimension. No mean subtraction (unlike LayerNorm).\n",
    "```\n",
    "\n",
    "### Rules\n",
    "- Do **NOT** use any built-in norm layers\n",
    "- Normalize over `dim=-1`\n",
    "- Must support autograd"
   ],
   "outputs": []
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# Install torch-judge in Colab (no-op in JupyterLab/Docker)\n",
    "try:\n",
    "    import google.colab\n",
    "    get_ipython().run_line_magic('pip', 'install -q torch-judge')\n",
    "except ImportError:\n",
    "    pass\n"
   ],
   "outputs": [],
   "execution_count": null
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "import torch"
   ],
   "outputs": [],
   "execution_count": null
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# ✏️ YOUR IMPLEMENTATION HERE\n",
    "\n",
    "def rms_norm(x, weight, eps=1e-6):\n",
    "    pass  # Replace this"
   ],
   "outputs": [],
   "execution_count": null
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "# 🧪 Debug\n",
    "x = torch.randn(2, 8)\n",
    "w = torch.ones(8)\n",
    "out = rms_norm(x, w)\n",
    "print(\"Output shape:\", out.shape)\n",
    "print(\"RMS of output:\", out.pow(2).mean(dim=-1).sqrt())  # should be ~1"
   ],
   "outputs": [],
   "execution_count": null
  },
  {
   "cell_type": "code",
   "metadata": {},
   "source": [
    "from torch_judge import check\n",
    "check('rmsnorm')"
   ],
   "outputs": [],
   "execution_count": null
  }
 ]
}