nested_where.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "4b5599b2",
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "\n",
    "os.environ[\"PYTORCH_ENABLE_MPS_FALLBACK\"] = \"1\"\n",
    "os.environ[\"SCIPY_ARRAY_API\"] = \"1\"\n",
    "\n",
    "import torch\n",
    "import numpy as np\n",
    "from array_api_compat import get_namespace\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "ef949524",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "The memory_profiler extension is already loaded. To reload it, use:\n",
      "  %reload_ext memory_profiler\n"
     ]
    }
   ],
   "source": [
    "%load_ext memory_profiler"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "6bb78031",
   "metadata": {},
   "outputs": [],
   "source": [
    "def nested_where(data):\n",
    "    xp = get_namespace(data)\n",
    "    return xp.where(\n",
    "        data > 0,\n",
    "        xp.where(data > 5, xp.where(data > 20, data * 3, data * 2), data + 10),\n",
    "        xp.where(data < 10, data - 10, xp.where(data < -20, data / 2, data)),\n",
    "    )\n",
    "\n",
    "nested_where_compiled = torch.compile(nested_where)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "913d8ec8",
   "metadata": {},
   "outputs": [],
   "source": [
    "rng = np.random.default_rng(42)\n",
    "data_np = rng.uniform(-50, 50, size=int(1e8)).astype(\"float32\")\n",
    "data_mps = torch.from_numpy(data_np).to(\"mps\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "f5447762",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "33"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import gc\n",
    "\n",
    "gc.collect()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "28a95751",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "peak memory: 1002.81 MiB, increment: 3.61 MiB\n"
     ]
    }
   ],
   "source": [
    "%%memit\n",
    "nested_where(data_mps)[:5].cpu()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "70c7af1e",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "181 ms ± 2.74 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "nested_where(data_mps)[:5].cpu()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "205ea02f",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "peak memory: 315.00 MiB, increment: 4.06 MiB\n"
     ]
    }
   ],
   "source": [
    "%%memit\n",
    "nested_where_compiled(data_mps)[:5].cpu()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "a11d85c8",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "180 ms ± 1.01 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "nested_where_compiled(data_mps)[:5].cpu()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "e77bccb3",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "peak memory: 2476.77 MiB, increment: 2165.73 MiB\n"
     ]
    }
   ],
   "source": [
    "%%memit\n",
    "nested_where(data_np)[:5]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "8875834a",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "707 ms ± 30.1 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "nested_where(data_np)[:5]"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "dev",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.13.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}