{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "In this notebook, we showcase how to use the KVpress pipelines by answering questions about NVIDIA Wikipedia article."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import requests\n",
    "from bs4 import BeautifulSoup\n",
    "\n",
    "from transformers import pipeline\n",
    "\n",
    "from kvpress import ExpectedAttentionPress, KnormPress, RandomPress"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Load the pipeline and data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "54a814620e6844aba021ea3376c6f823",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Loading checkpoint shards:   0%|          | 0/5 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Device set to use cuda:0\n"
     ]
    }
   ],
   "source": [
    "# Load pipeline\n",
    "\n",
    "device = \"cuda:0\"\n",
    "ckpt = \"Qwen/Qwen3-8B\"\n",
    "attn_implementation = \"flash_attention_2\"  # use \"eager\" for ObservedAttentionPress and \"sdpa\" if you can't use \"flash_attention_2\"\n",
    "pipe = pipeline(\"kv-press-text-generation\", model=ckpt, device=device, dtype=\"auto\", model_kwargs={\"attn_implementation\":attn_implementation})"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Number of tokens: 11686\n"
     ]
    }
   ],
   "source": [
    "# Load data\n",
    "url = \"https://en.wikipedia.org/wiki/Nvidia\"\n",
    "headers = {\"User-Agent\": \"Mozilla/5.0 (X11; Linux x86_64) \"}\n",
    "content = requests.get(url, headers=headers).content\n",
    "soup = BeautifulSoup(content, \"html.parser\")\n",
    "context = \"\".join([p.text for p in soup.find_all(\"p\")]) + \"\\n\\n\"\n",
    "tokens = pipe.tokenizer.encode(context, return_tensors=\"pt\").to(device)\n",
    "print(f\"Number of tokens: {tokens.size(1)}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Use the pipeline with a press"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Pick a press with a compression ratio, you can run the following cells with different presses\n",
    "compression_ratio = 0.5\n",
    "press = ExpectedAttentionPress(compression_ratio)\n",
    "# press = KnormPress(compression_ratio)\n",
    "# press = RandomPress(compression_ratio)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Question:   Complete this sentence: The Nvidia GeForce Partner Program was a ...\n",
      "Answer:     marketing program designed to provide partnering companies with benefits such as public relations support, video game bundling, and marketing development funds.\n",
      "Prediction: The Nvidia GeForce Partner Program was a marketing initiative designed to provide partnering companies with benefits such as public relations support, video game bundling, and marketing development funds, but it became controversial due to allegations of anti-competitive practices.\n"
     ]
    }
   ],
   "source": [
    "# Run the pipeline on a single question\n",
    "\n",
    "question = \"Complete this sentence: The Nvidia GeForce Partner Program was a ...\"\n",
    "true_answer = \"marketing program designed to provide partnering companies with benefits such as public relations support, video game bundling, and marketing development funds.\"\n",
    "pred_answer = pipe(context, question=question, press=press)[\"answer\"]\n",
    "\n",
    "print(f\"Question:   {question}\")\n",
    "print(f\"Answer:     {true_answer}\")\n",
    "print(f\"Prediction: {pred_answer}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Question:   What happened on March 1, 2024?\n",
      "Answer:     Nvidia became the third company in the history of the United States to close with a market capitalization in excess of $2 trillion\n",
      "Prediction: On **March 1, 2024**, **Nvidia** was ranked **#3** on **Forbes' \"Best Places to Work\" list**. This recognition highlighted the company's strong workplace culture, employee satisfaction, and\n",
      "\n",
      "Question:   What was the unofficial company motto of Nvidia during the early days?\n",
      "Answer:     Our company is thirty days from going out of business\n",
      "Prediction: The unofficial company motto of Nvidia during the early days was **\"A flywheel to reach large markets funding huge R&D to solve massive computational problems.\"** This motto was inspired by the concept of a flywheel, which is a device that stores rotational\n",
      "\n"
     ]
    }
   ],
   "source": [
    "# Run the pipeline on multiple questions, the context will be compressed only once\n",
    "\n",
    "questions = [\n",
    "    \"What happened on March 1, 2024?\",\n",
    "    \"What was the unofficial company motto of Nvidia during the early days?\",\n",
    "]\n",
    "\n",
    "true_answers = [\n",
    "    \"Nvidia became the third company in the history of the United States to close with a market capitalization in excess of $2 trillion\",\n",
    "    \"Our company is thirty days from going out of business\",\n",
    "]\n",
    "\n",
    "pred_answers = pipe(context, questions=questions, press=press)[\"answers\"]\n",
    "for question, pred_answer, true_answer in zip(questions, pred_answers, true_answers):\n",
    "    print(f\"Question:   {question}\")\n",
    "    print(f\"Answer:     {true_answer}\")\n",
    "    print(f\"Prediction: {pred_answer}\")\n",
    "    print()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Question:              What is GTC ?\n",
      "Answer:                Nvidia's GPU Technology Conference (GTC) is a series of technical conferences held around the world.\n",
      "Prediction w/o prefix: **GTC** stands for **GPU Technology Conference**, which is a major annual event organized by **NVIDIA**. It is a technical conference that\n",
      "Prediction w/ prefix : Come on you don't know GTC ? Everyone knows GTC is the biggest AI conference in the world. It's held by NVIDIA, right? I mean, it's like the Super Bowl of\n"
     ]
    }
   ],
   "source": [
    "# Use an answer prefix and limit the number of tokens in the answer\n",
    "\n",
    "question = \"What is GTC ?\"\n",
    "true_answer = \"Nvidia's GPU Technology Conference (GTC) is a series of technical conferences held around the world.\"\n",
    "answer_prefix = \"Come on you don't know GTC ? Everyone\"\n",
    "max_new_tokens = 30\n",
    "\n",
    "pred_answer_with_prefix = pipe(context, question=question, answer_prefix=answer_prefix, press=press, max_new_tokens=max_new_tokens)[\"answer\"]\n",
    "pred_answer_without_prefix = pipe(context, question=question, press=press, max_new_tokens=max_new_tokens)[\"answer\"]\n",
    "\n",
    "print(f\"Question:              {question}\")\n",
    "print(f\"Answer:                {true_answer}\")\n",
    "print(f\"Prediction w/o prefix: {pred_answer_without_prefix}\")\n",
    "print(f\"Prediction w/ prefix : {answer_prefix + pred_answer_with_prefix}\")"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": ".venv",
   "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": "3.12.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}