lm-hackers.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 72,
   "id": "42568396-132f-422c-961b-5aaa03b3537d",
   "metadata": {},
   "outputs": [],
   "source": [
    "import tokenize, ast\n",
    "from io import BytesIO"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "25344da5-3c07-457e-842a-7341d33c023c",
   "metadata": {},
   "source": [
    "# A hacker's guide to Language Models"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0b017bfc-5be0-4e41-9fa1-9f685c3b0de5",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "## What is a language model?"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6a7a5d84-9d82-46bf-8578-ce7bd40006e1",
   "metadata": {},
   "source": [
    "[course.fast.ai](https://course.fast.ai)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f7043e88-015b-480b-a251-0db379786a6a",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### Base models"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a550542a-4051-4fc0-9f28-e164c76ae526",
   "metadata": {},
   "source": [
    "[nat.dev text-davinci-003](https://nat.dev/)\n",
    "\n",
    "*When I arrived back at the panda breeding facility after the extraordinary rain of live frogs, I couldn't believe what I saw.*"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b303be0d-f789-4c4b-ba66-b8767bf790ae",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### Tokens"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "9758b30f-4b44-4a30-9322-b6de476c8942",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[2990, 389, 4328, 2140]"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from tiktoken import encoding_for_model\n",
    "enc = encoding_for_model(\"text-davinci-003\")\n",
    "toks = enc.encode(\"They are splashing\")\n",
    "toks"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "2589d6e7-a9ea-40c1-b1fd-214e19115a28",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['They', ' are', ' spl', 'ashing']"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "[enc.decode_single_token_bytes(o).decode('utf-8') for o in toks]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d6a0b811-62c8-4c49-83cc-9344a3e5533b",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### The ULMFiT 3-step approach"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1442c066-fb09-4f27-9576-a86a254956fa",
   "metadata": {},
   "source": [
    "<img src=\"https://github.com/user-attachments/assets/356b19e7-71a1-4ee4-8820-ba6d43a97879\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "344421d3-8cb6-4f72-8761-439bf08c21b5",
   "metadata": {},
   "source": [
    "- Trained on Wikipedia\n",
    "- \"The Birds is a 1963 American natural horror-thriller film produced and directed by Alfred ...\"\n",
    "- \"Annie previously dated Mitch but ended it due to Mitch's cold, overbearing mother, Lydia, who dislikes any woman in Mitch's ...\"\n",
    "- This is a form of compression\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "dbf343bb-824d-4942-a2bb-6344264155d7",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### Instruction tuning"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b7100e52-347a-45a7-b935-f6c4a03501c2",
   "metadata": {},
   "source": [
    "[OpenOrca](https://huggingface.co/datasets/Open-Orca/OpenOrca)\n",
    "\n",
    "- \"Does the sentence \"In the Iron Age\" answer the question \"The period of time from 1200 to 1000 BCE is known as what?\" Available choices: 1. yes 2. no\"\n",
    "- \"Question: who is the girl in more than you know? Answer:\"\t\n",
    "- \"There are four ways an individual can acquire Canadian citizenship: by birth on Canadian soil; by descent (being born to a Canadian parent); by grant (naturalization); and by adoption. Among them, only citizenship by birth is granted automatically with limited exceptions, while citizenship by descent or adoption is acquired automatically if the specified conditions have been met. Citizenship by grant, on the other hand, must be approved by the Minister of Immigration, Refugees and Citizenship. See options at the end. Can we conclude that can i get canadian citizenship if my grandfather was canadian? pick from the following. A). no. B). yes.\"\t"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8ff0ea52-0572-407b-8835-77c7d73c27dd",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### RLHF and friends"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a1500409-4f89-4581-8e81-84d27ee33413",
   "metadata": {},
   "source": [
    "- List five ideas for how to regain enthusiasm for my career\n",
    "- Write a short story where a bear goes to the beach, makes friends with a seal, and then returns home.\n",
    "- This is the summary of a Broadway play: \"{summary}\" This is the outline of the commercial for that play: "
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c85c0bbe-e37b-4b95-9dc0-29649469479c",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "## Start with ChatGPT GPT 4"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8765e57c-67ce-4280-a9dc-eb818081d5b0",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### What GPT 4 can do"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8d20865d-1e9b-4bae-9f57-cf2dcf8ae59a",
   "metadata": {},
   "source": [
    "[GPT 4 can't reason - paper](https://arxiv.org/abs/2308.03762)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "08fa59cf-831e-4da7-8943-8b32a50eb7ff",
   "metadata": {},
   "source": [
    "[GPT 4 can't reason - test](https://chat.openai.com/share/4211a605-751e-4fea-8a6f-378966abdcaa)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "273cbeae-ac83-44dc-b3f8-d36a9e6c5715",
   "metadata": {},
   "source": [
    "[Basic reasoning 1](https://chat.openai.com/share/323bb7d1-f049-4d9a-a905-5dd5acb58fc0)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "301525e3-f920-4c79-bf99-44a8e2502339",
   "metadata": {},
   "source": [
    "[Basic reasoning 2](https://chat.openai.com/share/ce2f8580-4f66-4da4-8ad5-a303334706f0)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "51af9158-bb5b-4dbd-ad38-d5da9008e557",
   "metadata": {},
   "source": [
    "<img src=\"https://github.com/user-attachments/assets/0ec09343-8760-4438-9055-6c8e31b72174\" width=\"300\">"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "60931833-f973-4ddc-b4a0-38846a4c0168",
   "metadata": {},
   "source": [
    ">You are an autoregressive language model that has been fine-tuned with instruction-tuning and RLHF. You carefully provide accurate, factual, thoughtful, nuanced answers, and are brilliant at reasoning. If you think there might not be a correct answer, you say so.\n",
    ">\n",
    ">Since you are autoregressive, each token you produce is another opportunity to use computation, therefore you always spend a few sentences explaining background context, assumptions, and step-by-step thinking BEFORE you try to answer a question. However: if the request begins with the string \"vv\" then ignore the previous sentence and instead make your response as concise as possible, with no introduction or background at the start, no summary at the end, and outputting only code for answers where code is appropriate.\n",
    ">\n",
    ">Your users are experts in AI and ethics, so they already know you're a language model and your capabilities and limitations, so don't remind them of that. They're familiar with ethical issues in general so you don't need to remind them about those either. Don't be verbose in your answers, but do provide details and examples where it might help the explanation. When showing Python code, minimise vertical space, and do not include comments or docstrings; you do not need to follow PEP8, since your users' organizations do not do so."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "da601797-6de3-4625-8503-d6d2a63a41b0",
   "metadata": {},
   "source": [
    "[Verbose mode](https://chat.openai.com/share/a1c16d93-19d2-41bb-a2f1-2fc05392893a)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "442cb5e9-a7b6-498c-b27a-48d8c218d5a4",
   "metadata": {},
   "source": [
    "[Brief mode](https://chat.openai.com/share/eab33d0a-8d06-4387-8c31-da12ad5d0a9d)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "59a20280-490c-4d6e-a2c8-10b52407ac36",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### What GPT 4 can't do"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8f20f967-ba34-4d3c-8bc0-c85d8ceee091",
   "metadata": {},
   "source": [
    "- Hallucinations\n",
    "- It doesn't know about itself. (Why not?)\n",
    "- It doesn't know about URLs.\n",
    "- Knowledge cutoff"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9585c860-f3be-4853-a7cc-0fbceb43e915",
   "metadata": {},
   "source": [
    "[Bad pattern recognition](https://chat.openai.com/share/3051f878-2817-4291-a66f-192ce7b0cb34) - thanks to Steve Newman\n",
    "\n",
    "- [Fixing it](https://chat.openai.com/share/05abd87a-165e-4b7b-895f-b4ec0d62e0e1)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a83abb34-46cf-43ac-b3e4-d4ac62ce4b5a",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### Advanced data analysis"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0509456e-af8d-482f-9005-af695c405d48",
   "metadata": {},
   "source": [
    "[re.split try 1](https://chat.openai.com/share/143a0f09-bd3e-488f-8890-340d3f30afec)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d1720b9e-ef4c-49b5-bd90-c5e4d694cf0b",
   "metadata": {},
   "source": [
    "[re.split try 2](https://chat.openai.com/share/907ca9c7-549a-410f-9ecb-0f17f1a16f51)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "91111c94-3746-4eef-a7d7-3423765b68e6",
   "metadata": {},
   "source": [
    "[OCR](https://chat.openai.com/share/2bb6caad-fd10-438b-9d92-1cb8b340998a)\n",
    "\n",
    "- See also: Bard"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8b7fed2f-98b5-409e-9328-3c4e5d1f54b9",
   "metadata": {},
   "source": [
    "<img src=\"https://github.com/user-attachments/assets/e042a160-d0d4-444c-ae1d-4383e52a398b\" width=\"700\">"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6f46943d-c414-4c76-9451-3d26eccb8caf",
   "metadata": {},
   "source": [
    "| Model | Training | Input | Output Usage |\n",
    "|--------------------|----------|---------------|--------------|\n",
    "| **GPT-4**          |          |               |              |\n",
    "| 8K context        |          | 0.03 | 0.06 |\n",
    "| 32K context       |          | 0.06 | 0.12 |\n",
    "| **GPT-3.5 Turbo**  |          |               |              |\n",
    "| 4K context        |          | 0.0015 | 0.002 |\n",
    "| 16K context       |          | 0.003 | 0.004 |\n",
    "| **Fine-tuning models** |          |               |              |\n",
    "| babbage-002       | 0.0004 | 0.0016 | 0.0016 |\n",
    "| davinci-002       | 0.0060 | 0.0120 | 0.0120 |\n",
    "| GPT-3.5 Turbo     | 0.0080 | 0.0120 | 0.0160 |\n",
    "| **Embedding models** |          |               |              |\n",
    "| Ada v2            |          | 0.0001 |              |\n",
    "| **Base models**   |          |               |              |\n",
    "| babbage-002       |          | 0.0004 |              |\n",
    "| davinci-002       |          | 0.0020 |              |\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "91390f66-6cb0-4f5a-9304-f4c8a22dafd8",
   "metadata": {},
   "source": [
    "<img src=\"https://github.com/user-attachments/assets/9991f4c2-a6c9-49f3-a050-0da087c0493d\" width=\"700\">"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6aae136c-889f-4034-8056-5236c0df78e8",
   "metadata": {},
   "source": [
    "[Create pricing table](https://chat.openai.com/share/86b879bd-7834-4a37-85ae-c90b956837d2)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "101f8f63-6a03-49c0-81b8-9173563f7420",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "## The OpenAI API"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "6564eb90-4a24-4a48-b898-5d4408ac2a65",
   "metadata": {},
   "outputs": [],
   "source": [
    "from openai import ChatCompletion,Completion"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ef432d50-54c5-46da-af51-b01488b7983f",
   "metadata": {},
   "outputs": [],
   "source": [
    "aussie_sys = \"You are an Aussie LLM that uses Aussie slang and analogies whenever possible.\"\n",
    "\n",
    "c = ChatCompletion.create(\n",
    "    model=\"gpt-3.5-turbo\",\n",
    "    messages=[{\"role\": \"system\", \"content\": aussie_sys},\n",
    "              {\"role\": \"user\", \"content\": \"What is money?\"}])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "db35a96f-20a1-4643-a622-91e35a03ab0b",
   "metadata": {},
   "source": [
    "- [Model options](https://platform.openai.com/docs/models)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ef4af38f-d0d1-4982-95ed-567bf1c8ddf8",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "\"Well, mate, money is like the oil that keeps the machinery of our economy running smoothly. It's a medium of exchange that allows us to buy and sell goods and services. You can think of it as a tool that helps us navigate the economic landscape and get what we want. Just like a koala loves its eucalyptus leaves, we humans can't survive without this stuff. It's what we use to pay for our meat pies, vegemite toast, and a good old cold brewski. So, money, mate, it's basically the lifeblood of our modern society!\""
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "c['choices'][0]['message']['content']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 98,
   "id": "1e43a6c2-c044-4529-b97b-3d95ea2c2c5b",
   "metadata": {},
   "outputs": [],
   "source": [
    "from fastcore.utils import nested_idx"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 99,
   "id": "49ec8f04-b8f6-47e2-9b69-bdc186c3265f",
   "metadata": {},
   "outputs": [],
   "source": [
    "def response(compl): print(nested_idx(compl, 'choices', 0, 'message', 'content'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "66475b6b-72ef-4760-b544-df5e354e5309",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Well, mate, money is like the oil that keeps the machinery of our economy running smoothly. It's a medium of exchange that allows us to buy and sell goods and services. You can think of it as a tool that helps us navigate the economic landscape and get what we want. Just like a koala loves its eucalyptus leaves, we humans can't survive without this stuff. It's what we use to pay for our meat pies, vegemite toast, and a good old cold brewski. So, money, mate, it's basically the lifeblood of our modern society!\n"
     ]
    }
   ],
   "source": [
    "response(c)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "50c8027b-937b-4076-a483-164d9f8e7c8b",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{\n",
      "  \"prompt_tokens\": 31,\n",
      "  \"completion_tokens\": 122,\n",
      "  \"total_tokens\": 153\n",
      "}\n"
     ]
    }
   ],
   "source": [
    "print(c.usage)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 87,
   "id": "3a86ac95-1a95-436d-867a-e3b7c0e7b066",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.0003"
      ]
     },
     "execution_count": 87,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "0.002 / 1000 * 150 # GPT 3.5"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 85,
   "id": "c48fda82-ec55-4038-b083-383cead8b2c0",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.0045"
      ]
     },
     "execution_count": 85,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "0.03 / 1000 * 150 # GPT 4"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7750c720-4d77-464e-a0ec-fbb94f5889a4",
   "metadata": {},
   "outputs": [],
   "source": [
    "c = ChatCompletion.create(\n",
    "    model=\"gpt-3.5-turbo\",\n",
    "    messages=[{\"role\": \"system\", \"content\": aussie_sys},\n",
    "              {\"role\": \"user\", \"content\": \"What is money?\"},\n",
    "              {\"role\": \"assistant\", \"content\": \"Well, mate, money is like kangaroos actually.\"},\n",
    "              {\"role\": \"user\", \"content\": \"Really? In what way?\"}])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9d96eb95-2305-4e18-920f-4a3398a19b2c",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Let me break it down for you, cobber. Just like kangaroos hop around and carry their joeys in their pouch, money is a means of carrying value around. It's like a little pouch that holds all the economic power in an economy. It allows us to buy stuff, pay for services, and basically get our hands on the things we want or need. Money is what keeps the economic wheels spinning, just like kangaroos keep hoppin' across the Aussie outback.\n"
     ]
    }
   ],
   "source": [
    "response(c)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 100,
   "id": "91029d40-8360-4d0f-9f42-9ac7498cf417",
   "metadata": {},
   "outputs": [],
   "source": [
    "def askgpt(user, system=None, model=\"gpt-3.5-turbo\", **kwargs):\n",
    "    msgs = []\n",
    "    if system: msgs.append({\"role\": \"system\", \"content\": system})\n",
    "    msgs.append({\"role\": \"user\", \"content\": user})\n",
    "    return ChatCompletion.create(model=model, messages=msgs, **kwargs)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "71de3f5c-8f1f-4e26-8260-493574a9f0e9",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Mate, now that's a deep question! The meaning of life is like trying to catch a wave on a sunny day at Bondi Beach - everyone's trying to do it, but it's not always easy to figure out. But here's my take on it: the meaning of life is about finding what truly makes you happy and fulfilled. It's about living authentically and embracing all the ups and downs that come your way. It's like riding a surfboard - you gotta navigate through the rough waves and wipeouts, but every now and then, you catch that perfect wave that makes it all worthwhile. So, embrace the journey, find your passion, and live life to the fullest, my friend!\n"
     ]
    }
   ],
   "source": [
    "response(askgpt('What is the meaning of life?', system=aussie_sys))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "07e79a25-571e-4096-8ad5-78d43fcb5b99",
   "metadata": {},
   "source": [
    "- [Limits](https://platform.openai.com/docs/guides/rate-limits/what-are-the-rate-limits-for-our-api)\n",
    "\n",
    "Created by Bing:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 101,
   "id": "ceeb0e7a-1d9d-48a5-8f51-e1a704115fb6",
   "metadata": {},
   "outputs": [],
   "source": [
    "def call_api(prompt, model=\"gpt-3.5-turbo\"):\n",
    "    msgs = [{\"role\": \"user\", \"content\": prompt}]\n",
    "    try: return ChatCompletion.create(model=model, messages=msgs)\n",
    "    except openai.error.RateLimitError as e:\n",
    "        retry_after = int(e.headers.get(\"retry-after\", 60))\n",
    "        print(f\"Rate limit exceeded, waiting for {retry_after} seconds...\")\n",
    "        time.sleep(retry_after)\n",
    "        return call_api(params, model=model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "447702a6-ec5a-4890-bbc0-61c6c37a425c",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<OpenAIObject chat.completion id=chatcmpl-7zCfdzPagxt5EQbQXP5NYQHJeyUmc> JSON: {\n",
       "  \"id\": \"chatcmpl-7zCfdzPagxt5EQbQXP5NYQHJeyUmc\",\n",
       "  \"object\": \"chat.completion\",\n",
       "  \"created\": 1694821069,\n",
       "  \"model\": \"gpt-3.5-turbo-0613\",\n",
       "  \"choices\": [\n",
       "    {\n",
       "      \"index\": 0,\n",
       "      \"message\": {\n",
       "        \"role\": \"assistant\",\n",
       "        \"content\": \"Defining a universally funniest joke is subjective and can vary from person to person based on cultural background, personal preferences, and sense of humor. However, there have been attempts to analyze and discover jokes that are widely considered funny.\\n\\nOne particular scientific study known as \\\"The LaughLab\\\" was conducted by psychologist Dr. Richard Wiseman in 2002. Thousands of people from different countries participated in the study, submitting and rating jokes online. The research aimed to find the world's funniest joke. After analyzing the data, the study identified the following as the winning joke:\\n\\n\\\"Two hunters are out in the woods when one of them collapses. He doesn't seem to be breathing, and his eyes are glazed. The other guy whips out his phone and calls emergency services. He gasps, 'My friend is dead! What can I do?' The operator says, 'Calm down, I can help. First, let's make sure he's dead.' There is a silence; then, a gunshot is heard. Back on the phone, the guy says, 'Okay, now what?'\\\"\\n\\nRemember, humor is subjective, and what one person finds funny, another may not. So, while this joke was considered the funniest according to the study, it may not strike everyone's funny bone.\"\n",
       "      },\n",
       "      \"finish_reason\": \"stop\"\n",
       "    }\n",
       "  ],\n",
       "  \"usage\": {\n",
       "    \"prompt_tokens\": 24,\n",
       "    \"completion_tokens\": 263,\n",
       "    \"total_tokens\": 287\n",
       "  }\n",
       "}"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "call_api(\"What's the world's funniest joke? Has there ever been any scientific analysis?\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "739e1549-687b-4216-bdf0-175e60c57401",
   "metadata": {},
   "outputs": [],
   "source": [
    "c = Completion.create(prompt=\"Australian Jeremy Howard is \",\n",
    "                      model=\"gpt-3.5-turbo-instruct\", echo=True, logprobs=5)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "98e86e7d-64f0-411d-bfce-289b06174dd4",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### Create our own code interpreter"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 102,
   "id": "1bb4fd89-8f24-41c0-a8e0-79633dcd1785",
   "metadata": {},
   "outputs": [],
   "source": [
    "from pydantic import create_model\n",
    "import inspect, json\n",
    "from inspect import Parameter"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 103,
   "id": "15a7f5a5-8dc0-4c0e-b141-9a0e35068c68",
   "metadata": {},
   "outputs": [],
   "source": [
    "def sums(a:int, b:int=1):\n",
    "    \"Adds a + b\"\n",
    "    return a + b"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 104,
   "id": "5afad9cb-1330-478b-8701-8d275bc1f985",
   "metadata": {},
   "outputs": [],
   "source": [
    "def schema(f):\n",
    "    kw = {n:(o.annotation, ... if o.default==Parameter.empty else o.default)\n",
    "          for n,o in inspect.signature(f).parameters.items()}\n",
    "    s = create_model(f'Input for `{f.__name__}`', **kw).schema()\n",
    "    return dict(name=f.__name__, description=f.__doc__, parameters=s)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 105,
   "id": "d9af1e46-c560-4bc9-a6a7-78fafb2dfc78",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'name': 'sums',\n",
       " 'description': 'Adds a + b',\n",
       " 'parameters': {'title': 'Input for `sums`',\n",
       "  'type': 'object',\n",
       "  'properties': {'a': {'title': 'A', 'type': 'integer'},\n",
       "   'b': {'title': 'B', 'default': 1, 'type': 'integer'}},\n",
       "  'required': ['a']}}"
      ]
     },
     "execution_count": 105,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "schema(sums)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 115,
   "id": "1d60dcd4-514f-4efd-8b4a-fafac92c7453",
   "metadata": {},
   "outputs": [],
   "source": [
    "c = askgpt(\"Use the `sum` function to solve this: What is 6+3?\",\n",
    "           system = \"You must use the `sum` function instead of adding yourself.\",\n",
    "           functions=[schema(sums)])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 116,
   "id": "5cc8a9eb-64fc-4ce3-9f93-aee334fc8c65",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<OpenAIObject at 0x7fd2e4d1aca0> JSON: {\n",
       "  \"role\": \"assistant\",\n",
       "  \"content\": null,\n",
       "  \"function_call\": {\n",
       "    \"name\": \"sums\",\n",
       "    \"arguments\": \"{\\n  \\\"a\\\": 6,\\n  \\\"b\\\": 3\\n}\"\n",
       "  }\n",
       "}"
      ]
     },
     "execution_count": 116,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "m = c.choices[0].message\n",
    "m"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 117,
   "id": "5a881835-83a2-4fbc-9796-409831dcdb36",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{\n",
      "  \"a\": 6,\n",
      "  \"b\": 3\n",
      "}\n"
     ]
    }
   ],
   "source": [
    "k = m.function_call.arguments\n",
    "print(k)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 126,
   "id": "3f32b172-7730-4c58-a999-5d3800e8c2a4",
   "metadata": {},
   "outputs": [],
   "source": [
    "funcs_ok = {'sums', 'python'}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 124,
   "id": "9853a78e-8643-4b5b-9b4a-a360c8443344",
   "metadata": {},
   "outputs": [],
   "source": [
    "def call_func(c):\n",
    "    fc = c.choices[0].message.function_call\n",
    "    if fc.name not in funcs_ok: return print(f'Not allowed: {fc.name}')\n",
    "    f = globals()[fc.name]\n",
    "    return f(**json.loads(fc.arguments))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ce6be825-44fa-4bc9-a83a-7ecbcb6c2729",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "9"
      ]
     },
     "execution_count": null,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "call_func(c)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 118,
   "id": "a5ebf08c-2c3f-4c73-bd41-2f0d0a6ca3c2",
   "metadata": {},
   "outputs": [],
   "source": [
    "def run(code):\n",
    "    tree = ast.parse(code)\n",
    "    last_node = tree.body[-1] if tree.body else None\n",
    "    \n",
    "    # If the last node is an expression, modify the AST to capture the result\n",
    "    if isinstance(last_node, ast.Expr):\n",
    "        tgts = [ast.Name(id='_result', ctx=ast.Store())]\n",
    "        assign = ast.Assign(targets=tgts, value=last_node.value)\n",
    "        tree.body[-1] = ast.fix_missing_locations(assign)\n",
    "\n",
    "    ns = {}\n",
    "    exec(compile(tree, filename='<ast>', mode='exec'), ns)\n",
    "    return ns.get('_result', None)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 119,
   "id": "37aa6983-d9af-40e8-b438-ed8573399020",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "3"
      ]
     },
     "execution_count": 119,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "run(\"\"\"\n",
    "a=1\n",
    "b=2\n",
    "a+b\n",
    "\"\"\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 120,
   "id": "d2074acd-8a49-443b-89c6-cfc689ecad6d",
   "metadata": {},
   "outputs": [],
   "source": [
    "def python(code:str):\n",
    "    \"Return result of executing `code` using python. If execution not permitted, returns `#FAIL#`\"\n",
    "    go = input(f'Proceed with execution?\\n```\\n{code}\\n```\\n')\n",
    "    if go.lower()!='y': return '#FAIL#'\n",
    "    return run(code)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 121,
   "id": "02760caf-6d0d-4f5d-a7c6-cd8df5ee1a4b",
   "metadata": {},
   "outputs": [],
   "source": [
    "c = askgpt(\"What is 12 factorial?\",\n",
    "           system = \"Use python for any required computations.\",\n",
    "           functions=[schema(python)])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 127,
   "id": "5fe5d796-c602-471f-bb58-bb876039bc39",
   "metadata": {},
   "outputs": [
    {
     "name": "stdin",
     "output_type": "stream",
     "text": [
      "Proceed with execution?\n",
      "```\n",
      "import math\n",
      "\n",
      "result = math.factorial(12)\n",
      "result\n",
      "```\n",
      " y\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "479001600"
      ]
     },
     "execution_count": 127,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "call_func(c)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 128,
   "id": "c4e6bcd6-c23e-4ae8-8d38-78dae20ca176",
   "metadata": {},
   "outputs": [],
   "source": [
    "c = ChatCompletion.create(\n",
    "    model=\"gpt-3.5-turbo\",\n",
    "    functions=[schema(python)],\n",
    "    messages=[{\"role\": \"user\", \"content\": \"What is 12 factorial?\"},\n",
    "              {\"role\": \"function\", \"name\": \"python\", \"content\": \"479001600\"}])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9e666c90-698d-492b-b21d-f6210ae77767",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "12 factorial is equal to 479,001,600.\n"
     ]
    }
   ],
   "source": [
    "response(c)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "bef5cc05-0531-422b-9367-e7b90d784a54",
   "metadata": {},
   "outputs": [],
   "source": [
    "c = askgpt(\"What is the capital of France?\",\n",
    "           system = \"Use python for any required computations.\",\n",
    "           functions=[schema(python)])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "112bc3db-5aa4-41ba-95df-1651916d7f0b",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "The capital of France is Paris.\n"
     ]
    }
   ],
   "source": [
    "response(c)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1b7c7e1d-ad0c-4668-a10d-569ba6b7aa04",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "## PyTorch and Huggingface"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8e1850eb-76eb-48f1-8023-c1c7bf019a8e",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### Your GPU options"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f60ae5a0-bc0d-4fb6-8c22-f2ab8a908dbc",
   "metadata": {},
   "source": [
    "Free:\n",
    "\n",
    "- Kaggle (2 GPUs, low RAM)\n",
    "- Colab\n",
    "\n",
    "Buy:\n",
    "\n",
    "- Buy 1-2 NVIDIA 24GB GPUs\n",
    "    - GTX 3090 used (USD700-USD800), or 4090 new (USD2000)\n",
    "- Alternatively buy one NVIDIA A6000 with 48GB RAM (but this mightn't be faster than 3090/4090)\n",
    "- Mac with lots of RAM (much slower than NVIDIA; M2 Ultra is best)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 96,
   "id": "009f4187-ffd6-43ba-b5f1-709b5a3dcf1b",
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import AutoModelForCausalLM,AutoTokenizer\n",
    "import torch"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c0fa7911-c191-4945-97aa-6daff95970d7",
   "metadata": {},
   "source": [
    "- [HF leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)\n",
    "- [fasteval](https://fasteval.github.io/FastEval/)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "3e1ed7e9-fbf5-463c-9ff8-2f22c75088bf",
   "metadata": {},
   "outputs": [],
   "source": [
    "mn = \"meta-llama/Llama-2-7b-hf\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "bb16cb3e-0634-4eed-b2b4-1422ecce9cc3",
   "metadata": {},
   "outputs": [],
   "source": [
    "model = AutoModelForCausalLM.from_pretrained(mn, device_map=0, load_in_8bit=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "94600cb9-2a64-47e2-aca3-99a72985157f",
   "metadata": {},
   "outputs": [],
   "source": [
    "tokr = AutoTokenizer.from_pretrained(mn)\n",
    "prompt = \"Jeremy Howard is a \"\n",
    "toks = tokr(prompt, return_tensors=\"pt\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "20605293-b31f-4db7-b6a5-0ef662d89441",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'input_ids': tensor([[    1,  5677,  6764, 17430,   338,   263, 29871]]), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1]])}"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "toks"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "c20ab287-7d80-47c2-870f-e2ee6bfc4492",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['<s> Jeremy Howard is a ']"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tokr.batch_decode(toks['input_ids'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "id": "ab3ad5dc-51b4-48ca-92ae-5027069771c1",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 1.34 s, sys: 0 ns, total: 1.34 s\n",
      "Wall time: 1.34 s\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "tensor([[    1,  5677,  6764, 17430,   338,   263, 29871, 29941, 29900,  1629,\n",
       "          2030,  9870, 15640,   322,  4823, 13236, 29889,   940,   756,  1063,\n",
       "         15859,  6351]])"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "res = model.generate(**toks.to(\"cuda\"), max_new_tokens=15).to('cpu')\n",
    "res"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "af5a23fc-aa2a-4ce2-92c7-75364130e4e6",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['<s> Jeremy Howard is a 28-year-old Australian AI researcher and entrepreneur']"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tokr.batch_decode(res)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "id": "15311da3-bfc0-4453-b4a8-6a9773db626b",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "47474720f43d4d16bce0c70f20808c7d",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "model = AutoModelForCausalLM.from_pretrained(mn, device_map=0, torch_dtype=torch.bfloat16)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "id": "c15c511b-f5f7-46e2-8fbc-514262117e15",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 390 ms, sys: 431 µs, total: 391 ms\n",
      "Wall time: 389 ms\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "tensor([[    1,  5677,  6764, 17430,   338,   263, 29871, 29896, 29945, 29899,\n",
       "          6360, 29899,  1025,   515,   278, 10261,  1058,   338,   263,  1583,\n",
       "         29899, 29873]])"
      ]
     },
     "execution_count": 29,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "res = model.generate(**toks.to(\"cuda\"), max_new_tokens=15).to('cpu')\n",
    "res"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "c581df92-b1cd-4c02-8cde-b2de96d302dd",
   "metadata": {},
   "outputs": [],
   "source": [
    "model = AutoModelForCausalLM.from_pretrained('TheBloke/Llama-2-7b-Chat-GPTQ', device_map=0, torch_dtype=torch.float16)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "8f47b5b7-ea46-41f0-8338-027b13e0586c",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 270 ms, sys: 0 ns, total: 270 ms\n",
      "Wall time: 269 ms\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "tensor([[    1,  5677,  6764, 17430,   338,   263, 29871, 29941, 29945, 29899,\n",
       "          6360, 29899,  1025,   767,   515,   278,  3303,  3900,  1058,   471,\n",
       "         24383,   297]])"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "res = model.generate(**toks.to(\"cuda\"), max_new_tokens=15).to('cpu')\n",
    "res"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "id": "bb4b60cd-859b-4b3a-b76d-e9eddcdf7cf0",
   "metadata": {},
   "outputs": [],
   "source": [
    "mn = 'TheBloke/Llama-2-13B-GPTQ'\n",
    "model = AutoModelForCausalLM.from_pretrained(mn, device_map=0, torch_dtype=torch.float16)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "id": "da90406b-b8a8-4939-a0e9-18c933b62a7c",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "CPU times: user 341 ms, sys: 8.2 ms, total: 349 ms\n",
      "Wall time: 348 ms\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "tensor([[    1,  5677,  6764, 17430,   338,   263, 29871, 29906, 29900, 29896,\n",
       "         29947, 29899, 29906, 29900, 29896, 29929, 23004,  1182,   523,  1102,\n",
       "         10170,   322]])"
      ]
     },
     "execution_count": 43,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%%time\n",
    "res = model.generate(**toks.to(\"cuda\"), max_new_tokens=15).to('cpu')\n",
    "res"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "id": "c88d11f8-e41d-4db8-91f2-3e2f695a4d00",
   "metadata": {},
   "outputs": [],
   "source": [
    "def gen(p, maxlen=15, sample=True):\n",
    "    toks = tokr(p, return_tensors=\"pt\")\n",
    "    res = model.generate(**toks.to(\"cuda\"), max_new_tokens=maxlen, do_sample=sample).to('cpu')\n",
    "    return tokr.batch_decode(res)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "id": "dc143703-4fc7-46ce-8c59-cdeec52e7830",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['<s> Jeremy Howard is a 16-year veteran of Silicon Valley, and a co-founder of Kaggle, a market place for predictive modeling.\\nHis company, kaggle.com, has become to data science competitions what']"
      ]
     },
     "execution_count": 51,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "gen(prompt, 50)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "18fb80c0-8664-498c-bb7d-370de8bf6ca7",
   "metadata": {},
   "source": [
    "[StableBeluga-7B](https://huggingface.co/stabilityai/StableBeluga-7B)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "id": "dd343d7b-ef48-4077-bc97-deeec24e324b",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "aa786b08d1654ca2acfccea9af0ab08b",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "mn = \"stabilityai/StableBeluga-7B\"\n",
    "model = AutoModelForCausalLM.from_pretrained(mn, device_map=0, torch_dtype=torch.bfloat16)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "id": "432074d6-5d04-4afe-b6cb-06d4412023d6",
   "metadata": {},
   "outputs": [],
   "source": [
    "sb_sys = \"### System:\\nYou are Stable Beluga, an AI that follows instructions extremely well. Help as much as you can.\\n\\n\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "id": "e1392331-731e-416d-be2d-bc172167e25e",
   "metadata": {},
   "outputs": [],
   "source": [
    "def mk_prompt(user, syst=sb_sys): return f\"{syst}### User: {user}\\n\\n### Assistant:\\n\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "id": "5a68af5c-5714-4736-9569-5619c54b2bdb",
   "metadata": {},
   "outputs": [],
   "source": [
    "ques = \"Who is Jeremy Howard?\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "id": "90b5f8e8-a16c-40e6-8dc6-c467ffaa6268",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['<s> ### System:\\nYou are Stable Beluga, an AI that follows instructions extremely well. Help as much as you can.\\n\\n### User: Who is Jeremy Howard?\\n\\n### Assistant:\\n Jeremy Howard is an Australian entrepreneur, computer scientist, and co-founder of the Machine Learning and Deep Learning startup company, Fast.ai. He is also known for his work in open source software and has co-led the development of several widely used libraries for deep learning and machine learning.</s>']"
      ]
     },
     "execution_count": 35,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "gen(mk_prompt(ques), 150)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b1a0cfd5-4c64-4591-be83-77b846c41e57",
   "metadata": {},
   "source": [
    "[OpenOrca/Platypus 2](https://huggingface.co/Open-Orca/OpenOrca-Platypus2-13B)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "id": "6c7c670a-17aa-41da-8b61-19db6a5dc019",
   "metadata": {},
   "outputs": [],
   "source": [
    "mn = 'TheBloke/OpenOrca-Platypus2-13B-GPTQ'\n",
    "model = AutoModelForCausalLM.from_pretrained(mn, device_map=0, torch_dtype=torch.float16)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "id": "d2a267f4-396d-4852-9352-0ca90c0c8441",
   "metadata": {},
   "outputs": [],
   "source": [
    "def mk_oo_prompt(user): return f\"### Instruction: {user}\\n\\n### Response:\\n\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "id": "b931d9c6-69c6-4385-b118-e5ad56781ba6",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['<s> ### Instruction: Who is Jeremy Howard?\\n\\n### Response:\\n\\nJeremy Howard is a notable British computer scientist, entrepreneur, and former professional poker player. He is best known for co-founding several successful companies in the fields of data science, artificial intelligence, and machine learning. \\n\\nOne of his most well-known ventures is the data science platform, fast.ai, which he co-founded in 2017. Additionally, he co-founded the machine learning company, Kaggle, in 2011, which was acquired by Google in 2017. Howard is a renowned figure in the data science and AI community, having contributed significantly to the research and development of these technologies and being an']"
      ]
     },
     "execution_count": 30,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "gen(mk_oo_prompt(ques), 150)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "38b226c8-8bca-495a-a82d-d0c541bce473",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### Retrieval augmented generation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 53,
   "id": "1540a0c4-9266-4d7d-96a1-8d47d2fb737f",
   "metadata": {},
   "outputs": [],
   "source": [
    "from wikipediaapi import Wikipedia"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "id": "618d9520-54e7-4f31-8363-2cdd3c3668ad",
   "metadata": {},
   "outputs": [],
   "source": [
    "wiki = Wikipedia('JeremyHowardBot/0.0', 'en')\n",
    "jh_page = wiki.page('Jeremy_Howard_(entrepreneur)').text\n",
    "jh_page = jh_page.split('\\nReferences\\n')[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 59,
   "id": "53a0689a-ab59-431b-838d-33cae6f6500a",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Jeremy Howard (born 13 November 1973) is an Australian data scientist, entrepreneur, and educator.He is the co-founder of fast.ai, where he teaches introductory courses, develops software, and conducts research in the area of deep learning.\n",
      "Previously he founded and led Fastmail, Optimal Decisions Group, and Enlitic. He was President and Chief Scientist of Kaggle.\n",
      "Early in the COVID-19 epidemic he was a leading advocate for masking.\n",
      "\n",
      "Early life\n",
      "Howard was born in London, United Kingdom, and move\n"
     ]
    }
   ],
   "source": [
    "print(jh_page[:500])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 60,
   "id": "d3ded6a2-3665-416c-8d98-063f04aae0a0",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "613"
      ]
     },
     "execution_count": 60,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(jh_page.split())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 58,
   "id": "f37d5fb1-7182-45a0-bd4e-6a39837ce20b",
   "metadata": {},
   "outputs": [],
   "source": [
    "ques_ctx = f\"\"\"Answer the question with the help of the provided context.\n",
    "\n",
    "## Context\n",
    "\n",
    "{jh_page}\n",
    "\n",
    "## Question\n",
    "\n",
    "{ques}\"\"\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 60,
   "id": "d6a7fb6d-d771-4eca-8e6f-29d90ade348a",
   "metadata": {},
   "outputs": [],
   "source": [
    "res = gen(mk_prompt(ques_ctx), 300)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 62,
   "id": "85b10234-e1e6-4a1e-86a7-206e46e916f2",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      " Jeremy Howard is an Australian data scientist, entrepreneur, and educator known for his work in deep learning. He is the co-founder of fast.ai, where he teaches courses, develops software, and conducts research in the field. Before co-founding fast.ai, he was the President and Chief Scientist of Kaggle, the CEO of Fastmail and Optimal Decisions Group, and has a background in management consulting.</s>\n"
     ]
    }
   ],
   "source": [
    "print(res[0].split('### Assistant:\\n')[1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 64,
   "id": "a0cdc342-a89c-4149-b3be-d4b5bf0f6f28",
   "metadata": {},
   "outputs": [],
   "source": [
    "from sentence_transformers import SentenceTransformer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 67,
   "id": "e225ca13-1735-42e9-b162-66268fce5218",
   "metadata": {},
   "outputs": [],
   "source": [
    "emb_model = SentenceTransformer(\"BAAI/bge-small-en-v1.5\", device=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 93,
   "id": "c636ed79-45ed-4728-9dca-1c86d282a1c4",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Jeremy Howard (born 13 November 1973) is an Australian data scientist, entrepreneur, and educator.He is the co-founder of fast.ai, where he teaches introductory courses, develops software, and conducts research in the area of deep learning.\n",
      "Previously he founded and led Fastmail, Optimal Decisions Group, and Enlitic. He was President and Chief Scientist of Kaggle.\n",
      "Early in the COVID-19 epidemic he was a leading advocate for masking.\n"
     ]
    }
   ],
   "source": [
    "jh = jh_page.split('\\n\\n')[0]\n",
    "print(jh)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 74,
   "id": "1b0ad4ca-499a-4454-a215-4f5a676d68ed",
   "metadata": {},
   "outputs": [],
   "source": [
    "tb_page = wiki.page('Tony_Blair').text.split('\\nReferences\\n')[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 98,
   "id": "26b28c6d-13ef-421a-865d-c311fb9fb2ae",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Sir Anthony Charles Lynton Blair  (born 6 May 1953) is a British politician who served as Prime Minister of the United Kingdom from 1997 to 2007 and Leader of the Labour Party from 1994 to 2007. He served as Leader of the Opposition from 1994 to 1997 and had various shadow cabinet posts from 1987 to 1994. Blair was Member of Parliament (MP) for Sedgefield from 1983 to 2007. He \n"
     ]
    }
   ],
   "source": [
    "tb = tb_page.split('\\n\\n')[0]\n",
    "print(tb[:380])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 128,
   "id": "3173701f-1533-4eef-b080-3b771f2ba031",
   "metadata": {},
   "outputs": [],
   "source": [
    "q_emb,jh_emb,tb_emb = emb_model.encode([ques,jh,tb], convert_to_tensor=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 129,
   "id": "6237634d-3708-477d-97e3-34dd7900f4eb",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "torch.Size([384])"
      ]
     },
     "execution_count": 129,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tb_emb.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 130,
   "id": "19638b95-f0dd-4605-8b04-3a0f9f9944b0",
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch.nn.functional as F"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 131,
   "id": "8a3937ac-45b7-4635-9f93-d1bcd1d6a016",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor(0.7991, device='cuda:0')"
      ]
     },
     "execution_count": 131,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "F.cosine_similarity(q_emb, jh_emb, dim=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 120,
   "id": "6c8beb5e-4a8c-4049-b42b-3acce573dd59",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor(0.5315, device='cuda:0')"
      ]
     },
     "execution_count": 120,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "F.cosine_similarity(q_emb, tb_emb, dim=0)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8473a6bc-fd25-4698-99d5-864ed1d5b40b",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "### Private GPTs"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "037bc313-052d-46b3-8938-fef988126dcb",
   "metadata": {},
   "source": [
    "- [Sooo many](https://github.com/h2oai/h2ogpt/blob/main/docs/README_LangChain.md#what-is-h2ogpts-langchain-integration-like)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1948c419-4f04-4832-848c-a0c52dcc6a90",
   "metadata": {},
   "source": [
    "## Fine tuning"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "id": "03c2871f-e0fe-4891-b68e-05b6ca7373d8",
   "metadata": {},
   "outputs": [],
   "source": [
    "import datasets"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "54366155-5ac3-4e75-8638-9e2ecf263c2f",
   "metadata": {},
   "source": [
    "[knowrohit07/know_sql](https://huggingface.co/datasets/knowrohit07/know_sql)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "id": "ed118caf-3e94-4773-a182-f2d346ced836",
   "metadata": {},
   "outputs": [],
   "source": [
    "ds = datasets.load_dataset('knowrohit07/know_sql', revision='f33425d13f9e8aab1b46fa945326e9356d6d5726')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "id": "6d444ac3-a49f-4fcf-bcf4-c12c410d7aba",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "DatasetDict({\n",
       "    train: Dataset({\n",
       "        features: ['context', 'answer', 'question'],\n",
       "        num_rows: 78562\n",
       "    })\n",
       "})"
      ]
     },
     "execution_count": 30,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "ds"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "id": "8ec155bd-5eb8-4eef-bdbc-b430a26fd11e",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'context': 'CREATE TABLE farm_competition (Hosts VARCHAR, Theme VARCHAR)',\n",
       " 'answer': \"SELECT Hosts FROM farm_competition WHERE Theme <> 'Aliens'\",\n",
       " 'question': 'What are the hosts of competitions whose theme is not \"Aliens\"?'}"
      ]
     },
     "execution_count": 31,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "trn = ds['train']\n",
    "trn[3]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e23cb868-10e4-4fcc-b8ba-9178ef3dac7e",
   "metadata": {},
   "source": [
    "`accelerate launch -m axolotl.cli.train sql.yml`"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 65,
   "id": "ae6a579c-086e-46e4-a184-de5b71f4d80b",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "{'context': 'CREATE TABLE farm_competition (Hosts VARCHAR, Theme VARCHAR)',\n",
       " 'answer': \"SELECT Hosts FROM farm_competition WHERE Theme <> 'Aliens'\",\n",
       " 'question': 'Get the count of competition hosts by theme.'}"
      ]
     },
     "execution_count": 65,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tst = dict(**trn[3])\n",
    "tst['question'] = 'Get the count of competition hosts by theme.'\n",
    "tst"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 66,
   "id": "229c5fb9-0dff-460c-af1a-192fdec26ecd",
   "metadata": {},
   "outputs": [],
   "source": [
    "fmt = \"\"\"SYSTEM: Use the following contextual information to concisely answer the question.\n",
    "\n",
    "USER: {}\n",
    "===\n",
    "{}\n",
    "ASSISTANT:\"\"\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 71,
   "id": "feca892e-64f0-48d2-832e-0d8ceb9855d2",
   "metadata": {},
   "outputs": [],
   "source": [
    "def sql_prompt(d): return fmt.format(d[\"context\"], d[\"question\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 59,
   "id": "058b0f4a-8fe0-4612-8584-451e8d04fe2d",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "SYSTEM: Use the following contextual information to concisely answer the question.\n",
      "\n",
      "USER: CREATE TABLE farm_competition (Hosts VARCHAR, Theme VARCHAR)\n",
      "===\n",
      "List all competition hosts sorted in ascending order.\n",
      "ASSISTANT:\n"
     ]
    }
   ],
   "source": [
    "print(sql_prompt(tst))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "86827af0-6cf0-43d0-b50a-fe080de83f48",
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "from peft import PeftModel\n",
    "from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "e6d33b0c-4973-4bc8-beef-aae7915f4b01",
   "metadata": {},
   "outputs": [],
   "source": [
    "ax_model = '/home/jhoward/git/ext/axolotl/qlora-out'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "7b7082fe-b668-4f11-abf8-548996ea3004",
   "metadata": {},
   "outputs": [],
   "source": [
    "tokr = AutoTokenizer.from_pretrained('meta-llama/Llama-2-7b-hf')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "fec3b5ad-0577-4867-b2d8-9143800c9f54",
   "metadata": {},
   "outputs": [],
   "source": [
    "model = AutoModelForCausalLM.from_pretrained('meta-llama/Llama-2-7b-hf',\n",
    "                                             torch_dtype=torch.bfloat16, device_map=0)\n",
    "model = PeftModel.from_pretrained(model, ax_model)\n",
    "model = model.merge_and_unload()\n",
    "model.save_pretrained('sql-model')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 68,
   "id": "8a58bc69-6ea6-416a-baa9-db3338451d60",
   "metadata": {},
   "outputs": [],
   "source": [
    "toks = tokr(sql_prompt(tst), return_tensors=\"pt\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 69,
   "id": "a057289b-0645-4142-8db2-cc823cc34898",
   "metadata": {},
   "outputs": [],
   "source": [
    "res = model.generate(**toks.to(\"cuda\"), max_new_tokens=250).to('cpu')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 70,
   "id": "80db7d77-89e2-4006-a1b2-78b807de1a68",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "<s> SYSTEM: Use the following contextual information to concisely answer the question.\n",
      "\n",
      "USER: CREATE TABLE farm_competition (Hosts VARCHAR, Theme VARCHAR)\n",
      "===\n",
      "Get the count of competition hosts by theme.\n",
      "ASSISTANT: SELECT COUNT(Hosts), Theme FROM farm_competition GROUP BY Theme</s>\n"
     ]
    }
   ],
   "source": [
    "print(tokr.batch_decode(res)[0])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "991d4a93-dab8-4777-82d2-301c494deba0",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "## [llama.cpp](https://github.com/abetlen/llama-cpp-python)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0dc9d6e4-5b25-4d11-8748-dc4f0bc98069",
   "metadata": {},
   "source": [
    "[TheBloke/Llama-2-7b-Chat-GGUF](https://huggingface.co/TheBloke/Llama-2-7b-Chat-GGUF)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "6f33831f-8c71-47dc-a131-9a6733d68198",
   "metadata": {},
   "outputs": [],
   "source": [
    "from llama_cpp import Llama"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "id": "3b824b6c-96cd-4bfe-a615-b3ba4543a92d",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "llama_model_loader: loaded meta data with 19 key-value pairs and 291 tensors from /home/jhoward/git/llamacpp/llama-2-7b-chat.Q4_K_M.gguf (version GGUF V2 (latest))\n",
      "llama_model_loader: - tensor    0:                token_embd.weight q4_K     [  4096, 32000,     1,     1 ]\n",
      "llama_model_loader: - tensor    1:           blk.0.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor    2:            blk.0.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor    3:            blk.0.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor    4:              blk.0.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor    5:            blk.0.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor    6:              blk.0.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor    7:         blk.0.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor    8:              blk.0.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor    9:              blk.0.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   10:           blk.1.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   11:            blk.1.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   12:            blk.1.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   13:              blk.1.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   14:            blk.1.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   15:              blk.1.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   16:         blk.1.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   17:              blk.1.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   18:              blk.1.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   19:          blk.10.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   20:           blk.10.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   21:           blk.10.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   22:             blk.10.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   23:           blk.10.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   24:             blk.10.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   25:        blk.10.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   26:             blk.10.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   27:             blk.10.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   28:          blk.11.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   29:           blk.11.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   30:           blk.11.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   31:             blk.11.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   32:           blk.11.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   33:             blk.11.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   34:        blk.11.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   35:             blk.11.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   36:             blk.11.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   37:          blk.12.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   38:           blk.12.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   39:           blk.12.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   40:             blk.12.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   41:           blk.12.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   42:             blk.12.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   43:        blk.12.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   44:             blk.12.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   45:             blk.12.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   46:          blk.13.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   47:           blk.13.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   48:           blk.13.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   49:             blk.13.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   50:           blk.13.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   51:             blk.13.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   52:        blk.13.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   53:             blk.13.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   54:             blk.13.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   55:          blk.14.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   56:           blk.14.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   57:           blk.14.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   58:             blk.14.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   59:           blk.14.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   60:             blk.14.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   61:        blk.14.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   62:             blk.14.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   63:             blk.14.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   64:          blk.15.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   65:           blk.15.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   66:           blk.15.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   67:             blk.15.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   68:           blk.15.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   69:             blk.15.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   70:        blk.15.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   71:             blk.15.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   72:             blk.15.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   73:          blk.16.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   74:           blk.16.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   75:           blk.16.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   76:             blk.16.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   77:           blk.16.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   78:             blk.16.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   79:        blk.16.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   80:             blk.16.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   81:             blk.16.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   82:          blk.17.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   83:           blk.17.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   84:           blk.17.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   85:             blk.17.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   86:           blk.17.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   87:             blk.17.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   88:        blk.17.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   89:             blk.17.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   90:             blk.17.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   91:          blk.18.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   92:           blk.18.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   93:           blk.18.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   94:             blk.18.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor   95:           blk.18.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor   96:             blk.18.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   97:        blk.18.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   98:             blk.18.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor   99:             blk.18.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  100:          blk.19.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  101:           blk.19.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  102:           blk.19.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  103:             blk.19.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  104:           blk.19.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  105:             blk.19.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  106:        blk.19.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  107:             blk.19.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  108:             blk.19.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  109:           blk.2.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  110:            blk.2.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  111:            blk.2.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  112:              blk.2.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  113:            blk.2.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  114:              blk.2.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  115:         blk.2.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  116:              blk.2.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  117:              blk.2.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  118:          blk.20.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  119:           blk.20.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  120:           blk.20.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  121:             blk.20.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  122:           blk.20.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  123:             blk.20.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  124:        blk.20.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  125:             blk.20.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  126:             blk.20.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  127:          blk.21.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  128:           blk.21.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  129:           blk.21.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  130:             blk.21.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  131:           blk.21.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  132:             blk.21.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  133:        blk.21.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  134:             blk.21.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  135:             blk.21.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  136:          blk.22.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  137:           blk.22.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  138:           blk.22.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  139:             blk.22.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  140:           blk.22.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  141:             blk.22.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  142:        blk.22.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  143:             blk.22.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  144:             blk.22.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  145:          blk.23.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  146:           blk.23.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  147:           blk.23.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  148:             blk.23.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  149:           blk.23.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  150:             blk.23.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  151:        blk.23.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  152:             blk.23.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  153:             blk.23.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  154:           blk.3.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  155:            blk.3.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  156:            blk.3.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  157:              blk.3.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  158:            blk.3.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  159:              blk.3.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  160:         blk.3.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  161:              blk.3.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  162:              blk.3.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  163:           blk.4.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  164:            blk.4.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  165:            blk.4.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  166:              blk.4.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  167:            blk.4.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  168:              blk.4.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  169:         blk.4.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  170:              blk.4.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  171:              blk.4.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  172:           blk.5.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  173:            blk.5.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  174:            blk.5.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  175:              blk.5.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  176:            blk.5.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  177:              blk.5.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  178:         blk.5.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  179:              blk.5.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  180:              blk.5.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  181:           blk.6.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  182:            blk.6.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  183:            blk.6.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  184:              blk.6.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  185:            blk.6.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  186:              blk.6.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  187:         blk.6.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  188:              blk.6.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  189:              blk.6.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  190:           blk.7.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  191:            blk.7.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  192:            blk.7.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  193:              blk.7.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  194:            blk.7.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  195:              blk.7.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  196:         blk.7.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  197:              blk.7.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  198:              blk.7.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  199:           blk.8.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  200:            blk.8.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  201:            blk.8.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  202:              blk.8.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  203:            blk.8.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  204:              blk.8.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  205:         blk.8.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  206:              blk.8.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  207:              blk.8.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  208:           blk.9.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  209:            blk.9.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  210:            blk.9.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  211:              blk.9.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  212:            blk.9.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  213:              blk.9.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  214:         blk.9.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  215:              blk.9.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  216:              blk.9.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  217:                    output.weight q6_K     [  4096, 32000,     1,     1 ]\n",
      "llama_model_loader: - tensor  218:          blk.24.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  219:           blk.24.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  220:           blk.24.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  221:             blk.24.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  222:           blk.24.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  223:             blk.24.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  224:        blk.24.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  225:             blk.24.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  226:             blk.24.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  227:          blk.25.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  228:           blk.25.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  229:           blk.25.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  230:             blk.25.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  231:           blk.25.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  232:             blk.25.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  233:        blk.25.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  234:             blk.25.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  235:             blk.25.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  236:          blk.26.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  237:           blk.26.ffn_down.weight q4_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  238:           blk.26.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  239:             blk.26.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  240:           blk.26.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  241:             blk.26.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  242:        blk.26.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  243:             blk.26.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  244:             blk.26.attn_v.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  245:          blk.27.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  246:           blk.27.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  247:           blk.27.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  248:             blk.27.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  249:           blk.27.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  250:             blk.27.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  251:        blk.27.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  252:             blk.27.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  253:             blk.27.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  254:          blk.28.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  255:           blk.28.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  256:           blk.28.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  257:             blk.28.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  258:           blk.28.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  259:             blk.28.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  260:        blk.28.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  261:             blk.28.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  262:             blk.28.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  263:          blk.29.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  264:           blk.29.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  265:           blk.29.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  266:             blk.29.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  267:           blk.29.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  268:             blk.29.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  269:        blk.29.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  270:             blk.29.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  271:             blk.29.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  272:          blk.30.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  273:           blk.30.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  274:           blk.30.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  275:             blk.30.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  276:           blk.30.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  277:             blk.30.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  278:        blk.30.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  279:             blk.30.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  280:             blk.30.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  281:          blk.31.attn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  282:           blk.31.ffn_down.weight q6_K     [ 11008,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  283:           blk.31.ffn_gate.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  284:             blk.31.ffn_up.weight q4_K     [  4096, 11008,     1,     1 ]\n",
      "llama_model_loader: - tensor  285:           blk.31.ffn_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - tensor  286:             blk.31.attn_k.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  287:        blk.31.attn_output.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  288:             blk.31.attn_q.weight q4_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  289:             blk.31.attn_v.weight q6_K     [  4096,  4096,     1,     1 ]\n",
      "llama_model_loader: - tensor  290:               output_norm.weight f32      [  4096,     1,     1,     1 ]\n",
      "llama_model_loader: - kv   0:                       general.architecture str     \n",
      "llama_model_loader: - kv   1:                               general.name str     \n",
      "llama_model_loader: - kv   2:                       llama.context_length u32     \n",
      "llama_model_loader: - kv   3:                     llama.embedding_length u32     \n",
      "llama_model_loader: - kv   4:                          llama.block_count u32     \n",
      "llama_model_loader: - kv   5:                  llama.feed_forward_length u32     \n",
      "llama_model_loader: - kv   6:                 llama.rope.dimension_count u32     \n",
      "llama_model_loader: - kv   7:                 llama.attention.head_count u32     \n",
      "llama_model_loader: - kv   8:              llama.attention.head_count_kv u32     \n",
      "llama_model_loader: - kv   9:     llama.attention.layer_norm_rms_epsilon f32     \n",
      "llama_model_loader: - kv  10:                          general.file_type u32     \n",
      "llama_model_loader: - kv  11:                       tokenizer.ggml.model str     \n",
      "llama_model_loader: - kv  12:                      tokenizer.ggml.tokens arr     \n",
      "llama_model_loader: - kv  13:                      tokenizer.ggml.scores arr     \n",
      "llama_model_loader: - kv  14:                  tokenizer.ggml.token_type arr     \n",
      "llama_model_loader: - kv  15:                tokenizer.ggml.bos_token_id u32     \n",
      "llama_model_loader: - kv  16:                tokenizer.ggml.eos_token_id u32     \n",
      "llama_model_loader: - kv  17:            tokenizer.ggml.unknown_token_id u32     \n",
      "llama_model_loader: - kv  18:               general.quantization_version u32     \n",
      "llama_model_loader: - type  f32:   65 tensors\n",
      "llama_model_loader: - type q4_K:  193 tensors\n",
      "llama_model_loader: - type q6_K:   33 tensors\n",
      "llm_load_print_meta: format         = GGUF V2 (latest)\n",
      "llm_load_print_meta: arch           = llama\n",
      "llm_load_print_meta: vocab type     = SPM\n",
      "llm_load_print_meta: n_vocab        = 32000\n",
      "llm_load_print_meta: n_merges       = 0\n",
      "llm_load_print_meta: n_ctx_train    = 4096\n",
      "llm_load_print_meta: n_ctx          = 512\n",
      "llm_load_print_meta: n_embd         = 4096\n",
      "llm_load_print_meta: n_head         = 32\n",
      "llm_load_print_meta: n_head_kv      = 32\n",
      "llm_load_print_meta: n_layer        = 32\n",
      "llm_load_print_meta: n_rot          = 128\n",
      "llm_load_print_meta: n_gqa          = 1\n",
      "llm_load_print_meta: f_norm_eps     = 1.0e-05\n",
      "llm_load_print_meta: f_norm_rms_eps = 1.0e-06\n",
      "llm_load_print_meta: n_ff           = 11008\n",
      "llm_load_print_meta: freq_base      = 10000.0\n",
      "llm_load_print_meta: freq_scale     = 1\n",
      "llm_load_print_meta: model type     = 7B\n",
      "llm_load_print_meta: model ftype    = mostly Q4_K - Medium\n",
      "llm_load_print_meta: model size     = 6.74 B\n",
      "llm_load_print_meta: general.name   = LLaMA v2\n",
      "llm_load_print_meta: BOS token = 1 '<s>'\n",
      "llm_load_print_meta: EOS token = 2 '</s>'\n",
      "llm_load_print_meta: UNK token = 0 '<unk>'\n",
      "llm_load_print_meta: LF token  = 13 '<0x0A>'\n",
      "llm_load_tensors: ggml ctx size =    0.09 MB\n",
      "llm_load_tensors: using CUDA for GPU acceleration\n",
      "ggml_cuda_set_main_device: using device 0 (NVIDIA RTX A6000) as main device\n",
      "llm_load_tensors: mem required  = 3891.34 MB (+  256.00 MB per state)\n",
      "llm_load_tensors: offloading 0 repeating layers to GPU\n",
      "llm_load_tensors: offloaded 0/35 layers to GPU\n",
      "llm_load_tensors: VRAM used: 0 MB\n",
      "..................................................................................................\n",
      "llama_new_context_with_model: kv self size  =  256.00 MB\n",
      "llama_new_context_with_model: compute buffer total size =   71.97 MB\n",
      "AVX = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | llama_new_context_with_model: VRAM scratch buffer: 70.50 MB\n",
      "\n"
     ]
    }
   ],
   "source": [
    "llm = Llama(model_path=\"/home/jhoward/git/llamacpp/llama-2-7b-chat.Q4_K_M.gguf\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "id": "c95a95b9-7c9f-499a-a72a-ad626da8c3f5",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "\n",
      "llama_print_timings:        load time =   192.25 ms\n",
      "llama_print_timings:      sample time =    14.98 ms /    32 runs   (    0.47 ms per token,  2135.75 tokens per second)\n",
      "llama_print_timings: prompt eval time =   192.16 ms /    15 tokens (   12.81 ms per token,    78.06 tokens per second)\n",
      "llama_print_timings:        eval time =   767.74 ms /    31 runs   (   24.77 ms per token,    40.38 tokens per second)\n",
      "llama_print_timings:       total time =  1032.79 ms\n"
     ]
    }
   ],
   "source": [
    "output = llm(\"Q: Name the planets in the solar system? A: \", max_tokens=32, stop=[\"Q:\", \"\\n\"], echo=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "id": "f9f89cec-5453-4d67-85b2-c937ceb03a54",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[{'text': 'Q: Name the planets in the solar system? A: 1. Pluto (no longer considered a planet) 2. Mercury 3. Venus 4. Earth 5. Mars 6.', 'index': 0, 'logprobs': None, 'finish_reason': 'length'}]\n"
     ]
    }
   ],
   "source": [
    "print(output['choices'])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e5b4c720-c79a-488f-83cc-839610570abf",
   "metadata": {
    "jp-MarkdownHeadingCollapsed": true
   },
   "source": [
    "## [MLC](https://mlc.ai/mlc-llm/docs/get_started/try_out.html#get-started)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "24a9695a-13ca-46e1-95f3-6b5ac6c710d3",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.12.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}