{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Intel MKL WARNING: Support of Intel(R) Streaming SIMD Extensions 4.2 (Intel(R) SSE4.2) enabled only processors has been deprecated. Intel oneAPI Math Kernel Library 2025.0 will require Intel(R) Advanced Vector Extensions (Intel(R) AVX) instructions.\n",
      "Intel MKL WARNING: Support of Intel(R) Streaming SIMD Extensions 4.2 (Intel(R) SSE4.2) enabled only processors has been deprecated. Intel oneAPI Math Kernel Library 2025.0 will require Intel(R) Advanced Vector Extensions (Intel(R) AVX) instructions.\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/Users/shreshth/anaconda3/envs/llm-test/lib/python3.11/site-packages/threadpoolctl.py:1214: RuntimeWarning: \n",
      "Found Intel OpenMP ('libiomp') and LLVM OpenMP ('libomp') loaded at\n",
      "the same time. Both libraries are known to be incompatible and this\n",
      "can cause random crashes or deadlocks on Linux when loaded in the\n",
      "same Python program.\n",
      "Using threadpoolctl may cause crashes or deadlocks. For more\n",
      "information and possible workarounds, please see\n",
      "    https://github.com/joblib/threadpoolctl/blob/master/multiple_openmp.md\n",
      "\n",
      "  warnings.warn(msg, RuntimeWarning)\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "{'name': 'nq',\n",
       " 't_bmodel': LogisticRegression(),\n",
       " 't_amodel': LogisticRegression(),\n",
       " 'sep_layer_range': (27, 32),\n",
       " 'ap_layer_range': (17, 22)}"
      ]
     },
     "execution_count": 1,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# test probe loading \n",
    "import pickle as pkl\n",
    "import numpy as np\n",
    "import sklearn \n",
    "from sklearn import linear_model\n",
    "import os\n",
    "os.environ[\"PYTORCH_ENABLE_MPS_FALLBACK\"] = \"1\"\n",
    "\n",
    "# load the probe data\n",
    "with open(\"./model/20240625-131035_demo.pkl\", \"rb\") as f:\n",
    "    probe_data = pkl.load(f)\n",
    "# take the NQ open one\n",
    "probe_data = probe_data[-2]\n",
    "probe_data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "probe = probe_data['t_bmodel']\n",
    "layer_range = probe_data['sep_layer_range']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "1c0e30b73cab48069e985203c598a9b0",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "import torch\n",
    "from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer\n",
    "\n",
    "model_id = \"meta-llama/Llama-2-7b-chat-hf\"\n",
    "model = AutoModelForCausalLM.from_pretrained(model_id, device_map=\"cpu\")\n",
    "tokenizer = AutoTokenizer.from_pretrained(model_id)\n",
    "tokenizer.use_default_system_prompt = False"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[    1,   518, 25580, 29962,  3532, 14816, 29903,  6778,    13,  3492,\n",
      "           526,   263,  8444, 20255, 29889,    13, 29966,   829, 14816, 29903,\n",
      "          6778,    13,    13,  5816,   338,   278,  7483,   310,  3444, 29973,\n",
      "           518, 29914, 25580, 29962]]) torch.Size([1, 34])\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)\n"
     ]
    }
   ],
   "source": [
    "from threading import Thread\n",
    "\n",
    "system_prompt = \"You are a helpful assistant.\"\n",
    "message = \"what is the capital of France?\"\n",
    "max_new_tokens = 100\n",
    "top_p = 0.9\n",
    "top_k = 50\n",
    "temperature = 0.7\n",
    "repetition_penalty = 1.2\n",
    "\n",
    "conversation = []\n",
    "\n",
    "conversation.append({\"role\": \"system\", \"content\": system_prompt})\n",
    "conversation.append({\"role\": \"user\", \"content\": message})\n",
    "input_ids = tokenizer.apply_chat_template(conversation, return_tensors=\"pt\")\n",
    "input_ids = input_ids.to(model.device)\n",
    "print(input_ids, input_ids.shape)\n",
    "streamer = TextIteratorStreamer(tokenizer, timeout=1000.0, skip_prompt=True, skip_special_tokens=True)\n",
    "generation_kwargs = dict(\n",
    "    input_ids=input_ids,\n",
    "    max_new_tokens=max_new_tokens,\n",
    "    do_sample=True,\n",
    "    top_p=top_p,\n",
    "    top_k=top_k,\n",
    "    temperature=temperature,\n",
    "    repetition_penalty=repetition_penalty,\n",
    "    streamer=streamer,\n",
    "    output_hidden_states=True,\n",
    "    return_dict_in_generate=True,\n",
    ")\n",
    "\n",
    "thread = Thread(target=model.generate, kwargs=generation_kwargs)\n",
    "thread.start()\n",
    "\n",
    "generated_text = \"\"\n",
    "highlighted_text = \"\"\n",
    "\n",
    "for new_text in streamer:\n",
    "    print(new_text)\n",
    "    generated_text += new_text\n",
    "    current_input_ids = tokenizer.encode(generated_text, return_tensors=\"pt\").to(model.device)\n",
    "    print(current_input_ids, current_input_ids.shape)\n",
    "    with torch.no_grad():\n",
    "        outputs = model(current_input_ids, output_hidden_states=True)\n",
    "        print(outputs)\n",
    "        hidden = outputs.hidden_states    \n",
    "        print(hidden.shape)\n",
    "        # Stack second last token embeddings from all layers \n",
    "        # if len(hidden) == 1:  # FIX: runtime error for mistral-7b on bioasq\n",
    "        #     sec_last_input = hidden[0]\n",
    "        # elif ((n_generated - 2) >= len(hidden)):\n",
    "        #     sec_last_input = hidden[-2]\n",
    "        # else:\n",
    "        #     sec_last_input = hidden[n_generated - 2]\n",
    "        # sec_last_token_embedding = torch.stack([layer[:, -1, :].cpu() for layer in sec_last_input])\n",
    "        # print(sec_last_token_embedding.shape)\n",
    "    last_hidden_state = outputs.hidden_states[-1][:, -1, :].cpu().numpy()\n",
    "    print(last_hidden_state.shape)  \n",
    "    # TODO potentially need to only compute uncertainty for the last token in sentence?\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# concat hidden states\n",
    "\n",
    "\n",
    "hidden_states = np.concatenate(np.array(hidden_states)[layer_range], axis=1)\n",
    "# predict with probe\n",
    "pred = probe.predict(hidden_states)\n",
    "print(pred)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "llm-test",
   "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": "3.11.4"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}