{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "e:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\tqdm\\auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
      "  from .autonotebook import tqdm as notebook_tqdm\n"
     ]
    }
   ],
   "source": [
    "import sys\n",
    "# import logging\n",
    "\n",
    "import datasets\n",
    "from datasets import load_dataset\n",
    "from peft import LoraConfig\n",
    "import torch\n",
    "import transformers\n",
    "from trl import SFTTrainer\n",
    "from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, BitsAndBytesConfig\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "training_config = {\n",
    "    \"bf16\": False,\n",
    "    \"do_eval\": False,\n",
    "    \"learning_rate\": 5.0e-06,\n",
    "    \"log_level\": \"info\",\n",
    "    \"logging_steps\": 50,\n",
    "    \"logging_strategy\": \"steps\",\n",
    "    \"lr_scheduler_type\": \"cosine\",\n",
    "    \"num_train_epochs\": 1,\n",
    "    \"max_steps\": -1,\n",
    "    \"output_dir\": \"./checkpoint_dir\",\n",
    "    \"overwrite_output_dir\": True,\n",
    "    \"per_device_eval_batch_size\": 4,  # Reduce batch size to lower memory usage\n",
    "    \"per_device_train_batch_size\": 8,  # Reduce batch size to lower memory usage\n",
    "    \"remove_unused_columns\": True,\n",
    "    \"save_steps\": 500,\n",
    "    \"save_total_limit\": 1,\n",
    "    \"seed\": 0,\n",
    "    \"gradient_checkpointing\": False,\n",
    "    \"gradient_checkpointing_kwargs\":{\"use_reentrant\": False},\n",
    "    \"gradient_accumulation_steps\": 1,\n",
    "    \"warmup_ratio\": 0.2,\n",
    "}\n",
    "\n",
    "peft_config = {\n",
    "    \"r\": 16,\n",
    "    \"lora_alpha\": 32,\n",
    "    \"lora_dropout\": 0.05,\n",
    "    \"bias\": \"none\",\n",
    "    \"task_type\": \"CAUSAL_LM\",\n",
    "    \"target_modules\": \"all-linear\",\n",
    "    \"modules_to_save\": None,\n",
    "}\n",
    "train_conf = TrainingArguments(**training_config)\n",
    "peft_conf = LoraConfig(**peft_config)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "You are attempting to use Flash Attention 2.0 with a model not initialized on GPU. Make sure to move the model to GPU after initializing it on CPU with `model.to('cuda')`.\n",
      "Loading checkpoint shards: 100%|██████████| 2/2 [05:34<00:00, 167.47s/it]\n",
      "Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "Phi3ForCausalLM(\n",
       "  (model): Phi3Model(\n",
       "    (embed_tokens): Embedding(32064, 3072, padding_idx=32000)\n",
       "    (embed_dropout): Dropout(p=0.0, inplace=False)\n",
       "    (layers): ModuleList(\n",
       "      (0-31): 32 x Phi3DecoderLayer(\n",
       "        (self_attn): Phi3FlashAttention2(\n",
       "          (o_proj): Linear(in_features=3072, out_features=3072, bias=False)\n",
       "          (qkv_proj): Linear(in_features=3072, out_features=9216, bias=False)\n",
       "          (rotary_emb): Phi3RotaryEmbedding()\n",
       "        )\n",
       "        (mlp): Phi3MLP(\n",
       "          (gate_up_proj): Linear(in_features=3072, out_features=16384, bias=False)\n",
       "          (down_proj): Linear(in_features=8192, out_features=3072, bias=False)\n",
       "          (activation_fn): SiLU()\n",
       "        )\n",
       "        (input_layernorm): Phi3RMSNorm()\n",
       "        (resid_attn_dropout): Dropout(p=0.0, inplace=False)\n",
       "        (resid_mlp_dropout): Dropout(p=0.0, inplace=False)\n",
       "        (post_attention_layernorm): Phi3RMSNorm()\n",
       "      )\n",
       "    )\n",
       "    (norm): Phi3RMSNorm()\n",
       "  )\n",
       "  (lm_head): Linear(in_features=3072, out_features=32064, bias=False)\n",
       ")"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "################\n",
    "# Model Loading\n",
    "################\n",
    "checkpoint_path = \"microsoft/Phi-3-mini-4k-instruct\"\n",
    "# checkpoint_path = \"microsoft/Phi-3-mini-128k-instruct\"\n",
    "model_kwargs = dict(\n",
    "    use_cache=False,\n",
    "    trust_remote_code=True,\n",
    "    attn_implementation=\"flash_attention_2\",  # loading the model with flash-attention support\n",
    "    torch_dtype=torch.float16,  # Changed to float16\n",
    "    device_map=None\n",
    ")\n",
    "model = AutoModelForCausalLM.from_pretrained(checkpoint_path, **model_kwargs)\n",
    "tokenizer = AutoTokenizer.from_pretrained(checkpoint_path)\n",
    "tokenizer.model_max_length = 2048\n",
    "tokenizer.pad_token = tokenizer.unk_token  # use unk rather than eos token to prevent endless generation\n",
    "tokenizer.pad_token_id = tokenizer.convert_tokens_to_ids(tokenizer.pad_token)\n",
    "tokenizer.padding_side = 'right'\n",
    "\n",
    "# Move the model to GPU\n",
    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
    "model.to(device)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "##################\n",
    "# Data Processing\n",
    "##################\n",
    "def apply_chat_template(example, tokenizer):\n",
    "    messages = example[\"messages\"]\n",
    "    # Add an empty system message if there is none\n",
    "    if messages[0][\"role\"] != \"system\":\n",
    "        messages.insert(0, {\"role\": \"system\", \"content\": \"\"})\n",
    "    example[\"text\"] = tokenizer.apply_chat_template(\n",
    "        messages, tokenize=False, add_generation_prompt=False)\n",
    "    return example\n",
    "\n",
    "raw_dataset = load_dataset(\"HuggingFaceH4/ultrachat_200k\")\n",
    "train_dataset = raw_dataset[\"train_sft\"]\n",
    "test_dataset = raw_dataset[\"test_sft\"]\n",
    "column_names = list(train_dataset.features)\n",
    "\n",
    "processed_train_dataset = train_dataset.map(\n",
    "    apply_chat_template,\n",
    "    fn_kwargs={\"tokenizer\": tokenizer},\n",
    "    num_proc=10,\n",
    "    remove_columns=column_names,\n",
    "    desc=\"Applying chat template to train_sft\",\n",
    ")\n",
    "\n",
    "processed_test_dataset = test_dataset.map(\n",
    "    apply_chat_template,\n",
    "    fn_kwargs={\"tokenizer\": tokenizer},\n",
    "    num_proc=10,\n",
    "    remove_columns=column_names,\n",
    "    desc=\"Applying chat template to test_sft\",\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "***** Running training *****\n",
      "  Num examples = 140,320\n",
      "  Num Epochs = 1\n",
      "  Instantaneous batch size per device = 8\n",
      "  Total train batch size (w. parallel, distributed & accumulation) = 8\n",
      "  Gradient Accumulation steps = 1\n",
      "  Total optimization steps = 17,540\n",
      "  Number of trainable parameters = 25,165,824\n",
      "  0%|          | 0/17540 [00:00<?, ?it/s]"
     ]
    },
    {
     "ename": "OutOfMemoryError",
     "evalue": "CUDA out of memory. Tried to allocate 288.00 MiB. GPU ",
     "output_type": "error",
     "traceback": [
      "\u001b[1;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[1;31mOutOfMemoryError\u001b[0m                          Traceback (most recent call last)",
      "Cell \u001b[1;32mIn[5], line 15\u001b[0m\n\u001b[0;32m      1\u001b[0m \u001b[38;5;66;03m###########\u001b[39;00m\n\u001b[0;32m      2\u001b[0m \u001b[38;5;66;03m# Training\u001b[39;00m\n\u001b[0;32m      3\u001b[0m \u001b[38;5;66;03m###########\u001b[39;00m\n\u001b[0;32m      4\u001b[0m trainer \u001b[38;5;241m=\u001b[39m SFTTrainer(\n\u001b[0;32m      5\u001b[0m     model\u001b[38;5;241m=\u001b[39mmodel,\n\u001b[0;32m      6\u001b[0m     args\u001b[38;5;241m=\u001b[39mtrain_conf,\n\u001b[1;32m   (...)\u001b[0m\n\u001b[0;32m     13\u001b[0m     packing\u001b[38;5;241m=\u001b[39m\u001b[38;5;28;01mTrue\u001b[39;00m\n\u001b[0;32m     14\u001b[0m )\n\u001b[1;32m---> 15\u001b[0m train_result \u001b[38;5;241m=\u001b[39m \u001b[43mtrainer\u001b[49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mtrain\u001b[49m\u001b[43m(\u001b[49m\u001b[43m)\u001b[49m\n\u001b[0;32m     16\u001b[0m metrics \u001b[38;5;241m=\u001b[39m train_result\u001b[38;5;241m.\u001b[39mmetrics\n\u001b[0;32m     17\u001b[0m trainer\u001b[38;5;241m.\u001b[39mlog_metrics(\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mtrain\u001b[39m\u001b[38;5;124m\"\u001b[39m, metrics)\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\trl\\trainer\\sft_trainer.py:361\u001b[0m, in \u001b[0;36mSFTTrainer.train\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m    358\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mneftune_noise_alpha \u001b[38;5;129;01mis\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m \u001b[38;5;28;01mNone\u001b[39;00m \u001b[38;5;129;01mand\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_trainer_supports_neftune:\n\u001b[0;32m    359\u001b[0m     \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mmodel \u001b[38;5;241m=\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_trl_activate_neftune(\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mmodel)\n\u001b[1;32m--> 361\u001b[0m output \u001b[38;5;241m=\u001b[39m \u001b[38;5;28msuper\u001b[39m()\u001b[38;5;241m.\u001b[39mtrain(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n\u001b[0;32m    363\u001b[0m \u001b[38;5;66;03m# After training we make sure to retrieve back the original forward pass method\u001b[39;00m\n\u001b[0;32m    364\u001b[0m \u001b[38;5;66;03m# for the embedding layer by removing the forward post hook.\u001b[39;00m\n\u001b[0;32m    365\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mneftune_noise_alpha \u001b[38;5;129;01mis\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m \u001b[38;5;28;01mNone\u001b[39;00m \u001b[38;5;129;01mand\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_trainer_supports_neftune:\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\transformers\\trainer.py:1885\u001b[0m, in \u001b[0;36mTrainer.train\u001b[1;34m(self, resume_from_checkpoint, trial, ignore_keys_for_eval, **kwargs)\u001b[0m\n\u001b[0;32m   1883\u001b[0m         hf_hub_utils\u001b[38;5;241m.\u001b[39menable_progress_bars()\n\u001b[0;32m   1884\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[1;32m-> 1885\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[43minner_training_loop\u001b[49m\u001b[43m(\u001b[49m\n\u001b[0;32m   1886\u001b[0m \u001b[43m        \u001b[49m\u001b[43margs\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43margs\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1887\u001b[0m \u001b[43m        \u001b[49m\u001b[43mresume_from_checkpoint\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mresume_from_checkpoint\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1888\u001b[0m \u001b[43m        \u001b[49m\u001b[43mtrial\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mtrial\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1889\u001b[0m \u001b[43m        \u001b[49m\u001b[43mignore_keys_for_eval\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mignore_keys_for_eval\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1890\u001b[0m \u001b[43m    \u001b[49m\u001b[43m)\u001b[49m\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\transformers\\trainer.py:2216\u001b[0m, in \u001b[0;36mTrainer._inner_training_loop\u001b[1;34m(self, batch_size, args, resume_from_checkpoint, trial, ignore_keys_for_eval)\u001b[0m\n\u001b[0;32m   2213\u001b[0m     \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mcontrol \u001b[38;5;241m=\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mcallback_handler\u001b[38;5;241m.\u001b[39mon_step_begin(args, \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mstate, \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mcontrol)\n\u001b[0;32m   2215\u001b[0m \u001b[38;5;28;01mwith\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39maccelerator\u001b[38;5;241m.\u001b[39maccumulate(model):\n\u001b[1;32m-> 2216\u001b[0m     tr_loss_step \u001b[38;5;241m=\u001b[39m \u001b[38;5;28;43mself\u001b[39;49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mtraining_step\u001b[49m\u001b[43m(\u001b[49m\u001b[43mmodel\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43minputs\u001b[49m\u001b[43m)\u001b[49m\n\u001b[0;32m   2218\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m (\n\u001b[0;32m   2219\u001b[0m     args\u001b[38;5;241m.\u001b[39mlogging_nan_inf_filter\n\u001b[0;32m   2220\u001b[0m     \u001b[38;5;129;01mand\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m is_torch_xla_available()\n\u001b[0;32m   2221\u001b[0m     \u001b[38;5;129;01mand\u001b[39;00m (torch\u001b[38;5;241m.\u001b[39misnan(tr_loss_step) \u001b[38;5;129;01mor\u001b[39;00m torch\u001b[38;5;241m.\u001b[39misinf(tr_loss_step))\n\u001b[0;32m   2222\u001b[0m ):\n\u001b[0;32m   2223\u001b[0m     \u001b[38;5;66;03m# if loss is nan or inf simply add the average of previous logged losses\u001b[39;00m\n\u001b[0;32m   2224\u001b[0m     tr_loss \u001b[38;5;241m+\u001b[39m\u001b[38;5;241m=\u001b[39m tr_loss \u001b[38;5;241m/\u001b[39m (\u001b[38;5;241m1\u001b[39m \u001b[38;5;241m+\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mstate\u001b[38;5;241m.\u001b[39mglobal_step \u001b[38;5;241m-\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_globalstep_last_logged)\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\transformers\\trainer.py:3238\u001b[0m, in \u001b[0;36mTrainer.training_step\u001b[1;34m(self, model, inputs)\u001b[0m\n\u001b[0;32m   3235\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m loss_mb\u001b[38;5;241m.\u001b[39mreduce_mean()\u001b[38;5;241m.\u001b[39mdetach()\u001b[38;5;241m.\u001b[39mto(\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39margs\u001b[38;5;241m.\u001b[39mdevice)\n\u001b[0;32m   3237\u001b[0m \u001b[38;5;28;01mwith\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mcompute_loss_context_manager():\n\u001b[1;32m-> 3238\u001b[0m     loss \u001b[38;5;241m=\u001b[39m \u001b[38;5;28;43mself\u001b[39;49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mcompute_loss\u001b[49m\u001b[43m(\u001b[49m\u001b[43mmodel\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[43minputs\u001b[49m\u001b[43m)\u001b[49m\n\u001b[0;32m   3240\u001b[0m \u001b[38;5;28;01mdel\u001b[39;00m inputs\n\u001b[0;32m   3241\u001b[0m torch\u001b[38;5;241m.\u001b[39mcuda\u001b[38;5;241m.\u001b[39mempty_cache()\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\transformers\\trainer.py:3264\u001b[0m, in \u001b[0;36mTrainer.compute_loss\u001b[1;34m(self, model, inputs, return_outputs)\u001b[0m\n\u001b[0;32m   3262\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[0;32m   3263\u001b[0m     labels \u001b[38;5;241m=\u001b[39m \u001b[38;5;28;01mNone\u001b[39;00m\n\u001b[1;32m-> 3264\u001b[0m outputs \u001b[38;5;241m=\u001b[39m model(\u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39minputs)\n\u001b[0;32m   3265\u001b[0m \u001b[38;5;66;03m# Save past state if it exists\u001b[39;00m\n\u001b[0;32m   3266\u001b[0m \u001b[38;5;66;03m# TODO: this needs to be fixed and made cleaner later.\u001b[39;00m\n\u001b[0;32m   3267\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39margs\u001b[38;5;241m.\u001b[39mpast_index \u001b[38;5;241m>\u001b[39m\u001b[38;5;241m=\u001b[39m \u001b[38;5;241m0\u001b[39m:\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\torch\\nn\\modules\\module.py:1532\u001b[0m, in \u001b[0;36mModule._wrapped_call_impl\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m   1530\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_compiled_call_impl(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)  \u001b[38;5;66;03m# type: ignore[misc]\u001b[39;00m\n\u001b[0;32m   1531\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[1;32m-> 1532\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_call_impl(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\torch\\nn\\modules\\module.py:1541\u001b[0m, in \u001b[0;36mModule._call_impl\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m   1536\u001b[0m \u001b[38;5;66;03m# If we don't have any hooks, we want to skip the rest of the logic in\u001b[39;00m\n\u001b[0;32m   1537\u001b[0m \u001b[38;5;66;03m# this function, and just call forward.\u001b[39;00m\n\u001b[0;32m   1538\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m (\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_backward_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_backward_pre_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_forward_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_forward_pre_hooks\n\u001b[0;32m   1539\u001b[0m         \u001b[38;5;129;01mor\u001b[39;00m _global_backward_pre_hooks \u001b[38;5;129;01mor\u001b[39;00m _global_backward_hooks\n\u001b[0;32m   1540\u001b[0m         \u001b[38;5;129;01mor\u001b[39;00m _global_forward_hooks \u001b[38;5;129;01mor\u001b[39;00m _global_forward_pre_hooks):\n\u001b[1;32m-> 1541\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m forward_call(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n\u001b[0;32m   1543\u001b[0m \u001b[38;5;28;01mtry\u001b[39;00m:\n\u001b[0;32m   1544\u001b[0m     result \u001b[38;5;241m=\u001b[39m \u001b[38;5;28;01mNone\u001b[39;00m\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\peft\\peft_model.py:1430\u001b[0m, in \u001b[0;36mPeftModelForCausalLM.forward\u001b[1;34m(self, input_ids, attention_mask, inputs_embeds, labels, output_attentions, output_hidden_states, return_dict, task_ids, **kwargs)\u001b[0m\n\u001b[0;32m   1428\u001b[0m     \u001b[38;5;28;01mwith\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_enable_peft_forward_hooks(\u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs):\n\u001b[0;32m   1429\u001b[0m         kwargs \u001b[38;5;241m=\u001b[39m {k: v \u001b[38;5;28;01mfor\u001b[39;00m k, v \u001b[38;5;129;01min\u001b[39;00m kwargs\u001b[38;5;241m.\u001b[39mitems() \u001b[38;5;28;01mif\u001b[39;00m k \u001b[38;5;129;01mnot\u001b[39;00m \u001b[38;5;129;01min\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mspecial_peft_forward_args}\n\u001b[1;32m-> 1430\u001b[0m         \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mbase_model(\n\u001b[0;32m   1431\u001b[0m             input_ids\u001b[38;5;241m=\u001b[39minput_ids,\n\u001b[0;32m   1432\u001b[0m             attention_mask\u001b[38;5;241m=\u001b[39mattention_mask,\n\u001b[0;32m   1433\u001b[0m             inputs_embeds\u001b[38;5;241m=\u001b[39minputs_embeds,\n\u001b[0;32m   1434\u001b[0m             labels\u001b[38;5;241m=\u001b[39mlabels,\n\u001b[0;32m   1435\u001b[0m             output_attentions\u001b[38;5;241m=\u001b[39moutput_attentions,\n\u001b[0;32m   1436\u001b[0m             output_hidden_states\u001b[38;5;241m=\u001b[39moutput_hidden_states,\n\u001b[0;32m   1437\u001b[0m             return_dict\u001b[38;5;241m=\u001b[39mreturn_dict,\n\u001b[0;32m   1438\u001b[0m             \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs,\n\u001b[0;32m   1439\u001b[0m         )\n\u001b[0;32m   1441\u001b[0m batch_size \u001b[38;5;241m=\u001b[39m _get_batch_size(input_ids, inputs_embeds)\n\u001b[0;32m   1442\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m attention_mask \u001b[38;5;129;01mis\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m \u001b[38;5;28;01mNone\u001b[39;00m:\n\u001b[0;32m   1443\u001b[0m     \u001b[38;5;66;03m# concat prompt attention mask\u001b[39;00m\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\torch\\nn\\modules\\module.py:1532\u001b[0m, in \u001b[0;36mModule._wrapped_call_impl\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m   1530\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_compiled_call_impl(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)  \u001b[38;5;66;03m# type: ignore[misc]\u001b[39;00m\n\u001b[0;32m   1531\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[1;32m-> 1532\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_call_impl(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\torch\\nn\\modules\\module.py:1541\u001b[0m, in \u001b[0;36mModule._call_impl\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m   1536\u001b[0m \u001b[38;5;66;03m# If we don't have any hooks, we want to skip the rest of the logic in\u001b[39;00m\n\u001b[0;32m   1537\u001b[0m \u001b[38;5;66;03m# this function, and just call forward.\u001b[39;00m\n\u001b[0;32m   1538\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m (\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_backward_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_backward_pre_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_forward_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_forward_pre_hooks\n\u001b[0;32m   1539\u001b[0m         \u001b[38;5;129;01mor\u001b[39;00m _global_backward_pre_hooks \u001b[38;5;129;01mor\u001b[39;00m _global_backward_hooks\n\u001b[0;32m   1540\u001b[0m         \u001b[38;5;129;01mor\u001b[39;00m _global_forward_hooks \u001b[38;5;129;01mor\u001b[39;00m _global_forward_pre_hooks):\n\u001b[1;32m-> 1541\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m forward_call(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n\u001b[0;32m   1543\u001b[0m \u001b[38;5;28;01mtry\u001b[39;00m:\n\u001b[0;32m   1544\u001b[0m     result \u001b[38;5;241m=\u001b[39m \u001b[38;5;28;01mNone\u001b[39;00m\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\peft\\tuners\\tuners_utils.py:179\u001b[0m, in \u001b[0;36mBaseTuner.forward\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m    178\u001b[0m \u001b[38;5;28;01mdef\u001b[39;00m \u001b[38;5;21mforward\u001b[39m(\u001b[38;5;28mself\u001b[39m, \u001b[38;5;241m*\u001b[39margs: Any, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs: Any):\n\u001b[1;32m--> 179\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mmodel\u001b[38;5;241m.\u001b[39mforward(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n",
      "File \u001b[1;32m~\\.cache\\huggingface\\modules\\transformers_modules\\Phi-3-mini-4k-instruct\\modeling_phi3.py:1286\u001b[0m, in \u001b[0;36mPhi3ForCausalLM.forward\u001b[1;34m(self, input_ids, attention_mask, position_ids, past_key_values, inputs_embeds, labels, use_cache, output_attentions, output_hidden_states, return_dict)\u001b[0m\n\u001b[0;32m   1283\u001b[0m return_dict \u001b[38;5;241m=\u001b[39m return_dict \u001b[38;5;28;01mif\u001b[39;00m return_dict \u001b[38;5;129;01mis\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m \u001b[38;5;28;01mNone\u001b[39;00m \u001b[38;5;28;01melse\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mconfig\u001b[38;5;241m.\u001b[39muse_return_dict\n\u001b[0;32m   1285\u001b[0m \u001b[38;5;66;03m# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)\u001b[39;00m\n\u001b[1;32m-> 1286\u001b[0m outputs \u001b[38;5;241m=\u001b[39m \u001b[38;5;28;43mself\u001b[39;49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mmodel\u001b[49m\u001b[43m(\u001b[49m\n\u001b[0;32m   1287\u001b[0m \u001b[43m    \u001b[49m\u001b[43minput_ids\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43minput_ids\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1288\u001b[0m \u001b[43m    \u001b[49m\u001b[43mattention_mask\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mattention_mask\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1289\u001b[0m \u001b[43m    \u001b[49m\u001b[43mposition_ids\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mposition_ids\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1290\u001b[0m \u001b[43m    \u001b[49m\u001b[43mpast_key_values\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mpast_key_values\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1291\u001b[0m \u001b[43m    \u001b[49m\u001b[43minputs_embeds\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43minputs_embeds\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1292\u001b[0m \u001b[43m    \u001b[49m\u001b[43muse_cache\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43muse_cache\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1293\u001b[0m \u001b[43m    \u001b[49m\u001b[43moutput_attentions\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43moutput_attentions\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1294\u001b[0m \u001b[43m    \u001b[49m\u001b[43moutput_hidden_states\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43moutput_hidden_states\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1295\u001b[0m \u001b[43m    \u001b[49m\u001b[43mreturn_dict\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mreturn_dict\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1296\u001b[0m \u001b[43m\u001b[49m\u001b[43m)\u001b[49m\n\u001b[0;32m   1298\u001b[0m hidden_states \u001b[38;5;241m=\u001b[39m outputs[\u001b[38;5;241m0\u001b[39m]\n\u001b[0;32m   1299\u001b[0m logits \u001b[38;5;241m=\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mlm_head(hidden_states)\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\torch\\nn\\modules\\module.py:1532\u001b[0m, in \u001b[0;36mModule._wrapped_call_impl\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m   1530\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_compiled_call_impl(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)  \u001b[38;5;66;03m# type: ignore[misc]\u001b[39;00m\n\u001b[0;32m   1531\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[1;32m-> 1532\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_call_impl(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\torch\\nn\\modules\\module.py:1541\u001b[0m, in \u001b[0;36mModule._call_impl\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m   1536\u001b[0m \u001b[38;5;66;03m# If we don't have any hooks, we want to skip the rest of the logic in\u001b[39;00m\n\u001b[0;32m   1537\u001b[0m \u001b[38;5;66;03m# this function, and just call forward.\u001b[39;00m\n\u001b[0;32m   1538\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m (\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_backward_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_backward_pre_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_forward_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_forward_pre_hooks\n\u001b[0;32m   1539\u001b[0m         \u001b[38;5;129;01mor\u001b[39;00m _global_backward_pre_hooks \u001b[38;5;129;01mor\u001b[39;00m _global_backward_hooks\n\u001b[0;32m   1540\u001b[0m         \u001b[38;5;129;01mor\u001b[39;00m _global_forward_hooks \u001b[38;5;129;01mor\u001b[39;00m _global_forward_pre_hooks):\n\u001b[1;32m-> 1541\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m forward_call(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n\u001b[0;32m   1543\u001b[0m \u001b[38;5;28;01mtry\u001b[39;00m:\n\u001b[0;32m   1544\u001b[0m     result \u001b[38;5;241m=\u001b[39m \u001b[38;5;28;01mNone\u001b[39;00m\n",
      "File \u001b[1;32m~\\.cache\\huggingface\\modules\\transformers_modules\\Phi-3-mini-4k-instruct\\modeling_phi3.py:1164\u001b[0m, in \u001b[0;36mPhi3Model.forward\u001b[1;34m(self, input_ids, attention_mask, position_ids, past_key_values, inputs_embeds, use_cache, output_attentions, output_hidden_states, return_dict)\u001b[0m\n\u001b[0;32m   1154\u001b[0m     layer_outputs \u001b[38;5;241m=\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_gradient_checkpointing_func(\n\u001b[0;32m   1155\u001b[0m         decoder_layer\u001b[38;5;241m.\u001b[39m\u001b[38;5;21m__call__\u001b[39m,\n\u001b[0;32m   1156\u001b[0m         hidden_states,\n\u001b[1;32m   (...)\u001b[0m\n\u001b[0;32m   1161\u001b[0m         use_cache,\n\u001b[0;32m   1162\u001b[0m     )\n\u001b[0;32m   1163\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[1;32m-> 1164\u001b[0m     layer_outputs \u001b[38;5;241m=\u001b[39m \u001b[43mdecoder_layer\u001b[49m\u001b[43m(\u001b[49m\n\u001b[0;32m   1165\u001b[0m \u001b[43m        \u001b[49m\u001b[43mhidden_states\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1166\u001b[0m \u001b[43m        \u001b[49m\u001b[43mattention_mask\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mattention_mask\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1167\u001b[0m \u001b[43m        \u001b[49m\u001b[43mposition_ids\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mposition_ids\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1168\u001b[0m \u001b[43m        \u001b[49m\u001b[43mpast_key_value\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mpast_key_values\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1169\u001b[0m \u001b[43m        \u001b[49m\u001b[43moutput_attentions\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43moutput_attentions\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1170\u001b[0m \u001b[43m        \u001b[49m\u001b[43muse_cache\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43muse_cache\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m   1171\u001b[0m \u001b[43m    \u001b[49m\u001b[43m)\u001b[49m\n\u001b[0;32m   1173\u001b[0m hidden_states \u001b[38;5;241m=\u001b[39m layer_outputs[\u001b[38;5;241m0\u001b[39m]\n\u001b[0;32m   1175\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m use_cache:\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\torch\\nn\\modules\\module.py:1532\u001b[0m, in \u001b[0;36mModule._wrapped_call_impl\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m   1530\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_compiled_call_impl(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)  \u001b[38;5;66;03m# type: ignore[misc]\u001b[39;00m\n\u001b[0;32m   1531\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[1;32m-> 1532\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_call_impl(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\torch\\nn\\modules\\module.py:1541\u001b[0m, in \u001b[0;36mModule._call_impl\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m   1536\u001b[0m \u001b[38;5;66;03m# If we don't have any hooks, we want to skip the rest of the logic in\u001b[39;00m\n\u001b[0;32m   1537\u001b[0m \u001b[38;5;66;03m# this function, and just call forward.\u001b[39;00m\n\u001b[0;32m   1538\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m (\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_backward_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_backward_pre_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_forward_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_forward_pre_hooks\n\u001b[0;32m   1539\u001b[0m         \u001b[38;5;129;01mor\u001b[39;00m _global_backward_pre_hooks \u001b[38;5;129;01mor\u001b[39;00m _global_backward_hooks\n\u001b[0;32m   1540\u001b[0m         \u001b[38;5;129;01mor\u001b[39;00m _global_forward_hooks \u001b[38;5;129;01mor\u001b[39;00m _global_forward_pre_hooks):\n\u001b[1;32m-> 1541\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m forward_call(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n\u001b[0;32m   1543\u001b[0m \u001b[38;5;28;01mtry\u001b[39;00m:\n\u001b[0;32m   1544\u001b[0m     result \u001b[38;5;241m=\u001b[39m \u001b[38;5;28;01mNone\u001b[39;00m\n",
      "File \u001b[1;32m~\\.cache\\huggingface\\modules\\transformers_modules\\Phi-3-mini-4k-instruct\\modeling_phi3.py:885\u001b[0m, in \u001b[0;36mPhi3DecoderLayer.forward\u001b[1;34m(self, hidden_states, attention_mask, position_ids, past_key_value, output_attentions, use_cache, **kwargs)\u001b[0m\n\u001b[0;32m    882\u001b[0m hidden_states \u001b[38;5;241m=\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39minput_layernorm(hidden_states)\n\u001b[0;32m    884\u001b[0m \u001b[38;5;66;03m# Self Attention\u001b[39;00m\n\u001b[1;32m--> 885\u001b[0m attn_outputs, self_attn_weights, present_key_value \u001b[38;5;241m=\u001b[39m \u001b[38;5;28;43mself\u001b[39;49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mself_attn\u001b[49m\u001b[43m(\u001b[49m\n\u001b[0;32m    886\u001b[0m \u001b[43m    \u001b[49m\u001b[43mhidden_states\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mhidden_states\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m    887\u001b[0m \u001b[43m    \u001b[49m\u001b[43mattention_mask\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mattention_mask\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m    888\u001b[0m \u001b[43m    \u001b[49m\u001b[43mposition_ids\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mposition_ids\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m    889\u001b[0m \u001b[43m    \u001b[49m\u001b[43mpast_key_value\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43mpast_key_value\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m    890\u001b[0m \u001b[43m    \u001b[49m\u001b[43moutput_attentions\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43moutput_attentions\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m    891\u001b[0m \u001b[43m    \u001b[49m\u001b[43muse_cache\u001b[49m\u001b[38;5;241;43m=\u001b[39;49m\u001b[43muse_cache\u001b[49m\u001b[43m,\u001b[49m\n\u001b[0;32m    892\u001b[0m \u001b[43m\u001b[49m\u001b[43m)\u001b[49m\n\u001b[0;32m    894\u001b[0m hidden_states \u001b[38;5;241m=\u001b[39m residual \u001b[38;5;241m+\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mresid_attn_dropout(attn_outputs)\n\u001b[0;32m    896\u001b[0m residual \u001b[38;5;241m=\u001b[39m hidden_states\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\torch\\nn\\modules\\module.py:1532\u001b[0m, in \u001b[0;36mModule._wrapped_call_impl\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m   1530\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_compiled_call_impl(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)  \u001b[38;5;66;03m# type: ignore[misc]\u001b[39;00m\n\u001b[0;32m   1531\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[1;32m-> 1532\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_call_impl(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\torch\\nn\\modules\\module.py:1541\u001b[0m, in \u001b[0;36mModule._call_impl\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m   1536\u001b[0m \u001b[38;5;66;03m# If we don't have any hooks, we want to skip the rest of the logic in\u001b[39;00m\n\u001b[0;32m   1537\u001b[0m \u001b[38;5;66;03m# this function, and just call forward.\u001b[39;00m\n\u001b[0;32m   1538\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m (\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_backward_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_backward_pre_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_forward_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_forward_pre_hooks\n\u001b[0;32m   1539\u001b[0m         \u001b[38;5;129;01mor\u001b[39;00m _global_backward_pre_hooks \u001b[38;5;129;01mor\u001b[39;00m _global_backward_hooks\n\u001b[0;32m   1540\u001b[0m         \u001b[38;5;129;01mor\u001b[39;00m _global_forward_hooks \u001b[38;5;129;01mor\u001b[39;00m _global_forward_pre_hooks):\n\u001b[1;32m-> 1541\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m forward_call(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n\u001b[0;32m   1543\u001b[0m \u001b[38;5;28;01mtry\u001b[39;00m:\n\u001b[0;32m   1544\u001b[0m     result \u001b[38;5;241m=\u001b[39m \u001b[38;5;28;01mNone\u001b[39;00m\n",
      "File \u001b[1;32m~\\.cache\\huggingface\\modules\\transformers_modules\\Phi-3-mini-4k-instruct\\modeling_phi3.py:473\u001b[0m, in \u001b[0;36mPhi3FlashAttention2.forward\u001b[1;34m(self, hidden_states, attention_mask, position_ids, past_key_value, output_attentions, use_cache, **kwargs)\u001b[0m\n\u001b[0;32m    469\u001b[0m     attention_mask \u001b[38;5;241m=\u001b[39m kwargs\u001b[38;5;241m.\u001b[39mpop(\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mpadding_mask\u001b[39m\u001b[38;5;124m\"\u001b[39m)\n\u001b[0;32m    471\u001b[0m bsz, q_len, _ \u001b[38;5;241m=\u001b[39m hidden_states\u001b[38;5;241m.\u001b[39msize()\n\u001b[1;32m--> 473\u001b[0m qkv \u001b[38;5;241m=\u001b[39m \u001b[38;5;28;43mself\u001b[39;49m\u001b[38;5;241;43m.\u001b[39;49m\u001b[43mqkv_proj\u001b[49m\u001b[43m(\u001b[49m\u001b[43mhidden_states\u001b[49m\u001b[43m)\u001b[49m\n\u001b[0;32m    474\u001b[0m query_pos \u001b[38;5;241m=\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mnum_heads \u001b[38;5;241m*\u001b[39m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39mhead_dim\n\u001b[0;32m    475\u001b[0m query_states \u001b[38;5;241m=\u001b[39m qkv[\u001b[38;5;241m.\u001b[39m\u001b[38;5;241m.\u001b[39m\u001b[38;5;241m.\u001b[39m, :query_pos]\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\torch\\nn\\modules\\module.py:1532\u001b[0m, in \u001b[0;36mModule._wrapped_call_impl\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m   1530\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_compiled_call_impl(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)  \u001b[38;5;66;03m# type: ignore[misc]\u001b[39;00m\n\u001b[0;32m   1531\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[1;32m-> 1532\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_call_impl(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\torch\\nn\\modules\\module.py:1541\u001b[0m, in \u001b[0;36mModule._call_impl\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m   1536\u001b[0m \u001b[38;5;66;03m# If we don't have any hooks, we want to skip the rest of the logic in\u001b[39;00m\n\u001b[0;32m   1537\u001b[0m \u001b[38;5;66;03m# this function, and just call forward.\u001b[39;00m\n\u001b[0;32m   1538\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m (\u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_backward_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_backward_pre_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_forward_hooks \u001b[38;5;129;01mor\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39m_forward_pre_hooks\n\u001b[0;32m   1539\u001b[0m         \u001b[38;5;129;01mor\u001b[39;00m _global_backward_pre_hooks \u001b[38;5;129;01mor\u001b[39;00m _global_backward_hooks\n\u001b[0;32m   1540\u001b[0m         \u001b[38;5;129;01mor\u001b[39;00m _global_forward_hooks \u001b[38;5;129;01mor\u001b[39;00m _global_forward_pre_hooks):\n\u001b[1;32m-> 1541\u001b[0m     \u001b[38;5;28;01mreturn\u001b[39;00m forward_call(\u001b[38;5;241m*\u001b[39margs, \u001b[38;5;241m*\u001b[39m\u001b[38;5;241m*\u001b[39mkwargs)\n\u001b[0;32m   1543\u001b[0m \u001b[38;5;28;01mtry\u001b[39;00m:\n\u001b[0;32m   1544\u001b[0m     result \u001b[38;5;241m=\u001b[39m \u001b[38;5;28;01mNone\u001b[39;00m\n",
      "File \u001b[1;32me:\\Users\\frink\\Documents\\GitHub\\LLM Things\\Phi-3-training-Low-Ram\\venv\\lib\\site-packages\\peft\\tuners\\lora\\layer.py:569\u001b[0m, in \u001b[0;36mLinear.forward\u001b[1;34m(self, x, *args, **kwargs)\u001b[0m\n\u001b[0;32m    566\u001b[0m x \u001b[38;5;241m=\u001b[39m x\u001b[38;5;241m.\u001b[39mto(lora_A\u001b[38;5;241m.\u001b[39mweight\u001b[38;5;241m.\u001b[39mdtype)\n\u001b[0;32m    568\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m \u001b[38;5;129;01mnot\u001b[39;00m \u001b[38;5;28mself\u001b[39m\u001b[38;5;241m.\u001b[39muse_dora[active_adapter]:\n\u001b[1;32m--> 569\u001b[0m     result \u001b[38;5;241m=\u001b[39m result \u001b[38;5;241m+\u001b[39m \u001b[43mlora_B\u001b[49m\u001b[43m(\u001b[49m\u001b[43mlora_A\u001b[49m\u001b[43m(\u001b[49m\u001b[43mdropout\u001b[49m\u001b[43m(\u001b[49m\u001b[43mx\u001b[49m\u001b[43m)\u001b[49m\u001b[43m)\u001b[49m\u001b[43m)\u001b[49m\u001b[43m \u001b[49m\u001b[38;5;241;43m*\u001b[39;49m\u001b[43m \u001b[49m\u001b[43mscaling\u001b[49m\n\u001b[0;32m    570\u001b[0m \u001b[38;5;28;01melse\u001b[39;00m:\n\u001b[0;32m    571\u001b[0m     x \u001b[38;5;241m=\u001b[39m dropout(x)\n",
      "\u001b[1;31mOutOfMemoryError\u001b[0m: CUDA out of memory. Tried to allocate 288.00 MiB. GPU "
     ]
    }
   ],
   "source": [
    "###########\n",
    "# Training\n",
    "###########\n",
    "trainer = SFTTrainer(\n",
    "    model=model,\n",
    "    args=train_conf,\n",
    "    peft_config=peft_conf,\n",
    "    train_dataset=processed_train_dataset,\n",
    "    eval_dataset=processed_test_dataset,\n",
    "    max_seq_length=2048,\n",
    "    dataset_text_field=\"text\",\n",
    "    tokenizer=tokenizer,\n",
    "    packing=True\n",
    ")\n",
    "train_result = trainer.train()\n",
    "metrics = train_result.metrics\n",
    "trainer.log_metrics(\"train\", metrics)\n",
    "trainer.save_metrics(\"train\", metrics)\n",
    "trainer.save_state()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "#############\n",
    "# Evaluation\n",
    "#############\n",
    "tokenizer.padding_side = 'left'\n",
    "metrics = trainer.evaluate()\n",
    "metrics[\"eval_samples\"] = len(processed_test_dataset)\n",
    "trainer.log_metrics(\"eval\", metrics)\n",
    "trainer.save_metrics(\"eval\", metrics)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "############\n",
    "# Save model\n",
    "############\n",
    "trainer.save_model(train_conf.output_dir)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "venv",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}