main.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "True"
      ]
     },
     "execution_count": 1,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import torch\n",
    "\n",
    "torch.cuda.is_available()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "import glob\n",
    "import math\n",
    "import sys\n",
    "import time\n",
    "from pathlib import Path\n",
    "from typing import Optional, Tuple, Union\n",
    "\n",
    "import lightning as L\n",
    "import torch\n",
    "from lightning.fabric.loggers import CSVLogger\n",
    "from lightning.fabric.strategies import FSDPStrategy\n",
    "from torch.utils.data import DataLoader\n",
    "\n",
    "# # support running without installing as a package\n",
    "# wd = Path(__file__).parent.parent.resolve()\n",
    "# sys.path.append(str(wd))\n",
    "\n",
    "from tsai_gpt.model import GPT, Block, Config\n",
    "from tsai_gpt.packed_dataset import CombinedDataset, PackedDataset\n",
    "from tsai_gpt.speed_monitor import SpeedMonitorBase, estimate_flops, measure_flops\n",
    "from tsai_gpt.speed_monitor import SpeedMonitorFabric as SpeedMonitor\n",
    "from tsai_gpt.utils import (\n",
    "    chunked_cross_entropy,\n",
    "    get_default_supported_precision,\n",
    "    num_parameters,\n",
    "    load_checkpoint,\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_name = \"pythia-160m\"\n",
    "name = \"redpajama\"\n",
    "out_dir = Path(\"out\") / name\n",
    "save_interval = 1000\n",
    "eval_interval = 1000\n",
    "eval_iters = 100\n",
    "log_interval = 100"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Hyperparameters\n",
    "learning_rate = 6e-3\n",
    "batch_size = 32\n",
    "micro_batch_size = 8\n",
    "gradient_accumulation_steps = batch_size // micro_batch_size\n",
    "assert gradient_accumulation_steps > 0\n",
    "# max_iters = 600000  # num_epochs * (epoch_size // micro_batch_size) // devices\n",
    "max_iters = 15000\n",
    "weight_decay = 1e-1\n",
    "beta1 = 0.9\n",
    "beta2 = 0.95\n",
    "grad_clip = 1.0\n",
    "decay_lr = True\n",
    "warmup_iters = 2000\n",
    "lr_decay_iters = max_iters\n",
    "min_lr = 6e-6"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Data proportions from https://arxiv.org/pdf/2302.13971.pdf Table 1\n",
    "data_config = [\n",
    "    (\"arxiv\", 2.5),\n",
    "    (\"book\", 4.5),\n",
    "    (\"c4\", 15.0),\n",
    "    (\"cc\", 67.0),\n",
    "    (\"github\", 4.5),\n",
    "    (\"stackexchange\", 2.0),\n",
    "    (\"wikipedia\", 4.5),\n",
    "]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "hparams = {\n",
    "    k: v\n",
    "    for k, v in locals().items()\n",
    "    if isinstance(v, (int, float, str)) and not k.startswith(\"_\")\n",
    "}\n",
    "logger = CSVLogger(\"out\", name, flush_logs_every_n_steps=log_interval)\n",
    "\n",
    "\n",
    "def setup(\n",
    "    devices: int = 4,\n",
    "    train_data_dir: Path = Path(\"data/redpajama_sample\"),\n",
    "    val_data_dir: Optional[Path] = None,\n",
    "    precision: Optional[str] = None,\n",
    "    resume: Union[bool, Path] = False,\n",
    ") -> None:\n",
    "    precision = precision or get_default_supported_precision(training=True)\n",
    "\n",
    "    if devices > 1:\n",
    "        strategy = FSDPStrategy(\n",
    "            auto_wrap_policy={Block},\n",
    "            activation_checkpointing_policy={Block},\n",
    "            state_dict_type=\"full\",\n",
    "            limit_all_gathers=True,\n",
    "            cpu_offload=False,\n",
    "        )\n",
    "    else:\n",
    "        strategy = \"auto\"\n",
    "\n",
    "    fabric = L.Fabric(\n",
    "        devices=devices, strategy=strategy, precision=precision, loggers=logger\n",
    "    )\n",
    "    fabric.print(hparams)\n",
    "    fabric.launch(main, train_data_dir, val_data_dir, resume)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "model_copy = None"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "def main(\n",
    "    fabric: L.Fabric,\n",
    "    train_data_dir: Path,\n",
    "    val_data_dir: Path,\n",
    "    resume: Union[bool, Path],\n",
    ") -> None:\n",
    "    global model_copy\n",
    "    speed_monitor = SpeedMonitor(fabric, window_size=50, time_unit=\"seconds\")\n",
    "\n",
    "    if fabric.global_rank == 0:\n",
    "        out_dir.mkdir(parents=True, exist_ok=True)\n",
    "\n",
    "    config = Config.from_name(model_name)\n",
    "\n",
    "    train_dataloader, val_dataloader = create_dataloaders(\n",
    "        batch_size=micro_batch_size,\n",
    "        block_size=config.block_size,\n",
    "        fabric=fabric,\n",
    "        train_data_dir=train_data_dir,\n",
    "        val_data_dir=val_data_dir,\n",
    "        seed=(1337 + fabric.global_rank),\n",
    "    )\n",
    "    if val_dataloader is None:\n",
    "        train_dataloader = fabric.setup_dataloaders(train_dataloader)\n",
    "    else:\n",
    "        train_dataloader, val_dataloader = fabric.setup_dataloaders(\n",
    "            train_dataloader, val_dataloader\n",
    "        )\n",
    "\n",
    "    fabric.seed_everything(1337)  # same seed for every process to init model (FSDP)\n",
    "\n",
    "    fabric.print(f\"Loading model with {config.__dict__}\")\n",
    "    t0 = time.perf_counter()\n",
    "    import torch\n",
    "    import torch.nn as nn\n",
    "\n",
    "    def _init_weights(module: nn.Module) -> None:\n",
    "        \"\"\"Meant to be used with `gpt.apply(gpt._init_weights)`.\"\"\"\n",
    "        if isinstance(module, nn.Linear):\n",
    "            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)\n",
    "            if module.bias is not None:\n",
    "                torch.nn.init.zeros_(module.bias)\n",
    "        elif isinstance(module, nn.Embedding):\n",
    "            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)\n",
    "\n",
    "    with fabric.init_module(empty_init=True):\n",
    "        model = GPT(config)\n",
    "        model.apply(_init_weights)\n",
    "    model.apply(_init_weights)\n",
    "\n",
    "    # checkpoint_path = Path(\"out/redpajama/iter-000999-ckpt.pth\")\n",
    "\n",
    "    # load_checkpoint(fabric, model, checkpoint_path)\n",
    "\n",
    "    # print(model.transformer.h[0].mlp.fc.weight)\n",
    "\n",
    "    fabric.print(f\"Time to instantiate model: {time.perf_counter() - t0:.02f} seconds.\")\n",
    "    fabric.print(f\"Total parameters {num_parameters(model):,}\")\n",
    "\n",
    "    model = fabric.setup(model)\n",
    "    optimizer = torch.optim.AdamW(\n",
    "        model.parameters(),\n",
    "        lr=learning_rate,\n",
    "        weight_decay=weight_decay,\n",
    "        betas=(beta1, beta2),\n",
    "        foreach=False,\n",
    "    )\n",
    "\n",
    "    # model_copy = model\n",
    "\n",
    "    optimizer = fabric.setup_optimizers(optimizer)\n",
    "\n",
    "    state = {\n",
    "        \"model\": model,\n",
    "        \"optimizer\": optimizer,\n",
    "        \"hparams\": hparams,\n",
    "        \"iter_num\": 0,\n",
    "        \"step_count\": 0,\n",
    "    }\n",
    "\n",
    "    if resume is True:\n",
    "        resume = max(out_dir.glob(\"*.pth\"), key=lambda p: int(p.name.split(\"-\")[1]))\n",
    "    if resume:\n",
    "        fabric.print(f\"Resuming training from {resume}\")\n",
    "        fabric.load(resume, state)\n",
    "\n",
    "    train_time = time.perf_counter()\n",
    "    train(fabric, state, train_dataloader, val_dataloader, speed_monitor)\n",
    "    fabric.print(f\"Training time: {(time.perf_counter()-train_time):.2f}s\")\n",
    "    if fabric.device.type == \"cuda\":\n",
    "        fabric.print(f\"Memory used: {torch.cuda.max_memory_allocated() / 1e9:.02f} GB\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "def train(\n",
    "    fabric: L.Fabric,\n",
    "    state: dict,\n",
    "    train_dataloader: DataLoader,\n",
    "    val_dataloader: DataLoader,\n",
    "    speed_monitor: SpeedMonitorBase,\n",
    ") -> None:\n",
    "    model = state[\"model\"]\n",
    "    optimizer = state[\"optimizer\"]\n",
    "\n",
    "    if val_dataloader is not None:\n",
    "        validate(fabric, model, val_dataloader)  # sanity check\n",
    "\n",
    "    with torch.device(\"meta\"):\n",
    "        meta_model = GPT(model.config)\n",
    "        # \"estimated\" is not as precise as \"measured\". Estimated is optimistic but widely used in the wild.\n",
    "        # When comparing MFU or FLOP numbers with other projects that use estimated FLOPs,\n",
    "        # consider passing `SpeedMonitor(flops_per_batch=estimated_flops)` instead\n",
    "        estimated_flops = estimate_flops(meta_model) * micro_batch_size\n",
    "        fabric.print(\n",
    "            f\"Estimated TFLOPs: {estimated_flops * fabric.world_size / 1e12:.2f}\"\n",
    "        )\n",
    "        x = torch.randint(0, 1, (micro_batch_size, model.max_seq_length))\n",
    "        measured_flops = measure_flops(meta_model, x)\n",
    "        fabric.print(\n",
    "            f\"Measured TFLOPs: {measured_flops * fabric.world_size / 1e12:.2f}\"\n",
    "        )\n",
    "        del meta_model, x\n",
    "\n",
    "    total_lengths = 0\n",
    "    total_t0 = time.perf_counter()\n",
    "\n",
    "    for state[\"iter_num\"], train_data in enumerate(train_dataloader, state[\"iter_num\"]):\n",
    "        if state[\"iter_num\"] >= max_iters:\n",
    "            checkpoint_path = out_dir / f\"iter-{state['iter_num']:06d}-ckpt.pth\"\n",
    "            fabric.print(f\"Saving checkpoint to {str(checkpoint_path)!r}\")\n",
    "            fabric.save(checkpoint_path, state)\n",
    "            break\n",
    "\n",
    "        # determine and set the learning rate for this iteration\n",
    "        lr = get_lr(state[\"iter_num\"]) if decay_lr else learning_rate\n",
    "        for param_group in optimizer.param_groups:\n",
    "            param_group[\"lr\"] = lr\n",
    "\n",
    "        iter_t0 = time.perf_counter()\n",
    "\n",
    "        input_ids = train_data[:, 0 : model.max_seq_length].contiguous()\n",
    "        targets = train_data[:, 1 : model.max_seq_length + 1].contiguous()\n",
    "\n",
    "        is_accumulating = (state[\"iter_num\"] + 1) % gradient_accumulation_steps != 0\n",
    "        with fabric.no_backward_sync(model, enabled=is_accumulating):\n",
    "            logits = model(input_ids)\n",
    "            loss = chunked_cross_entropy(logits, targets, chunk_size=0)\n",
    "            fabric.backward(loss / gradient_accumulation_steps)\n",
    "\n",
    "        # return\n",
    "\n",
    "        if not is_accumulating:\n",
    "            fabric.clip_gradients(model, optimizer, max_norm=grad_clip)\n",
    "            optimizer.step()\n",
    "            optimizer.zero_grad()\n",
    "            state[\"step_count\"] += 1\n",
    "\n",
    "        t1 = time.perf_counter()\n",
    "        total_lengths += input_ids.size(1)\n",
    "        speed_monitor.on_train_batch_end(\n",
    "            (state[\"iter_num\"] + 1) * micro_batch_size,\n",
    "            t1 - total_t0,\n",
    "            # this assumes that device FLOPs are the same and that all devices have the same batch size\n",
    "            fabric.world_size,\n",
    "            flops_per_batch=measured_flops,\n",
    "            lengths=total_lengths,\n",
    "        )\n",
    "        if state[\"iter_num\"] % log_interval == 0:\n",
    "            fabric.print(\n",
    "                f\"iter {state['iter_num']} step {state['step_count']}: loss {loss.item():.4f}, LR: {lr:.6f}, iter time:\"\n",
    "                f\" {(t1 - iter_t0) * 1000:.2f}ms{' (optimizer.step)' if not is_accumulating else ''}\"\n",
    "            )\n",
    "\n",
    "        if (\n",
    "            val_dataloader is not None\n",
    "            and not is_accumulating\n",
    "            and state[\"step_count\"] % eval_interval == 0\n",
    "        ):\n",
    "            t0 = time.perf_counter()\n",
    "            val_loss = validate(fabric, model, val_dataloader)\n",
    "            t1 = time.perf_counter() - t0\n",
    "            speed_monitor.eval_end(t1)\n",
    "            fabric.print(\n",
    "                f\"step {state['iter_num']}: val loss {val_loss.item():.4f}, val time: {t1 * 1000:.2f}ms\"\n",
    "            )\n",
    "            fabric.barrier()\n",
    "        if not is_accumulating and state[\"step_count\"] % save_interval == 0:\n",
    "            checkpoint_path = out_dir / f\"iter-{state['iter_num']:06d}-ckpt.pth\"\n",
    "            fabric.print(f\"Saving checkpoint to {str(checkpoint_path)!r}\")\n",
    "            fabric.save(checkpoint_path, state)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "@torch.inference_mode()\n",
    "def validate(\n",
    "    fabric: L.Fabric, model: torch.nn.Module, val_dataloader: DataLoader\n",
    ") -> torch.Tensor:\n",
    "    fabric.print(\"Validating ...\")\n",
    "    model.eval()\n",
    "\n",
    "    losses = torch.zeros(eval_iters, device=fabric.device)\n",
    "    for k, val_data in enumerate(val_dataloader):\n",
    "        input_ids = val_data[:, 0 : model.max_seq_length].contiguous()\n",
    "        targets = val_data[:, 1 : model.max_seq_length + 1].contiguous()\n",
    "        logits = model(input_ids)\n",
    "        losses[k] = chunked_cross_entropy(logits, targets, chunk_size=0)\n",
    "    out = losses.mean()\n",
    "\n",
    "    model.train()\n",
    "    return out"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "def create_dataloader(\n",
    "    batch_size: int,\n",
    "    block_size: int,\n",
    "    data_dir: Path,\n",
    "    fabric: L.Fabric,\n",
    "    shuffle: bool = True,\n",
    "    seed: int = 12345,\n",
    ") -> DataLoader:\n",
    "    datasets = []\n",
    "    for prefix, _ in data_config:\n",
    "        filenames = glob.glob(str(data_dir / f\"{prefix}*\"))\n",
    "        dataset = PackedDataset(\n",
    "            filenames,\n",
    "            n_chunks=4,\n",
    "            block_size=block_size,\n",
    "            shuffle=shuffle,\n",
    "            seed=seed,\n",
    "            num_processes=fabric.world_size,\n",
    "            process_rank=fabric.global_rank,\n",
    "        )\n",
    "        datasets.append(dataset)\n",
    "\n",
    "    if not datasets:\n",
    "        raise RuntimeError(\n",
    "            f\"No data found at {data_dir}. Make sure you ran prepare_redpajama.py to create the dataset.\"\n",
    "        )\n",
    "\n",
    "    weights = [weight for _, weight in data_config]\n",
    "    sum_weights = sum(weights)\n",
    "    weights = [el / sum_weights for el in weights]\n",
    "\n",
    "    combined_dataset = CombinedDataset(datasets=datasets, seed=seed, weights=weights)\n",
    "\n",
    "    return DataLoader(\n",
    "        combined_dataset, batch_size=batch_size, shuffle=False, pin_memory=True\n",
    "    )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "def create_dataloaders(\n",
    "    batch_size: int,\n",
    "    block_size: int,\n",
    "    fabric: L.Fabric,\n",
    "    train_data_dir: Path = Path(\"data/redpajama_sample\"),\n",
    "    val_data_dir: Optional[Path] = None,\n",
    "    seed: int = 12345,\n",
    ") -> Tuple[DataLoader, DataLoader]:\n",
    "    # Increase by one because we need the next word as well\n",
    "    effective_block_size = block_size + 1\n",
    "    train_dataloader = create_dataloader(\n",
    "        batch_size=batch_size,\n",
    "        block_size=effective_block_size,\n",
    "        fabric=fabric,\n",
    "        data_dir=train_data_dir,\n",
    "        shuffle=True,\n",
    "        seed=seed,\n",
    "    )\n",
    "    val_dataloader = (\n",
    "        create_dataloader(\n",
    "            batch_size=batch_size,\n",
    "            block_size=effective_block_size,\n",
    "            fabric=fabric,\n",
    "            data_dir=val_data_dir,\n",
    "            shuffle=False,\n",
    "            seed=seed,\n",
    "        )\n",
    "        if val_data_dir\n",
    "        else None\n",
    "    )\n",
    "    return train_dataloader, val_dataloader"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_lr(it: int) -> float:\n",
    "    # 1) linear warmup for warmup_iters steps\n",
    "    if it < warmup_iters:\n",
    "        return learning_rate * it / warmup_iters\n",
    "    # 2) if it > lr_decay_iters, return min learning rate\n",
    "    if it > lr_decay_iters:\n",
    "        return min_lr\n",
    "    # 3) in between, use cosine decay down to min learning rate\n",
    "    decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)\n",
    "    assert 0 <= decay_ratio <= 1\n",
    "    coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1\n",
    "    return min_lr + coeff * (learning_rate - min_lr)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using bfloat16 Automatic Mixed Precision (AMP)\n",
      "Seed set to 1337\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'model_name': 'pythia-160m', 'name': 'redpajama', 'save_interval': 1000, 'eval_interval': 1000, 'eval_iters': 100, 'log_interval': 100, 'learning_rate': 0.006, 'batch_size': 32, 'micro_batch_size': 8, 'gradient_accumulation_steps': 4, 'max_iters': 15000, 'weight_decay': 0.1, 'beta1': 0.9, 'beta2': 0.95, 'grad_clip': 1.0, 'decay_lr': True, 'warmup_iters': 2000, 'lr_decay_iters': 15000, 'min_lr': 6e-06}\n",
      "Loading model with {'name': 'pythia-160m', 'hf_config': {'org': 'EleutherAI', 'name': 'pythia-160m-deduped'}, 'block_size': 2048, 'vocab_size': 50254, 'padding_multiple': 128, 'padded_vocab_size': 50304, 'n_layer': 12, 'n_head': 12, 'n_embd': 768, 'rotary_percentage': 0.25, 'parallel_residual': True, 'bias': True, 'lm_head_bias': False, 'n_query_groups': 12, 'shared_attention_norm': False, '_norm_class': 'LayerNorm', 'norm_eps': 1e-05, '_mlp_class': 'GptNeoxMLP', 'gelu_approximate': 'none', 'intermediate_size': 3072, 'rope_condense_ratio': 1, 'rope_base': 10000, 'head_size': 64, 'rope_n_elem': 16}\n",
      "Time to instantiate model: 1.99 seconds.\n",
      "Total parameters 162,322,944\n",
      "Estimated TFLOPs: 22.14\n",
      "Measured TFLOPs: 15.86\n",
      "iter 0 step 0: loss 11.0478, LR: 0.000000, iter time: 1312.30ms\n",
      "iter 100 step 25: loss 7.3711, LR: 0.000300, iter time: 282.00ms\n",
      "iter 200 step 50: loss 5.9653, LR: 0.000600, iter time: 293.93ms\n",
      "iter 300 step 75: loss 6.1456, LR: 0.000900, iter time: 290.72ms\n",
      "iter 400 step 100: loss 6.4233, LR: 0.001200, iter time: 291.77ms\n",
      "iter 500 step 125: loss 5.8922, LR: 0.001500, iter time: 292.98ms\n",
      "iter 600 step 150: loss 5.7330, LR: 0.001800, iter time: 292.54ms\n",
      "iter 700 step 175: loss 5.2412, LR: 0.002100, iter time: 293.18ms\n",
      "iter 800 step 200: loss 4.7973, LR: 0.002400, iter time: 291.61ms\n",
      "iter 900 step 225: loss 5.4157, LR: 0.002700, iter time: 292.85ms\n",
      "iter 1000 step 250: loss 5.1732, LR: 0.003000, iter time: 292.74ms\n",
      "iter 1100 step 275: loss 5.1144, LR: 0.003300, iter time: 291.97ms\n",
      "iter 1200 step 300: loss 4.6204, LR: 0.003600, iter time: 291.41ms\n",
      "iter 1300 step 325: loss 5.2649, LR: 0.003900, iter time: 292.33ms\n",
      "iter 1400 step 350: loss 5.3906, LR: 0.004200, iter time: 291.61ms\n",
      "iter 1500 step 375: loss 5.1544, LR: 0.004500, iter time: 292.87ms\n",
      "iter 1600 step 400: loss 5.2281, LR: 0.004800, iter time: 291.19ms\n",
      "iter 1700 step 425: loss 4.6215, LR: 0.005100, iter time: 290.65ms\n",
      "iter 1800 step 450: loss 5.1470, LR: 0.005400, iter time: 291.07ms\n",
      "iter 1900 step 475: loss 5.1262, LR: 0.005700, iter time: 291.85ms\n",
      "iter 2000 step 500: loss 4.7982, LR: 0.006000, iter time: 291.74ms\n",
      "iter 2100 step 525: loss 4.7870, LR: 0.005999, iter time: 291.40ms\n",
      "iter 2200 step 550: loss 4.6758, LR: 0.005997, iter time: 291.24ms\n",
      "iter 2300 step 575: loss 4.2770, LR: 0.005992, iter time: 290.94ms\n",
      "iter 2400 step 600: loss 4.9993, LR: 0.005986, iter time: 290.82ms\n",
      "iter 2500 step 625: loss 4.7006, LR: 0.005978, iter time: 291.72ms\n",
      "iter 2600 step 650: loss 4.4606, LR: 0.005969, iter time: 291.41ms\n",
      "iter 2700 step 675: loss 4.2507, LR: 0.005957, iter time: 291.65ms\n",
      "iter 2800 step 700: loss 4.2737, LR: 0.005944, iter time: 298.98ms\n",
      "iter 2900 step 725: loss 3.2729, LR: 0.005929, iter time: 291.06ms\n",
      "iter 3000 step 750: loss 3.6851, LR: 0.005913, iter time: 290.95ms\n",
      "iter 3100 step 775: loss 4.3133, LR: 0.005895, iter time: 291.41ms\n",
      "iter 3200 step 800: loss 4.0082, LR: 0.005875, iter time: 290.55ms\n",
      "iter 3300 step 825: loss 4.4818, LR: 0.005853, iter time: 291.40ms\n",
      "iter 3400 step 850: loss 4.0966, LR: 0.005830, iter time: 291.75ms\n",
      "iter 3500 step 875: loss 3.3417, LR: 0.005805, iter time: 291.56ms\n",
      "iter 3600 step 900: loss 3.3930, LR: 0.005779, iter time: 291.98ms\n",
      "iter 3700 step 925: loss 3.9926, LR: 0.005751, iter time: 291.38ms\n",
      "iter 3800 step 950: loss 4.4130, LR: 0.005721, iter time: 290.98ms\n",
      "iter 3900 step 975: loss 4.2273, LR: 0.005690, iter time: 290.82ms\n",
      "Saving checkpoint to 'out/redpajama/iter-003999-ckpt.pth'\n",
      "iter 4000 step 1000: loss 4.1836, LR: 0.005657, iter time: 289.39ms\n",
      "iter 4100 step 1025: loss 3.8898, LR: 0.005622, iter time: 290.57ms\n",
      "iter 4200 step 1050: loss 3.2994, LR: 0.005586, iter time: 290.66ms\n",
      "iter 4300 step 1075: loss 3.5536, LR: 0.005549, iter time: 291.97ms\n",
      "iter 4400 step 1100: loss 4.0568, LR: 0.005510, iter time: 290.74ms\n",
      "iter 4500 step 1125: loss 4.0688, LR: 0.005469, iter time: 291.51ms\n",
      "iter 4600 step 1150: loss 3.9602, LR: 0.005428, iter time: 291.69ms\n",
      "iter 4700 step 1175: loss 3.9015, LR: 0.005384, iter time: 291.05ms\n",
      "iter 4800 step 1200: loss 3.9838, LR: 0.005340, iter time: 290.89ms\n",
      "iter 4900 step 1225: loss 4.1498, LR: 0.005294, iter time: 291.43ms\n",
      "iter 5000 step 1250: loss 3.9890, LR: 0.005246, iter time: 292.04ms\n",
      "iter 5100 step 1275: loss 3.7998, LR: 0.005198, iter time: 291.67ms\n",
      "iter 5200 step 1300: loss 4.3898, LR: 0.005148, iter time: 292.07ms\n",
      "iter 5300 step 1325: loss 3.8301, LR: 0.005096, iter time: 291.71ms\n",
      "iter 5400 step 1350: loss 3.9250, LR: 0.005044, iter time: 291.87ms\n",
      "iter 5500 step 1375: loss 3.4592, LR: 0.004990, iter time: 292.45ms\n",
      "iter 5600 step 1400: loss 3.9057, LR: 0.004936, iter time: 292.48ms\n",
      "iter 5700 step 1425: loss 3.4640, LR: 0.004880, iter time: 292.17ms\n",
      "iter 5800 step 1450: loss 3.5189, LR: 0.004823, iter time: 291.53ms\n",
      "iter 5900 step 1475: loss 3.8723, LR: 0.004765, iter time: 291.76ms\n",
      "iter 6000 step 1500: loss 3.5505, LR: 0.004705, iter time: 291.40ms\n",
      "iter 6100 step 1525: loss 2.7599, LR: 0.004645, iter time: 290.44ms\n",
      "iter 6200 step 1550: loss 4.0639, LR: 0.004584, iter time: 290.73ms\n",
      "iter 6300 step 1575: loss 3.9124, LR: 0.004522, iter time: 290.77ms\n",
      "iter 6400 step 1600: loss 3.7831, LR: 0.004459, iter time: 290.48ms\n",
      "iter 6500 step 1625: loss 3.6439, LR: 0.004396, iter time: 291.02ms\n",
      "iter 6600 step 1650: loss 3.6231, LR: 0.004331, iter time: 293.27ms\n",
      "iter 6700 step 1675: loss 3.4389, LR: 0.004266, iter time: 291.11ms\n",
      "iter 6800 step 1700: loss 3.5385, LR: 0.004200, iter time: 290.80ms\n",
      "iter 6900 step 1725: loss 3.4988, LR: 0.004133, iter time: 291.01ms\n",
      "iter 7000 step 1750: loss 3.8966, LR: 0.004066, iter time: 290.56ms\n",
      "iter 7100 step 1775: loss 3.6816, LR: 0.003998, iter time: 290.93ms\n",
      "iter 7200 step 1800: loss 3.4510, LR: 0.003929, iter time: 291.20ms\n",
      "iter 7300 step 1825: loss 3.9102, LR: 0.003860, iter time: 292.28ms\n",
      "iter 7400 step 1850: loss 3.6360, LR: 0.003790, iter time: 291.56ms\n",
      "iter 7500 step 1875: loss 3.8664, LR: 0.003720, iter time: 290.58ms\n",
      "iter 7600 step 1900: loss 3.6073, LR: 0.003650, iter time: 291.40ms\n",
      "iter 7700 step 1925: loss 2.9199, LR: 0.003579, iter time: 290.78ms\n",
      "iter 7800 step 1950: loss 2.7844, LR: 0.003508, iter time: 290.67ms\n",
      "iter 7900 step 1975: loss 3.1176, LR: 0.003436, iter time: 291.73ms\n",
      "Saving checkpoint to 'out/redpajama/iter-007999-ckpt.pth'\n",
      "iter 8000 step 2000: loss 3.7936, LR: 0.003364, iter time: 290.49ms\n",
      "iter 8100 step 2025: loss 3.6197, LR: 0.003292, iter time: 290.46ms\n",
      "iter 8200 step 2050: loss 3.7480, LR: 0.003220, iter time: 291.78ms\n",
      "iter 8300 step 2075: loss 3.6900, LR: 0.003148, iter time: 291.11ms\n",
      "iter 8400 step 2100: loss 2.8864, LR: 0.003075, iter time: 291.39ms\n",
      "iter 8500 step 2125: loss 3.6963, LR: 0.003003, iter time: 291.51ms\n",
      "iter 8600 step 2150: loss 3.7093, LR: 0.002931, iter time: 291.80ms\n",
      "iter 8700 step 2175: loss 3.3042, LR: 0.002858, iter time: 290.53ms\n",
      "iter 8800 step 2200: loss 3.0944, LR: 0.002786, iter time: 290.83ms\n",
      "iter 8900 step 2225: loss 3.4312, LR: 0.002714, iter time: 290.81ms\n",
      "iter 9000 step 2250: loss 3.5048, LR: 0.002642, iter time: 290.99ms\n",
      "iter 9100 step 2275: loss 3.2803, LR: 0.002570, iter time: 291.00ms\n",
      "iter 9200 step 2300: loss 3.5930, LR: 0.002498, iter time: 292.10ms\n",
      "iter 9300 step 2325: loss 2.2495, LR: 0.002427, iter time: 290.29ms\n",
      "iter 9400 step 2350: loss 2.9088, LR: 0.002356, iter time: 290.19ms\n",
      "iter 9500 step 2375: loss 2.6597, LR: 0.002286, iter time: 291.29ms\n",
      "iter 9600 step 2400: loss 3.6206, LR: 0.002216, iter time: 291.64ms\n",
      "iter 9700 step 2425: loss 2.3134, LR: 0.002146, iter time: 289.83ms\n",
      "iter 9800 step 2450: loss 2.4301, LR: 0.002077, iter time: 289.59ms\n",
      "iter 9900 step 2475: loss 2.4800, LR: 0.002008, iter time: 290.77ms\n",
      "iter 10000 step 2500: loss 2.2368, LR: 0.001940, iter time: 290.11ms\n",
      "iter 10100 step 2525: loss 3.1508, LR: 0.001873, iter time: 291.03ms\n",
      "iter 10200 step 2550: loss 3.2954, LR: 0.001806, iter time: 291.14ms\n",
      "iter 10300 step 2575: loss 3.0130, LR: 0.001740, iter time: 291.20ms\n",
      "iter 10400 step 2600: loss 3.0044, LR: 0.001675, iter time: 290.75ms\n",
      "iter 10500 step 2625: loss 2.8596, LR: 0.001610, iter time: 290.14ms\n",
      "iter 10600 step 2650: loss 2.0126, LR: 0.001547, iter time: 290.53ms\n",
      "iter 10700 step 2675: loss 3.0040, LR: 0.001484, iter time: 292.51ms\n",
      "iter 10800 step 2700: loss 3.4691, LR: 0.001422, iter time: 290.79ms\n",
      "iter 10900 step 2725: loss 3.3719, LR: 0.001361, iter time: 291.21ms\n",
      "iter 11000 step 2750: loss 2.9904, LR: 0.001301, iter time: 292.52ms\n",
      "iter 11100 step 2775: loss 2.7121, LR: 0.001241, iter time: 291.23ms\n",
      "iter 11200 step 2800: loss 3.2472, LR: 0.001183, iter time: 291.06ms\n",
      "iter 11300 step 2825: loss 3.3517, LR: 0.001126, iter time: 291.27ms\n",
      "iter 11400 step 2850: loss 3.2715, LR: 0.001070, iter time: 292.14ms\n",
      "iter 11500 step 2875: loss 3.4200, LR: 0.001016, iter time: 290.81ms\n",
      "iter 11600 step 2900: loss 3.4924, LR: 0.000962, iter time: 291.75ms\n",
      "iter 11700 step 2925: loss 2.2736, LR: 0.000910, iter time: 290.48ms\n",
      "iter 11800 step 2950: loss 3.1776, LR: 0.000858, iter time: 291.91ms\n",
      "iter 11900 step 2975: loss 3.1710, LR: 0.000808, iter time: 291.62ms\n",
      "Saving checkpoint to 'out/redpajama/iter-011999-ckpt.pth'\n",
      "iter 12000 step 3000: loss 3.6688, LR: 0.000760, iter time: 290.94ms\n",
      "iter 12100 step 3025: loss 3.0179, LR: 0.000712, iter time: 290.84ms\n",
      "iter 12200 step 3050: loss 3.2257, LR: 0.000666, iter time: 291.06ms\n",
      "iter 12300 step 3075: loss 3.1653, LR: 0.000622, iter time: 292.47ms\n",
      "iter 12400 step 3100: loss 3.4042, LR: 0.000578, iter time: 291.42ms\n",
      "iter 12500 step 3125: loss 3.1884, LR: 0.000537, iter time: 290.93ms\n",
      "iter 12600 step 3150: loss 3.4705, LR: 0.000496, iter time: 291.49ms\n",
      "iter 12700 step 3175: loss 3.5805, LR: 0.000457, iter time: 291.72ms\n",
      "iter 12800 step 3200: loss 2.8953, LR: 0.000420, iter time: 292.49ms\n",
      "iter 12900 step 3225: loss 3.3408, LR: 0.000384, iter time: 297.87ms\n",
      "iter 13000 step 3250: loss 3.0779, LR: 0.000349, iter time: 298.95ms\n",
      "iter 13100 step 3275: loss 2.5973, LR: 0.000316, iter time: 291.06ms\n",
      "iter 13200 step 3300: loss 3.5901, LR: 0.000285, iter time: 291.16ms\n",
      "iter 13300 step 3325: loss 2.4544, LR: 0.000255, iter time: 290.62ms\n",
      "iter 13400 step 3350: loss 2.9969, LR: 0.000227, iter time: 290.56ms\n",
      "iter 13500 step 3375: loss 3.1975, LR: 0.000201, iter time: 291.62ms\n",
      "iter 13600 step 3400: loss 2.8946, LR: 0.000176, iter time: 290.60ms\n",
      "iter 13700 step 3425: loss 3.4701, LR: 0.000153, iter time: 291.61ms\n",
      "iter 13800 step 3450: loss 2.6274, LR: 0.000131, iter time: 289.90ms\n",
      "iter 13900 step 3475: loss 3.3881, LR: 0.000111, iter time: 291.66ms\n",
      "iter 14000 step 3500: loss 3.0832, LR: 0.000093, iter time: 291.88ms\n",
      "iter 14100 step 3525: loss 3.2224, LR: 0.000077, iter time: 291.17ms\n",
      "iter 14200 step 3550: loss 3.5854, LR: 0.000062, iter time: 290.77ms\n",
      "iter 14300 step 3575: loss 3.3620, LR: 0.000049, iter time: 292.27ms\n",
      "iter 14400 step 3600: loss 3.5590, LR: 0.000037, iter time: 291.91ms\n",
      "iter 14500 step 3625: loss 3.2781, LR: 0.000028, iter time: 290.50ms\n",
      "iter 14600 step 3650: loss 3.4279, LR: 0.000020, iter time: 291.54ms\n",
      "iter 14700 step 3675: loss 2.8695, LR: 0.000014, iter time: 291.52ms\n",
      "iter 14800 step 3700: loss 2.8212, LR: 0.000009, iter time: 291.34ms\n",
      "iter 14900 step 3725: loss 3.3649, LR: 0.000007, iter time: 292.48ms\n",
      "Saving checkpoint to 'out/redpajama/iter-015000-ckpt.pth'\n",
      "Training time: 4615.15s\n",
      "Memory used: 21.58 GB\n"
     ]
    }
   ],
   "source": [
    "torch.set_float32_matmul_precision(\"medium\")\n",
    "setup(devices=1, train_data_dir=Path(\"data/lit-redpajama-sample\"))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "base",
   "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": "3.10.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}