trainer_slurm_fsdp_v1.py

# coding=utf-8
#
# Copyright 2020 Heinrich Heine University Duesseldorf
#
# Part of this code is based on the source code of BERT-DST
# (arXiv:1907.03040)
# Part of this code is based on the source code of Transformers
# (arXiv:1910.03771)
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import glob
import json
import logging
import os
import subprocess
import sys
from typing import Dict, Union

import hydra
import numpy as np
import torch
from fairscale.nn.data_parallel.fully_sharded_data_parallel import FullyShardedDataParallel as FullyShardedDDP
from fairscale.optim.grad_scaler import ShardedGradScaler
from omegaconf import DictConfig, OmegaConf
from torch import distributed as dist
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler)
from torch.utils.data.distributed import DistributedSampler
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm, trange
from transformers import (get_linear_schedule_with_warmup, AutoTokenizer, PreTrainedTokenizer)

from general_util.logger import setting_logger
from general_util.training_utils import batch_to_device, unwrap_model, set_seed, note_best_checkpoint, initialize_optimizer, \
    load_and_cache_examples

logger: logging.Logger

torch.backends.cuda.matmul.allow_tf32 = True


def save_model(model: Union[torch.nn.Module, FullyShardedDDP], cfg: DictConfig, output_dir: str, tokenizer: PreTrainedTokenizer = None):
    # Save model checkpoint.
    if cfg.local_rank != -1:
        state_dict = model.state_dict()
        if cfg.local_rank == 0:
            unwrap_model(model).save_pretrained(output_dir, state_dict=state_dict)
    else:
        model.save_pretrained(output_dir)

    # Save tokenizer and training args.
    if cfg.local_rank in [-1, 0]:
        if tokenizer is not None:
            tokenizer.save_pretrained(output_dir)
        OmegaConf.save(cfg, os.path.join(output_dir, "training_config.yaml"))
        logger.info("Saving model checkpoint to %s", output_dir)


def forward_step(model, inputs: Dict[str, torch.Tensor], cfg, scaler, return_outputs: bool = False):
    if cfg.fp16:
        with torch.cuda.amp.autocast(dtype=(torch.bfloat16 if getattr(cfg, "fp16_bfloat16", False) else torch.float16)):
            outputs = model(**inputs)
    else:
        outputs = model(**inputs)

    if isinstance(outputs, tuple):
        loss = outputs[0]
    else:
        loss = outputs["loss"]

    if cfg.n_gpu > 1:
        loss = loss.mean()  # mean() to average on multi-gpu parallel (not distributed) training
    if cfg.gradient_accumulation_steps > 1:
        loss = loss / cfg.gradient_accumulation_steps

    if cfg.fp16:
        scaler.scale(loss).backward()
    else:
        loss.backward()

    if return_outputs:
        return loss.item(), outputs

    return loss.item()


def train(cfg, train_dataset, model, tokenizer, continue_from_global_step=0):
    """ Train the model """
    if cfg.local_rank in [-1, 0]:
        _dir_splits = cfg.output_dir.split('/')
        _log_dir = '/'.join([_dir_splits[0], 'runs'] + _dir_splits[1:])
        tb_writer = SummaryWriter(log_dir=_log_dir)
        tb_helper = hydra.utils.instantiate(cfg.summary_helper,
                                            writer=tb_writer) if "summary_helper" in cfg and cfg.summary_helper else None
    else:
        tb_writer = None
        tb_helper = None

    cfg.train_batch_size = cfg.per_gpu_train_batch_size * max(1, cfg.n_gpu)
    train_sampler = RandomSampler(train_dataset) if cfg.local_rank == -1 else DistributedSampler(train_dataset)
    train_collator = hydra.utils.instantiate(cfg.collator) if "collator" in cfg and cfg.collator else None
    train_dataloader = DataLoader(dataset=train_dataset,
                                  sampler=train_sampler,
                                  batch_size=cfg.train_batch_size,
                                  collate_fn=train_collator,
                                  num_workers=cfg.num_workers,
                                  pin_memory=True,
                                  prefetch_factor=cfg.prefetch_factor)

    if "extended_vocab" in cfg and cfg.extended_vocab:
        logger.info(f"Extended extra vocab size: {cfg.extended_vocab}")
        model.resize_token_embeddings(model.config.vocab_size + cfg.extended_vocab)

    if "with_lightseq" in cfg and cfg.with_lightseq:
        logger.info(f"Enabling Lightseq.")

        from general_util.lightseq_utils import inject_ls_roberta_enc_layer

        inject_ls_roberta_enc_layer(model, cfg, model.config)

    if cfg.max_steps > 0:
        t_total = cfg.max_steps
        cfg.num_train_epochs = cfg.max_steps // (len(train_dataloader) // cfg.gradient_accumulation_steps) + 1
    else:
        t_total = len(train_dataloader) // cfg.gradient_accumulation_steps * cfg.num_train_epochs

    num_warmup_steps = int(t_total * cfg.warmup_proportion) if cfg.warmup_proportion else cfg.warmup_steps

    optimizer = scheduler = None
    # Prepare optimizer and schedule (linear warmup and decay)
    if cfg.local_rank == -1:
        optimizer = initialize_optimizer(cfg, model=model)
        scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=t_total)

    if cfg.fp16:
        if cfg.local_rank != -1:
            scaler = ShardedGradScaler()
        else:
            from torch.cuda.amp.grad_scaler import GradScaler

            scaler = GradScaler()
    else:
        scaler = None

    # Distributed training (should be after apex fp16 initialization)
    if cfg.local_rank != -1:
        model = hydra.utils.instantiate(cfg.fairscale_config, model=model, device=cfg.device)
        optimizer = initialize_optimizer(cfg, model=model)
        scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=t_total)

    logger.info(optimizer)

    # Train!
    logger.info("***** Running training *****")
    logger.info("  Num examples = %d", len(train_dataset))
    logger.info("  Num Epochs = %d", cfg.num_train_epochs)
    logger.info("  Instantaneous batch size per GPU = %d", cfg.per_gpu_train_batch_size)
    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
                cfg.train_batch_size * cfg.gradient_accumulation_steps * (dist.get_world_size() if cfg.local_rank != -1 else 1))
    logger.info("  Gradient Accumulation steps = %d", cfg.gradient_accumulation_steps)
    logger.info("  Total optimization steps = %d", t_total)
    logger.info("  Warmup steps = %d", num_warmup_steps)

    if continue_from_global_step > 0:
        logger.info("Fast forwarding to global step %d to resume training from latest checkpoint...", continue_from_global_step)

    global_step = 0
    tr_loss, logging_loss = 0.0, 0.0
    model.zero_grad()
    train_iterator = trange(int(cfg.num_train_epochs), desc="Epoch", disable=cfg.local_rank not in [-1, 0])
    set_seed(cfg)  # Added here for reproducibility (even between python 2 and 3)

    for epoch in train_iterator:
        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=cfg.local_rank not in [-1, 0], dynamic_ncols=True)
        if cfg.local_rank != -1:
            train_dataloader.sampler.set_epoch(epoch)

        for step, batch in enumerate(epoch_iterator):
            # If training is continued from a checkpoint, fast forward
            # to the state of that checkpoint.
            if global_step < continue_from_global_step:
                if (step + 1) % cfg.gradient_accumulation_steps == 0:
                    scheduler.step()  # Update learning rate schedule
                    global_step += 1
                continue

            model.train()
            batch = batch_to_device(batch, cfg.device)

            last_outputs = None
            if (step + 1) % cfg.gradient_accumulation_steps != 0 and cfg.local_rank != -1:
                # Avoid unnecessary DDP synchronization since there will be no backward pass on this example.
                with model.no_sync():
                    loss = forward_step(model, batch, cfg, scaler)
            else:
                loss, last_outputs = forward_step(model, batch, cfg, scaler, return_outputs=True)

            tr_loss += loss
            if (step + 1) % cfg.gradient_accumulation_steps == 0:
                if cfg.fp16:
                    scaler.unscale_(optimizer)

                if cfg.max_grad_norm and not ("optimizer" in cfg and cfg.optimizer and "lamb" in cfg.optimizer):
                    if hasattr(optimizer, "clip_grad_norm"):
                        optimizer.clip_grad_norm(cfg.max_grad_norm)
                    elif hasattr(model, "clip_grad_norm_"):
                        model.clip_grad_norm_(cfg.max_grad_norm)
                    else:
                        torch.nn.utils.clip_grad_norm_(model.parameters(), cfg.max_grad_norm)

                if cfg.fp16:
                    scaler.step(optimizer)
                    scaler.update()
                else:
                    optimizer.step()

                scheduler.step()  # Update learning rate schedule
                model.zero_grad(set_to_none=True)
                global_step += 1

                # Log metrics
                if cfg.local_rank in [-1, 0] and cfg.logging_steps > 0 and global_step % cfg.logging_steps == 0:
                    tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
                    tb_writer.add_scalar('loss', (tr_loss - logging_loss) / cfg.logging_steps, global_step)
                    logging_loss = tr_loss

                    if tb_helper:
                        tb_helper(step=global_step, last_batch=batch, last_outputs=last_outputs)

                # Save model checkpoint
                if cfg.save_steps > 0 and global_step % cfg.save_steps == 0:
                    output_dir = os.path.join(cfg.output_dir, 'checkpoint-{}'.format(global_step))
                    if cfg.local_rank in [-1, 0] and not os.path.exists(output_dir):
                        os.makedirs(output_dir)
                    save_model(model, cfg, output_dir, tokenizer)

                # Evaluation
                if cfg.evaluate_during_training and cfg.eval_steps > 0 and global_step % cfg.eval_steps == 0:
                    state_dict = model.state_dict()

                    if cfg.ddp_eval or cfg.local_rank in [-1, 0]:
                        results = evaluate(cfg, model, tokenizer, prefix=str(global_step), _split="dev")

                        if cfg.local_rank in [-1, 0]:
                            for key, value in results.items():
                                tb_writer.add_scalar(f"eval/{key}", value, global_step)

                            sub_path = os.path.join(cfg.output_dir, 'checkpoint-{}'.format(global_step))
                            flag = note_best_checkpoint(cfg, results, sub_path)
                            if cfg.save_best and flag:
                                if cfg.local_rank == 0:
                                    unwrap_model(model).save_pretrained(cfg.output_dir, state_dict=state_dict)
                                else:
                                    model.save_pretrained(cfg.output_dir)

                                tokenizer.save_pretrained(cfg.output_dir)
                                OmegaConf.save(cfg, os.path.join(cfg.output_dir, "training_config.yaml"))
                                logger.info("Saving best model checkpoint to %s", cfg.output_dir)

            if 0 < cfg.max_steps < global_step:
                epoch_iterator.close()
                break

        if 0 < cfg.max_steps < global_step:
            train_iterator.close()
            break

    if cfg.local_rank in [-1, 0]:
        tb_writer.close()

    return global_step, tr_loss / global_step


def evaluate(cfg, model, tokenizer: PreTrainedTokenizer, prefix="", _split="dev"):
    dataset = load_and_cache_examples(cfg, tokenizer, _split=_split)

    if cfg.local_rank in [-1, 0] and not os.path.exists(os.path.join(cfg.output_dir, prefix)):
        os.makedirs(os.path.join(cfg.output_dir, prefix))

    cfg.eval_batch_size = cfg.per_gpu_eval_batch_size
    if _split == 'dev' and cfg.ddp_eval and cfg.local_rank != -1:
        eval_sampler = DistributedSampler(dataset, shuffle=False)
    else:
        eval_sampler = SequentialSampler(dataset)  # Note that DistributedSampler samples randomly
    eval_collator = hydra.utils.instantiate(cfg.collator) if "collator" in cfg and cfg.collator else None
    eval_dataloader = DataLoader(dataset,
                                 sampler=eval_sampler,
                                 batch_size=cfg.eval_batch_size,
                                 collate_fn=eval_collator)

    single_model_gpu = unwrap_model(model)
    single_model_gpu.get_eval_log(reset=True)
    # Eval!
    torch.cuda.empty_cache()
    logger.info("***** Running evaluation {}.{} *****".format(_split, prefix))
    logger.info("  Num examples = %d", len(dataset))
    logger.info("  Batch size = %d", cfg.eval_batch_size)
    # Seems FSDP does not need to unwrap the model for evaluating.
    model.eval()
    pred_list = []
    prob_list = []
    for batch in tqdm(eval_dataloader, desc="Evaluating", disable=cfg.local_rank not in [-1, 0], dynamic_ncols=True):
        batch = batch_to_device(batch, cfg.device)
        if cfg.fp16:
            with torch.cuda.amp.autocast(dtype=(torch.bfloat16 if getattr(cfg, "fp16_bfloat16", False) else torch.float16)):
                with torch.no_grad():
                    outputs = model(**batch)
                    probs = outputs["logits"].softmax(dim=-1).detach().float().cpu()
                    # TODO: Doesn't work under ddp evaluation.
                    prob, pred = probs.max(dim=-1)
                    pred_list.extend(pred.tolist())
                    prob_list.extend(prob.tolist())
        else:
            with torch.no_grad():
                outputs = model(**batch)
                probs = outputs["logits"].softmax(dim=-1).detach().float().cpu()
                prob, pred = probs.max(dim=-1)
                pred_list.extend(pred.tolist())
                prob_list.extend(prob.tolist())

    metric_log, results = single_model_gpu.get_eval_log(reset=True, ddp=(_split == 'dev' and cfg.ddp_eval and cfg.local_rank != -1),
                                                        device=cfg.device)
    logger.info("****** Evaluation Results ******")
    logger.info(f"Global Steps: {prefix}")
    logger.info(metric_log)

    if cfg.local_rank in [-1, 0]:
        prediction_file = os.path.join(cfg.output_dir, prefix, "eval_predictions.npy")
        np.save(prediction_file, pred_list)
        json.dump(prob_list, open(os.path.join(cfg.output_dir, prefix, "eval_probs.json"), "w"))

    torch.cuda.empty_cache()

    return results


def setup_slurm_distributed(cfg: DictConfig, backend="nccl", port=None):
    """
    Most code are copied from https://github.com/BIGBALLON/distribuuuu/blob/master/tutorial/mnmc_ddp_slurm.py.
    """
    num_gpus = torch.cuda.device_count()
    if num_gpus <= 1 or cfg.no_cuda:
        cfg.local_rank = -1
        cfg.device = str(torch.device("cuda" if torch.cuda.is_available() else "cpu"))
        cfg.n_gpu = num_gpus
        cfg.ddp_eval = False
        return

    proc_id = int(os.environ["SLURM_PROCID"])
    n_tasks = int(os.environ["SLURM_NTASKS"])
    node_list = os.environ["SLURM_NODELIST"]

    torch.cuda.set_device(proc_id % num_gpus)

    addr = subprocess.getoutput(f"scontrol show hostname {node_list} | head -n1")
    # specify master port
    if port is not None:
        os.environ["MASTER_PORT"] = str(port)
    elif "MASTER_PORT" not in os.environ:
        os.environ["MASTER_PORT"] = "29500"
    if "MASTER_ADDR" not in os.environ:
        os.environ["MASTER_ADDR"] = addr

    os.environ["WORLD_SIZE"] = str(n_tasks)
    os.environ["LOCAL_RANK"] = str(proc_id % num_gpus)
    os.environ["RANK"] = str(proc_id)

    cfg.n_gpu = 1
    cfg.local_rank = int(os.environ["LOCAL_RANK"])
    # cfg.local_rank = int(os.environ["RANK"])
    cfg.world_size = int(os.environ["WORLD_SIZE"])
    cfg.device = str(torch.device("cuda", cfg.local_rank))

    dist.init_process_group(backend=backend, world_size=int(os.environ["WORLD_SIZE"]), rank=int(os.environ["RANK"]))

    # print(cfg.n_gpu, cfg.local_rank, cfg.world_size, cfg.device)
    # print(cfg.local_rank)
    cfg.local_rank = dist.get_rank()
    # print(cfg.local_rank)


@hydra.main(config_path="conf", config_name="config")
def main(cfg: DictConfig):
    setup_slurm_distributed(cfg)

    global logger
    logger = setting_logger(cfg.output_dir, local_rank=cfg.local_rank)
    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
                   cfg.local_rank, cfg.device, cfg.n_gpu, bool(cfg.local_rank != -1), cfg.fp16)

    # Set seed
    set_seed(cfg)

    # Load pre-trained model and tokenizer
    if cfg.local_rank not in [-1, 0]:
        dist.barrier()  # Make sure only the first process in distributed training will download model & vocab

    if cfg.pretrain:
        pretrain_state_dict = torch.load(cfg.pretrain, map_location='cpu')
    else:
        pretrain_state_dict = None

    tokenizer = AutoTokenizer.from_pretrained(cfg.model_name_or_path)
    model = hydra.utils.call(cfg.model, cfg.model_name_or_path, state_dict=pretrain_state_dict)

    if cfg.local_rank == 0:
        dist.barrier()  # Make sure only the first process in distributed training will download model & vocab

    if cfg.local_rank == -1:  # For FullyShardedDDP, place the model on cpu first.
        model.to(cfg.device)

    # logger.info("Training/evaluation parameters %s", OmegaConf.to_yaml(cfg))
    if cfg.local_rank in [-1, 0] and cfg.do_train:
        if not os.path.exists(cfg.output_dir):
            os.makedirs(cfg.output_dir)
        OmegaConf.save(cfg, os.path.join(cfg.output_dir, "training_config.yaml"))

    # Training
    if cfg.do_train:
        # TODO: Add option for continuously training from checkpoint.
        #  The operation should be introduced in ``train`` method since both the state dict
        #  of schedule and optimizer (and scaler, if any) should be loaded.
        # If output files already exists, assume to continue training from latest checkpoint (unless overwrite_output_dir is set)
        continue_from_global_step = 0  # If set to 0, start training from the beginning
        # if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
        #     checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/*/' + WEIGHTS_NAME, recursive=True)))
        #     if len(checkpoints) > 0:
        #         checkpoint = checkpoints[-1]
        #         logger.info("Resuming training from the latest checkpoint: %s", checkpoint)
        #         continue_from_global_step = int(checkpoint.split('-')[-1])
        #         model = model_class.from_pretrained(checkpoint)
        #         model.to(args.device)

        train_dataset = load_and_cache_examples(cfg, tokenizer, _split="train")

        if getattr(cfg, "do_preprocess", False):
            exit(0)

        global_step, tr_loss = train(cfg, train_dataset, model, tokenizer, continue_from_global_step)
        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)

    # Test
    results = {}
    if cfg.do_eval and cfg.local_rank in [-1, 0]:
        if cfg.local_rank == 0:
            cfg.ddp_eval = False  # Canceling distributed evaluation since other progresses have already existed.
            cfg.local_rank = -1
        checkpoints = [cfg.output_dir]
        if cfg.save_best:
            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
        elif cfg.prediction_cfg.best_checkpoint and os.path.exists(cfg.prediction_cfg.best_checkpoint):
            checkpoints = [cfg.prediction_cfg.best_checkpoint]
            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
        elif cfg.eval_sub_path:
            checkpoints = list(
                os.path.dirname(c) for c in
                sorted(glob.glob(cfg.output_dir + f"/{cfg.eval_sub_path}/" + "pytorch_model.bin", recursive=True))
            )
            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
        logger.info(" the following checkpoints: %s", checkpoints)
        for checkpoint in checkpoints:
            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
            prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""
            split = "dev"

            model = hydra.utils.call(cfg.model, checkpoint)
            model.to(cfg.device)

            if cfg.test_file:
                prefix = f'test' + (f'-{prefix}' if prefix != "" else "")
                split = "test"

            result = evaluate(cfg, model, tokenizer, prefix=prefix, _split=split)
            result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
            results.update(result)

    return results


if __name__ == "__main__":
    hydra_formatted_args = []
    # convert the cli params added by torch.distributed.launch into Hydra format
    for arg in sys.argv:
        if arg.startswith("--"):
            hydra_formatted_args.append(arg[len("--"):])
        else:
            hydra_formatted_args.append(arg)
    sys.argv = hydra_formatted_args

    main()