Dev gzf deepspeed (#1844)

zhifu gao

2024-06-24 abb33d6b2097e5b0643326bc1b376a63cdc2f967

 funasr/train_utils/trainer.py

@@ -85,7 +85,12 @@
        self.batch_total = 0
        self.use_fp16 = use_fp16
        self.save_checkpoint_interval = kwargs.get("save_checkpoint_interval", 5000)
        self.validate_interval = kwargs.get("validate_interval", 5000)
        self.validate_interval = kwargs.get("validate_interval", -1)
        if self.validate_interval < 0:
            self.validate_interval = self.save_checkpoint_interval
        assert (
            self.save_checkpoint_interval == self.validate_interval
        ), f"save_checkpoint_interval must equal to validate_interval"
        self.keep_nbest_models = kwargs.get("keep_nbest_models", 500)
        self.avg_keep_nbest_models_type = kwargs.get("avg_keep_nbest_models_type", "acc")
        self.avg_nbest_model = kwargs.get("avg_nbest_model", 10)
@@ -116,6 +121,7 @@
        self.reset_gpu_cache = kwargs.get("reset_gpu_cache", False)
        self.start_data_split_i = 0
        self.start_step = 0
        self.step_in_epoch = 0
        self.use_wandb = kwargs.get("use_wandb", False)
        if self.use_wandb:
            wandb.login(key=kwargs.get("wandb_token"))
@@ -137,6 +143,8 @@
        optim=None,
        scheduler=None,
        scaler=None,
        step_in_epoch=None,
        **kwargs,
    ):
        """
        Saves a checkpoint containing the model's state, the optimizer's state,
@@ -147,6 +155,7 @@
            epoch (int): The epoch number at which the checkpoint is being saved.
        """

        step_in_epoch = None if step is None else step_in_epoch
        if self.rank == 0:
            logging.info(f"Save checkpoint: {epoch}, rank: {self.local_rank}\n")
            # self.step_or_epoch += 1
@@ -161,7 +170,14 @@
                "best_step_or_epoch": self.best_step_or_epoch,
                "avg_keep_nbest_models_type": self.avg_keep_nbest_models_type,
                "step": step,
                "step_in_epoch": step_in_epoch,
                "data_split_i": kwargs.get("data_split_i", 0),
                "data_split_num": kwargs.get("data_split_num", 1),
                "batch_total": self.batch_total,
                "train_loss_avg": kwargs.get("train_loss_avg", 0),
                "train_acc_avg": kwargs.get("train_acc_avg", 0),
            }
            step = step_in_epoch
            if hasattr(model, "module"):
                state["state_dict"] = model.module.state_dict()

@@ -195,7 +211,7 @@
                    )
                else:
                    logging.info(
                        f"No improvement in acc: {self.val_acc_step_or_eoch[ckpt_name]:.4f} < {self.val_acc_step_or_eoch[self.best_step_or_epoch]:.4f}"
                        f"No improvement in acc: {self.val_acc_step_or_eoch[ckpt_name]:.4f} < {self.val_acc_step_or_eoch[self.best_step_or_epoch]:.4f}, {os.path.join(self.output_dir, self.best_step_or_epoch)}"
                    )
            elif self.avg_keep_nbest_models_type == "loss":
                if (
@@ -210,7 +226,7 @@
                    )
                else:
                    logging.info(
                        f"No improvement in loss: {self.val_loss_step_or_eoch[ckpt_name]:.4f} > {self.val_loss_step_or_eoch[self.best_step_or_epoch]:.4f}"
                        f"No improvement in loss: {self.val_loss_step_or_eoch[ckpt_name]:.4f} > {self.val_loss_step_or_eoch[self.best_step_or_epoch]:.4f}, {os.path.join(self.output_dir, self.best_step_or_epoch)}"
                    )
            else:
                print("Undo")
@@ -251,7 +267,7 @@
            ckpt = os.path.join(self.output_dir, "model.pt")
            if os.path.isfile(ckpt):
                checkpoint = torch.load(ckpt, map_location="cpu")
                self.start_epoch = checkpoint["epoch"] + 1
                self.start_epoch = checkpoint["epoch"]
                # self.model.load_state_dict(checkpoint['state_dict'])
                src_state = checkpoint["state_dict"]
                dst_state = model.state_dict()
@@ -288,12 +304,22 @@
                    checkpoint["best_step_or_epoch"] if "best_step_or_epoch" in checkpoint else ""
                )
                self.start_data_split_i = (
                    checkpoint["start_data_split_i"] if "start_data_split_i" in checkpoint else 0
                    checkpoint["data_split_i"] if "data_split_i" in checkpoint else 0
                )
                self.batch_total = checkpoint["batch_total"] if "batch_total" in checkpoint else 0
                self.start_step = checkpoint["step"] if "step" in checkpoint else 0
                self.start_step = 0 if self.start_step is None else self.start_step

                self.step_in_epoch = (
                    checkpoint["step_in_epoch"] if "step_in_epoch" in checkpoint else 0
                )
                self.step_in_epoch = 0 if self.step_in_epoch is None else self.step_in_epoch
                print(checkpoint["train_acc_avg"])
                self.train_acc_avg = (
                    checkpoint["train_acc_avg"] if "train_acc_avg" in checkpoint else 0
                )
                self.train_loss_avg = (
                    checkpoint["train_loss_avg"] if "train_loss_avg" in checkpoint else 0
                )
                model.to(self.device)
                print(f"Checkpoint loaded successfully from '{ckpt}'")
            else:
@@ -321,7 +347,7 @@
        """
        if self.use_ddp or self.use_fsdp:
            dist.barrier()
        logging.info(f"Train epoch: {epoch}, rank: {self.local_rank}\n")
        logging.info(f"Train epoch: {epoch}, rank: {self.rank}\n")
        model.train()

        # Set the number of steps for gradient accumulation
@@ -336,11 +362,12 @@
        time_beg = time.perf_counter()
        time5 = time_beg
        for batch_idx, batch in enumerate(dataloader_train):
            if self.use_ddp or self.use_fsdp:
                dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
                if iterator_stop > 0:
                    break
            # if self.use_ddp or self.use_fsdp:
            #     dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
            #     if iterator_stop > 0:
            #         break
            self.batch_total += 1
            self.step_in_epoch += 1
            time1 = time.perf_counter()
            speed_stats["data_load"] = f"{time1-time_beg:0.3f}"

@@ -354,14 +381,12 @@
                with maybe_autocast(self.use_fp16):
                    retval = model(**batch)

                    if (
                        self.reset_gpu_cache
                        and (torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024) > 70
                    ):
                        torch.cuda.empty_cache()
                    # if (
                    #     self.reset_gpu_cache
                    #     and (torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024) > 70
                    # ):
                    #     torch.cuda.empty_cache()

                time3 = time.perf_counter()
                speed_stats["forward_time"] = f"{time3 - time2:0.3f}"
                loss, stats, weight = retval
                stats = {k: v for k, v in stats.items() if v is not None}
                if self.use_ddp or self.use_fsdp:
@@ -376,33 +401,28 @@
                    # Multiply world_size because DistributedDataParallel
                    # automatically normalizes the gradient by world_size.
                    loss *= self.world_size
                # loss *= self.world_size
                # Scale the loss since we're not updating for every mini-batch
                loss = loss / accum_grad

                time3 = time.perf_counter()
                speed_stats["forward_time"] = f"{time3 - time2:0.3f}"
                if self.use_fp16:
                    scaler.scale(loss).backward()
                else:
                    loss.backward()
                time4 = time.perf_counter()
                speed_stats["backward_time"] = f"{time4 - time3:0.3f}"
                speed_stats["backward_and_AllReaduce_time"] = f"{time4 - time3:0.3f}"

                self.train_loss_avg = (
                    self.train_loss_avg * batch_idx + loss.detach().cpu().item()
                ) / (batch_idx + 1)
                    self.train_loss_avg * (batch_idx + kwargs.get("start_step", 0))
                    + loss.detach().cpu().item()
                ) / (batch_idx + kwargs.get("start_step", 0) + 1)
                if "acc" in stats:
                    self.train_acc_avg = (
                        self.train_acc_avg * batch_idx + stats["acc"].detach().cpu().item()
                    ) / (batch_idx + 1)
                if self.use_ddp or self.use_fsdp:
                    train_loss_avg = torch.tensor(self.train_loss_avg, dtype=torch.float32).to(
                        self.device
                    )
                    train_acc_avg = torch.tensor(self.train_acc_avg, dtype=torch.float32).to(
                        self.device
                    )
                    dist.all_reduce(train_loss_avg, op=dist.ReduceOp.SUM)
                    dist.all_reduce(train_acc_avg, op=dist.ReduceOp.SUM)
                    self.train_loss_avg = train_loss_avg.detach().cpu().item() / self.world_size
                    self.train_acc_avg = train_acc_avg.detach().cpu().item() / self.world_size
                        self.train_acc_avg * (batch_idx + kwargs.get("start_step", 0))
                        + stats["acc"].detach().cpu().item()
                    ) / (batch_idx + kwargs.get("start_step", 0) + 1)

            # Perform an optimizer step only after accumulating enough gradients
            if (batch_idx + 1) % accum_grad == 0:
@@ -431,8 +451,22 @@
                scheduler.step()
                # Clear gradients for the next accumulation stage
                optim.zero_grad(set_to_none=True)
                total_time = f"{time.perf_counter() - time5:0.3f}"

                if self.use_ddp or self.use_fsdp:
                    train_loss_avg = torch.tensor(self.train_loss_avg, dtype=torch.float32).to(
                        self.device
                    )
                    train_acc_avg = torch.tensor(self.train_acc_avg, dtype=torch.float32).to(
                        self.device
                    )
                    dist.all_reduce(train_loss_avg, op=dist.ReduceOp.SUM)
                    dist.all_reduce(train_acc_avg, op=dist.ReduceOp.SUM)
                    self.train_loss_avg = train_loss_avg.detach().cpu().item() / self.world_size
                    self.train_acc_avg = train_acc_avg.detach().cpu().item() / self.world_size

                total_time = f"{(time.perf_counter() - time5)/accum_grad:0.3f}"
                time5 = time.perf_counter()

                speed_stats["optim_time"] = f"{time5 - time4:0.3f}"

                speed_stats["total_time"] = total_time
@@ -443,9 +477,11 @@
                self.log(
                    epoch,
                    batch_idx,
                    log_step=batch_idx + kwargs.get("start_step", 0),
                    step_in_epoch=self.step_in_epoch,
                    batch_num_epoch=batch_num_epoch,
                    lr=lr,
                    loss=loss.detach().cpu().item(),
                    loss=accum_grad * loss.detach().cpu().item(),
                    speed_stats=speed_stats,
                    stats=stats,
                    writer=writer,
@@ -454,16 +490,17 @@
                    data_split_num=kwargs.get("data_split_num", 1),
                )

            if (batch_idx + 1) % self.validate_interval == 0:
            if self.step_in_epoch % self.validate_interval == 0:
                self.validate_epoch(
                    model=model,
                    dataloader_val=dataloader_val,
                    epoch=epoch,
                    writer=writer,
                    step=batch_idx + 1,
                    step_in_epoch=self.step_in_epoch,
                )

            if (batch_idx + 1) % self.save_checkpoint_interval == 0:
            if self.step_in_epoch % self.save_checkpoint_interval == 0:
                self.save_checkpoint(
                    epoch,
                    model=model,
@@ -471,17 +508,22 @@
                    scheduler=scheduler,
                    scaler=scaler,
                    step=batch_idx + 1,
                    step_in_epoch=self.step_in_epoch,
                    data_split_i=kwargs.get("data_split_i", 0),
                    data_split_num=kwargs.get("data_split_num", 1),
                    train_loss_avg=self.train_loss_avg,
                    train_acc_avg=self.train_acc_avg,
                )

            time_beg = time.perf_counter()
        else:
            if self.use_ddp or self.use_fsdp:
                iterator_stop.fill_(1)
                dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)
        # else:
        #     if self.use_ddp or self.use_fsdp:
        #         iterator_stop.fill_(1)
        #         dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)

        if self.use_ddp or self.use_fsdp:
            dist.barrier()
            iterator_stop = torch.tensor(0).to(self.device)
            # iterator_stop = torch.tensor(0).to(self.device)

    def validate_epoch(
        self,
@@ -500,7 +542,7 @@
        """
        if self.use_ddp or self.use_fsdp:
            dist.barrier()
        logging.info(f"Validate epoch: {epoch}, rank: {self.local_rank}\n")
        logging.info(f"Validate epoch: {epoch}, rank: {self.rank}\n")
        model.eval()

        with torch.no_grad():
@@ -578,10 +620,10 @@
                    iterator_stop.fill_(1)
                    dist.all_reduce(iterator_stop, dist.ReduceOp.SUM)

        if kwargs.get("step", None) is None:
        if kwargs.get("step_in_epoch", None) is None:
            ckpt_name = f"model.pt.ep{epoch}"
        else:
            ckpt_name = f'model.pt.ep{epoch}.{kwargs.get("step")}'
            ckpt_name = f'model.pt.ep{epoch}.{kwargs.get("step_in_epoch")}'
        self.val_acc_step_or_eoch[ckpt_name] = self.val_acc_avg
        self.val_loss_step_or_eoch[ckpt_name] = self.val_loss_avg
        model.train()
@@ -594,6 +636,7 @@
        self,
        epoch=0,
        batch_idx=0,
        step_in_epoch=0,
        batch_num_epoch=-1,
        lr=0.0,
        loss=0.0,
@@ -603,11 +646,12 @@
        tag="train",
        data_split_i=0,
        data_split_num=1,
        log_step=None,
        **kwargs,
    ):

        if (batch_idx + 1) % self.log_interval == 0:

            batch_idx = log_step if log_step is not None else batch_idx
            gpu_info = (
                "GPU, memory: usage: {:.3f} GB, "
                "peak: {:.3f} GB, "
@@ -627,11 +671,11 @@
                f"rank: {self.rank}, "
                f"epoch: {epoch}/{self.max_epoch}, "
                f"data_slice: {data_split_i}/{data_split_num}, "
                f"step: {batch_idx + 1}/{batch_num_epoch}, total step: {self.batch_total}, "
                f"step_in_slice: {batch_idx + 1}/{batch_num_epoch}, step_in_epoch: {step_in_epoch}, total step: {self.batch_total}, "
                f"(loss_avg_rank: {loss:.3f}), "
                f"(loss_avg_epoch: {loss_avg_epoch:.3f}), "
                f"(ppl_avg_epoch: {math.exp(loss_avg_epoch):.3e}), "
                f"(acc_avg_epoch: {acc_avg_epoch:.3f}), "
                f"(loss_avg_slice: {loss_avg_epoch:.3f}), "
                f"(ppl_avg_slice: {math.exp(loss_avg_epoch):.3e}), "
                f"(acc_avg_slice: {acc_avg_epoch:.3f}), "
                f"(lr: {lr:.3e}), "
                f"{[(k, round(v.detach().cpu().item(), 3)) for k, v in stats.items()]}, "
                f"{speed_stats}, "