From 72a77ec2cb0bcb3c10a65a405fca5beebaa61a68 Mon Sep 17 00:00:00 2001 From: KungYork <30741085+KungYork@users.noreply.github.com> Date: Wed, 26 Jul 2023 20:04:49 +0800 Subject: [PATCH] Update trainer_adapter.py Update Kunlunxin bert trainer_adapter.py to fix time collecting bug under 1x1 scenario. --- training/kunlunxin/bert-pytorch/extern/trainer_adapter.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/training/kunlunxin/bert-pytorch/extern/trainer_adapter.py b/training/kunlunxin/bert-pytorch/extern/trainer_adapter.py index 418deb131..8b153739b 100644 --- a/training/kunlunxin/bert-pytorch/extern/trainer_adapter.py +++ b/training/kunlunxin/bert-pytorch/extern/trainer_adapter.py @@ -7,8 +7,9 @@ from torch.cuda.amp import GradScaler from torch.nn.parallel import DistributedDataParallel as DDP from torch.optim import Optimizer -from torch_xmlir.optimizer import FusedLAMB +import torch_xmlir +from torch_xmlir.optimizer import FusedLAMB import torch_xmlir.core.xpu_model as xm import utils @@ -62,7 +63,7 @@ def backward(step: int, optimizer: Optimizer, grad_scaler: GradScaler = None): loss.backward() - + torch_xmlir.xpu.xpu_synchronize() update_step = step % config.gradient_accumulation_steps == 0 if update_step: update_model_params(loss, optimizer, grad_scaler)