github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-09 04:50:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

[npu] change device to accelerator api (#5239)

Browse Source 
* update accelerator

* fix timer

* fix amp

* update

* fix

* update bug

* add error raise

* fix autocast

* fix set device

* remove doc accelerator

* update doc

* update doc

* update doc

* use nullcontext

* update cpu

* update null context

* change time limit for example

* udpate

* update

* update

* update

* [npu] polish accelerator code

---------

Co-authored-by: Xuanlei Zhao <xuanlei.zhao@gmail.com>
Co-authored-by: zxl <43881818+oahzxl@users.noreply.github.com>
This commit is contained in:
Hongxin Liu
2024-01-09 10:20:05 +08:00
committed by GitHub
parent dd2c28a323
commit d202cc28c0
128 changed files with 1773 additions and 868 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
colossalai/legacy/engine/schedule/_pipeline_schedule.py
View File

@@ -7,12 +7,12 @@ from typing import Callable, List, Tuple, Union
import torch.cuda
import colossalai.legacy.communication as comm
from colossalai.accelerator import get_accelerator
from colossalai.legacy.amp.naive_amp import NaiveAMPModel
from colossalai.legacy.context.parallel_mode import ParallelMode
from colossalai.legacy.core import global_context as gpc
from colossalai.legacy.utils import switch_virtual_pipeline_parallel_rank
from colossalai.logging import get_dist_logger
from colossalai.utils.device import get_current_device
from ._base_schedule import BaseSchedule
@@ -352,7 +352,7 @@ class PipelineSchedule(BaseSchedule):
output_objs = []
return_tensors = []
if return_loss and gpc.is_pipeline_last_stage(ignore_virtual=True):
accum_loss = torch.zeros(1, device=get_current_device())
accum_loss = torch.zeros(1, device=get_accelerator().get_current_device())
else:
accum_loss = None
# Used for tensor meta information communication
@@ -584,7 +584,7 @@ class InterleavedPipelineSchedule(PipelineSchedule):
if not forward_only:
output_obj_grads = [[] for _ in range(len(model))]
if return_loss and gpc.is_pipeline_last_stage(ignore_virtual=True):
accum_loss = torch.zeros(1, device=get_current_device())
accum_loss = torch.zeros(1, device=get_accelerator().get_current_device())
else:
accum_loss = None