github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-16 06:30:41 +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

4
colossalai/pipeline/schedule/generate.py
View File

@@ -7,10 +7,10 @@ import torch.cuda
from torch.nn import Module
from torch.utils._pytree import tree_map
from colossalai.accelerator import get_accelerator
from colossalai.inference.engine.microbatch_manager import MicroBatchManager, Status
from colossalai.pipeline.p2p import PipelineP2PCommunication
from colossalai.pipeline.stage_manager import PipelineStageManager
from colossalai.utils.device import get_current_device
from ._utils import get_batch_size, get_micro_batch, model_forward, to_device
from .base import PipelineSchedule
@@ -86,7 +86,7 @@ class GenerateSchedule(PipelineSchedule):
"""
micro_batch = get_micro_batch(self.batch, self.microbatch_offset, self.microbatch_size)
self.microbatch_offset += self.microbatch_size
return tree_map(partial(to_device, device=get_current_device()), micro_batch)
return tree_map(partial(to_device, device=get_accelerator().get_current_device()), micro_batch)
def _prepare_inputs_for_interval_stage(self):
"""

6
colossalai/pipeline/schedule/interleaved_pp.py
View File

@@ -6,10 +6,10 @@ import torch.cuda
from torch.nn import Module
from torch.utils._pytree import tree_map
from colossalai.accelerator import get_accelerator
from colossalai.interface import OptimizerWrapper
from colossalai.pipeline.p2p import PipelineP2PCommunication
from colossalai.pipeline.stage_manager import PipelineStageManager
from colossalai.utils.device import get_current_device
from ._utils import detach, get_batch_size, get_micro_batch, merge_batch, model_forward, retain_grad, to_device
from .base import PipelineSchedule
@@ -56,7 +56,7 @@ class InterleavedSchedule(PipelineSchedule):
"""
micro_batch = get_micro_batch(self.batch, self.microbatch_offset[model_chunk_id], self.microbatch_size)
self.microbatch_offset[model_chunk_id] += self.microbatch_size
return tree_map(partial(to_device, device=get_current_device()), micro_batch)
return tree_map(partial(to_device, device=get_accelerator().get_current_device()), micro_batch)
def get_model_chunk_id(self, microbatch_id: int, forward: bool) -> int:
"""Helper method to get the model chunk ID given the iteration number.
@@ -292,7 +292,7 @@ class InterleavedSchedule(PipelineSchedule):
outputs = [] if return_outputs and self.stage_manager.is_last_stage() else None
if return_loss and self.stage_manager.is_last_stage():
accum_loss = torch.zeros(1, device=get_current_device())
accum_loss = torch.zeros(1, device=get_accelerator().get_current_device())
else:
accum_loss = None

6
colossalai/pipeline/schedule/one_f_one_b.py
View File

@@ -6,10 +6,10 @@ import torch.cuda
from torch.nn import Module
from torch.utils._pytree import tree_map
from colossalai.accelerator import get_accelerator
from colossalai.interface import ModelWrapper, OptimizerWrapper
from colossalai.pipeline.p2p import PipelineP2PCommunication
from colossalai.pipeline.stage_manager import PipelineStageManager
from colossalai.utils.device import get_current_device
from ._utils import (
detach,
@@ -80,7 +80,7 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
"""
micro_batch = get_micro_batch(self.batch, self.microbatch_offset, self.microbatch_size)
self.microbatch_offset += self.microbatch_size
return tree_map(partial(to_device, device=get_current_device()), micro_batch)
return tree_map(partial(to_device, device=get_accelerator().get_current_device()), micro_batch)
def recv_forward(self, prev_rank: int = None) -> Any:
"""Copy the forward output from the previous stage in pipeline as the input tensor of this stage.
@@ -297,7 +297,7 @@ class OneForwardOneBackwardSchedule(PipelineSchedule):
outputs = [] if return_outputs and self.stage_manager.is_last_stage() else None
if return_loss and self.stage_manager.is_last_stage():
accum_loss = torch.zeros(1, device=get_current_device())
accum_loss = torch.zeros(1, device=get_accelerator().get_current_device())
else:
accum_loss = None