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

12
colossalai/legacy/nn/layer/parallel_sequence/_operation.py
View File

@@ -5,11 +5,11 @@ import torch
from torch import distributed as dist
from torch.cuda.amp import custom_bwd, custom_fwd
from colossalai.accelerator import get_accelerator
from colossalai.legacy.communication import ring_forward
from colossalai.legacy.context.parallel_mode import ParallelMode
from colossalai.legacy.core import global_context as gpc
from colossalai.legacy.nn.layer.parallel_sequence._utils import _calc_current_device_range, _calc_incoming_device_range
from colossalai.utils import get_current_device
class RingQK(torch.autograd.Function):
@@ -30,7 +30,7 @@ class RingQK(torch.autograd.Function):
sub_seq_length,
sub_seq_length * gpc.get_world_size(ParallelMode.SEQUENCE),
dtype=sub_q.dtype,
device=get_current_device(),
device=get_accelerator().get_current_device(),
)
# compute local QK^T
@@ -71,7 +71,7 @@ class RingQK(torch.autograd.Function):
grad_q = torch.zeros_like(
sub_q,
dtype=sub_q.dtype,
device=get_current_device(),
device=get_accelerator().get_current_device(),
)
# compute with local sub_k
@@ -105,7 +105,7 @@ class RingAV(torch.autograd.Function):
batch_size * num_attention_heads,
sub_seq_length,
attention_head_size,
device=get_current_device(),
device=get_accelerator().get_current_device(),
dtype=attention_score.dtype,
)
@@ -142,7 +142,9 @@ class RingAV(torch.autograd.Function):
grad_v /= local_world_size
# calculate gradient for attention score
grad_attention_score = torch.zeros_like(attention_scores, dtype=grad_output.dtype, device=get_current_device())
grad_attention_score = torch.zeros_like(
attention_scores, dtype=grad_output.dtype, device=get_accelerator().get_current_device()
)
# compute with local sub_k
grad_attention_score[:, :, local_start_idx:local_end_idx] += torch.matmul(grad_output, sub_v.transpose(2, 1))