Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
PaddlePaddle
Paddle
提交
e83f5f33
P
Paddle
项目概览
PaddlePaddle
/
Paddle
1 年多 前同步成功
通知
2302
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
未验证
提交
e83f5f33
编写于
12月 09, 2022
作者:
姜
姜永久
提交者:
GitHub
12月 09, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
remove xpu eager guard tests (#48786)
上级
25dafc58
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
176 addition
and
192 deletion
+176
-192
python/paddle/fluid/tests/unittests/xpu/parallel_dygraph_gradient_check_in_eager_mode.py
...ests/xpu/parallel_dygraph_gradient_check_in_eager_mode.py
+45
-47
python/paddle/fluid/tests/unittests/xpu/process_group_bkcl.py
...on/paddle/fluid/tests/unittests/xpu/process_group_bkcl.py
+131
-145
未找到文件。
python/paddle/fluid/tests/unittests/xpu/parallel_dygraph_gradient_check_in_eager_mode.py
浏览文件 @
e83f5f33
...
@@ -19,7 +19,6 @@ import numpy as np
...
@@ -19,7 +19,6 @@ import numpy as np
import
paddle
import
paddle
import
paddle.distributed
as
dist
import
paddle.distributed
as
dist
import
paddle.fluid
as
fluid
import
paddle.fluid
as
fluid
from
paddle.fluid.framework
import
_test_eager_guard
from
paddle.nn
import
Linear
from
paddle.nn
import
Linear
paddle
.
seed
(
1024
)
paddle
.
seed
(
1024
)
...
@@ -69,58 +68,57 @@ class SimpleNet(fluid.Layer):
...
@@ -69,58 +68,57 @@ class SimpleNet(fluid.Layer):
class
TestDistTraning
(
unittest
.
TestCase
):
class
TestDistTraning
(
unittest
.
TestCase
):
def
test_multiple_xpus
(
self
):
def
test_multiple_xpus
(
self
):
self
.
trainer_id
=
dist
.
get_rank
()
self
.
trainer_id
=
dist
.
get_rank
()
with
_test_eager_guard
():
self
.
pg
=
dist
.
init_parallel_env
()
self
.
pg
=
dist
.
init_parallel_env
()
model_a
=
SimpleNet
(
self
.
trainer_id
)
model_a
=
SimpleNet
(
self
.
trainer_id
)
model_b
=
SimpleNet
(
self
.
trainer_id
)
model_b
=
SimpleNet
(
self
.
trainer_id
)
state_dict
=
model_a
.
state_dict
()
state_dict
=
model_a
.
state_dict
()
model_b
.
set_state_dict
(
state_dict
)
model_b
.
set_state_dict
(
state_dict
)
model_a
=
paddle
.
DataParallel
(
model_a
=
paddle
.
DataParallel
(
model_a
,
find_unused_parameters
=
True
,
group
=
self
.
pg
model_a
,
find_unused_parameters
=
True
,
group
=
self
.
pg
)
model_b
=
paddle
.
DataParallel
(
model_b
,
find_unused_parameters
=
True
,
group
=
self
.
pg
)
ones_input
=
paddle
.
ones
(
shape
=
(
batch
,
in_dim
))
ones_input
.
stop_gradient
=
True
w1_grad_sum
=
np
.
zeros
((
in_dim
,
out_dim
),
dtype
=
'float32'
)
w2_grad_sum
=
np
.
zeros
((
in_dim
,
out_dim
),
dtype
=
'float32'
)
for
step_id
in
range
(
5
):
random_input
=
paddle
.
rand
(
shape
=
(
batch
,
in_dim
))
random_input
.
stop_gradient
=
True
if
step_id
%
2
==
0
:
out_a
=
model_a
(
random_input
)
out_b
=
model_b
(
random_input
)
else
:
out_a
=
model_a
(
ones_input
)
out_b
=
model_b
(
ones_input
)
out_a
.
sum
().
backward
()
out_b
.
sum
().
backward
()
self
.
check_gradient
(
model_a
.
parameters
())
self
.
check_gradient
(
model_b
.
parameters
())
# test acc gradient
w1_grad_sum
=
self
.
check_acc
(
model_a
.
_layers
.
w1
.
grad
,
w1_grad_sum
,
model_b
.
_layers
.
w1
.
grad
,
)
)
model_b
=
paddle
.
DataParallel
(
w2_grad_sum
=
self
.
check_acc
(
model_b
,
find_unused_parameters
=
True
,
group
=
self
.
pg
model_a
.
_layers
.
w2
.
grad
,
w2_grad_sum
,
model_b
.
_layers
.
w2
.
grad
,
)
)
ones_input
=
paddle
.
ones
(
shape
=
(
batch
,
in_dim
))
model_a
.
clear_gradients
()
ones_input
.
stop_gradient
=
True
w1_grad_sum
=
np
.
zeros
((
in_dim
,
out_dim
),
dtype
=
'float32'
)
w2_grad_sum
=
np
.
zeros
((
in_dim
,
out_dim
),
dtype
=
'float32'
)
for
step_id
in
range
(
5
):
random_input
=
paddle
.
rand
(
shape
=
(
batch
,
in_dim
))
random_input
.
stop_gradient
=
True
if
step_id
%
2
==
0
:
out_a
=
model_a
(
random_input
)
out_b
=
model_b
(
random_input
)
else
:
out_a
=
model_a
(
ones_input
)
out_b
=
model_b
(
ones_input
)
out_a
.
sum
().
backward
()
out_b
.
sum
().
backward
()
self
.
check_gradient
(
model_a
.
parameters
())
self
.
check_gradient
(
model_b
.
parameters
())
# test acc gradient
w1_grad_sum
=
self
.
check_acc
(
model_a
.
_layers
.
w1
.
grad
,
w1_grad_sum
,
model_b
.
_layers
.
w1
.
grad
,
)
w2_grad_sum
=
self
.
check_acc
(
model_a
.
_layers
.
w2
.
grad
,
w2_grad_sum
,
model_b
.
_layers
.
w2
.
grad
,
)
model_a
.
clear_gradients
()
def
check_acc
(
self
,
grad
,
grad_sum
,
acc_grad
):
def
check_acc
(
self
,
grad
,
grad_sum
,
acc_grad
):
if
grad
is
not
None
:
if
grad
is
not
None
:
...
...
python/paddle/fluid/tests/unittests/xpu/process_group_bkcl.py
浏览文件 @
e83f5f33
...
@@ -21,7 +21,6 @@ import numpy as np
...
@@ -21,7 +21,6 @@ import numpy as np
import
paddle
import
paddle
import
paddle.distributed
as
dist
import
paddle.distributed
as
dist
from
paddle.fluid.dygraph.parallel
import
ParallelEnv
from
paddle.fluid.dygraph.parallel
import
ParallelEnv
from
paddle.fluid.framework
import
_test_eager_guard
def
init_process_group
(
strategy
=
None
):
def
init_process_group
(
strategy
=
None
):
...
@@ -45,150 +44,137 @@ class TestProcessGroupFp32(unittest.TestCase):
...
@@ -45,150 +44,137 @@ class TestProcessGroupFp32(unittest.TestCase):
self
.
shape
=
(
2
,
10
,
5
)
self
.
shape
=
(
2
,
10
,
5
)
def
test_create_process_group_bkcl
(
self
):
def
test_create_process_group_bkcl
(
self
):
with
_test_eager_guard
():
device_id
=
paddle
.
distributed
.
ParallelEnv
().
dev_id
device_id
=
paddle
.
distributed
.
ParallelEnv
().
dev_id
paddle
.
set_device
(
'xpu:%d'
%
device_id
)
paddle
.
set_device
(
'xpu:%d'
%
device_id
)
pg
=
init_process_group
()
pg
=
init_process_group
()
sys
.
stdout
.
write
(
sys
.
stdout
.
write
(
"rank {}: size {} name {}
\n
"
.
format
(
pg
.
rank
(),
pg
.
size
(),
pg
.
name
())
"rank {}: size {} name {}
\n
"
.
format
(
)
pg
.
rank
(),
pg
.
size
(),
pg
.
name
()
sys
.
stdout
.
write
(
"rank {}: test new group api ok
\n
"
.
format
(
pg
.
rank
()))
)
)
# test allreduce sum
sys
.
stdout
.
write
(
# rank 0
"rank {}: test new group api ok
\n
"
.
format
(
pg
.
rank
())
x
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
)
tensor_x
=
paddle
.
to_tensor
(
x
)
# rank 1
# test allreduce sum
y
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
# rank 0
tensor_y
=
paddle
.
to_tensor
(
y
)
x
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
tensor_x
=
paddle
.
to_tensor
(
x
)
sum_result
=
tensor_x
+
tensor_y
# rank 1
if
pg
.
rank
()
==
0
:
y
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
task
=
dist
.
all_reduce
(
tensor_x
)
tensor_y
=
paddle
.
to_tensor
(
y
)
assert
np
.
array_equal
(
tensor_x
,
sum_result
)
else
:
sum_result
=
tensor_x
+
tensor_y
task
=
dist
.
all_reduce
(
tensor_y
)
if
pg
.
rank
()
==
0
:
assert
np
.
array_equal
(
tensor_y
,
sum_result
)
task
=
dist
.
all_reduce
(
tensor_x
)
assert
np
.
array_equal
(
tensor_x
,
sum_result
)
sys
.
stdout
.
write
(
else
:
"rank {}: test allreduce sum api ok
\n
"
.
format
(
pg
.
rank
())
task
=
dist
.
all_reduce
(
tensor_y
)
)
assert
np
.
array_equal
(
tensor_y
,
sum_result
)
# TODO
sys
.
stdout
.
write
(
# test allreduce max/min/prod
"rank {}: test allreduce sum api ok
\n
"
.
format
(
pg
.
rank
())
)
# test broadcast
# rank 0
# TODO
x
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
# test allreduce max/min/prod
tensor_x
=
paddle
.
to_tensor
(
x
)
# rank 1
# test broadcast
y
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
# rank 0
tensor_y
=
paddle
.
to_tensor
(
y
)
x
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
tensor_x
=
paddle
.
to_tensor
(
x
)
broadcast_result
=
paddle
.
assign
(
tensor_x
)
# rank 1
if
pg
.
rank
()
==
0
:
y
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
# XPU don't support event query by now, so just use sync op here
tensor_y
=
paddle
.
to_tensor
(
y
)
task
=
dist
.
broadcast
(
tensor_x
,
0
)
paddle
.
device
.
xpu
.
synchronize
()
broadcast_result
=
paddle
.
assign
(
tensor_x
)
assert
np
.
array_equal
(
broadcast_result
,
tensor_x
)
if
pg
.
rank
()
==
0
:
else
:
# XPU don't support event query by now, so just use sync op here
task
=
dist
.
broadcast
(
tensor_y
,
0
)
task
=
dist
.
broadcast
(
tensor_x
,
0
)
paddle
.
device
.
xpu
.
synchronize
()
paddle
.
device
.
xpu
.
synchronize
()
assert
np
.
array_equal
(
broadcast_result
,
tensor_y
)
assert
np
.
array_equal
(
broadcast_result
,
tensor_x
)
else
:
sys
.
stdout
.
write
(
"rank {}: test broadcast api ok
\n
"
.
format
(
pg
.
rank
()))
task
=
dist
.
broadcast
(
tensor_y
,
0
)
paddle
.
device
.
xpu
.
synchronize
()
# test barrier
assert
np
.
array_equal
(
broadcast_result
,
tensor_y
)
# rank 0
if
pg
.
rank
()
==
0
:
sys
.
stdout
.
write
(
pg
.
barrier
(
device_id
)
"rank {}: test broadcast api ok
\n
"
.
format
(
pg
.
rank
())
# rank 1
)
else
:
task
=
pg
.
barrier
(
device_id
)
# test barrier
task
.
wait
()
# rank 0
if
pg
.
rank
()
==
0
:
sys
.
stdout
.
write
(
"rank {}: test barrier api ok
\n
"
.
format
(
pg
.
rank
()))
pg
.
barrier
(
device_id
)
# rank 1
# test allgather
else
:
# rank 0
task
=
pg
.
barrier
(
device_id
)
x
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
task
.
wait
()
y
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
tensor_x
=
paddle
.
to_tensor
(
x
)
sys
.
stdout
.
write
(
"rank {}: test barrier api ok
\n
"
.
format
(
pg
.
rank
()))
tensor_y
=
paddle
.
to_tensor
(
y
)
out_shape
=
list
(
self
.
shape
)
# test allgather
out_shape
[
0
]
*=
2
# rank 0
out
=
np
.
random
.
random
(
out_shape
).
astype
(
self
.
dtype
)
x
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
tensor_out
=
paddle
.
to_tensor
(
out
)
y
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
if
pg
.
rank
()
==
0
:
tensor_x
=
paddle
.
to_tensor
(
x
)
task
=
pg
.
all_gather
(
tensor_x
,
tensor_out
)
tensor_y
=
paddle
.
to_tensor
(
y
)
task
.
wait
()
out_shape
=
list
(
self
.
shape
)
paddle
.
device
.
xpu
.
synchronize
()
out_shape
[
0
]
*=
2
# rank 1
out
=
np
.
random
.
random
(
out_shape
).
astype
(
self
.
dtype
)
else
:
tensor_out
=
paddle
.
to_tensor
(
out
)
tensor_out_list
=
[
if
pg
.
rank
()
==
0
:
paddle
.
empty_like
(
tensor_x
),
task
=
pg
.
all_gather
(
tensor_x
,
tensor_out
)
paddle
.
empty_like
(
tensor_x
),
task
.
wait
()
]
paddle
.
device
.
xpu
.
synchronize
()
task
=
dist
.
all_gather
(
tensor_out_list
,
tensor_y
)
# rank 1
paddle
.
device
.
xpu
.
synchronize
()
else
:
tensor_out
=
paddle
.
concat
(
tensor_out_list
)
tensor_out_list
=
[
out_1
=
paddle
.
slice
(
tensor_out
,
[
0
],
[
0
],
[
out_shape
[
0
]
//
2
])
paddle
.
empty_like
(
tensor_x
),
out_2
=
paddle
.
slice
(
paddle
.
empty_like
(
tensor_x
),
tensor_out
,
[
0
],
[
out_shape
[
0
]
//
2
],
[
out_shape
[
0
]]
]
)
task
=
dist
.
all_gather
(
tensor_out_list
,
tensor_y
)
assert
np
.
array_equal
(
tensor_x
,
out_1
)
paddle
.
device
.
xpu
.
synchronize
()
assert
np
.
array_equal
(
tensor_y
,
out_2
)
tensor_out
=
paddle
.
concat
(
tensor_out_list
)
sys
.
stdout
.
write
(
"rank {}: test allgather api ok
\n
"
.
format
(
pg
.
rank
()))
out_1
=
paddle
.
slice
(
tensor_out
,
[
0
],
[
0
],
[
out_shape
[
0
]
//
2
])
out_2
=
paddle
.
slice
(
if
pg
.
rank
()
==
0
:
tensor_out
,
[
0
],
[
out_shape
[
0
]
//
2
],
[
out_shape
[
0
]]
task
=
pg
.
all_gather
(
tensor_x
,
tensor_out
)
)
task
.
wait
()
assert
np
.
array_equal
(
tensor_x
,
out_1
)
paddle
.
device
.
xpu
.
synchronize
()
assert
np
.
array_equal
(
tensor_y
,
out_2
)
# rank 1
sys
.
stdout
.
write
(
else
:
"rank {}: test allgather api ok
\n
"
.
format
(
pg
.
rank
())
tensor_out_list
=
[]
)
task
=
dist
.
all_gather
(
tensor_out_list
,
tensor_y
)
paddle
.
device
.
xpu
.
synchronize
()
if
pg
.
rank
()
==
0
:
tensor_out
=
paddle
.
concat
(
tensor_out_list
)
task
=
pg
.
all_gather
(
tensor_x
,
tensor_out
)
out_1
=
paddle
.
slice
(
tensor_out
,
[
0
],
[
0
],
[
out_shape
[
0
]
//
2
])
task
.
wait
()
out_2
=
paddle
.
slice
(
paddle
.
device
.
xpu
.
synchronize
()
tensor_out
,
[
0
],
[
out_shape
[
0
]
//
2
],
[
out_shape
[
0
]]
# rank 1
)
else
:
assert
np
.
array_equal
(
tensor_x
,
out_1
)
tensor_out_list
=
[]
assert
np
.
array_equal
(
tensor_y
,
out_2
)
task
=
dist
.
all_gather
(
tensor_out_list
,
tensor_y
)
sys
.
stdout
.
write
(
"rank {}: test allgather api2 ok
\n
"
.
format
(
pg
.
rank
()))
paddle
.
device
.
xpu
.
synchronize
()
tensor_out
=
paddle
.
concat
(
tensor_out_list
)
# test Reduce
out_1
=
paddle
.
slice
(
tensor_out
,
[
0
],
[
0
],
[
out_shape
[
0
]
//
2
])
# rank 0
out_2
=
paddle
.
slice
(
x
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
tensor_out
,
[
0
],
[
out_shape
[
0
]
//
2
],
[
out_shape
[
0
]]
y
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
)
tensor_x
=
paddle
.
to_tensor
(
x
)
assert
np
.
array_equal
(
tensor_x
,
out_1
)
tensor_y
=
paddle
.
to_tensor
(
y
)
assert
np
.
array_equal
(
tensor_y
,
out_2
)
sum_result
=
tensor_x
+
tensor_y
sys
.
stdout
.
write
(
if
pg
.
rank
()
==
0
:
"rank {}: test allgather api2 ok
\n
"
.
format
(
pg
.
rank
())
task
=
dist
.
reduce
(
tensor_x
,
0
,
sync_op
=
True
)
)
paddle
.
device
.
xpu
.
synchronize
()
# rank 1
# test Reduce
else
:
# rank 0
task
=
dist
.
reduce
(
tensor_y
,
0
,
sync_op
=
False
)
x
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
task
.
wait
()
y
=
np
.
random
.
random
(
self
.
shape
).
astype
(
self
.
dtype
)
paddle
.
device
.
xpu
.
synchronize
()
tensor_x
=
paddle
.
to_tensor
(
x
)
if
pg
.
rank
()
==
0
:
tensor_y
=
paddle
.
to_tensor
(
y
)
assert
np
.
array_equal
(
tensor_x
,
sum_result
)
sum_result
=
tensor_x
+
tensor_y
sys
.
stdout
.
write
(
"rank {}: test reduce sum api ok
\n
"
.
format
(
pg
.
rank
()))
if
pg
.
rank
()
==
0
:
task
=
dist
.
reduce
(
tensor_x
,
0
,
sync_op
=
True
)
paddle
.
device
.
xpu
.
synchronize
()
# rank 1
else
:
task
=
dist
.
reduce
(
tensor_y
,
0
,
sync_op
=
False
)
task
.
wait
()
paddle
.
device
.
xpu
.
synchronize
()
if
pg
.
rank
()
==
0
:
assert
np
.
array_equal
(
tensor_x
,
sum_result
)
sys
.
stdout
.
write
(
"rank {}: test reduce sum api ok
\n
"
.
format
(
pg
.
rank
())
)
class
TestProcessGroupFp16
(
TestProcessGroupFp32
):
class
TestProcessGroupFp16
(
TestProcessGroupFp32
):
...
...
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录