Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 f53e5a04 编写于 作者: 姜 姜永久 提交者： GitHub
浏览文件
操作

rm multinode eager guard tests (#48766) 

* rm multinode eager guard tests

* remove unwanted tests

* reset process_mpi test
上级 7216d9bf
展开全部 隐藏空白更改
内联 并排
Showing with 599 addition and 642 deletion
python/paddle/fluid/tests/unittests/collective/multinode/mn_dygraph_group_sharded_stage3.py
浏览文件 @ f53e5a04
......@@ -34,7 +34,6 @@ from paddle.distributed.fleet.meta_parallel.sharding.group_sharded_stage3 import
from paddle.distributed.fleet.meta_parallel.sharding.group_sharded_utils import (
GroupShardedScaler,
)
from paddle.fluid.framework import _test_eager_guard
from paddle.nn import Linear
epoch = 10
......@@ -331,5 +330,4 @@ def test_stage2_stage3():
if __name__ == '__main__':
with _test_eager_guard():
test_stage2_stage3()
test_stage2_stage3()
python/paddle/fluid/tests/unittests/collective/multinode/mn_dygraph_sharding_stage2.py
浏览文件 @ f53e5a04
......@@ -29,7 +29,6 @@ from paddle.distributed.fleet.meta_optimizers.dygraph_optimizer.sharding_optimiz
from paddle.distributed.fleet.meta_parallel.sharding.sharding_stage2 import (
ShardingStage2,
)
from paddle.fluid.framework import _test_eager_guard
from paddle.nn import Linear
seed = 2022
......@@ -248,7 +247,5 @@ def test_dp_stage2():
if __name__ == '__main__':
with _test_eager_guard():
pass
fleet.init(is_collective=True, strategy=strategy)
test_dp_stage2()
python/paddle/fluid/tests/unittests/collective/process_group_gloo.py
浏览文件 @ f53e5a04
......@@ -20,7 +20,6 @@ import numpy as np
import paddle
from paddle.fluid import core
from paddle.fluid.dygraph.parallel import ParallelEnv
from paddle.fluid.framework import _test_eager_guard
class TestProcessGroupFp32(unittest.TestCase):
......@@ -35,154 +34,151 @@ class TestProcessGroupFp32(unittest.TestCase):
self.shape = (2, 10, 5)
def test_create_process_group_gloo(self):
with _test_eager_guard():
nranks = ParallelEnv().nranks
rank = ParallelEnv().local_rank
is_master = True if rank == 0 else False
store = paddle.fluid.core.TCPStore(
"127.0.0.1", 6272, is_master, nranks, 30
)
pg = paddle.fluid.core.ProcessGroupGloo.create(store, rank, nranks)
# test allreduce sum
# rank 0
paddle.device.set_device('cpu')
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
sum_result = x + y
if rank == 0:
task = pg.allreduce(tensor_x)
task.wait()
np.testing.assert_equal(tensor_x, sum_result)
else:
task = pg.allreduce(tensor_y)
task.wait()
np.testing.assert_equal(tensor_y, sum_result)
print("test allreduce sum api ok")
# test allreduce max
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
max_result = paddle.maximum(tensor_x, tensor_y)
if rank == 0:
task = pg.allreduce(tensor_x, core.ReduceOp.MAX)
task.wait()
assert np.array_equal(tensor_x, max_result)
else:
task = pg.allreduce(tensor_y, core.ReduceOp.MAX)
task.wait()
assert np.array_equal(tensor_y, max_result)
print("test allreduce max api ok")
# test broadcast
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
broadcast_result = paddle.assign(tensor_x)
if rank == 0:
task = pg.broadcast(tensor_x, 0)
assert np.array_equal(broadcast_result, tensor_x)
else:
task = pg.broadcast(tensor_y, 0)
assert np.array_equal(broadcast_result, tensor_y)
print("test broadcast api ok")
# test barrier
# rank 0
if pg.rank() == 0:
task = pg.barrier()
task.wait()
# rank 1
else:
task = pg.barrier()
task.wait()
print("test barrier api ok\n")
# test allgather
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
y = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
tensor_y = paddle.to_tensor(y)
out_shape = list(self.shape)
out_shape[0] *= 2
out = np.random.random(out_shape).astype(self.dtype)
tensor_out = paddle.to_tensor(out)
if pg.rank() == 0:
task = pg.all_gather(tensor_x, tensor_out)
task.wait()
paddle.device.cuda.synchronize()
# rank 1
else:
task = pg.all_gather(tensor_y, tensor_out)
task.wait()
out_1 = paddle.slice(tensor_out, [0], [0], [out_shape[0] // 2])
out_2 = paddle.slice(
tensor_out, [0], [out_shape[0] // 2], [out_shape[0]]
)
assert np.array_equal(tensor_x, out_1)
assert np.array_equal(tensor_y, out_2)
print("test allgather api ok\n")
# test Reduce
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
y = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
tensor_y = paddle.to_tensor(y)
sum_result = tensor_x + tensor_y
if pg.rank() == 0:
task = pg.reduce(tensor_x, 0)
task.wait()
# rank 1
else:
task = pg.reduce(tensor_y, 0)
task.wait()
if pg.rank() == 0:
assert np.array_equal(tensor_x, sum_result)
print("test reduce sum api ok\n")
# test Scatter
# rank 0
in_shape = list(self.shape)
in_shape[0] *= 2
x = np.random.random(in_shape).astype(self.dtype)
y = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
tensor_y = paddle.to_tensor(y)
if pg.rank() == 0:
task = pg.scatter(tensor_x, tensor_y, 0)
task.wait()
# rank 1
else:
task = pg.scatter(tensor_x, tensor_y, 0)
task.wait()
out1 = paddle.slice(tensor_x, [0], [0], [self.shape[0]])
out2 = paddle.slice(
tensor_x, [0], [self.shape[0]], [self.shape[0] * 2]
)
if pg.rank() == 0:
assert np.array_equal(tensor_y, out1)
else:
assert np.array_equal(tensor_y, out2)
print("test scatter api ok\n")
nranks = ParallelEnv().nranks
rank = ParallelEnv().local_rank
is_master = True if rank == 0 else False
store = paddle.fluid.core.TCPStore(
"127.0.0.1", 6272, is_master, nranks, 30
)
pg = paddle.fluid.core.ProcessGroupGloo.create(store, rank, nranks)
# test allreduce sum
# rank 0
paddle.device.set_device('cpu')
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
sum_result = x + y
if rank == 0:
task = pg.allreduce(tensor_x)
task.wait()
np.testing.assert_equal(tensor_x, sum_result)
else:
task = pg.allreduce(tensor_y)
task.wait()
np.testing.assert_equal(tensor_y, sum_result)
print("test allreduce sum api ok")
# test allreduce max
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
max_result = paddle.maximum(tensor_x, tensor_y)
if rank == 0:
task = pg.allreduce(tensor_x, core.ReduceOp.MAX)
task.wait()
assert np.array_equal(tensor_x, max_result)
else:
task = pg.allreduce(tensor_y, core.ReduceOp.MAX)
task.wait()
assert np.array_equal(tensor_y, max_result)
print("test allreduce max api ok")
# test broadcast
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
# rank 1
y = np.random.random(self.shape).astype(self.dtype)
tensor_y = paddle.to_tensor(y)
broadcast_result = paddle.assign(tensor_x)
if rank == 0:
task = pg.broadcast(tensor_x, 0)
assert np.array_equal(broadcast_result, tensor_x)
else:
task = pg.broadcast(tensor_y, 0)
assert np.array_equal(broadcast_result, tensor_y)
print("test broadcast api ok")
# test barrier
# rank 0
if pg.rank() == 0:
task = pg.barrier()
task.wait()
# rank 1
else:
task = pg.barrier()
task.wait()
print("test barrier api ok\n")
# test allgather
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
y = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
tensor_y = paddle.to_tensor(y)
out_shape = list(self.shape)
out_shape[0] *= 2
out = np.random.random(out_shape).astype(self.dtype)
tensor_out = paddle.to_tensor(out)
if pg.rank() == 0:
task = pg.all_gather(tensor_x, tensor_out)
task.wait()
paddle.device.cuda.synchronize()
# rank 1
else:
task = pg.all_gather(tensor_y, tensor_out)
task.wait()
out_1 = paddle.slice(tensor_out, [0], [0], [out_shape[0] // 2])
out_2 = paddle.slice(
tensor_out, [0], [out_shape[0] // 2], [out_shape[0]]
)
assert np.array_equal(tensor_x, out_1)
assert np.array_equal(tensor_y, out_2)
print("test allgather api ok\n")
# test Reduce
# rank 0
x = np.random.random(self.shape).astype(self.dtype)
y = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
tensor_y = paddle.to_tensor(y)
sum_result = tensor_x + tensor_y
if pg.rank() == 0:
task = pg.reduce(tensor_x, 0)
task.wait()
# rank 1
else:
task = pg.reduce(tensor_y, 0)
task.wait()
if pg.rank() == 0:
assert np.array_equal(tensor_x, sum_result)
print("test reduce sum api ok\n")
# test Scatter
# rank 0
in_shape = list(self.shape)
in_shape[0] *= 2
x = np.random.random(in_shape).astype(self.dtype)
y = np.random.random(self.shape).astype(self.dtype)
tensor_x = paddle.to_tensor(x)
tensor_y = paddle.to_tensor(y)
if pg.rank() == 0:
task = pg.scatter(tensor_x, tensor_y, 0)
task.wait()
# rank 1
else:
task = pg.scatter(tensor_x, tensor_y, 0)
task.wait()
out1 = paddle.slice(tensor_x, [0], [0], [self.shape[0]])
out2 = paddle.slice(tensor_x, [0], [self.shape[0]], [self.shape[0] * 2])
if pg.rank() == 0:
assert np.array_equal(tensor_y, out1)
else:
assert np.array_equal(tensor_y, out2)
print("test scatter api ok\n")
if __name__ == "__main__":
......
python/paddle/fluid/tests/unittests/collective/process_group_mpi.py
浏览文件 @ f53e5a04
......@@ -28,7 +28,7 @@ from paddle.distributed.collective import (
_set_group_map_by_name,
)
from paddle.fluid import core
from paddle.fluid.framework import _set_expected_place, _test_eager_guard
from paddle.fluid.framework import _set_expected_place
ctypes.CDLL("libmpi.so", mode=ctypes.RTLD_GLOBAL)
......@@ -444,51 +444,49 @@ class TestProcessGroup(unittest.TestCase):
self.shape = (2, 10, 5)
def test_create_process_group_mpi(self):
with _test_eager_guard():
group = init_process_group()
pg = group.process_group
group = init_process_group()
pg = group.process_group
# test allreduce sum
test_allreduce_sum(pg, self.shape, self.dtype)
# test allreduce sum
test_allreduce_sum(pg, self.shape, self.dtype)
# test allreduce max
test_allreduce_max(pg, self.shape, self.dtype)
# test allreduce max
test_allreduce_max(pg, self.shape, self.dtype)
# test allreduce min
test_allreduce_min(pg, self.shape, self.dtype)
# test allreduce min
test_allreduce_min(pg, self.shape, self.dtype)
# test allreduce prod
test_allreduce_prod(pg, self.shape, self.dtype)
# test allreduce prod
test_allreduce_prod(pg, self.shape, self.dtype)
# test broadcast
test_broadcast(pg, self.shape, self.dtype)
# test broadcast
test_broadcast(pg, self.shape, self.dtype)
# test barrier
test_barrair(pg)
# test barrier
test_barrair(pg)
# test allgather
test_allgather(pg, self.shape, self.dtype)
# test allgather
test_allgather(pg, self.shape, self.dtype)
# test alltoall
test_all2all(pg, self.shape, self.dtype)
# test alltoall
test_all2all(pg, self.shape, self.dtype)
# test Reduce
test_reduce_sum(pg, self.shape, self.dtype)
# test Reduce
test_reduce_sum(pg, self.shape, self.dtype)
# test reduce max
test_reduce_max(pg, self.shape, self.dtype)
# test reduce max
test_reduce_max(pg, self.shape, self.dtype)
# test reduce min
test_reduce_min(pg, self.shape, self.dtype)
# test reduce min
test_reduce_min(pg, self.shape, self.dtype)
# test reduce product
test_reduce_prod(pg, self.shape, self.dtype)
# test reduce product
test_reduce_prod(pg, self.shape, self.dtype)
# test Scatter
test_scatter(pg, self.shape, self.dtype)
# test Scatter
test_scatter(pg, self.shape, self.dtype)
# test send recv.
test_send_recv(pg, group, self.shape, self.dtype)
# test send recv.
test_send_recv(pg, group, self.shape, self.dtype)
if __name__ == "__main__":
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或