Add startup bash files of test_ascend_group. (#30645)

Add startup bash files of test_ascend_group

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -39,6 +39,7 @@ list(APPEND MIXED_DIST_TEST_OPS test_fleet_run_random_port)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_launch_async)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_launch_cloud)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_launch_ascend)
+list(APPEND MIXED_DIST_TEST_OPS test_ascend_group)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_launch_nproc)
 list(APPEND MIXED_DIST_TEST_OPS test_fleet_api_input)
 list(APPEND MIXED_DIST_TEST_OPS test_collective_optimizer)
@@ -524,6 +525,7 @@ if(WITH_DISTRIBUTE)
         bash_test_modules(test_fleet_launch_cloud START_BASH test_fleet_launch_cloud.sh ENVS PADDLE_BINARY_DIR=${PADDLE_BINARY_DIR})
         bash_test_modules(test_fleet_launch_ascend START_BASH test_fleet_launch_ascend.sh ENVS PADDLE_BINARY_DIR=${PADDLE_BINARY_DIR})
         bash_test_modules(test_fleet_launch_nproc START_BASH test_fleet_launch_nproc.sh ENVS PADDLE_BINARY_DIR=${PADDLE_BINARY_DIR})
+        bash_test_modules(test_ascend_group START_BASH test_ascend_group.sh ENVS PADDLE_BINARY_DIR=${PADDLE_BINARY_DIR})
 
         # port range (20000, 23000) is reserved for dist-ops
         set(dist_ut_port 20001)

python/paddle/fluid/tests/unittests/ascend_group.py

+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import sys
+import time
+import paddle.fluid as fluid
+from paddle.fluid import unique_name
+import paddle.fluid.core as core
+import paddle
+from paddle.fluid.layer_helper import LayerHelper
+from paddle.distributed import fleet
+
+paddle.enable_static()
+
+OpRole = core.op_proto_and_checker_maker.OpRole
+OP_ROLE_KEY = core.op_proto_and_checker_maker.kOpRoleAttrName()
+OP_ROLE_VAR_KEY = core.op_proto_and_checker_maker.kOpRoleVarAttrName()
+
+role = fleet.PaddleCloudRoleMaker(is_collective=True)
+fleet.init(role)
+
+def init_communicator(startup_program, main_program, current_endpoint, endpoints, ring_id):
+    nranks = len(endpoints)
+    other_endpoints = endpoints[:]
+    other_endpoints.remove(current_endpoint)
+    group_rank=endpoints.index(current_endpoint)
+    assert group_rank >=0
+
+    block = startup_program.global_block()
+    nccl_id_var = block.create_var(
+        name=unique_name.generate('nccl_id'),
+        persistable=True,
+        type=core.VarDesc.VarType.RAW)
+    block.append_op(
+        type='c_gen_nccl_id',
+        inputs={},
+        outputs={'Out': nccl_id_var},
+        attrs={
+            'rank': group_rank,
+            'endpoint': current_endpoint,
+            'other_endpoints': other_endpoints,
+            OP_ROLE_KEY: OpRole.Forward,
+        })
+    block.append_op(
+        type='c_comm_init',
+        inputs={'X': nccl_id_var},
+        outputs={},
+        attrs={
+            'nranks': nranks,
+            'rank': group_rank,
+            'ring_id': ring_id,
+            OP_ROLE_KEY: OpRole.Forward,
+        })
+    block.create_var(
+        name="data",
+        persistable=True,
+        dtype='float32')
+
+    with fluid.program_guard(main_program):
+        op_type="c_allreduce_sum"
+        data=fluid.layers.fill_constant(shape=[1], dtype='float32', value=2.5)
+        helper = LayerHelper(op_type, **locals())
+        helper.append_op(
+            type=op_type,
+            inputs={'X': [data]},
+            outputs={'Out': [data]},
+            attrs={'ring_id': ring_id,
+                   'use_calc_stream': True})
+
+def train(world_endpoints, world_device_ids, local_device_ids,local_rank):
+    startup_programs=[]
+    main_programs=[]
+
+    #trainer_endpoints=["127.0.0.1:6071","127.0.0.1:6072","127.0.0.1:6073","127.0.0.1:6074"]
+    trainer_endpoints=world_endpoints
+    groups=[[], [], []]
+    groups[0]=[trainer_endpoints[0], trainer_endpoints[1]]
+    groups[1]=[trainer_endpoints[2], trainer_endpoints[3]]
+    groups[2]=[trainer_endpoints[0], trainer_endpoints[2]]
+
+    for i in range(len(trainer_endpoints)):
+        startup_programs.append(fluid.Program())
+        main_programs.append(fluid.Program())
+
+    for idx, group in enumerate(groups):
+        for te in group:
+            te_idx = trainer_endpoints.index(te)
+            startup_program = startup_programs[te_idx]
+            main_program=main_programs[te_idx]
+            init_communicator(startup_program, main_program, te, group, idx)
+
+    print(len(startup_programs))
+    print(startup_programs[local_rank])
+    print(main_programs[local_rank])
+
+worker_endpoints=fleet.worker_endpoints()
+world_device_ids=fleet.world_device_ids()
+local_device_ids=fleet.local_device_ids()
+local_rank=int(fleet.local_rank())
+
+print("worker_endpoints:", worker_endpoints)
+print("world_device_ids:", world_device_ids)
+print("local_device_ids:", local_device_ids)
+print("local_rank:", local_rank)
+
+train(worker_endpoints, world_device_ids,local_device_ids,local_rank)

python/paddle/fluid/tests/unittests/test_ascend_group.sh

+#!/bin/bash
+
+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+set -e
+
+cluster_node_ips="127.0.0.1"
+export PADDLE_TRAINERS_NUM=4
+export POD_IP=127.0.0.1
+export PADDLE_TRAINERS=127.0.0.1
+export PADDLE_TRAINER_ID=0
+
+export PADDLE_PORT=35789
+export TRAINER_PORTS_NUM=4
+
+distributed_args="--ips=${cluster_node_ips} --ascend_npus=0,1,2,3 --log_dir=testlog"
+python -m paddle.distributed.fleet.launch ${distributed_args} \
+  ascend_group.py fleetascendgroup
\ No newline at end of file