Fix paddle env variables. (#11564)

benchmark/fluid/fluid_benchmark.py

@@ -97,7 +97,7 @@ def dist_transpile(trainer_id, args):
         return train_program, fluid.default_startup_program()
     else:
         raise ValueError(
-            'TRAINING_ROLE environment variable must be either TRAINER or PSERVER'
+            'PADDLE_TRAINING_ROLE environment variable must be either TRAINER or PSERVER'
         )
 
 

benchmark/fluid/kube_gen_job.py

@@ -108,10 +108,10 @@ def gen_job():
     tn_container["ports"][0]["containerPort"] = spreadport
 
     envs.append({"name": "PADDLE_JOB_NAME", "value": args.jobname})
-    envs.append({"name": "TRAINERS", "value": str(args.trainers)})
+    envs.append({"name": "PADDLE_TRAINERS", "value": str(args.trainers)})
     envs.append({"name": "PSERVERS", "value": str(args.pservers)})
     envs.append({"name": "ENTRY", "value": args.entry})
-    envs.append({"name": "PADDLE_INIT_PORT", "value": str(args.port)})
+    envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)})
     envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)})
     # NOTE: these directories below are cluster specific, please modify
     # this settings before you run on your own cluster.
@@ -167,16 +167,22 @@ def gen_job():
     tn_container["volumeMounts"] = volumeMounts
 
     ps_container["env"] = envs
-    ps_container["env"].append({"name": "TRAINING_ROLE", "value": "PSERVER"})
+    ps_container["env"].append({
+        "name": "PADDLE_TRAINING_ROLE",
+        "value": "PSERVER"
+    })
     tn_container["env"] = envs
     if args.disttype == "pserver":
         tn_container["env"].append({
-            "name": "TRAINING_ROLE",
+            "name": "PADDLE_TRAINING_ROLE",
             "value": "TRAINER"
         })
     elif args.disttype == "nccl2" or args.disttype == "local":
         # NCCL2 have no training role, set to plain WORKER
-        tn_container["env"].append({"name": "TRAINING_ROLE", "value": "WORKER"})
+        tn_container["env"].append({
+            "name": "PADDLE_TRAINING_ROLE",
+            "value": "WORKER"
+        })
 
     os.mkdir(args.jobname)
     if args.disttype == "pserver":

doc/fluid/howto/cluster/fluid_cluster_train_cn.md

@@ -168,13 +168,13 @@ cd /paddle/python/paddle/fluid/tests/book
 
 第二步，启动Parameter Server：
 ```bash
-PADDLE_INIT_PORT=6174 PADDLE_INIT_PSERVERS=192.168.1.2 TRAINERS=2 POD_IP=192.168.1.2 PADDLE_INIT_TRAINER_ID=1 TRAINING_ROLE=PSERVER python test_fit_a_line.py
+PADDLE_PSERVER_PORT=6174 PADDLE_PSERVER_IPS=192.168.1.2 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=192.168.1.2 PADDLE_TRAINER_ID=1 PADDLE_TRAINING_ROLE=PSERVER python test_fit_a_line.py
 ```
 执行命令后请等待出现提示： ```Server listening on 192.168.1.2:6174 ```, 表示Paramter Server已经正常启动。
 
 第三步，启动Trainer：
 ```bash
-PADDLE_INIT_PORT=6174 PADDLE_INIT_PSERVERS=192.168.1.3 TRAINERS=2 POD_IP=192.168.1.3 PADDLE_INIT_TRAINER_ID=1 TRAINING_ROLE=TRAINER python test_fit_a_line.py
+PADDLE_PSERVER_PORT=6174 PADDLE_PSERVER_IPS=192.168.1.3 PADDLE_TRAINERS=2 PADDLE_CURRENT_IPP=192.168.1.3 PADDLE_TRAINER_ID=1 PADDLE_TRAINING_ROLE=TRAINER python test_fit_a_line.py
 ```
 由于我们定义的Trainer的数量是2个，因此需要在另外一个计算节点上再启动一个Trainer。
 

doc/fluid/howto/cluster/fluid_recordio.md

@@ -114,8 +114,8 @@ def gen_train_list(file_pattern, trainers, trainer_id):
            ret_list.append(f)
    return ret_list
 
-trainers = int(os.getenv("TRAINERS"))
-trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
+trainers = int(os.getenv("PADDLE_TRAINERS"))
+trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
 data_file = fluid.layers.io.open_files(
     filenames=gen_train_list("./mnist-[0-9]*.recordio", 2, 0),
     thread_num=1,

python/paddle/fluid/tests/book/notest_understand_sentiment.py

@@ -194,16 +194,16 @@ def train(word_dict,
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

python/paddle/fluid/tests/book/test_fit_a_line.py

@@ -69,16 +69,16 @@ def train(use_cuda, save_dirname, is_local):
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

python/paddle/fluid/tests/book/test_image_classification.py

@@ -178,16 +178,16 @@ def train(net_type, use_cuda, save_dirname, is_local):
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

python/paddle/fluid/tests/book/test_label_semantic_roles.py

@@ -209,16 +209,16 @@ def train(use_cuda, save_dirname=None, is_local=True):
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

python/paddle/fluid/tests/book/test_machine_translation.py

@@ -200,16 +200,16 @@ def train_main(use_cuda, is_sparse, is_local=True):
     if is_local:
         train_loop(framework.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

python/paddle/fluid/tests/book/test_recognize_digits.py

@@ -151,16 +151,16 @@ def train(nn_type,
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

python/paddle/fluid/tests/book/test_recommender_system.py

@@ -220,16 +220,16 @@ def train(use_cuda, save_dirname, is_local=True):
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":

python/paddle/fluid/tests/book/test_word2vec.py

@@ -125,16 +125,16 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True):
     if is_local:
         train_loop(fluid.default_main_program())
     else:
-        port = os.getenv("PADDLE_INIT_PORT", "6174")
-        pserver_ips = os.getenv("PADDLE_INIT_PSERVERS")  # ip,ip...
+        port = os.getenv("PADDLE_PSERVER_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVER_IPS")  # ip,ip...
         eplist = []
         for ip in pserver_ips.split(","):
             eplist.append(':'.join([ip, port]))
         pserver_endpoints = ",".join(eplist)  # ip:port,ip:port...
-        trainers = int(os.getenv("TRAINERS"))
+        trainers = int(os.getenv("PADDLE_TRAINERS"))
         current_endpoint = os.getenv("POD_IP") + ":" + port
-        trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
-        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+        training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
         t = fluid.DistributeTranspiler()
         t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
         if training_role == "PSERVER":