def local_mpi_engine(args):
print("launch cluster engine with cluster to run model: {}".format(
args.model))
from paddlerec.core.engine.local_mpi import LocalMPIEngine
print("use 1X1 MPI ClusterTraining at localhost to run model: {}".format(
args.model))
mpi = util.run_which("mpirun")
if not mpi:
raise RuntimeError("can not find mpirun, please check environment")
_envs = envs.load_yaml(args.model)
run_extras = get_all_inters_from_yaml(args.model, ["train.", "runner."])
trainer_class = run_extras.get("runner." + _envs["mode"] + ".runner_class",
None)
executor_mode = "train"
distributed_strategy = run_extras.get(
"runner." + _envs["mode"] + ".distribute_strategy", "async")
fleet_mode = run_extras.get("runner." + _envs["mode"] + ".fleet_mode",
"ps")
if trainer_class:
trainer = trainer_class
else:
trainer = "GeneralTrainer"
cluster_envs = {}
cluster_envs["mpirun"] = mpi
cluster_envs["train.trainer.trainer"] = trainer
cluster_envs["log_dir"] = "logs"
cluster_envs["train.trainer.engine"] = "local_cluster"
cluster_envs["train.trainer.executor_mode"] = executor_mode
cluster_envs["fleet_mode"] = fleet_mode
cluster_envs["train.trainer.strategy"] = distributed_strategy
cluster_envs["train.trainer.threads"] = "2"
cluster_envs["train.trainer.engine"] = "local_cluster"
cluster_envs["train.trainer.platform"] = envs.get_platform()
set_runtime_envs(cluster_envs, args.model)
launch = LocalMPIEngine(cluster_envs, args.model)
return launch
def local_mpi_engine(args):
print("launch cluster engine with cluster to run model: {}".format(
args.model))
from paddlerec.core.engine.local_mpi import LocalMPIEngine
print("use 1X1 MPI ClusterTraining at localhost to run model: {}".format(
args.model))
mpi = util.run_which("mpirun")
if not mpi:
raise RuntimeError("can not find mpirun, please check environment")
cluster_envs = {}
cluster_envs["mpirun"] = mpi
cluster_envs["train.trainer.trainer"] = "CtrCodingTrainer"
cluster_envs["log_dir"] = "logs"
cluster_envs["train.trainer.engine"] = "local_cluster"
cluster_envs["train.trainer.platform"] = envs.get_platform()
set_runtime_envs(cluster_envs, args.model)
launch = LocalMPIEngine(cluster_envs, args.model)
return launch
