def lazy_system_assign(ref,
value,
validate_shape=None,
use_locking=None,
name=None):
op_conf = _SystemAssignOpConf(ref, value, name=name)
device_tag, machine_device_ids = parallel_conf_util.GetDeviceTagAndMachineDeviceIds(
ref.parallel_conf)
with oneflow.scope.placement(device_tag, machine_device_ids):
interpret_util.Forward(op_conf)
return ref
def LazyConsistentWatch(blob_watched, handler):
handler_uuid = str(uuid.uuid1())
op_conf = op_conf_util.OperatorConf()
op_conf.name = id_util.UniqueStr("ForeignWatch_")
setattr(op_conf.foreign_watch_conf, "in", blob_watched.unique_name)
op_conf.foreign_watch_conf.handler_uuid = handler_uuid
device_name = blob_watched.parallel_conf.device_name[0]
tag_and_dev_ids = parallel_conf_util.GetDeviceTagAndMachineDeviceIds(
blob_watched.parallel_conf)
with oneflow.scope.placement(*tag_and_dev_ids):
compile_context.CurJobAddOp(op_conf)
watcher_util.BindUuidAndHandler(handler_uuid, blob_watched, handler)
def InterpretScope(session, function_desc, config_proto):
job_conf = function_desc.job_config_proto
job_conf.job_name = function_desc.job_func.__name__
placement_scope = function_desc.function_attribute.default_placement_scope
if placement_scope is None:
tag_and_dev_ids = placement_util.GetDefaultMachineDeviceIds(
oneflow.env.current_resource())
placement_scope = placement_util.GetPlacementScope(*tag_and_dev_ids)
distribute_strategy = function_desc.function_attribute.default_distribute_strategy
if distribute_strategy is None:
distribute_strategy = distribute_util.DistributeConsistentStrategy()
is_mirrored = isinstance(distribute_strategy,
distribute_util.DistributeMirroredStrategy)
tag_and_dev_ids = parallel_conf_util.GetDeviceTagAndMachineDeviceIds(
placement_scope.default_parallel_conf)
scope = MakeInitialScope(job_conf, *tag_and_dev_ids, is_mirrored)
with _JobBuildAndInferCtx(
job_conf.job_name), placement_scope, distribute_strategy:
c_api_util.CurJobBuildAndInferCtx_SetJobConf(job_conf)
with runtime_mode.ModeScope(runtime_mode.GLOBAL_MODE):
with _SessionInitialScope(session, scope):
yield
def BuildWithNewParallelConf(self, instruction_builder, parallel_conf):
tag_and_dev_ids = parallel_conf_util.GetDeviceTagAndMachineDeviceIds(
parallel_conf
)
return self.BuildWithNewParallelDesc(instruction_builder, *tag_and_dev_ids)
