def api_gpu_device_num(val: int) -> None:
"""Set number of GPUs on each machine to run oneflow on.
Args:
val (int): number of GPUs. It is identical on every machine. In other words,
you can't specify different number of GPUs you would like to use on each machine.
"""
if oneflow._oneflow_internal.flags.with_cuda():
return enable_if.unique([gpu_device_num, do_nothing])(val)
else:
print(
"INFO: for CPU-only OneFlow, oneflow.config.gpu_device_num is equivalent to oneflow.config.cpu_device_num"
)
print(traceback.format_stack()[-2])
return enable_if.unique([cpu_device_num, do_nothing])(val)
def api_env_init() -> bool:
"""Init environment for job
Returns:
bool: [description]
"""
return enable_if.unique([env_init, do_nothing])()
def api_enable_eager_execution(val: bool = True) -> None:
"""If True, job will execute in eager mode, else use lazy mode(static graph).
Args:
val (bool, optional): Whether eager execution or not. Defaults to True.
"""
return enable_if.unique([enable_eager_environment])(val)
def api_logtostderr(val: int) -> None:
"""Set whether log messages go to stderr instead of logfiles
Args:
val (int): [description]
"""
return enable_if.unique([logtostderr, do_nothing])(val)
def api_log_dir(val: str) -> None:
"""Specify a dir to store OneFlow's logging files. If not specified, it is `./log` by default.
Args:
val (str): string , log file path
"""
return enable_if.unique([log_dir, do_nothing])(val)
def api_cpu_device_num(val: int) -> None:
"""Set number of CPUs on each machine to run oneflow on. Usually you don't need to set this.
Args:
val (int): number of CPUs. It is identical on every machine.
"""
return enable_if.unique([cpu_device_num, do_nothing])(val)
def api_enable_fusion(val: bool = True) -> None:
"""Whether or not allow fusion the operators
Args:
val (bool, optional): True or False. Defaults to True.
"""
return enable_if.unique([enable_fusion, do_nothing])(val=val)
def api_reserved_device_mem_mbyte(val: int) -> None:
"""Set up the memory size of reserved device
Args:
val (int): memory size, e.g. 1024(mb)
"""
return enable_if.unique([reserved_device_mem_mbyte, do_nothing])(val)
def api_compute_thread_pool_size(val: int) -> None:
"""Set up the size of compute thread pool
Args:
val (int): size of thread pool
"""
return enable_if.unique([compute_thread_pool_size, do_nothing])(val)
def api_enable_mem_chain_merge(val: bool = True) -> None:
"""Whether or not to enable MemChain merge.
Args:
val (bool, optional): True or False. Defaults to True.
"""
return enable_if.unique([enable_mem_chain_merge, do_nothing])(val=val)
def api_ctrl_port(val: int) -> None:
"""Set port number used to control the execution across multiple machines. Same on every machine.
Args:
val: a port number accessible to peer machines
"""
return enable_if.unique([ctrl_port, do_nothing])(val)
def api_enable_model_io_v2(val):
"""Whether or not use version2 of model input/output function.
Args:
val ([type]): True or False
"""
return enable_if.unique([enable_model_io_v2, do_nothing])(val)
def api_data_port(val: int) -> None:
"""Set port number used to data transfer among multiple machines. Same on every machine.
Args:
val: a port number accessible to peer machines
"""
return enable_if.unique([data_port, do_nothing])(val)
def api_enable_legacy_model_io(val: bool = True):
"""Whether or not use legacy model io.
Args:
val ([type]): True or False
"""
return enable_if.unique([enable_legacy_model_io, do_nothing])(val)
def api_nccl_num_streams(val: int) -> None:
"""Set up the number of nccl parallel streams while use boxing
Args:
val (int): number of streams
"""
return enable_if.unique([nccl_num_streams, do_nothing])(val)
def api_nccl_enable_mixed_fusion(val: bool) -> None:
"""Whether or not use nccl mixed fusion
Args:
val (bool): True or False
"""
return enable_if.unique([nccl_enable_mixed_fusion, do_nothing])(val)
def api_nccl_fusion_threshold_mb(val: int) -> None:
"""Set up threshold for oprators fusion
Args:
val (int): int number, e.g. 10(mb)
"""
return enable_if.unique([nccl_fusion_threshold_mb, do_nothing])(val)
def api_load_library_now(val: str) -> None:
"""Load necessary library for job now
Args:
val (str): path to shared object file
"""
return enable_if.unique([load_library_now, do_nothing])(val)
def api_nccl_enable_all_to_all(val: bool) -> None:
"""Whether or not use nccl all2all during s2s boxing
Args:
val (bool): True or False
"""
return enable_if.unique([nccl_enable_all_to_all, do_nothing])(val)
def api_nccl_fusion_broadcast(val: bool) -> None:
"""Whether or not use nccl fusion during broadcast progress
Args:
val (bool): True or False
"""
return enable_if.unique([nccl_fusion_broadcast, do_nothing])(val)
def api_nccl_fusion_max_ops(val: int) -> None:
"""Maximum number of ops for nccl fusion.
Args:
val (int): Maximum number of ops
"""
return enable_if.unique([nccl_fusion_max_ops, do_nothing])(val)
def api_enable_debug_mode(val: bool) -> None:
"""Whether use debug mode or not.
Args:
val (bool): True or False
"""
return enable_if.unique([enable_debug_mode, do_nothing])(val)
def api_nccl_fusion_all_gather(val: bool) -> None:
"""Whether or not use nccl fusion during all gather progress
Args:
val (bool): True or False
"""
return enable_if.unique([nccl_fusion_all_gather, do_nothing])(val)
def api_machine_num(val: int) -> None:
"""Set available number of machine/node for running job .
Args:
val (int): available number of machines
"""
return enable_if.unique([machine_num, do_nothing])(val)
def api_max_mdsave_worker_num(val: int) -> None:
"""Set up max number of workers for mdsave process.
Args:
val (int): max number of workers
"""
return enable_if.unique([max_mdsave_worker_num, do_nothing])(val)
def api_nccl_fusion_all_reduce_use_buffer(val: bool) -> None:
"""Whether or not use buffer during nccl fusion progress
Args:
val (bool): True or False
"""
return enable_if.unique([nccl_fusion_all_reduce_use_buffer,
do_nothing])(val)
def api_enable_tensor_float_32_compute(val: bool = True) -> None:
"""Whether or not to enable Tensor-float-32 on supported GPUs
Args:
val (bool, optional): True or False. Defaults to True.
"""
return enable_if.unique([enable_tensor_float_32_compute,
do_nothing])(val=val)
def api_num_callback_threads(val: int) -> None:
"""Set up number of callback threads for boxing process.
Boxing is used to convert between different parallel properties of logical tensor
Args:
val (int): number of callback threads
"""
return enable_if.unique([num_callback_threads, do_nothing])(val)
def api_comm_net_worker_num(val: int) -> None:
"""Set up the workers number in epoll mode network,
If use RDMA mode network, then doesn't need.
Args:
val (int): number of workers
"""
return enable_if.unique([comm_net_worker_num, do_nothing])(val)
def api_enable_cudnn_fused_normalization_add_relu(val: bool) -> None:
"""Whether enable cudnn_fused_normalization_add_relu.
Args:
val (bool): whether enable or not
"""
return enable_if.unique(
[enable_cudnn_fused_normalization_add_relu, do_nothing])(val)
