Python create_orttrainer_and_save_checkpoint示例

示例#1

def distributed_zero_mixed_precision_lamb(world_rank, world_size, device,
                                          checkpoint_dir):
    opts = {
        'device': {
            'id': device
        },
        'mixed_precision': {
            'enabled': True
        },
        'distributed': {
            'world_rank': world_rank,
            'world_size': world_size,
            'allreduce_post_accumulation': True,
            'deepspeed_zero_optimization': {
                'stage': 1
            }
        },
        'debug': {
            'deterministic_compute': True
        }
    }
    create_orttrainer_and_save_checkpoint(device,
                                          opts,
                                          checkpoint_dir,
                                          state_dict_key_name='state_dict_' +
                                          str(world_rank))

示例#2

文件： orttraining_test_save_checkpoint.py 项目： kshama-msft/onnxruntime

def distributed_zero_mixed_precision_lamb(world_rank, world_size, device,
                                          checkpoint_dir):
    opts = {
        "device": {
            "id": device
        },
        "mixed_precision": {
            "enabled": True
        },
        "distributed": {
            "world_rank": world_rank,
            "world_size": world_size,
            "allreduce_post_accumulation": True,
            "deepspeed_zero_optimization": {
                "stage": 1
            },
        },
        "debug": {
            "deterministic_compute": True
        },
    }
    create_orttrainer_and_save_checkpoint(device,
                                          opts,
                                          checkpoint_dir,
                                          state_dict_key_name="state_dict_" +
                                          str(world_rank))

示例#3

文件： orttraining_test_save_checkpoint.py 项目： yyqgood/onnxruntime

def single_node_mixed_precision(device = 'cuda', checkpoint_dir = 'checkpoint_dir/single_node/mixed_precision/'):
    opts = {
                'device' : {'id' : device},
                'mixed_precision':
                {
                    'enabled': True
                },
                'debug' : {'deterministic_compute': True}
            }
    create_orttrainer_and_save_checkpoint(device, opts, checkpoint_dir)

示例#4

文件： orttraining_test_save_checkpoint.py 项目： yyqgood/onnxruntime

def data_parallelism_full_precision(world_rank, world_size, device, checkpoint_dir = 'checkpoint_dir/data_parallelism/full_precision/'):
    opts = {
                'device' : {'id' : device},
                'distributed' :
                {
                    'world_rank' : world_rank,
                    'world_size' : world_size,
                    'allreduce_post_accumulation' : True
                },
                'debug' : {'deterministic_compute': True}
            }
    create_orttrainer_and_save_checkpoint(device, opts, checkpoint_dir if world_rank == 0 else None)

示例#5

文件： orttraining_test_save_checkpoint.py 项目： kshama-msft/onnxruntime

def single_node_mixed_precision(checkpoint_dir, device="cuda"):
    opts = {
        "device": {
            "id": device
        },
        "mixed_precision": {
            "enabled": True
        },
        "debug": {
            "deterministic_compute": True
        }
    }
    create_orttrainer_and_save_checkpoint(device, opts, checkpoint_dir)

示例#6

文件： orttraining_test_save_checkpoint.py 项目： yyqgood/onnxruntime

def distributed_zero_full_precision_adam(world_rank, world_size, device, checkpoint_dir = 'checkpoint_dir/distributed_zero/full_precision/adam/'):
    opts = {
                'device' : {'id' : device},
                'distributed' :
                {
                    'world_rank' : world_rank,
                    'world_size' : world_size,
                    'allreduce_post_accumulation' : True,
                    'deepspeed_zero_optimization':
                    {
                        'stage': 1
                    }
                },
                'debug' : {'deterministic_compute': True}
            }
    create_orttrainer_and_save_checkpoint(device, opts, checkpoint_dir, state_dict_key_name='state_dict_'+str(world_rank), use_lamb=False)

示例#7

文件： orttraining_test_save_checkpoint.py 项目： kshama-msft/onnxruntime

def data_parallelism_full_precision(world_rank, world_size, device,
                                    checkpoint_dir):
    opts = {
        "device": {
            "id": device
        },
        "distributed": {
            "world_rank": world_rank,
            "world_size": world_size,
            "allreduce_post_accumulation": True
        },
        "debug": {
            "deterministic_compute": True
        },
    }
    create_orttrainer_and_save_checkpoint(
        device, opts, checkpoint_dir if world_rank == 0 else None)

示例#8

def single_node_full_precision(checkpoint_dir, device='cuda'):
    opts = {'device': {'id': device}, 'debug': {'deterministic_compute': True}}
    create_orttrainer_and_save_checkpoint(device, opts, checkpoint_dir)