torch.distributed

Initialization

表1 Initialization

序号

API名称

是否支持

1

torch.distributed.is_available

2

torch.distributed.init_process_group

3

torch.distributed.is_initialized

4

torch.distributed.is_mpi_available

5

torch.distributed.is_nccl_available

6

torch.distributed.is_torchelastic_launched

Post-Initialization

表2 Post-Initialization

序号

API名称

是否支持

1

torch.distributed.Backend

2

torch.distributed.get_backend

3

torch.distributed.get_rank

4

torch.distributed.get_world_size

Distributed Key-Value Store

表3 Distributed Key-Value Store

序号

API名称

是否支持

1

torch.distributed.Store

2

torch.distributed.TCPStore

3

torch.distributed.HashStore

4

torch.distributed.FileStore

5

torch.distributed.PrefixStore

6

torch.distributed.Store.set

7

torch.distributed.Store.get

8

torch.distributed.Store.add

9

torch.distributed.Store.compare_set

10

torch.distributed.Store.wait

11

torch.distributed.Store.num_keys

12

torch.distributed.Store.delete_key

13

torch.distributed.Store.set_timeout

Groups

表4 Groups

序号

API名称

是否支持

1

torch.distributed.new_group

Point-to-point communication

表5 Point-to-point communication

序号

API名称

是否支持

1

torch.distributed.send

2

torch.distributed.recv

3

torch.distributed.isend

4

torch.distributed.irecv

Synchronous and asynchronous collective operations

表6 Synchronous and asynchronous collective operations

序号

API名称

是否支持

1

is_completed

2

wait

3

get_future

Collective functions

表7 Collective functions

序号

API名称

是否支持

1

torch.distributed.broadcast

2

torch.distributed.broadcast_object_list

3

torch.distributed.all_reduce

4

torch.distributed.reduce

5

torch.distributed.all_gather

6

torch.distributed.all_gather_object

7

torch.distributed.gather

8

torch.distributed.gather_object

9

torch.distributed.scatter

10

torch.distributed.scatter_object_list

11

torch.distributed.reduce_scatter

12

torch.distributed.all_to_all

13

torch.distributed.barrier

14

torch.distributed.monitored_barrier

15

torch.distributed.ReduceOp

16

torch.distributed.reduce_op

Profiling Collective Communication

表8 Profiling Collective Communication

序号

API名称

是否支持

1

torch.profiler

2

torch.autograd.profiler

Multi-GPU collective functions

表9 Multi-GPU collective functions

序号

API名称

是否支持

1

torch.distributed.broadcast_multigpu

2

torch.distributed.all_reduce_multigpu

3

torch.distributed.reduce_multigpu

4

torch.distributed.all_gather_multigpu

5

torch.distributed.reduce_scatter_multigpu

Launch utility

表10 Launch utility

序号

API名称

是否支持

1

torch.distributed.launch

Spawn utility

表11 Spawn utility

序号

API名称

是否支持

1

torch.multiprocessing.spawn