torch.distributed.fsdp
Note
若API“是否支持”为“是”，“限制与说明”为“-”，说明此API和原生API支持度保持一致。
| API名称 | 是否支持 | 限制与说明 |
|---|---|---|
| torch.distributed.fsdp.FullyShardedDataParallel | 是 | 支持bf16，fp16，fp32 在昇腾NPU场景中使用FSDP，推荐传入“device_id=torch.device("npu:0")”设备相关参数 |
| torch.distributed.fsdp.FullyShardedDataParallel.apply | 是 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.check_is_root | 是 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.clip_grad_norm_ | 是 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.flatten_sharded_optim_state_dict | 否 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.forward | 是 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.fsdp_modules | 是 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.full_optim_state_dict | 是 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.get_state_dict_type | 否 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.module | 否 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.named_buffers | 是 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.named_parameters | 是 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.no_sync | 是 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.optim_state_dict | 否 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.optim_state_dict_to_load | 是 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.register_comm_hook | 是 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.rekey_optim_state_dict | 否 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.scatter_full_optim_state_dict | 否 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.set_state_dict_type | 否 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.shard_full_optim_state_dict | 否 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.sharded_optim_state_dict | 否 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.state_dict_type | 是 | - |
| torch.distributed.fsdp.FullyShardedDataParallel.summon_full_params | 否 | - |
| torch.distributed.fsdp.BackwardPrefetch | 是 | - |
| torch.distributed.fsdp.ShardingStrategy | 是 | - |
| torch.distributed.fsdp.MixedPrecision | 是 | - |
| torch.distributed.fsdp.CPUOffload | 是 | - |
| torch.distributed.fsdp.StateDictConfig | 是 | - |
| torch.distributed.fsdp.FullStateDictConfig | 是 | - |
| torch.distributed.fsdp.ShardedStateDictConfig | 是 | - |
| torch.distributed.fsdp.LocalStateDictConfig | 是 | - |
| torch.distributed.fsdp.OptimStateDictConfig | 是 | - |
| torch.distributed.fsdp.FullOptimStateDictConfig | 是 | - |
| torch.distributed.fsdp.ShardedOptimStateDictConfig | 是 | - |
| torch.distributed.fsdp.LocalOptimStateDictConfig | 是 | - |
| torch.distributed.fsdp.StateDictSettings | 是 | - |