def _auto_wrap_smoke_test(self, enable_mixed_precision): device = torch.device("cuda") torch.cuda.set_device(0) # Random port in case the next test run quickly, same port would cause conflict. os.environ["MASTER_ADDR"] = "localhost" os.environ["MASTER_PORT"] = str(random.randint(2000, 3000)) torch.distributed.init_process_group(backend="nccl", rank=0, world_size=1) try: with enable_wrap(wrapper_cls=FSDP, mixed_precision=enable_mixed_precision): sequential = nn.Sequential( nn.Linear(5, 5), nn.Linear(5, 5), nn.Sequential(nn.Linear(5, 5), nn.Linear(5, 5)) ) my_auto_wrap_policy = functools.partial(default_auto_wrap_policy, min_num_params=40) model = auto_wrap(sequential, auto_wrap_policy=my_auto_wrap_policy) model.to(device) input = torch.rand((1, 5), dtype=torch.float).to(device) with autocast(enabled=enable_mixed_precision): output = model(input) loss = F.mse_loss(input, output) loss.backward() finally: torch.distributed.destroy_process_group() del os.environ["MASTER_ADDR"] del os.environ["MASTER_PORT"]
def auto_wrap_big_layers(module: nn.Module, fsdp_config: AttrDict): """ Automatically wrap the bigger layer in the module """ with enable_wrap(auto_wrap_policy=_BigConvAutoWrapPolicy( fsdp_config.AUTO_WRAP_THRESHOLD), wrapper_cls=_FSDP_WRAPPER, **fsdp_config): return auto_wrap(module)
def test_auto_wrap_preset_exclude_wrap_include_children(self): """ Test to ensure excluded modules are not wrapped, but children are if param size is greater than min_num_params """ with enable_wrap(wrapper_cls=FSDP, process_group=self.process_group, flatten_parameters=False): sequential = nn.ModuleList([nn.Linear(10, 10)]) my_auto_wrap_policy = functools.partial(default_auto_wrap_policy, min_num_params=40) model = auto_wrap(sequential, auto_wrap_policy=my_auto_wrap_policy) assert isinstance(model, nn.ModuleList) assert isinstance(model[0], FSDP)
def test_auto_wrap_preset_force_leaf(self): """ Test to ensure force-leaf modules are not wrapped, and children are not wrapped. """ with enable_wrap(wrapper_cls=FSDP, process_group=self.process_group, flatten_parameters=False): sequential = nn.Sequential(nn.Linear(10, 10), nn.MultiheadAttention(100, 1)) my_auto_wrap_policy = functools.partial(default_auto_wrap_policy, min_num_params=40) model = auto_wrap(sequential, auto_wrap_policy=my_auto_wrap_policy) assert isinstance(model.module[0], FSDP) # Assert children of multihead attention are not wrapped assert isinstance(model.module[1], nn.MultiheadAttention) assert isinstance(model.module[1].out_proj, nn.Linear)
def test_auto_wrap_preset_exclude_wrap(self): """ Test to ensure excluded modules are not wrapped, regardless if the total param size is greater than the min_num_params. """ with enable_wrap(wrapper_cls=FSDP, process_group=self.process_group, flatten_parameters=False): sequential = nn.ModuleList([nn.Linear(5, 5), nn.Linear(5, 5)]) my_auto_wrap_policy = functools.partial(default_auto_wrap_policy, min_num_params=40) model = auto_wrap(sequential, auto_wrap_policy=my_auto_wrap_policy) assert isinstance(model, nn.ModuleList) assert isinstance(model[0], nn.Linear) assert isinstance(model[1], nn.Linear)
def test_auto_wrap_preset_blocklist(self): """ Test to ensure blocklisted modules are not wrapped, and children are not wrapped. """ with enable_wrap(process_group=self.process_group, flatten_parameters=False): sequential = nn.Sequential(nn.Linear(10, 10), nn.MultiheadAttention(100, 1)) model = auto_wrap(sequential, min_num_params=40) assert isinstance(model.module[0], FSDP) # Assert children of multihead attention are not wrapped assert isinstance(model.module[1], nn.MultiheadAttention) assert isinstance(model.module[1].out_proj, nn.Linear)
def test_auto_wrap_preset_blocklist_custom(self): """ Test to ensure blocklisted modules are not wrapped. """ with enable_wrap(module_blocklist=[nn.Linear], process_group=self.process_group, flatten_parameters=False): sequential = nn.Sequential(nn.Linear(10, 10), nn.ModuleList([nn.Linear(10, 10)])) model = auto_wrap(sequential, min_num_params=40) # Model was wrapped in FSDP as no inner modules were wrapped. assert isinstance(model, FSDP) assert isinstance(model.module[0], nn.Linear) assert isinstance(model.module[1], nn.ModuleList)
def test_auto_wrap(self): """ Test to ensure with auto wrap, we wrap child modules correctly based on the min_num_params. ``nn.Linear(5, 5)`` does not exceed the bucket size, but combined they do. """ with enable_wrap(wrapper_cls=FSDP, process_group=self.process_group, flatten_parameters=False): sequential = nn.Sequential( nn.Linear(5, 5), nn.Linear(5, 5), nn.Sequential(nn.Linear(5, 5), nn.Linear(5, 5)) ) my_auto_wrap_policy = functools.partial(default_auto_wrap_policy, min_num_params=40) model = auto_wrap(sequential, auto_wrap_policy=my_auto_wrap_policy) assert isinstance(model, FSDP) assert isinstance(model.module[0], nn.Linear) assert isinstance(model.module[1], nn.Linear) assert isinstance(model.module[2], FSDP) assert isinstance(model.module[2].module[0], nn.Linear) assert isinstance(model.module[2].module[1], nn.Linear)
def test_auto_wrap_preset_force_leaf_custom(self): """ Test to ensure force-leaf modules are not wrapped. """ my_auto_wrap_policy = functools.partial( default_auto_wrap_policy, min_num_params=40, force_leaf_modules=default_auto_wrap_policy.FORCE_LEAF_MODULES.union({nn.Linear}), ) with enable_wrap( auto_wrap_policy=my_auto_wrap_policy, wrapper_cls=FSDP, process_group=self.process_group, flatten_parameters=False, ): sequential = nn.Sequential(nn.Linear(10, 10), nn.ModuleList([nn.Linear(10, 10)])) model = auto_wrap(sequential) # Model was wrapped in FSDP as no inner modules were wrapped. assert isinstance(model, FSDP) assert isinstance(model.module[0], nn.Linear) assert isinstance(model.module[1], nn.ModuleList)
def _auto_wrap_smoke_test(self, enable_mixed_precision): from torch.cuda.amp import autocast device = torch.device("cuda") torch.cuda.set_device(0) torch.distributed.init_process_group(backend="nccl", rank=0, world_size=1) with enable_wrap(mixed_precision=enable_mixed_precision): sequential = nn.Sequential( nn.Linear(5, 5), nn.Linear(5, 5), nn.Sequential(nn.Linear(5, 5), nn.Linear(5, 5))) model = auto_wrap(sequential, min_num_params=40) model.to(device) input = torch.rand((1, 5), dtype=torch.float).to(device) with autocast(enabled=enable_mixed_precision): output = model(input) loss = F.mse_loss(input, output) loss.backward() torch.distributed.destroy_process_group()