MCPcopy
hub / github.com/meta-pytorch/opacus / demo_basic

Function demo_basic

opacus/tests/multigpu_fsdpcheck.py:70–146  ·  view source on GitHub ↗
(rank, weight, world_size, grad_sample_mode, mixed_precision)

Source from the content-addressed store, hash-verified

68
69
70def demo_basic(rank, weight, world_size, grad_sample_mode, mixed_precision):
71 torch.manual_seed(world_size)
72 batch_size = 32
73 torch.cuda.set_device(rank)
74 setup(rank, world_size)
75
76 # create model and move it to GPU with id rank
77 model = ToyModel().to(rank)
78 model.net1.weight.data.zero_()
79
80 # create dataset
81 labels = torch.randn(2 * batch_size, 5).to(rank)
82 data = torch.randn(2 * batch_size, 10)
83 dataset = TensorDataset(data, labels)
84 max_grad_norm = 1
85 # we set the seed to be same for all workers, so the noise generated on rank 0 for DP-DDP should match the noise generated on all the workers for FSDP
86 noise_multiplier = 5.0
87
88 if grad_sample_mode == "ghost":
89 dp_model = DPDDP(model)
90 else:
91 if not mixed_precision:
92 dp_model = FSDP2Wrapper(model)
93 else:
94 dp_model = FSDP2Wrapper(
95 model,
96 mp_policy=dist.fsdp.MixedPrecisionPolicy(
97 param_dtype=torch.bfloat16, reduce_dtype=torch.float32
98 ),
99 opacus_high_precision_layers=(nn.LayerNorm,),
100 )
101
102 optimizer = optim.SGD(model.parameters(), lr=1)
103
104 privacy_engine = PrivacyEngine()
105
106 sampler = DistributedSampler(
107 dataset, num_replicas=world_size, rank=rank, shuffle=False
108 )
109 data_loader = DataLoader(dataset, batch_size=batch_size, sampler=sampler)
110
111 dp_model, optimizer, loss_fn, data_loader = privacy_engine.make_private(
112 module=dp_model,
113 optimizer=optimizer,
114 criterion=nn.CrossEntropyLoss(),
115 data_loader=data_loader,
116 noise_multiplier=noise_multiplier,
117 max_grad_norm=max_grad_norm,
118 poisson_sampling=False,
119 grad_sample_mode=grad_sample_mode,
120 )
121 if grad_sample_mode == "ghost" and mixed_precision is True:
122 for x, y in data_loader:
123 with torch.amp.autocast("cuda", dtype=torch.bfloat16):
124 outputs = dp_model(x.to(rank))
125 assert outputs.dtype == torch.bfloat16
126 loss = loss_fn(outputs, y)
127 optimizer.zero_grad()

Callers

nothing calls this directly

Calls 10

make_privateMethod · 0.95
FSDP2WrapperFunction · 0.90
PrivacyEngineClass · 0.90
toMethod · 0.80
setupFunction · 0.70
ToyModelClass · 0.70
cleanupFunction · 0.70
zero_gradMethod · 0.45
backwardMethod · 0.45
stepMethod · 0.45

Tested by

no test coverage detected