def test_basic_checkpoint_end_to_end(self, cpu_offload, offload_activations):
global _save_on_cpu_called
with patch_save_on_cpu(get_patched_save_on_cpu()):
seq = TestFSDPCheckpoint.SequentialModule().to(torch.cuda.current_device())
# Runs FSDP with no checkpointing
fsdp_only_seq = FSDP(deepcopy(seq), cpu_offload=cpu_offload)
# Runs checkpoint-wrapped FSDP
checkpointed_fsdp = checkpoint_wrapper(
FSDP(deepcopy(seq), cpu_offload=cpu_offload),
offload_to_cpu=offload_activations,
)
# Runs FSDP-wrapped checkpointed module
fsdp_wrapped_checkpoint = FSDP(
checkpoint_wrapper(deepcopy(seq), offload_to_cpu=offload_activations),
cpu_offload=cpu_offload,
)
# Runs FSDP with manual calls to checkpoint.
fsdp_call_checkpoint = FSDP(deepcopy(seq), cpu_offload=cpu_offload)
# note that reentrant-based checkpointing requires inputs to have grad
# flag set.
inp = torch.randn(10, 3, device=torch.cuda.current_device(), requires_grad=True)
models = [
fsdp_only_seq,
checkpointed_fsdp,
fsdp_wrapped_checkpoint,
fsdp_call_checkpoint,
]
# Ensure _save_on_cpu is not yet called
self.assertFalse(_save_on_cpu_called)
for i in range(6):
losses = []
outputs = []
for m in models:
check_offload = m != fsdp_only_seq and i == 0 and offload_activations
if m == fsdp_call_checkpoint:
# _save_on_cpu should not be called yet
self.assertFalse(_save_on_cpu_called)
offload_ctx = (
get_patched_save_on_cpu()(pin_memory=True)
if offload_activations
else contextlib.suppress()
)
with offload_ctx:
out = checkpoint(m, inp)
else:
# _save_on_cpu should not be called yet
self.assertFalse(_save_on_cpu_called)
out = m(inp)
if check_offload:
self.assertTrue(_save_on_cpu_called)
loss = out.sum()
loss.backward()
losses.append(loss)
outputs.append(out)
_save_on_cpu_called = False
self._verify_parity(losses, outputs, models)
def test_auto_wrap_with_ignored_modules(self, wrap_method: WrapMethod):
sequential = TestFSDPWrap.NestedSequentialModel.get_model(cuda=False)
ignored_modules = [sequential[1], sequential[2][0]]
my_auto_wrap_policy = functools.partial(
size_based_auto_wrap_policy,
min_num_params=40,
)
fsdp_kwargs = {
"process_group": self.process_group,
"auto_wrap_policy": my_auto_wrap_policy,
"ignored_modules": ignored_modules,
}
if wrap_method == WrapMethod.FSDP_CTOR:
model = FSDP(sequential, **fsdp_kwargs)
elif wrap_method == WrapMethod.WRAP_API:
with enable_wrap(wrapper_cls=FSDP, **fsdp_kwargs):
model = wrap(sequential)
else:
assert 0, f"Unsupported wrap method: {wrap_method}"
# Since the 2nd linear (`sequential[1]`) is ignored, the wrapping
# policy does not exceed the parameter threshold before the inner
# sequential (`sequential[2]`) anymore; hence, it flattens
# `sequential[0]` and `sequential[2][0]` into `model` and leaves
# `sequential[1]` and `sequential[2][1]` as-is since they are ignored
self.assertTrue(isinstance(model, FSDP))
self.assertTrue(isinstance(model.module[0], nn.Linear))
self.assertTrue(isinstance(model.module[1], nn.Linear))
self.assertTrue(isinstance(model.module[2], nn.Sequential))
self.assertTrue(isinstance(model.module[2][0], nn.Linear))
self.assertTrue(isinstance(model.module[2][1], nn.Linear))
def _run_fsdp_one_iteration(self, norm_type, nested_fsdp, cpu_offload):
"""Test FSDP with clip grad norm."""
fsdp_model = DeterministicModel(nested_fsdp, cpu_offload=cpu_offload)
local_model = DeterministicModel(False)
input = torch.rand(14, 2, device=self.rank)
fsdp_model = FSDP(fsdp_model, cpu_offload=cpu_offload)
self.assertTrue(len(input) >= self.world_size)
out = local_model(input[:self.world_size])
out.sum().backward()
in_data = torch.tensor(input[self.rank], device=self.rank)
out_fsdp = fsdp_model(in_data)
out_fsdp.sum().backward()
total_norms_fsdp = _collect_total_grad_norm_fsdp(
fsdp_model, norm_type, self.rank)
total_norms_local = _collect_total_grad_norm_local(
local_model, norm_type)
total_norms_local /= self.world_size
norm_cap = total_norms_fsdp / 2.0
self.assertEqual(total_norms_local, total_norms_fsdp)
fsdp_model.clip_grad_norm_(norm_cap, norm_type=norm_type)
nn_utils.clip_grad_norm_(local_model.parameters(),
norm_cap,
norm_type=norm_type)
total_norms_after_clip_fsdp = _collect_total_grad_norm_fsdp(
fsdp_model, norm_type, self.rank)
total_norms_after_clip_local = _collect_total_grad_norm_local(
local_model, norm_type)
self.assertTrue(total_norms_after_clip_fsdp <= norm_cap)
self.assertEqual(total_norms_after_clip_local,
total_norms_after_clip_fsdp)
def test_fsdp_calc_grad_norm(self, norm_type, nested_fsdp):
"""Test grad norm cal API."""
model = FSDP(DeterministicModel(nested_fsdp))
input = torch.rand(15, 2, device=self.rank)
out = model(input)
out.sum().backward()
total_norm = _calc_grad_norm(model.params_with_grad, norm_type)
total_norm_expected = _collect_total_grad_norm_local(model, norm_type)
self.assertEqual(total_norm, total_norm_expected)
def test_bn_always_wrapped_individually(self):
"""
Ensures that by using _or_policy with _wrap_batchnorm_individually, even
if the other policy results in a module containing a BN unit being
wrapped, the contained BN unit will still be individually wrapped.
"""
class MyModule(nn.Module):
def __init__(self):
super().__init__()
self.bn_container = BatchNormNet()
def wrap_bn_container(module, recurse, *args, **kwargs):
if recurse:
return True
return isinstance(module, BatchNormNet)
my_policy = functools.partial(
_or_policy,
policies=[wrap_bn_container, _wrap_batchnorm_individually])
mod = MyModule()
fsdp = FSDP(mod, auto_wrap_policy=my_policy)
# Wrapping should be FSDP(FSDP(BatchNormNet(FSDP(BN))))
# and not FSDP(FSDP(BatchNormNet(BN))) (in the latter the inner
# BN is not individually wrapped.)
for bn in [
fsdp.bn_container.bn1, fsdp.bn_container.bn2,
fsdp.bn_container.bn3, fsdp.bn_container.sync_bn
]:
self.assertTrue(isinstance(bn, FSDP))
# if we just wrapped BN container, individual batchnorms are not
# wrapped.
mod = MyModule()
fsdp = FSDP(mod, auto_wrap_policy=wrap_bn_container)
self.assertTrue(isinstance(mod.bn_container, FSDP))
for bn in [
fsdp.bn_container.bn1, fsdp.bn_container.bn2,
fsdp.bn_container.bn3, fsdp.bn_container.sync_bn
]:
self.assertFalse(isinstance(bn, FSDP))
def __init__(self, nested):
super().__init__()
# TODO: test the various init modes.
move_to_cuda = cuda_init_mode == CUDAInitMode.CUDA_BEFORE
# if nested=True, the FSDP module will be nested one layer deep
# and we should pick that up.
if nested:
self.lin1 = nn.Sequential(
_maybe_cuda(fn_self._get_linear(1, 1), move_to_cuda),
FSDP(
_maybe_cuda(fn_self._get_linear(1, 1),
move_to_cuda)),
)
else:
self.lin1 = FSDP(
_maybe_cuda(fn_self._get_linear(1, 1), move_to_cuda))
self.lin2 = FSDP(
_maybe_cuda(fn_self._get_linear(1, 1), move_to_cuda))
self.lin3 = FSDP(
_maybe_cuda(fn_self._get_linear(1, 1), move_to_cuda))
def test_always_wrap(self):
"""
Test to ensure that if `always_wrap_policy` is
passed into FSDP, all submodules are wrapped.
"""
seq = TestFSDPWrap.NestedSequentialModel.get_model(cuda=True)
model = FSDP(seq,
process_group=self.process_group,
auto_wrap_policy=always_wrap_policy)
TestFSDPWrap.NestedSequentialModel.verify_model_all_wrapped(
self, model)
def test_error_already_wrapped(self, nested, fsdp_init_mode):
"""
Test that an error is raised if we attempt to wrap when submodules are
already FSDP.
"""
wrapped_fsdp = self._get_already_wrapped_fsdp(
nested=nested, fsdp_init_mode=fsdp_init_mode)
if fsdp_init_mode == FSDPInitMode.CUDA_AFTER:
wrapped_fsdp = wrapped_fsdp.cuda()
with self.assertRaisesRegex(ValueError,
"to NOT be FullyShardedDataParallel"):
mod = FSDP(wrapped_fsdp, auto_wrap_policy=default_auto_wrap_policy)
def test_auto_wrap_api(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.
"""
sequential = TestFSDPWrap.NestedSequentialModel.get_model(cuda=False)
my_auto_wrap_policy = functools.partial(size_based_auto_wrap_policy,
min_num_params=40)
model = FSDP(sequential,
process_group=self.process_group,
auto_wrap_policy=my_auto_wrap_policy)
TestFSDPWrap.NestedSequentialModel.verify_model(self, model)
def test_wrap(self, wrap_method):
if wrap_method == WrapMethod.WRAP_API:
with enable_wrap(wrapper_cls=FSDP,
process_group=self.process_group):
layer = wrap(nn.Linear(5, 5))
else:
assert wrap_method == WrapMethod.FSDP_CTOR
layer = FSDP(nn.Linear(5, 5),
process_group=self.process_group,
auto_wrap_policy=functools.partial(
size_based_auto_wrap_policy, min_num_params=1))
self.assertTrue(isinstance(layer, FSDP))
self.assertEqual(layer.rank, self.process_group.rank())
self.assertEqual(layer.world_size, self.process_group.size())
def test_auto_wrap_smoke_test(self, cuda_init_mode, cpu_offload,
use_device_id):
# CPU offload and CUDA after don't work together as expected.
if (cpu_offload.offload_params
and cuda_init_mode == CUDAInitMode.CUDA_AFTER):
return
device = torch.device("cuda")
torch.cuda.set_device(0)
device_id = (torch.device("cuda", torch.cuda.current_device())
if use_device_id else None)
# Random port in case the next test run quickly, same port would cause conflict.
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = str(find_free_port())
file_name = tempfile.NamedTemporaryFile(delete=False).name
torch.distributed.init_process_group(
backend="nccl",
init_method=f"{FILE_SCHEMA}_{file_name}",
rank=0,
world_size=1,
)
# NOTE: We move model to CUDA after init with FSDP to simulate real use
# cases where full model cannot be loaded onto GPU, but their shards can.
cuda_after_init = cuda_init_mode == CUDAInitMode.CUDA_AFTER
try:
sequential = TestFSDPWrap.NestedSequentialModel.get_model(
cuda=(not cuda_after_init))
my_auto_wrap_policy = functools.partial(
size_based_auto_wrap_policy, min_num_params=40)
model = FSDP(sequential,
cpu_offload=cpu_offload,
auto_wrap_policy=my_auto_wrap_policy,
device_id=device_id)
TestFSDPWrap.NestedSequentialModel.verify_model(self, model)
if cuda_after_init:
model = model.cuda()
input = torch.rand((1, 5), dtype=torch.float).to(device)
output = model(input)
loss = F.mse_loss(input, output)
loss.backward()
finally:
torch.distributed.destroy_process_group()
try:
os.remove(file_name)
except FileNotFoundError:
pass
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
"""
sequential = nn.ModuleList([nn.Linear(10, 10)])
my_auto_wrap_policy = functools.partial(size_based_auto_wrap_policy,
min_num_params=40)
model = FSDP(sequential,
process_group=self.process_group,
auto_wrap_policy=my_auto_wrap_policy)
self.assertTrue(isinstance(model, FSDP))
self.assertTrue(isinstance(model[0], FSDP))
def test_transformer_auto_wrap_policy(self):
model = TransformerWithSharedParams(group=self.process_group)
my_auto_wrap_policy = functools.partial(transformer_auto_wrap_policy,
transformer_layer_cls={
TransformerEncoderLayer,
TransformerDecoderLayer
})
fsdp_model = FSDP(model,
process_group=self.process_group,
auto_wrap_policy=my_auto_wrap_policy)
self.assertTrue(isinstance(fsdp_model, FSDP))
for layer in fsdp_model.module.module.transformer.encoder.layers:
self.assertTrue(isinstance(layer, FSDP))
for layer in fsdp_model.module.module.transformer.decoder.layers:
self.assertTrue(isinstance(layer, FSDP))
def test_wrap_batchnorm_individually(self, use_or_policy):
def never_wrap_policy(*args, **kwargs):
return False
policy = (functools.partial(
_or_policy,
policies=[never_wrap_policy, _wrap_batchnorm_individually])
if use_or_policy else _wrap_batchnorm_individually)
model = BatchNormNet()
fsdp = FSDP(model, auto_wrap_policy=policy)
# Batchnorms should be wrapped
for layer in [fsdp.bn1, fsdp.bn2, fsdp.bn3, fsdp.sync_bn]:
self.assertTrue(isinstance(layer, FSDP))
self.assertFalse(isinstance(fsdp.lin, FSDP))
def test_auto_wrap_preset_force_leaf(self):
"""
Test to ensure force-leaf modules are not wrapped, and children are not wrapped. The
size_based_auto_wrap_policy forces leaf modules of type {nn.MultiheadAttention} to not be wrapped
"""
sequential = nn.Sequential(nn.Linear(10, 10),
nn.MultiheadAttention(100, 1))
my_auto_wrap_policy = functools.partial(size_based_auto_wrap_policy,
min_num_params=40)
model = FSDP(sequential,
process_group=self.process_group,
auto_wrap_policy=my_auto_wrap_policy)
self.assertTrue(isinstance(model.module[0], FSDP))
# Assert children of multihead attention are not wrapped
self.assertTrue(isinstance(model.module[1], nn.MultiheadAttention))
self.assertTrue(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. the size_based_auto_wrap_policy excludes wrapping for {nn.ModuleList, nn.ModuleDict}
"""
sequential = nn.ModuleList([nn.Linear(5, 5), nn.Linear(5, 5)])
my_auto_wrap_policy = functools.partial(size_based_auto_wrap_policy,
min_num_params=40)
model = FSDP(sequential,
process_group=self.process_group,
auto_wrap_policy=my_auto_wrap_policy)
self.assertTrue(isinstance(model, FSDP))
self.assertTrue(isinstance(model[0], nn.Linear))
self.assertTrue(isinstance(model[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(
size_based_auto_wrap_policy,
min_num_params=40,
force_leaf_modules=size_based_auto_wrap_policy.FORCE_LEAF_MODULES.
union({nn.Linear}),
)
sequential = nn.Sequential(nn.Linear(10, 10),
nn.ModuleList([nn.Linear(10, 10)]))
model = FSDP(sequential,
process_group=self.process_group,
auto_wrap_policy=my_auto_wrap_policy)
# Model was wrapped in FSDP as no inner modules were wrapped.
self.assertTrue(isinstance(model, FSDP))
self.assertTrue(isinstance(model.module[0], nn.Linear))
self.assertTrue(isinstance(model.module[1], nn.ModuleList))
def test_always_wrap_with_ignored_modules(self, wrap_method: WrapMethod):
sequential = TestFSDPWrap.NestedSequentialModel.get_model(cuda=False)
ignored_modules = [sequential[1], sequential[2][0]]
fsdp_kwargs = {
"process_group": self.process_group,
"auto_wrap_policy": always_wrap_policy,
"ignored_modules": ignored_modules,
}
if wrap_method == WrapMethod.FSDP_CTOR:
model = FSDP(sequential, **fsdp_kwargs)
elif wrap_method == WrapMethod.WRAP_API:
with enable_wrap(wrapper_cls=FSDP, **fsdp_kwargs):
model = wrap(sequential)
else:
assert 0, f"Unsupported wrap method: {wrap_method}"
# All non-ignored modules should be wrapped with FSDP
self.assertTrue(isinstance(model, FSDP))
self.assertTrue(isinstance(model.module[0], FSDP))
self.assertTrue(isinstance(model.module[1], nn.Linear))
self.assertTrue(isinstance(model.module[2], FSDP))
self.assertTrue(isinstance(model.module[2].module[0], nn.Linear))
self.assertTrue(isinstance(model.module[2].module[1], FSDP))
def test_basic_checkpoint_end_to_end(self, cpu_offload,
offload_activations):
seq = TestFSDPCheckpoint.SequentialModule().to(
torch.cuda.current_device())
# Runs FSDP with no checkpointing
fsdp_only_seq = FSDP(deepcopy(seq), cpu_offload=cpu_offload)
# Runs checkpoint-wrapped FSDP
checkpointed_fsdp = checkpoint_wrapper(
FSDP(deepcopy(seq), cpu_offload=cpu_offload),
offload_to_cpu=offload_activations,
)
# Runs FSDP-wrapped checkpointed module
fsdp_wrapped_checkpoint = FSDP(
checkpoint_wrapper(deepcopy(seq),
offload_to_cpu=offload_activations),
cpu_offload=cpu_offload,
)
# Runs FSDP with manual calls to checkpoint.
fsdp_call_checkpoint = FSDP(deepcopy(seq), cpu_offload=cpu_offload)
# note that reentrant-based checkpointing requires inputs to have grad
# flag set.
inp = torch.randn(10,
3,
device=torch.cuda.current_device(),
requires_grad=True)
models = [
fsdp_only_seq,
checkpointed_fsdp,
fsdp_wrapped_checkpoint,
fsdp_call_checkpoint,
]
offload_to_cpu_event = "Memcpy DtoH" if torch.version.cuda else "CopyDeviceToHost"
for i in range(6):
losses = []
outputs = []
for m in models:
check_offload = m != fsdp_only_seq and i == 0 and offload_activations
profiler_ctx = (torch.profiler.profile(
use_cuda=True) if check_offload else contextlib.suppress())
with profiler_ctx as prof:
if m == fsdp_call_checkpoint:
offload_ctx = (torch.autograd.graph.save_on_cpu(
pin_memory=True) if offload_activations else
contextlib.suppress())
with offload_ctx:
out = checkpoint(m, inp)
else:
out = m(inp)
if check_offload:
event_names = [event.name for event in prof.events()]
offload_occured = any(offload_to_cpu_event in name
for name in event_names)
self.assertTrue(offload_occured)
loss = out.sum()
loss.backward()
losses.append(loss)
outputs.append(out)
self._verify_parity(losses, outputs, models)
def test_main_wrap_api(self, cpu_offload, backward_prefetch,
forward_prefetch, cuda_init_mode):
if cuda_init_mode == CUDAInitMode.CUDA_AFTER and cpu_offload.offload_params:
# they don't work together, expected
return
move_to_cuda = cuda_init_mode == CUDAInitMode.CUDA_BEFORE
class Nested(nn.Module):
def __init__(self):
super().__init__()
self.nested_lin = _maybe_cuda(nn.Linear(1, 1, bias=False),
move_to_cuda)
def forward(self, input):
return self.nested_lin(input)
class MyModel(nn.Module):
def __init__(self):
super().__init__()
self.lin1 = _maybe_cuda(nn.Linear(1, 1, bias=False),
move_to_cuda)
self.lin2 = _maybe_cuda(nn.Linear(1, 1, bias=False),
move_to_cuda)
self.lin3 = _maybe_cuda(nn.Linear(1, 1, bias=False),
move_to_cuda)
self.lin4 = Nested()
def forward(self, input):
return self.lin4(self.lin3(self.lin2(self.lin1(input))))
model = MyModel()
wrapped_model = FSDP(
model,
auto_wrap_policy=functools.partial(
size_based_auto_wrap_policy,
min_num_params=0, # wrap all modules
),
cpu_offload=cpu_offload,
backward_prefetch=backward_prefetch,
forward_prefetch=forward_prefetch,
)
if cuda_init_mode == CUDAInitMode.CUDA_AFTER:
wrapped_model = wrapped_model.cuda()
modules_in_fsdp_graph_order = [
wrapped_model.module.lin1, wrapped_model.module.lin2,
wrapped_model.module.lin3,
wrapped_model.module.lin4.module.nested_lin,
wrapped_model.module.lin4, wrapped_model
]
for module in modules_in_fsdp_graph_order:
self.assertTrue(isinstance(module, FSDP))
self._check_cpu_offload(module, cpu_offload)
self._check_backward_prefetch(module, backward_prefetch)
self._check_forward_prefetch(module, forward_prefetch)
# Run model a few times for sanity check.
optim = torch.optim.SGD(wrapped_model.parameters(),
lr=1e-2,
momentum=0.9)
inp = torch.ones(1).cuda()
for _ in range(6):
optim.zero_grad()
loss = wrapped_model(inp).sum()
loss.backward()
optim.step()
# Since we ran with backward prefetch, verify backward prefetch related
# data.
for i, module in enumerate(modules_in_fsdp_graph_order):
self.assertEqual(i, module._my_fsdp_idx_in_graph)
self.assertTrue(
module._fsdp_graph_order == modules_in_fsdp_graph_order)
