class TestFSDPWrap(FSDPTest):
"""
Tests main API for wrapping FSDP, which is to pass auto_wrap_policy into
FSDP constructor.
"""
def setUp(self) -> None:
super().setUp()
class NestedSequentialModel:
@staticmethod
def get_model(cuda=True):
sequential = nn.Sequential(
nn.Linear(5, 5),
nn.Linear(5, 5),
nn.Sequential(nn.Linear(5, 5), nn.Linear(5, 5)),
)
if cuda:
sequential = sequential.cuda()
return sequential
@staticmethod
def verify_model(cls, model):
cls.assertTrue(isinstance(model, FSDP))
cls.assertTrue(isinstance(model.module[0], nn.Linear))
cls.assertTrue(isinstance(model.module[1], nn.Linear))
cls.assertTrue(isinstance(model.module[2], FSDP))
# following modules were not wrapped by the policy.
cls.assertTrue(isinstance(model.module[2].module[0], nn.Linear))
cls.assertTrue(isinstance(model.module[2].module[1], nn.Linear))
def _get_linear(self, fin, fout):
return nn.Linear(fin, fout, bias=False)
def _get_already_wrapped_fsdp(
self, fsdp_init_mode=FSDPInitMode.CUDA_BEFORE, nested=False
) -> FSDP:
fn_self = self
class MyModel(nn.Module):
def __init__(self, nested):
super().__init__()
# TODO: test the various init modes.
move_to_cuda = fsdp_init_mode == FSDPInitMode.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 forward(self, input: torch.Tensor) -> torch.Tensor:
return self.lin3(self.lin2(self.lin1(input)))
model = MyModel(nested=nested)
return model
@skip_if_lt_x_gpu(2)
@parametrize("nested", [True, False])
@parametrize("fsdp_init_mode", [FSDPInitMode.CUDA_AFTER, FSDPInitMode.CUDA_BEFORE])
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, fsdp_auto_wrap_policy=default_auto_wrap_policy)
@skip_if_lt_x_gpu(2)
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=False), CPUOffload(offload_params=True)]
)
@parametrize(
"fsdp_init_mode",
[FSDPInitMode.CUDA_AFTER, FSDPInitMode.CUDA_BEFORE]
)
def test_main_wrap_api(self, cpu_offload, fsdp_init_mode):
if fsdp_init_mode == FSDPInitMode.CUDA_AFTER and cpu_offload.offload_params:
# they don't work together, expected
return
move_to_cuda = fsdp_init_mode == FSDPInitMode.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,
fsdp_auto_wrap_policy=functools.partial(
default_auto_wrap_policy,
min_num_params=0, # wrap all modules
),
cpu_offload=cpu_offload,
)
if fsdp_init_mode == FSDPInitMode.CUDA_AFTER:
wrapped_model = wrapped_model.cuda()
modules = [
wrapped_model,
wrapped_model.module.lin1,
wrapped_model.module.lin2,
wrapped_model.module.lin3,
wrapped_model.module.lin4,
# Nested FSDP
wrapped_model.module.lin4.module.nested_lin,
]
for module in modules:
self.assertTrue(isinstance(module, FSDP))
self._check_cpu_offload(module, cpu_offload)
# 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()
class TestAutoWrap(TestCase):
def setUp(self) -> None:
super().setUp()
# For all the tests here, we use a fake group
self.process_group = DummyProcessGroup(rank=0, size=1)
@parametrize("wrap_method", [WrapMethod.FSDP_CTOR, WrapMethod.WRAP_API])
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,
fsdp_auto_wrap_policy=functools.partial(default_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_wrap_disabled_outside_context(self):
layer = wrap(nn.Linear(5, 5))
self.assertTrue(isinstance(layer, nn.Linear))
def test_wrap_override_defaults(self):
new_process_group = DummyProcessGroup(rank=0, size=2)
with enable_wrap(wrapper_cls=FSDP, process_group=self.process_group):
layer = wrap(nn.Linear(5, 5), process_group=new_process_group)
self.assertTrue(isinstance(layer, FSDP))
self.assertEqual(layer.rank, 0)
self.assertEqual(layer.world_size, 2)
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(
default_auto_wrap_policy, min_num_params=40
)
model = FSDP(
sequential,
process_group=self.process_group,
fsdp_auto_wrap_policy=my_auto_wrap_policy
)
TestFSDPWrap.NestedSequentialModel.verify_model(self, model)
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 default_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(
default_auto_wrap_policy, min_num_params=40
)
model = FSDP(
sequential,
process_group=self.process_group,
fsdp_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_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(
default_auto_wrap_policy, min_num_params=40
)
model = FSDP(sequential, process_group=self.process_group, fsdp_auto_wrap_policy=my_auto_wrap_policy)
self.assertTrue(isinstance(model, FSDP))
self.assertTrue(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. The
default_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(
default_auto_wrap_policy, min_num_params=40
)
model = FSDP(sequential, process_group=self.process_group, fsdp_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_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}
),
)
sequential = nn.Sequential(
nn.Linear(10, 10), nn.ModuleList([nn.Linear(10, 10)])
)
model = FSDP(sequential, process_group=self.process_group, fsdp_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))
@unittest.skipIf(not torch.cuda.is_available(), "Test Requires CUDA")
@parametrize("fsdp_init_mode", [FSDPInitMode.CUDA_BEFORE, FSDPInitMode.CUDA_AFTER])
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=False), CPUOffload(offload_params=True)]
)
def test_auto_wrap_smoke_test(self, fsdp_init_mode, cpu_offload):
# CPU offload and CUDA after don't work together as expected.
if (
cpu_offload.offload_params and fsdp_init_mode == FSDPInitMode.CUDA_AFTER
):
return
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(find_free_port())
torch.distributed.init_process_group(backend="nccl", 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 = fsdp_init_mode == FSDPInitMode.CUDA_AFTER
try:
sequential = TestFSDPWrap.NestedSequentialModel.get_model(cuda=(not cuda_after_init))
my_auto_wrap_policy = functools.partial(
default_auto_wrap_policy, min_num_params=40
)
model = FSDP(sequential, cpu_offload=cpu_offload, fsdp_auto_wrap_policy=my_auto_wrap_policy)
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()
del os.environ["MASTER_ADDR"]
del os.environ["MASTER_PORT"]
class TestParityWithDDP(FSDPTest):
"""
Compare losses and parameter values after several updates when using
PyTorch DDP vs. FullyShardedDataParallel.
"""
def _get_init_modes_for_test(self, cpu_offload):
modes = [FSDPInitMode.CUDA_AFTER, FSDPInitMode.CUDA_BEFORE]
# Note that FSDPInitMode.CUDA_NEVER works currently only with CPU
# offload as we explicitly bring the param back to CUDA device. In
# general, it will not work since we try to all_gather p.data which is
# on CPU but NCCL only supports GPU.
if cpu_offload.offload_params:
modes.append(FSDPInitMode.CUDA_NEVER)
return modes
@skip_if_lt_x_gpu(2)
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=True),
CPUOffload(offload_params=False)])
def test_nested_wrapped_model(self, cpu_offload):
init_modes = self._get_init_modes_for_test(cpu_offload)
for fsdp_init_mode in init_modes:
with self.subTest(fsdp_init_mode=fsdp_init_mode):
self._test_identical_outputs(NestedWrappedModule,
fsdp_init_mode=fsdp_init_mode,
cpu_offload=cpu_offload)
@skip_if_lt_x_gpu(2)
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=True),
CPUOffload(offload_params=False)])
def test_nested_all_wrapped_model(self, cpu_offload):
init_modes = self._get_init_modes_for_test(cpu_offload)
for fsdp_init_mode in init_modes:
with self.subTest(fsdp_init_mode=fsdp_init_mode):
model_fn = functools.partial(NestedWrappedModule,
wrap_everything=True)
self._test_identical_outputs(model_fn,
fsdp_init_mode=fsdp_init_mode,
cpu_offload=cpu_offload)
@skip_if_lt_x_gpu(2)
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=True),
CPUOffload(offload_params=False)])
def test_transformer_parameterized(self, cpu_offload):
init_modes = self._get_init_modes_for_test(cpu_offload)
for fsdp_init_mode in init_modes:
with self.subTest(fsdp_init_mode=fsdp_init_mode):
self._test_identical_outputs(TransformerWithSharedParams,
fsdp_init_mode=fsdp_init_mode,
cpu_offload=cpu_offload)
@skip_if_lt_x_gpu(2)
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=True),
CPUOffload(offload_params=False)])
def test_delayed_optim_step(self, cpu_offload):
# We use a model with a long CUDA delay right before the optimizer step.
# This tests our streams logic, and that we don't start the allgather
# until after the optimization step completes.
init_modes = self._get_init_modes_for_test(cpu_offload)
for fsdp_init_mode in init_modes:
with self.subTest(fsdp_init_mode=fsdp_init_mode):
model_fn = functools.partial(NestedWrappedModuleWithDelay,
delay_after_loss_ms=250)
self._test_identical_outputs(model_fn,
fsdp_init_mode=fsdp_init_mode,
cpu_offload=cpu_offload)
@skip_if_lt_x_gpu(2)
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=True),
CPUOffload(offload_params=False)])
def test_delayed_reduce_scatter(self, cpu_offload):
# We insert a delay in the torch.distributed._reduce_scatter_base op, so that
# the post_backward_stream takes much longer than the backward pass.
# This tests that we properly block at the end of the backward pass for
# the reductions to finish.
init_modes = self._get_init_modes_for_test(cpu_offload)
for fsdp_init_mode in init_modes:
with self.subTest(fsdp_init_mode=fsdp_init_mode):
model_fn = functools.partial(NestedWrappedModuleWithDelay,
delay_before_reduction_ms=250)
self._test_identical_outputs(model_fn,
fsdp_init_mode=fsdp_init_mode,
cpu_offload=cpu_offload)
def _dummy_ddp_fn(self, model):
return DummyDDP(model)
@skip_if_lt_x_gpu(2)
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=True),
CPUOffload(offload_params=False)])
def test_mixture_of_experts(self, cpu_offload):
init_modes = self._get_init_modes_for_test(cpu_offload)
for fsdp_init_mode in init_modes:
with self.subTest(fsdp_init_mode=fsdp_init_mode):
self._test_identical_outputs(
MixtureOfExperts,
# MixtureOfExperts implements custom reduce logic, so the reference
# behavior should use that logic instead of PyTorch DDP.
ref_ddp_fn=self._dummy_ddp_fn,
fsdp_init_mode=fsdp_init_mode,
cpu_offload=cpu_offload,
)
@skip_if_lt_x_gpu(2)
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=True),
CPUOffload(offload_params=False)])
def test_mixture_of_experts_with_delay_before_free(self, cpu_offload):
init_modes = self._get_init_modes_for_test(cpu_offload)
for fsdp_init_mode in init_modes:
with self.subTest(fsdp_init_mode=fsdp_init_mode):
model_fn = functools.partial(MixtureOfExperts,
delay_before_free_ms=250)
self._test_identical_outputs(
model_fn,
ref_ddp_fn=self._dummy_ddp_fn,
fsdp_init_mode=fsdp_init_mode,
cpu_offload=cpu_offload,
)
def _test_identical_outputs(self,
model_init_fn,
*args,
ref_ddp_fn=None,
num_steps=2,
fsdp_init_mode=FSDPInitMode.CUDA_AFTER,
lr=0.01,
cpu_offload=CPUOffload(),
**kwargs):
group = dist.distributed_c10d._get_default_group()
rank = group.rank()
# Establish reference behavior with PyTorch DDP (+ optionally autocast).
model = model_init_fn(group=group, wrap_fsdp=False).cuda()
if ref_ddp_fn is None:
model = nn.parallel.DistributedDataParallel(model,
device_ids=[rank],
output_device=rank)
else:
model = ref_ddp_fn(model)
ref_loss = self._train_for_several_steps(model,
num_steps,
autocast=False,
lr=lr,
fsdp_cpu_offload=cpu_offload)
ref_full_params = list(model.parameters())
# Confirm we get the same behavior using FullyShardedDataParallel.
try:
model = model_init_fn(group=group,
wrap_fsdp=True,
fsdp_init_mode=fsdp_init_mode,
cpu_offload=cpu_offload)
except Exception as e:
raise ValueError(
f"model_Init_fn {model_init_fn} got error {str(e)}")
cpu_offload = cpu_offload or CPUOffload() # disabled if not specified.
model = FullyShardedDataParallel(model, cpu_offload=cpu_offload)
# Call model.cuda() after init FSDP if specified.
if fsdp_init_mode == FSDPInitMode.CUDA_AFTER:
model = model.cuda()
# Note that we don't do this check for FSDPInitMode.CUDA_AFTER since we
# expect FSDP code to raise error that we check below, in the case of
# offload params.
if fsdp_init_mode != FSDPInitMode.CUDA_AFTER and cpu_offload.offload_params:
for p in model.parameters():
# Should be on CPU regardless of if param is sharded.
self.assertEqual(p.device, torch.device("cpu"),
f"Mismatch, cpu offload is {cpu_offload}")
only_check_err = fsdp_init_mode == FSDPInitMode.CUDA_AFTER and cpu_offload.offload_params
ctx = (self.assertRaisesRegex(AssertionError,
"Expected param to be on CPU")
if only_check_err else suppress())
with ctx:
shard_loss = self._train_for_several_steps(
model,
num_steps,
autocast=False,
lr=lr,
fsdp_cpu_offload=cpu_offload,
)
# We only check for errors in the case we have the following setup:
# model = FSDP(model, cpu_offload=True)
# model = model.cuda()
# so skip the rest of this logic.
if only_check_err:
return
# If CPU offload, next call will change model params to GPU. Sanity
# check that params are on CPU before.
if cpu_offload.offload_params:
device_set = {p.device for p in model.parameters()}
self.assertEqual({torch.device("cpu")}, device_set,
f"Got device set {device_set}")
get_full_params(model)
shard_full_params = list(model.parameters())
if cpu_offload.offload_params:
shard_loss = shard_loss.cuda()
torch.testing.assert_allclose(ref_loss, shard_loss)
self.assertEqual(
ref_full_params,
shard_full_params,
exact_device=True,
msg="FullyShardedDataParallel didn't match PyTorch DDP",
)
class TestFSDPCheckpoint(FSDPTest):
class SequentialModule(nn.Module):
def __init__(self, checkpoint_layer=False, wrap_fsdp=False, *fsdp_args, **fsdp_kwargs):
torch.manual_seed(0)
torch.cuda.manual_seed(0)
super().__init__()
l1 = nn.Linear(3, 3).cuda()
l2 = nn.Linear(3, 3).cuda()
l3 = nn.Linear(3, 3).cuda()
if checkpoint_layer:
l1 = checkpoint_wrapper(l1)
l2 = checkpoint_wrapper(l2)
l3 = checkpoint_wrapper(l3)
fsdp_wrapper = partial(
_maybe_wrap_fsdp,
wrap_fsdp=wrap_fsdp,
*fsdp_args,
**fsdp_kwargs
)
self.ffn = nn.Sequential(
fsdp_wrapper(l1),
fsdp_wrapper(l2),
fsdp_wrapper(l3),
)
def forward(self, x):
return self.ffn(x)
def _verify_parity(self, losses, outputs, models):
assert losses
assert outputs
assert models
for (l, o) in zip(losses[1:], outputs[1:]):
self.assertEqual(losses[0], l)
self.assertEqual(outputs[0], o)
# Verify grads
ref_model = models[0]
ref_grads = [p.grad for p in ref_model.parameters()]
for m in models[1:]:
grads = [p.grad for p in m.parameters()]
for ref_g, g in zip(ref_grads, grads):
self.assertEqual(ref_g, g)
@skip_if_lt_x_gpu(2)
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=True), CPUOffload(offload_params=False)]
)
def test_checkpoint_fsdp_wrapping(self, cpu_offload):
# Test checkpoint(FSDP(layer1), FSDP(layer2), ....)
ckpt_sequential_wrapped_fsdp = checkpoint_wrapper(
TestFSDPCheckpoint.SequentialModule(
wrap_fsdp=True, cpu_offload=cpu_offload
)
)
# Test FSDP(checkpoint(layer1)), FSDP(checkpoint(layer2)), ....
inner_ckpt = TestFSDPCheckpoint.SequentialModule(
checkpoint_layer=True, wrap_fsdp=True, cpu_offload=cpu_offload
)
baseline = TestFSDPCheckpoint.SequentialModule(
wrap_fsdp=True, 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 = [
ckpt_sequential_wrapped_fsdp,
inner_ckpt,
baseline
]
for _ in range(2):
losses = []
outputs = []
for m in models:
out = m(inp)
loss = out.sum()
loss.backward()
losses.append(loss)
outputs.append(out)
self._verify_parity(losses, outputs, models)
@skip_if_lt_x_gpu(2)
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=True), CPUOffload(offload_params=False)]
)
def test_basic_checkpoint_end_to_end(self, cpu_offload):
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))
# Runs FSDP-wrapped checkpointed module
fsdp_wrapped_checkpoint = FSDP(checkpoint_wrapper(deepcopy(seq)), 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
]
for _ in range(6):
losses = []
outputs = []
for m in models:
if m == fsdp_call_checkpoint:
out = checkpoint(m, inp)
else:
out = m(inp)
loss = out.sum()
loss.backward()
losses.append(loss)
outputs.append(out)
self._verify_parity(losses, outputs, models)
class TestFSDPCheckpoint(FSDPTest):
class SequentialModule(nn.Module):
def __init__(
self,
checkpoint_layer=False,
offload_activations=False,
wrap_fsdp=False,
*fsdp_args,
**fsdp_kwargs,
):
torch.manual_seed(0)
torch.cuda.manual_seed(0)
super().__init__()
l1 = nn.Linear(3, 3).cuda()
l2 = nn.Linear(3, 3).cuda()
l3 = nn.Linear(3, 3).cuda()
if checkpoint_layer:
ckpt_wrapper = partial(
checkpoint_wrapper, offload_to_cpu=offload_activations
)
l1 = ckpt_wrapper(l1)
l2 = ckpt_wrapper(l2)
l3 = ckpt_wrapper(l3)
fsdp_wrapper = partial(
_maybe_wrap_fsdp, wrap_fsdp=wrap_fsdp, *fsdp_args, **fsdp_kwargs
)
self.ffn = nn.Sequential(
fsdp_wrapper(l1),
fsdp_wrapper(l2),
fsdp_wrapper(l3),
)
def forward(self, x):
return self.ffn(x)
def _verify_parity(self, losses, outputs, models):
assert losses
assert outputs
assert models
for (l, o) in zip(losses[1:], outputs[1:]):
self.assertEqual(losses[0], l)
self.assertEqual(outputs[0], o)
# Verify grads
ref_model = models[0]
ref_grads = [p.grad for p in ref_model.parameters()]
for m in models[1:]:
grads = [p.grad for p in m.parameters()]
for ref_g, g in zip(ref_grads, grads):
self.assertEqual(ref_g, g)
@skip_if_lt_x_gpu(2)
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=True), CPUOffload(offload_params=False)],
)
@parametrize("offload_activations", [True, False])
def test_checkpoint_fsdp_wrapping(self, cpu_offload, offload_activations):
# Test checkpoint(FSDP(layer1), FSDP(layer2), ....)
ckpt_sequential_wrapped_fsdp = checkpoint_wrapper(
TestFSDPCheckpoint.SequentialModule(
wrap_fsdp=True, cpu_offload=cpu_offload
),
offload_to_cpu=offload_activations,
)
# Test FSDP(checkpoint(layer1)), FSDP(checkpoint(layer2)), ....
inner_ckpt = TestFSDPCheckpoint.SequentialModule(
checkpoint_layer=True,
offload_activations=offload_activations,
wrap_fsdp=True,
cpu_offload=cpu_offload,
)
baseline = TestFSDPCheckpoint.SequentialModule(
wrap_fsdp=True, 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 = [ckpt_sequential_wrapped_fsdp, inner_ckpt, baseline]
offload_to_cpu_event = "Memcpy DtoH"
for i in range(2):
losses = []
outputs = []
for m in models:
check_offload = m != baseline 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:
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)
@skip_if_lt_x_gpu(2)
@parametrize(
"cpu_offload",
[CPUOffload(offload_params=True), CPUOffload(offload_params=False)],
)
@parametrize("offload_activations", [True, False])
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"
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)