Skip to content

Fix FSDP CPU offloading when shards are already on CPU #21805

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
zhixiangli wants to merge 1 commit into
base: master
Choose a base branch
from
+18 −8
Open

Fix FSDP CPU offloading when shards are already on CPU #21805

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 5 additions & 2 deletions src/lightning/fabric/strategies/fsdp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -895,8 +895,11 @@ def _get_full_state_dict_context(
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from torch.distributed.fsdp.api import FullOptimStateDictConfig

state_dict_config = FullStateDictConfig(offload_to_cpu=True, rank0_only=rank0_only)
optim_state_dict_config = FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=rank0_only)
# Offloading to CPU when FSDP is on CPU triggers a use-after-free in PyTorch's FlatParamHandle.to_cpu().
param = next(module.parameters(), None)
offload_to_cpu = param is None or param.device.type != "cpu"
state_dict_config = FullStateDictConfig(offload_to_cpu=offload_to_cpu, rank0_only=rank0_only)
optim_state_dict_config = FullOptimStateDictConfig(offload_to_cpu=offload_to_cpu, rank0_only=rank0_only)
state_dict_type_context = FSDP.state_dict_type(
module=module,
state_dict_type=StateDictType.FULL_STATE_DICT,
Expand Down
19 changes: 13 additions & 6 deletions tests/tests_fabric/strategies/test_fsdp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -433,17 +433,24 @@ def test_set_timeout(init_process_group_mock):

@mock.patch("torch.distributed.fsdp.fully_sharded_data_parallel.FullyShardedDataParallel.set_state_dict_type")
def test_get_full_state_dict_context_offload(set_type_mock, monkeypatch):
"""Test that the state dict context manager handles CPU offloading."""

with _get_full_state_dict_context(module=Mock(spec=FullyShardedDataParallel), world_size=1):
"""Test that the state dict context manager only offloads to CPU when the shards live on an accelerator."""
# Shards on accelerator: offload to CPU.
accelerator_param = Mock()
accelerator_param.device = torch.device("cuda", 0)
module = Mock(spec=FullyShardedDataParallel)
module.parameters = Mock(return_value=iter([accelerator_param]))
with _get_full_state_dict_context(module=module, world_size=4):
assert set_type_mock.call_args_list[0][0][2].offload_to_cpu # model config
assert set_type_mock.call_args_list[0][0][3].offload_to_cpu # optim config

set_type_mock.reset_mock()

with _get_full_state_dict_context(module=Mock(spec=FullyShardedDataParallel), world_size=4):
assert set_type_mock.call_args_list[0][0][2].offload_to_cpu # model config
assert set_type_mock.call_args_list[0][0][3].offload_to_cpu # optim config
# Shards on CPU: do not offload (prevents PyTorch use-after-free).
module = Mock(spec=FullyShardedDataParallel)
module.parameters = Mock(return_value=iter([torch.nn.Parameter(torch.zeros(1))]))
with _get_full_state_dict_context(module=module, world_size=4):
assert not set_type_mock.call_args_list[0][0][2].offload_to_cpu # model config
assert not set_type_mock.call_args_list[0][0][3].offload_to_cpu # optim config


def test_device_mesh_type_annotation():
Expand Down
Loading