[Bugfix] Preserve unloaded non-persistent buffers during layerwise reload (#44371)
Signed-off-by: Joan Velja <joan.velja22@gmail.com> Co-authored-by: Dakai An <77474977+andakai@users.noreply.github.com>
This commit is contained in:
@@ -71,6 +71,18 @@ class _ParentAliasedChildBufferLayer(torch.nn.Module):
|
||||
)
|
||||
|
||||
|
||||
class _ChildAliasOnlyBufferLayer(torch.nn.Module):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.conv1d = torch.nn.Linear(3, 2, bias=False)
|
||||
self.conv1d.weight.data.copy_(
|
||||
torch.arange(6, dtype=torch.float32).reshape(2, 3)
|
||||
)
|
||||
self.register_buffer(
|
||||
"conv_weights", self.conv1d.weight.detach().view(-1), persistent=False
|
||||
)
|
||||
|
||||
|
||||
class _AliasedBufferWithUninitializedChildLayer(_AliasedBufferLayer):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
@@ -80,6 +92,13 @@ class _AliasedBufferWithUninitializedChildLayer(_AliasedBufferLayer):
|
||||
)
|
||||
|
||||
|
||||
class _NonPersistentBufferLayer(torch.nn.Module):
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.weight = torch.nn.Parameter(torch.ones(2, 2))
|
||||
self.register_buffer("scale", torch.tensor(0.25), persistent=False)
|
||||
|
||||
|
||||
def test_move_metatensors():
|
||||
tensor = torch.empty((1, 2, 3))
|
||||
meta_tensor = to_meta_tensor(tensor)
|
||||
@@ -289,6 +308,8 @@ def test_layerwise_reload_skips_non_persistent_parameter_alias_buffers(monkeypat
|
||||
assert layer.weight_view.untyped_storage().data_ptr() == (
|
||||
layer.weight.untyped_storage().data_ptr()
|
||||
)
|
||||
assert "weight_view" in layer._non_persistent_buffers_set
|
||||
assert "0.weight_view" not in model.state_dict()
|
||||
|
||||
|
||||
def test_capture_layer_to_meta_skips_uninitialized_parameter_storage_ptrs():
|
||||
@@ -332,6 +353,109 @@ def test_layerwise_reload_skips_child_parameter_alias_buffers(monkeypatch):
|
||||
assert layer.conv_weights.untyped_storage().data_ptr() == (
|
||||
layer.conv1d.weight.untyped_storage().data_ptr()
|
||||
)
|
||||
assert "conv_weights" in layer._non_persistent_buffers_set
|
||||
assert "0.conv_weights" not in model.state_dict()
|
||||
|
||||
|
||||
def test_layerwise_reload_restores_alias_buffer_on_zero_size_layer(monkeypatch):
|
||||
layer = _ChildAliasOnlyBufferLayer()
|
||||
model = torch.nn.Sequential(layer)
|
||||
loaded_conv = torch.full_like(layer.conv1d.weight, 7.0)
|
||||
|
||||
def materialize_with_sentinel(meta_tensor):
|
||||
tensor = torch.empty_strided(
|
||||
size=tuple(meta_tensor.size()),
|
||||
stride=tuple(meta_tensor.stride()),
|
||||
dtype=meta_tensor.dtype,
|
||||
requires_grad=False,
|
||||
)
|
||||
tensor.fill_(-123.0)
|
||||
tensor.__class__ = meta_tensor.__class__
|
||||
tensor.__dict__ = meta_tensor.__dict__.copy()
|
||||
return tensor
|
||||
|
||||
monkeypatch.setattr(
|
||||
reload_meta, "materialize_meta_tensor", materialize_with_sentinel
|
||||
)
|
||||
|
||||
record_metadata_for_reloading(model)
|
||||
initialize_layerwise_reload(model)
|
||||
layer.conv1d.weight.weight_loader(layer.conv1d.weight, loaded_conv)
|
||||
finalize_layerwise_reload(model, model_config=None)
|
||||
|
||||
assert torch.equal(layer.conv_weights, loaded_conv.view(-1))
|
||||
assert layer.conv_weights.untyped_storage().data_ptr() == (
|
||||
layer.conv1d.weight.untyped_storage().data_ptr()
|
||||
)
|
||||
assert "conv_weights" in layer._non_persistent_buffers_set
|
||||
assert "0.conv_weights" not in model.state_dict()
|
||||
|
||||
|
||||
def test_layerwise_reload_preserves_unloaded_non_persistent_buffers(monkeypatch):
|
||||
layer = _NonPersistentBufferLayer()
|
||||
model = torch.nn.Sequential(layer)
|
||||
loaded_weight = torch.full_like(layer.weight, 7.0)
|
||||
original_scale = layer.scale.clone()
|
||||
|
||||
def materialize_with_sentinel(meta_tensor):
|
||||
tensor = torch.empty_strided(
|
||||
size=tuple(meta_tensor.size()),
|
||||
stride=tuple(meta_tensor.stride()),
|
||||
dtype=meta_tensor.dtype,
|
||||
requires_grad=False,
|
||||
)
|
||||
tensor.fill_(-123.0)
|
||||
tensor.__class__ = meta_tensor.__class__
|
||||
tensor.__dict__ = meta_tensor.__dict__.copy()
|
||||
return tensor
|
||||
|
||||
monkeypatch.setattr(
|
||||
reload_meta, "materialize_meta_tensor", materialize_with_sentinel
|
||||
)
|
||||
|
||||
record_metadata_for_reloading(model)
|
||||
initialize_layerwise_reload(model)
|
||||
layer.weight.weight_loader(layer.weight, loaded_weight)
|
||||
finalize_layerwise_reload(model, model_config=None)
|
||||
|
||||
assert torch.equal(layer.weight, loaded_weight)
|
||||
assert torch.equal(layer.scale, original_scale)
|
||||
assert "scale" in layer._non_persistent_buffers_set
|
||||
assert "0.scale" not in model.state_dict()
|
||||
|
||||
|
||||
def test_layerwise_reload_updates_loaded_non_persistent_buffers(monkeypatch):
|
||||
layer = _NonPersistentBufferLayer()
|
||||
model = torch.nn.Sequential(layer)
|
||||
loaded_weight = torch.full_like(layer.weight, 7.0)
|
||||
loaded_scale = torch.full_like(layer.scale, 0.5)
|
||||
|
||||
def materialize_with_sentinel(meta_tensor):
|
||||
tensor = torch.empty_strided(
|
||||
size=tuple(meta_tensor.size()),
|
||||
stride=tuple(meta_tensor.stride()),
|
||||
dtype=meta_tensor.dtype,
|
||||
requires_grad=False,
|
||||
)
|
||||
tensor.fill_(-123.0)
|
||||
tensor.__class__ = meta_tensor.__class__
|
||||
tensor.__dict__ = meta_tensor.__dict__.copy()
|
||||
return tensor
|
||||
|
||||
monkeypatch.setattr(
|
||||
reload_meta, "materialize_meta_tensor", materialize_with_sentinel
|
||||
)
|
||||
|
||||
record_metadata_for_reloading(model)
|
||||
initialize_layerwise_reload(model)
|
||||
layer.weight.weight_loader(layer.weight, loaded_weight)
|
||||
layer.scale.weight_loader(layer.scale, loaded_scale)
|
||||
finalize_layerwise_reload(model, model_config=None)
|
||||
|
||||
assert torch.equal(layer.weight, loaded_weight)
|
||||
assert torch.equal(layer.scale, loaded_scale)
|
||||
assert "scale" in layer._non_persistent_buffers_set
|
||||
assert "0.scale" not in model.state_dict()
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
|
||||
Reference in New Issue
Block a user