Skip to content

[LoRA] Handle DoRA better #9547

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

Merged
merged 14 commits into from
Oct 8, 2024
+46 −13
Merged

[LoRA] Handle DoRA better #9547

Show file tree
Hide file tree
Changes from 7 commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
9d3d68f
handle dora.
sayakpaul Sep 28, 2024
1b71c5c
print test
sayakpaul Sep 28, 2024
7a32ee2
debug
sayakpaul Sep 28, 2024
9d7e3a2
fix
sayakpaul Sep 28, 2024
6fcb40b
fix-copies
sayakpaul Sep 28, 2024
97d13a5
update logits
sayakpaul Sep 28, 2024
114ceff
Merge branch 'main' into handle-dora
sayakpaul Oct 6, 2024
a8fdc76
add warning in the test.
sayakpaul Oct 6, 2024
c97e190
Merge branch 'main' into handle-dora
sayakpaul Oct 7, 2024
e3c250e
Merge branch 'main' into handle-dora
sayakpaul Oct 8, 2024
e08cf74
make is_dora check consistent.
sayakpaul Oct 8, 2024
ddf9f91
Merge branch 'main' into handle-dora
sayakpaul Oct 8, 2024
1c253e2
fix-copies
sayakpaul Oct 8, 2024
ba7087f
Merge branch 'main' into handle-dora
sayakpaul Oct 8, 2024
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
38 changes: 33 additions & 5 deletions src/diffusers/loaders/lora_pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -99,7 +99,7 @@ def load_lora_weights(
# First, ensure that the checkpoint is a compatible one and can be successfully loaded.
state_dict, network_alphas = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)

is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
is_correct_format = all("lora" in key for key in state_dict.keys())
if not is_correct_format:
raise ValueError("Invalid LoRA checkpoint.")

Expand Down Expand Up @@ -211,6 +211,11 @@ def lora_state_dict(
user_agent=user_agent,
allow_pickle=allow_pickle,
)
is_dora_scale_present = any("dora_scale" in k for k in state_dict)
if is_dora_scale_present:
warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
logger.warning(warn_msg)
state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

network_alphas = None
# TODO: replace it with a method from `state_dict_utils`
Expand Down Expand Up @@ -562,7 +567,8 @@ def load_lora_weights(
unet_config=self.unet.config,
**kwargs,
)
is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())

is_correct_format = all("lora" in key for key in state_dict.keys())
if not is_correct_format:
raise ValueError("Invalid LoRA checkpoint.")

Expand Down Expand Up @@ -684,6 +690,11 @@ def lora_state_dict(
user_agent=user_agent,
allow_pickle=allow_pickle,
)
is_dora_scale_present = any("dora_scale" in k for k in state_dict)
if is_dora_scale_present:
warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
logger.warning(warn_msg)
state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

network_alphas = None
# TODO: replace it with a method from `state_dict_utils`
Expand Down Expand Up @@ -1125,7 +1136,13 @@ def load_lora_weights(
# First, ensure that the checkpoint is a compatible one and can be successfully loaded.
state_dict = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)

is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
is_dora_scale_present = any("dora_scale" in k for k in state_dict)
if is_dora_scale_present:
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I see that we sometimes add this code to lora_state_dict() and sometimes on load_lora_weights. Is there any reason we do this differently?

Copy link
Member Author

@sayakpaul sayakpaul Oct 8, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Update: I see what you mean. Will update accordingly.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In e08cf74, I have tried to make this more consistent by moving all the "dora_scale" check related code to lora_state_dict().

LMK if that works for you, @yiyixuxu.

warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
logger.warning(warn_msg)
state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

is_correct_format = all("lora" in key for key in state_dict.keys())
if not is_correct_format:
raise ValueError("Invalid LoRA checkpoint.")

Expand Down Expand Up @@ -1587,6 +1604,11 @@ def lora_state_dict(
user_agent=user_agent,
allow_pickle=allow_pickle,
)
is_dora_scale_present = any("dora_scale" in k for k in state_dict)
if is_dora_scale_present:
warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
logger.warning(warn_msg)
state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

# TODO (sayakpaul): to a follow-up to clean and try to unify the conditions.

Expand Down Expand Up @@ -1659,7 +1681,7 @@ def load_lora_weights(
pretrained_model_name_or_path_or_dict, return_alphas=True, **kwargs
)

is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
is_correct_format = all("lora" in key for key in state_dict.keys())
if not is_correct_format:
raise ValueError("Invalid LoRA checkpoint.")

Expand Down Expand Up @@ -2405,7 +2427,13 @@ def load_lora_weights(
# First, ensure that the checkpoint is a compatible one and can be successfully loaded.
state_dict = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)

is_correct_format = all("lora" in key or "dora_scale" in key for key in state_dict.keys())
is_dora_scale_present = any("dora_scale" in k for k in state_dict)
if is_dora_scale_present:
warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
logger.warning(warn_msg)
state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

is_correct_format = all("lora" in key for key in state_dict.keys())
if not is_correct_format:
raise ValueError("Invalid LoRA checkpoint.")

Expand Down
2 changes: 1 addition & 1 deletion tests/lora/test_lora_layers_sdxl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -628,6 +628,6 @@ def test_integration_logits_for_dora_lora(self):
).images

predicted_slice = images[0, -3:, -3:, -1].flatten()
expected_slice_scale = np.array([0.3932, 0.3742, 0.4429, 0.3737, 0.3504, 0.433, 0.3948, 0.3769, 0.4516])
expected_slice_scale = np.array([0.1817, 0.0697, 0.2346, 0.0900, 0.1261, 0.2279, 0.1767, 0.1991, 0.2886])
max_diff = numpy_cosine_similarity_distance(expected_slice_scale, predicted_slice)
assert max_diff < 1e-3
Loading