Skip to content

Fix gradient checkpointing bugs in freezing part of models (requires_grad=False) #3404

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 5 commits into from
May 17, 2023
Merged

Fix gradient checkpointing bugs in freezing part of models (requires_grad=False) #3404

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
187f9fc
gradient checkpointing bug fix
IrisRainbowNeko May 12, 2023
26d7831
Merge branch 'main' into grad_ckpt_part
patrickvonplaten May 16, 2023
ab6522a
bug fix; changes for reviews
IrisRainbowNeko May 16, 2023
94c8dca
reformat
IrisRainbowNeko May 16, 2023
eb5fc40
reformat
IrisRainbowNeko May 16, 2023
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
173 changes: 137 additions & 36 deletions src/diffusers/models/unet_2d_blocks.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,7 @@
import torch.nn.functional as F
from torch import nn

from ..utils import is_torch_version
from .attention import AdaGroupNorm
from .attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0
from .dual_transformer_2d import DualTransformer2DModel
Expand Down Expand Up @@ -866,13 +867,27 @@ def custom_forward(*inputs):

return custom_forward

hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(attn, return_dict=False),
hidden_states,
encoder_hidden_states,
cross_attention_kwargs,
)[0]
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(attn, return_dict=False),
hidden_states,
encoder_hidden_states,
cross_attention_kwargs,
use_reentrant=False,
)[0]
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(attn, return_dict=False),
hidden_states,
encoder_hidden_states,
cross_attention_kwargs,
)[0]
else:
hidden_states = resnet(hidden_states, temb)
hidden_states = attn(
Expand Down Expand Up @@ -957,7 +972,14 @@ def custom_forward(*inputs):

return custom_forward

hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb)

Expand Down Expand Up @@ -1361,7 +1383,14 @@ def custom_forward(*inputs):

return custom_forward

hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb)

Expand Down Expand Up @@ -1558,7 +1587,14 @@ def custom_forward(*inputs):

return custom_forward

hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb)

Expand Down Expand Up @@ -1653,14 +1689,29 @@ def custom_forward(*inputs):

return custom_forward

hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(attn, return_dict=False),
hidden_states,
encoder_hidden_states,
attention_mask,
cross_attention_kwargs,
)
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(attn, return_dict=False),
hidden_states,
encoder_hidden_states,
attention_mask,
cross_attention_kwargs,
use_reentrant=False,
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(attn, return_dict=False),
hidden_states,
encoder_hidden_states,
attention_mask,
cross_attention_kwargs,
)
else:
hidden_states = resnet(hidden_states, temb)
hidden_states = attn(
Expand Down Expand Up @@ -1874,13 +1925,27 @@ def custom_forward(*inputs):

return custom_forward

hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(attn, return_dict=False),
hidden_states,
encoder_hidden_states,
cross_attention_kwargs,
)[0]
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(attn, return_dict=False),
hidden_states,
encoder_hidden_states,
cross_attention_kwargs,
use_reentrant=False,
)[0]
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(attn, return_dict=False),
hidden_states,
encoder_hidden_states,
cross_attention_kwargs,
)[0]
else:
hidden_states = resnet(hidden_states, temb)
hidden_states = attn(
Expand Down Expand Up @@ -1960,7 +2025,14 @@ def custom_forward(*inputs):

return custom_forward

hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb)

Expand Down Expand Up @@ -2388,7 +2460,14 @@ def custom_forward(*inputs):

return custom_forward

hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb)

Expand Down Expand Up @@ -2593,7 +2672,14 @@ def custom_forward(*inputs):

return custom_forward

hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
else:
hidden_states = resnet(hidden_states, temb)

Expand Down Expand Up @@ -2714,14 +2800,29 @@ def custom_forward(*inputs):

return custom_forward

hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(attn, return_dict=False),
hidden_states,
encoder_hidden_states,
attention_mask,
cross_attention_kwargs,
)[0]
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(attn, return_dict=False),
hidden_states,
encoder_hidden_states,
attention_mask,
cross_attention_kwargs,
use_reentrant=False,
)[0]
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(attn, return_dict=False),
hidden_states,
encoder_hidden_states,
attention_mask,
cross_attention_kwargs,
)[0]
else:
hidden_states = resnet(hidden_states, temb)
hidden_states = attn(
Expand Down
46 changes: 34 additions & 12 deletions src/diffusers/models/vae.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,7 +18,7 @@
import torch
import torch.nn as nn

from ..utils import BaseOutput, randn_tensor
from ..utils import BaseOutput, is_torch_version, randn_tensor
from .unet_2d_blocks import UNetMidBlock2D, get_down_block, get_up_block


Expand Down Expand Up @@ -117,11 +117,20 @@ def custom_forward(*inputs):
return custom_forward

# down
for down_block in self.down_blocks:
sample = torch.utils.checkpoint.checkpoint(create_custom_forward(down_block), sample)

# middle
sample = torch.utils.checkpoint.checkpoint(create_custom_forward(self.mid_block), sample)
if is_torch_version(">=", "1.11.0"):
for down_block in self.down_blocks:
sample = torch.utils.checkpoint.checkpoint(
create_custom_forward(down_block), sample, use_reentrant=False
)
# middle
sample = torch.utils.checkpoint.checkpoint(
create_custom_forward(self.mid_block), sample, use_reentrant=False
)
else:
for down_block in self.down_blocks:
sample = torch.utils.checkpoint.checkpoint(create_custom_forward(down_block), sample)
# middle
sample = torch.utils.checkpoint.checkpoint(create_custom_forward(self.mid_block), sample)

else:
# down
Expand Down Expand Up @@ -221,13 +230,26 @@ def custom_forward(*inputs):

return custom_forward

# middle
sample = torch.utils.checkpoint.checkpoint(create_custom_forward(self.mid_block), sample)
sample = sample.to(upscale_dtype)
if is_torch_version(">=", "1.11.0"):
# middle
sample = torch.utils.checkpoint.checkpoint(
create_custom_forward(self.mid_block), sample, use_reentrant=False
)
sample = sample.to(upscale_dtype)

# up
for up_block in self.up_blocks:
sample = torch.utils.checkpoint.checkpoint(create_custom_forward(up_block), sample)
# up
for up_block in self.up_blocks:
sample = torch.utils.checkpoint.checkpoint(
create_custom_forward(up_block), sample, use_reentrant=False
)
else:
# middle
sample = torch.utils.checkpoint.checkpoint(create_custom_forward(self.mid_block), sample)
sample = sample.to(upscale_dtype)

# up
for up_block in self.up_blocks:
sample = torch.utils.checkpoint.checkpoint(create_custom_forward(up_block), sample)
else:
# middle
sample = self.mid_block(sample)
Expand Down
Loading