Skip to content
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

refine_spatio_temporal_oflow #755

Merged
merged 1 commit into from
Mar 25, 2024
Merged

refine_spatio_temporal_oflow #755

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
129 changes: 45 additions & 84 deletions src/infer_compiler_registry/register_diffusers/spatio_temporal_oflow.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,79 +37,38 @@
diffusers_0260_v = version.parse("0.26.0")

if diffusers_version >= diffusers_0260_v:
DiffusersUNetSpatioTemporalConditionModel = transformed_diffusers.models.unets.unet_spatio_temporal_condition.UNetSpatioTemporalConditionModel
DiffusersTransformerSpatioTemporalModel = transformed_diffusers.models.transformers.transformer_temporal.TransformerSpatioTemporalModel
DiffusersUNetSpatioTemporalConditionModel = (
transformed_diffusers.models.unets.unet_spatio_temporal_condition.UNetSpatioTemporalConditionModel
)
DiffusersTransformerSpatioTemporalModel = (
transformed_diffusers.models.transformers.transformer_temporal.TransformerSpatioTemporalModel
)

else:
DiffusersUNetSpatioTemporalConditionModel = transformed_diffusers.models.unet_spatio_temporal_condition.UNetSpatioTemporalConditionModel
DiffusersTransformerSpatioTemporalModel = transformed_diffusers.models.transformer_temporal.TransformerSpatioTemporalModel

DiffusersSpatioTemporalResBlock = transformed_diffusers.models.resnet.SpatioTemporalResBlock
DiffusersTemporalBasicTransformerBlock = transformed_diffusers.models.attention.TemporalBasicTransformerBlock


class TemporalDecoder(nn.Module):
def __init__(
self,
in_channels: int = 4,
out_channels: int = 3,
block_out_channels: Tuple[int] = (128, 256, 512, 512),
layers_per_block: int = 2,
):
super().__init__()
self.layers_per_block = layers_per_block

self.conv_in = nn.Conv2d(
in_channels, block_out_channels[-1], kernel_size=3, stride=1, padding=1
)
self.mid_block = MidBlockTemporalDecoder(
num_layers=self.layers_per_block,
in_channels=block_out_channels[-1],
out_channels=block_out_channels[-1],
attention_head_dim=block_out_channels[-1],
)

# up
self.up_blocks = nn.ModuleList([])
reversed_block_out_channels = list(reversed(block_out_channels))
output_channel = reversed_block_out_channels[0]
for i in range(len(block_out_channels)):
prev_output_channel = output_channel
output_channel = reversed_block_out_channels[i]

is_final_block = i == len(block_out_channels) - 1
up_block = UpBlockTemporalDecoder(
num_layers=self.layers_per_block + 1,
in_channels=prev_output_channel,
out_channels=output_channel,
add_upsample=not is_final_block,
)
self.up_blocks.append(up_block)
prev_output_channel = output_channel

self.conv_norm_out = nn.GroupNorm(
num_channels=block_out_channels[0], num_groups=32, eps=1e-6
)

self.conv_act = nn.SiLU()
self.conv_out = torch.nn.Conv2d(
in_channels=block_out_channels[0],
out_channels=out_channels,
kernel_size=3,
padding=1,
)

conv_out_kernel_size = (3, 1, 1)
padding = [int(k // 2) for k in conv_out_kernel_size]
self.time_conv_out = torch.nn.Conv3d(
in_channels=out_channels,
out_channels=out_channels,
kernel_size=conv_out_kernel_size,
padding=padding,
)

self.gradient_checkpointing = False

DiffusersUNetSpatioTemporalConditionModel = (
transformed_diffusers.models.unet_spatio_temporal_condition.UNetSpatioTemporalConditionModel
)
DiffusersTransformerSpatioTemporalModel = (
transformed_diffusers.models.transformer_temporal.TransformerSpatioTemporalModel
)

if diffusers_version >= version.parse("0.25.00"):
DiffusersTemporalDecoder = (
transformed_diffusers.models.autoencoders.autoencoder_kl_temporal_decoder.TemporalDecoder
)
else:
DiffusersTemporalDecoder = (
transformed_diffusers.models.autoencoder_kl_temporal_decoder.TemporalDecoder
)

DiffusersSpatioTemporalResBlock = (
transformed_diffusers.models.resnet.SpatioTemporalResBlock
)
DiffusersTemporalBasicTransformerBlock = (
transformed_diffusers.models.attention.TemporalBasicTransformerBlock
)

class TemporalDecoder(DiffusersTemporalDecoder):
def forward(
self,
sample: torch.FloatTensor,
Expand Down Expand Up @@ -150,18 +109,24 @@ def custom_forward(*inputs):
else:
# middle
sample = torch.utils.checkpoint.checkpoint(
create_custom_forward(self.mid_block), sample, image_only_indicator,
create_custom_forward(self.mid_block),
sample,
image_only_indicator,
)
sample = sample.to(upscale_dtype)

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

# up
Expand All @@ -186,7 +151,6 @@ def custom_forward(*inputs):

return sample


# VideoResBlock
class SpatioTemporalResBlock(DiffusersSpatioTemporalResBlock):
def forward(
Expand All @@ -209,9 +173,9 @@ def forward(
# )
#
# Dynamic shape for VAE divide chunks
hidden_states_mix = hidden_states.unflatten(0, shape=(batch_size, -1)).permute(
0, 2, 1, 3, 4
)
hidden_states_mix = hidden_states.unflatten(
0, shape=(batch_size, -1)
).permute(0, 2, 1, 3, 4)
hidden_states = hidden_states.unflatten(0, shape=(batch_size, -1)).permute(
0, 2, 1, 3, 4
)
Expand All @@ -232,7 +196,6 @@ def forward(
hidden_states = hidden_states.permute(0, 2, 1, 3, 4).flatten(0, 1)
return hidden_states


class TransformerSpatioTemporalModel(DiffusersTransformerSpatioTemporalModel):
def forward(
self,
Expand Down Expand Up @@ -272,9 +235,9 @@ def forward(
# batch_size, num_frames, -1, time_context.shape[-1]
# )[:, 0]
# Rewrite for onediff SVD dynamic shape
time_context_first_timestep = time_context.unflatten(0, shape=(batch_size, -1))[
:, 0
]
time_context_first_timestep = time_context.unflatten(
0, shape=(batch_size, -1)
)[:, 0]
# time_context = time_context_first_timestep[None, :].broadcast_to(
# height * width, batch_size, 1, time_context.shape[-1]
# )
Expand Down Expand Up @@ -361,7 +324,6 @@ def forward(

return TransformerTemporalModelOutput(sample=output)


class TemporalBasicTransformerBlock(DiffusersTemporalBasicTransformerBlock):
def forward(
self,
Expand Down Expand Up @@ -434,7 +396,6 @@ def forward(

return hidden_states


class UNetSpatioTemporalConditionModel(DiffusersUNetSpatioTemporalConditionModel):
def forward(
self,
Expand Down
Loading