Skip to content

Make CogVideoX RoPE implementation consistent #9963

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 2 commits into from
Nov 19, 2024
Merged

Make CogVideoX RoPE implementation consistent #9963

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8eb7ff0
update cogvideox rope implementation
a-r-r-o-w Nov 19, 2024
9808532
apply suggestions from review
a-r-r-o-w Nov 19, 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
39 changes: 27 additions & 12 deletions src/diffusers/pipelines/cogvideo/pipeline_cogvideox.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -446,19 +446,34 @@ def _prepare_rotary_positional_embeddings(
p = self.transformer.config.patch_size
p_t = self.transformer.config.patch_size_t or 1

base_size_width = self.transformer.config.sample_width // p
base_size_height = self.transformer.config.sample_height // p
base_num_frames = (num_frames + p_t - 1) // p_t
if p_t is None:
# CogVideoX 1.0
base_size_width = self.transformer.config.sample_width // p
base_size_height = self.transformer.config.sample_height // p

grid_crops_coords = get_resize_crop_region_for_grid(
(grid_height, grid_width), base_size_width, base_size_height
)
freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
embed_dim=self.transformer.config.attention_head_dim,
crops_coords=grid_crops_coords,
grid_size=(grid_height, grid_width),
temporal_size=base_num_frames,
)
grid_crops_coords = get_resize_crop_region_for_grid(
(grid_height, grid_width), base_size_width, base_size_height
)
freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
embed_dim=self.transformer.config.attention_head_dim,
crops_coords=grid_crops_coords,
grid_size=(grid_height, grid_width),
temporal_size=num_frames,
)
else:
# CogVideoX 1.5
base_size_width = self.transformer.config.sample_width // p
base_size_height = self.transformer.config.sample_height // p
base_num_frames = (num_frames + p_t - 1) // p_t

freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
embed_dim=self.transformer.config.attention_head_dim,
crops_coords=None,
grid_size=(grid_height, grid_width),
temporal_size=base_num_frames,
grid_type="slice",
max_size=(base_size_height, base_size_width),
)

freqs_cos = freqs_cos.to(device=device)
freqs_sin = freqs_sin.to(device=device)
Expand Down
39 changes: 27 additions & 12 deletions src/diffusers/pipelines/cogvideo/pipeline_cogvideox_fun_control.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -492,19 +492,34 @@ def _prepare_rotary_positional_embeddings(
p = self.transformer.config.patch_size
p_t = self.transformer.config.patch_size_t or 1

base_size_width = self.transformer.config.sample_width // p
base_size_height = self.transformer.config.sample_height // p
base_num_frames = (num_frames + p_t - 1) // p_t
if p_t is None:
# CogVideoX 1.0
base_size_width = self.transformer.config.sample_width // p
base_size_height = self.transformer.config.sample_height // p

grid_crops_coords = get_resize_crop_region_for_grid(
(grid_height, grid_width), base_size_width, base_size_height
)
freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
embed_dim=self.transformer.config.attention_head_dim,
crops_coords=grid_crops_coords,
grid_size=(grid_height, grid_width),
temporal_size=base_num_frames,
)
grid_crops_coords = get_resize_crop_region_for_grid(
(grid_height, grid_width), base_size_width, base_size_height
)
freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
embed_dim=self.transformer.config.attention_head_dim,
crops_coords=grid_crops_coords,
grid_size=(grid_height, grid_width),
temporal_size=num_frames,
)
else:
# CogVideoX 1.5
base_size_width = self.transformer.config.sample_width // p
base_size_height = self.transformer.config.sample_height // p
base_num_frames = (num_frames + p_t - 1) // p_t

freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
embed_dim=self.transformer.config.attention_head_dim,
crops_coords=None,
grid_size=(grid_height, grid_width),
temporal_size=base_num_frames,
grid_type="slice",
max_size=(base_size_height, base_size_width),
)

freqs_cos = freqs_cos.to(device=device)
freqs_sin = freqs_sin.to(device=device)
Expand Down
7 changes: 4 additions & 3 deletions src/diffusers/pipelines/cogvideo/pipeline_cogvideox_image2video.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -528,6 +528,7 @@ def unfuse_qkv_projections(self) -> None:
self.transformer.unfuse_qkv_projections()
self.fusing_transformer = False

# Copied from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline._prepare_rotary_positional_embeddings
def _prepare_rotary_positional_embeddings(
self,
height: int,
Expand All @@ -539,10 +540,10 @@ def _prepare_rotary_positional_embeddings(
grid_width = width // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)

p = self.transformer.config.patch_size
p_t = self.transformer.config.patch_size_t
p_t = self.transformer.config.patch_size_t or 1

if p_t is None:
# CogVideoX 1.0 I2V
# CogVideoX 1.0
base_size_width = self.transformer.config.sample_width // p
base_size_height = self.transformer.config.sample_height // p

Expand All @@ -556,7 +557,7 @@ def _prepare_rotary_positional_embeddings(
temporal_size=num_frames,
)
else:
# CogVideoX 1.5 I2V
# CogVideoX 1.5
base_size_width = self.transformer.config.sample_width // p
base_size_height = self.transformer.config.sample_height // p
base_num_frames = (num_frames + p_t - 1) // p_t
Expand Down
39 changes: 27 additions & 12 deletions src/diffusers/pipelines/cogvideo/pipeline_cogvideox_video2video.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -522,19 +522,34 @@ def _prepare_rotary_positional_embeddings(
p = self.transformer.config.patch_size
p_t = self.transformer.config.patch_size_t or 1

base_size_width = self.transformer.config.sample_width // p
base_size_height = self.transformer.config.sample_height // p
base_num_frames = (num_frames + p_t - 1) // p_t
if p_t is None:
# CogVideoX 1.0
base_size_width = self.transformer.config.sample_width // p
base_size_height = self.transformer.config.sample_height // p

grid_crops_coords = get_resize_crop_region_for_grid(
(grid_height, grid_width), base_size_width, base_size_height
)
freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
embed_dim=self.transformer.config.attention_head_dim,
crops_coords=grid_crops_coords,
grid_size=(grid_height, grid_width),
temporal_size=base_num_frames,
)
grid_crops_coords = get_resize_crop_region_for_grid(
(grid_height, grid_width), base_size_width, base_size_height
)
freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
embed_dim=self.transformer.config.attention_head_dim,
crops_coords=grid_crops_coords,
grid_size=(grid_height, grid_width),
temporal_size=num_frames,
)
else:
# CogVideoX 1.5
base_size_width = self.transformer.config.sample_width // p
base_size_height = self.transformer.config.sample_height // p
base_num_frames = (num_frames + p_t - 1) // p_t

freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
embed_dim=self.transformer.config.attention_head_dim,
crops_coords=None,
grid_size=(grid_height, grid_width),
temporal_size=base_num_frames,
grid_type="slice",
max_size=(base_size_height, base_size_width),
)

freqs_cos = freqs_cos.to(device=device)
freqs_sin = freqs_sin.to(device=device)
Expand Down
Loading