Skip to content
GitLab
Menu
Projects
Groups
Snippets
Loading...
Help
Help
Support
Community forum
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in / Register
Toggle navigation
Menu
Open sidebar
xuwx1
LightX2V
Commits
f7665abb
Unverified
Commit
f7665abb
authored
Nov 27, 2025
by
Yang Yong (雍洋)
Committed by
GitHub
Nov 27, 2025
Browse files
Fix rope for parallel (#530)
parent
2479e81d
Changes
1
Hide whitespace changes
Inline
Side-by-side
Showing
1 changed file
with
22 additions
and
0 deletions
+22
-0
lightx2v/models/schedulers/wan/scheduler.py
lightx2v/models/schedulers/wan/scheduler.py
+22
-0
No files found.
lightx2v/models/schedulers/wan/scheduler.py
View file @
f7665abb
...
@@ -2,6 +2,8 @@ from typing import List, Optional, Union
...
@@ -2,6 +2,8 @@ from typing import List, Optional, Union
import
numpy
as
np
import
numpy
as
np
import
torch
import
torch
import
torch.distributed
as
dist
from
torch.nn
import
functional
as
F
from
lightx2v.models.schedulers.scheduler
import
BaseScheduler
from
lightx2v.models.schedulers.scheduler
import
BaseScheduler
from
lightx2v.utils.utils
import
masks_like
from
lightx2v.utils.utils
import
masks_like
...
@@ -14,6 +16,10 @@ class WanScheduler(BaseScheduler):
...
@@ -14,6 +16,10 @@ class WanScheduler(BaseScheduler):
self
.
infer_steps
=
self
.
config
[
"infer_steps"
]
self
.
infer_steps
=
self
.
config
[
"infer_steps"
]
self
.
target_video_length
=
self
.
config
[
"target_video_length"
]
self
.
target_video_length
=
self
.
config
[
"target_video_length"
]
self
.
sample_shift
=
self
.
config
[
"sample_shift"
]
self
.
sample_shift
=
self
.
config
[
"sample_shift"
]
if
self
.
config
[
"seq_parallel"
]:
self
.
seq_p_group
=
self
.
config
.
get
(
"device_mesh"
).
get_group
(
mesh_dim
=
"seq_p"
)
else
:
self
.
seq_p_group
=
None
self
.
patch_size
=
(
1
,
2
,
2
)
self
.
patch_size
=
(
1
,
2
,
2
)
self
.
shift
=
1
self
.
shift
=
1
self
.
num_train_timesteps
=
1000
self
.
num_train_timesteps
=
1000
...
@@ -87,8 +93,24 @@ class WanScheduler(BaseScheduler):
...
@@ -87,8 +93,24 @@ class WanScheduler(BaseScheduler):
cos_half
=
cos_sin
.
real
.
contiguous
()
cos_half
=
cos_sin
.
real
.
contiguous
()
sin_half
=
cos_sin
.
imag
.
contiguous
()
sin_half
=
cos_sin
.
imag
.
contiguous
()
cos_sin
=
torch
.
cat
([
cos_half
,
sin_half
],
dim
=-
1
)
cos_sin
=
torch
.
cat
([
cos_half
,
sin_half
],
dim
=-
1
)
if
self
.
seq_p_group
is
not
None
:
world_size
=
dist
.
get_world_size
(
self
.
seq_p_group
)
cur_rank
=
dist
.
get_rank
(
self
.
seq_p_group
)
seqlen
=
cos_sin
.
shape
[
0
]
padding_size
=
(
world_size
-
(
seqlen
%
world_size
))
%
world_size
if
padding_size
>
0
:
cos_sin
=
F
.
pad
(
cos_sin
,
(
0
,
0
,
0
,
padding_size
))
cos_sin
=
torch
.
chunk
(
cos_sin
,
world_size
,
dim
=
0
)[
cur_rank
]
else
:
else
:
cos_sin
=
cos_sin
.
reshape
(
seq_len
,
1
,
-
1
)
cos_sin
=
cos_sin
.
reshape
(
seq_len
,
1
,
-
1
)
if
self
.
seq_p_group
is
not
None
:
world_size
=
dist
.
get_world_size
(
self
.
seq_p_group
)
cur_rank
=
dist
.
get_rank
(
self
.
seq_p_group
)
seqlen
=
cos_sin
.
shape
[
0
]
padding_size
=
(
world_size
-
(
seqlen
%
world_size
))
%
world_size
if
padding_size
>
0
:
cos_sin
=
F
.
pad
(
cos_sin
,
(
0
,
0
,
0
,
0
,
0
,
padding_size
))
cos_sin
=
torch
.
chunk
(
cos_sin
,
world_size
,
dim
=
0
)[
cur_rank
]
return
cos_sin
return
cos_sin
def
prepare_latents
(
self
,
seed
,
latent_shape
,
dtype
=
torch
.
float32
):
def
prepare_latents
(
self
,
seed
,
latent_shape
,
dtype
=
torch
.
float32
):
...
...
Write
Preview
Markdown
is supported
0%
Try again
or
attach a new file
.
Attach a file
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Cancel
Please
register
or
sign in
to comment