Skip to content

fix device dismatch issue for pe_audio_video model parallelism #42917

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

Open
kaixuanliu wants to merge 9 commits into
base: main
Choose a base branch
from
+19 −8
Open

fix device dismatch issue for pe_audio_video model parallelism #42917

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
5efe73c
fix device dismatch issue for pe_audio_video model parallelism
kaixuanliu Dec 17, 2025
bcae7f0
skip the model parallelism unit test
kaixuanliu Dec 18, 2025
56f7b59
update
kaixuanliu Dec 19, 2025
b8a5ea2
Set all padding_mask_videos to 1 to avoid NaN values in outputs
kaixuanliu Dec 19, 2025
e5fb246
fix bug
kaixuanliu Dec 19, 2025
58e78ea
keep at least 1 valid frame
kaixuanliu Dec 19, 2025
63373d3
fix format issue
kaixuanliu Dec 19, 2025
e9aa6b6
use random.seed
kaixuanliu Dec 19, 2025
f8b0df3
update code
kaixuanliu Dec 19, 2025
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
3 changes: 2 additions & 1 deletion src/transformers/models/timm_wrapper/modeling_timm_wrapper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -84,7 +84,8 @@ class TimmWrapperPreTrainedModel(PreTrainedModel):
main_input_name = "pixel_values"
input_modalities = ("image",)
config: TimmWrapperConfig
_no_split_modules = []
# add WA here as `timm` does not support model parallelism
_no_split_modules = ["TimmWrapperModel"]
model_tags = ["timm"]

# used in Trainer to avoid passing `loss_kwargs` to model forward
Expand Down
3 changes: 2 additions & 1 deletion tests/models/pe_audio/test_modeling_pe_audio.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -101,7 +101,8 @@ def seq_length(self):

def prepare_config_and_inputs(self):
input_values = floats_tensor([self.batch_size, self.num_channels, self.audio_seq_length])
valid_lengths = ids_tensor([self.batch_size], self.audio_seq_length)
# Generate valid_lengths in range [1, self.audio_seq_length] to ensure at least one valid frame
valid_lengths = ids_tensor([self.batch_size], self.audio_seq_length - 1) + 1
padding_mask = torch.arange(self.audio_seq_length, device=torch_device)[None, :] < valid_lengths[:, None]
padding_mask = padding_mask.int()
config = self.get_config()
Expand Down
10 changes: 8 additions & 2 deletions tests/models/pe_audio_video/test_modeling_pe_audio_video.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -149,7 +149,8 @@ def seq_length(self):

def prepare_config_and_inputs(self):
input_values = floats_tensor([self.batch_size, self.num_audio_channels, self.audio_seq_length])
valid_audio_lengths = ids_tensor([self.batch_size], self.audio_seq_length)
# Generate valid_lengths in range [1, self.audio_seq_length] to ensure at least one valid frame
valid_audio_lengths = ids_tensor([self.batch_size], self.audio_seq_length - 1) + 1
padding_mask = torch.arange(self.audio_seq_length, device=torch_device)[None, :] < valid_audio_lengths[:, None]
padding_mask = padding_mask.int()

Expand All @@ -162,7 +163,8 @@ def prepare_config_and_inputs(self):
self.config_kwargs["video_config"]["vision_config"]["model_args"]["img_size"][1],
]
)
valid_video_lengths = ids_tensor([self.batch_size], self.num_frames)
# Generate valid_lengths in range [1, self.num_frames] to ensure at least one valid frame
valid_video_lengths = ids_tensor([self.batch_size], self.num_frames - 1) + 1
padding_mask_videos = (
torch.arange(self.num_frames, device=torch_device)[None, :] < valid_video_lengths[:, None]
)
Expand Down Expand Up @@ -258,6 +260,10 @@ def test_feed_forward_chunking(self):
def test_save_load(self):
pass

@unittest.skip(reason="TimmWrapperModel does not support model parallelism")
def test_model_parallelism(self):
pass

Comment on lines +263 to +266
Copy link
Member

@zucchini-nlp zucchini-nlp Dec 18, 2025

Choose a reason for hiding this comment

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

no, no, I meant to keep the changes and skip the tests. With the proposed diff, we can support model parallelism but the tests fail because of the way it is designed
Can you revert the prev diff and "move no_split_module under a timm's PreTrainedModel" instead of PE?

Copy link
Contributor Author

@kaixuanliu kaixuanliu Dec 19, 2025

Choose a reason for hiding this comment

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

Is it OK now?

@unittest.skip(reason="@eustlb this is not really expected")
def test_batching_equivalence(self):
pass
Expand Down
11 changes: 7 additions & 4 deletions tests/models/pe_video/test_modeling_pe_video.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -103,10 +103,9 @@ def prepare_config_and_inputs(self):
self.config_kwargs["vision_config"]["model_args"]["img_size"][1],
]
)
valid_lengths = ids_tensor([self.batch_size], self.num_frames)
padding_mask_videos = (
torch.ones([self.batch_size, self.num_frames], device=torch_device) < valid_lengths[:, None]
)
# Generate valid_lengths in range [1, num_frames] to ensure at least one valid frame
valid_lengths = ids_tensor([self.batch_size], self.num_frames - 1) + 1
padding_mask_videos = torch.arange(self.num_frames, device=torch_device).unsqueeze(0) < valid_lengths[:, None]
Copy link
Contributor Author

@kaixuanliu kaixuanliu Dec 19, 2025

Choose a reason for hiding this comment

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

Here we should use arange instead of ones

padding_mask_videos = padding_mask_videos.int()
config = self.get_config()

Expand Down Expand Up @@ -187,6 +186,10 @@ def test_feed_forward_chunking(self):
def test_save_load(self):
pass

@unittest.skip(reason="TimmWrapperModel does not support model parallelism")
def test_model_parallelism(self):
pass

@unittest.skip(reason="@eustlb this is not really expected")
def test_batching_equivalence(self):
pass
Expand Down