Skip to content

Fix MaskFormerSwin attention mask dtype to follow hidden states #49032

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
vasqu merged 4 commits into from
Sep 25, 2026
+5 −3
Merged

Fix MaskFormerSwin attention mask dtype to follow hidden states #49032

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
c1e65c9
Fix MaskFormerSwin attention mask dtype to follow hidden states
kaixuanliu Sep 23, 2026
be23e14
update
kaixuanliu Sep 23, 2026
ffcb608
reverted the change
kaixuanliu Sep 23, 2026
ca08ff0
update
kaixuanliu Sep 24, 2026
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
8 changes: 5 additions & 3 deletions src/transformers/models/maskformer/modeling_maskformer_swin.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -457,7 +457,7 @@ def __init__(self, config, dim, input_resolution, num_heads, drop_path_rate=0.0,
self.intermediate = MaskFormerSwinIntermediate(config, dim)
self.output = MaskFormerSwinOutput(config, dim)

def get_attn_mask(self, input_resolution, device=None):
def get_attn_mask(self, input_resolution, device=None, dtype=torch.float32):
"""Build the cyclic-shift attention mask for shifted-window MSA; returns None when shift_size is 0.

Each (h, w) position belongs to one of 9 cyclic-shift regions (3 along each axis), encoded
Expand All @@ -476,7 +476,7 @@ def get_attn_mask(self, input_resolution, device=None):
w_idx = torch.arange(width, device=device)
h_region = (h_idx >= height - self.window_size).long() + (h_idx >= height - self.shift_size).long()
w_region = (w_idx >= width - self.window_size).long() + (w_idx >= width - self.shift_size).long()
img_mask = (h_region[None, :, None, None] * 3 + w_region[None, None, :, None]).float()
img_mask = (h_region[None, :, None, None] * 3 + w_region[None, None, :, None]).to(dtype=dtype)
mask_windows = window_partition(img_mask, self.window_size)
mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
Expand Down Expand Up @@ -511,7 +511,9 @@ def forward(self, hidden_states, input_dimensions, output_attentions=False):
# partition windows
hidden_states_windows = window_partition(shifted_hidden_states, self.window_size)
hidden_states_windows = hidden_states_windows.view(-1, self.window_size * self.window_size, channels)
attn_mask = self.get_attn_mask((height_pad, width_pad), device=hidden_states_windows.device)
attn_mask = self.get_attn_mask(
(height_pad, width_pad), device=hidden_states_windows.device, dtype=hidden_states_windows.dtype
)

self_attention_outputs = self.attention(hidden_states_windows, attn_mask, output_attentions=output_attentions)

Expand Down
Loading