Skip to content

Fix 3D attn_mask head/batch layout in T2V/V2T cross-attention (repeat → repeat_interleave) #77

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
h-munakata merged 1 commit into from
Aug 30, 2026
Merged

Fix 3D attn_mask head/batch layout in T2V/V2T cross-attention (repeat → repeat_interleave) #77

Show file tree
Hide file tree
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
4 changes: 2 additions & 2 deletions lighthouse/common/CIM.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -700,7 +700,7 @@ def forward_post(self,
# print(src_key_padding_mask[:, 76:].shape) # torch.Size([32, 26])

qmask, kmask = src_key_padding_mask[:, 1:video_length + 1].unsqueeze(2), src_key_padding_mask[:, video_length + 1:].unsqueeze(1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat(self.nhead, 1, 1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat_interleave(self.nhead, dim=0) # batch-major layout for nn.MultiheadAttention 3D attn_mask
# print(attn_mask.shape)
# print(attn_mask[0][0])
# print(q.shape) 75 32 256
Expand Down Expand Up @@ -789,7 +789,7 @@ def forward_post(self,
global_token, q, k, v = src[0].unsqueeze(0), pos_src[video_length + 1: ], pos_src[1:video_length + 1], src[1:video_length + 1]

qmask, kmask = src_key_padding_mask[:, video_length + 1:].unsqueeze(2), src_key_padding_mask[:, 1:video_length + 1].unsqueeze(1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat(self.nhead, 1, 1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat_interleave(self.nhead, dim=0) # batch-major layout for nn.MultiheadAttention 3D attn_mask
# print(attn_mask.shape)
# print(attn_mask[0][0])
# print(q.shape) 23 32 256
Expand Down
2 changes: 1 addition & 1 deletion lighthouse/common/cg_detr_transformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -576,7 +576,7 @@ def forward_post(self,

qmask, kmask = src_key_padding_mask[:, :video_length].unsqueeze(2), src_key_padding_mask[:,
video_length:].unsqueeze(1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat(self.nhead, 1, 1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat_interleave(self.nhead, dim=0) # batch-major layout for nn.MultiheadAttention 3D attn_mask

# - key_padding_mask: :math:`(S)` or :math:`(N, S)` where N is the batch size, S is the source sequence length.
# If a FloatTensor is provided, it will be directly added to the value.
Expand Down
2 changes: 1 addition & 1 deletion lighthouse/common/qd_detr_transformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -476,7 +476,7 @@ def forward_post(self,
# print(src_key_padding_mask[:, 76:].shape) # torch.Size([32, 26])

qmask, kmask = src_key_padding_mask[:, 1:video_length + 1].unsqueeze(2), src_key_padding_mask[:, video_length + 1:].unsqueeze(1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat(self.nhead, 1, 1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat_interleave(self.nhead, dim=0) # batch-major layout for nn.MultiheadAttention 3D attn_mask
# print(attn_mask.shape)
# print(attn_mask[0][0])
# print(q.shape) 75 32 256
Expand Down
2 changes: 1 addition & 1 deletion lighthouse/common/taskweave_transformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -591,7 +591,7 @@ def forward_post(self,
q, k, v = pos_src[:video_length], pos_src[video_length:], src[video_length:]

qmask, kmask = src_key_padding_mask[:, :video_length].unsqueeze(2), src_key_padding_mask[:, video_length :].unsqueeze(1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat(self.nhead, 1, 1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat_interleave(self.nhead, dim=0) # batch-major layout for nn.MultiheadAttention 3D attn_mask

src2 = self.self_attn(q, k, value=v, attn_mask=attn_mask,
key_padding_mask=src_key_padding_mask[:, video_length:])[0]
Expand Down
4 changes: 2 additions & 2 deletions lighthouse/common/tr_detr_transformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -434,7 +434,7 @@ def forward_post(self,
# print(src_key_padding_mask[:, 76:].shape) # torch.Size([32, 26])
# print(src_key_padding_mask[:, 1:video_length+1].shape) # torch.Size([32, 97]) 切片会保留维度torch.Size([32, 75])，只有坐标便不会
qmask, kmask = src_key_padding_mask[:, :video_length].unsqueeze(2), src_key_padding_mask[:, video_length:].unsqueeze(1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat(self.nhead, 1, 1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat_interleave(self.nhead, dim=0) # batch-major layout for nn.MultiheadAttention 3D attn_mask
# print(attn_mask.shape)
# print(attn_mask[0][0])
# print(q.shape) 75 32 256
Expand Down Expand Up @@ -525,7 +525,7 @@ def forward_post(self,
# print(src_key_padding_mask[:, 76:].shape) # torch.Size([32, 26])

qmask, kmask = src_key_padding_mask[:, 1:video_length + 1].unsqueeze(2), src_key_padding_mask[:, video_length + 1:].unsqueeze(1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat(self.nhead, 1, 1)
attn_mask = torch.matmul(qmask.float(), kmask.float()).bool().repeat_interleave(self.nhead, dim=0) # batch-major layout for nn.MultiheadAttention 3D attn_mask
# print(attn_mask.shape)
# print(attn_mask[0][0])
# print(q.shape) 75 32 256
Expand Down