Skip to content

fix(ernie-image): pass attn_mask=None when text is unpadded #13802

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
Ace3Z wants to merge 1 commit into
base: main
Choose a base branch
from
+14 −9
Open

fix(ernie-image): pass attn_mask=None when text is unpadded #13802

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
18 changes: 9 additions & 9 deletions src/diffusers/models/transformers/transformer_ernie_image.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -408,15 +408,15 @@ def forward(
)
rotary_pos_emb = self.pos_embed(torch.cat([image_ids, text_ids], dim=1))

# Attention mask: True = valid (attend), False = padding (mask out), matches sdpa bool convention
valid_text = (
torch.arange(Tmax, device=device).view(1, Tmax) < text_lens.view(B, 1)
if Tmax > 0
else torch.zeros((B, 0), device=device, dtype=torch.bool)
)
attention_mask = torch.cat([torch.ones((B, N_img), device=device, dtype=torch.bool), valid_text], dim=1)[
:, None, None, :
]
# Attention mask: True = valid (attend), False = padding (mask out), matches sdpa bool convention.
# Skip building it when there's no padding so flash-attn (which rejects non-None masks) can run.
if Tmax > 0 and not bool((text_lens == Tmax).all()):
valid_text = torch.arange(Tmax, device=device).view(1, Tmax) < text_lens.view(B, 1)
attention_mask = torch.cat([torch.ones((B, N_img), device=device, dtype=torch.bool), valid_text], dim=1)[
:, None, None, :
]
else:
attention_mask = None

# AdaLN
sample = self.time_proj(timestep)
Expand Down
5 changes: 5 additions & 0 deletions tests/models/transformers/test_models_transformer_ernie_image.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@

from ...testing_utils import torch_device
from ..testing_utils import (
AttentionBackendTesterMixin,
BaseModelTesterConfig,
ModelTesterMixin,
TorchCompileTesterMixin,
Expand Down Expand Up @@ -130,3 +131,7 @@ def test_compile_works_with_aot(self, tmp_path):
@pytest.mark.skip(reason="Fullgraph is broken.")
def test_compile_on_different_shapes(self):
super().test_compile_on_different_shapes()


class TestErnieImageTransformerAttentionBackend(ErnieImageTransformerTesterConfig, AttentionBackendTesterMixin):
"""Attention backend tests for ErnieImage Transformer."""
Loading