Skip to content

Cache RoPE freqs on device to avoid repeated CPU-GPU copy in QwenImage #13406

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
akshan-main wants to merge 4 commits into
base: main
Choose a base branch
from
+25 −10
Open

Cache RoPE freqs on device to avoid repeated CPU-GPU copy in QwenImage #13406

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
1994ea3
Cache RoPE freqs on device to avoid repeated CPU-GPU copy in QwenImage
akshan-main Apr 3, 2026
073b5b1
Apply style fixes
github-actions[bot] Apr 9, 2026
6875e95
use lru_cache_unless_export
akshan-main Apr 9, 2026
4e8de47
Merge branch 'main' into fix-qwenimage-rope-sync
dg845 Apr 9, 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
35 changes: 25 additions & 10 deletions src/diffusers/models/transformers/transformer_qwenimage.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -233,6 +233,11 @@ def rope_params(self, index, dim, theta=10000):
freqs = torch.polar(torch.ones_like(freqs), freqs)
return freqs

@lru_cache_unless_export(maxsize=None)
def _get_device_freqs(self, device: torch.device) -> tuple[torch.Tensor, torch.Tensor]:
"""Return pos_freqs and neg_freqs on the given device."""
return self.pos_freqs.to(device), self.neg_freqs.to(device)

def forward(
self,
video_fhw: tuple[int, int, int, list[tuple[int, int, int]]],
Expand Down Expand Up @@ -300,8 +305,9 @@ def forward(
max_vid_index = max(height, width, max_vid_index)

max_txt_seq_len_int = int(max_txt_seq_len)
# Create device-specific copy for text freqs without modifying self.pos_freqs
txt_freqs = self.pos_freqs.to(device)[max_vid_index : max_vid_index + max_txt_seq_len_int, ...]
# Use cached device-transferred freqs to avoid CPU→GPU sync every forward call
pos_freqs_device, _ = self._get_device_freqs(device)
txt_freqs = pos_freqs_device[max_vid_index : max_vid_index + max_txt_seq_len_int, ...]
vid_freqs = torch.cat(vid_freqs, dim=0)

return vid_freqs, txt_freqs
Expand All @@ -311,8 +317,9 @@ def _compute_video_freqs(
self, frame: int, height: int, width: int, idx: int = 0, device: torch.device = None
) -> torch.Tensor:
seq_lens = frame * height * width
pos_freqs = self.pos_freqs.to(device) if device is not None else self.pos_freqs
neg_freqs = self.neg_freqs.to(device) if device is not None else self.neg_freqs
pos_freqs, neg_freqs = (
self._get_device_freqs(device) if device is not None else (self.pos_freqs, self.neg_freqs)
)

freqs_pos = pos_freqs.split([x // 2 for x in self.axes_dim], dim=1)
freqs_neg = neg_freqs.split([x // 2 for x in self.axes_dim], dim=1)
Expand Down Expand Up @@ -367,6 +374,11 @@ def rope_params(self, index, dim, theta=10000):
freqs = torch.polar(torch.ones_like(freqs), freqs)
return freqs

@lru_cache_unless_export(maxsize=None)
def _get_device_freqs(self, device: torch.device) -> tuple[torch.Tensor, torch.Tensor]:
"""Return pos_freqs and neg_freqs on the given device."""
return self.pos_freqs.to(device), self.neg_freqs.to(device)

def forward(
self,
video_fhw: tuple[int, int, int, list[tuple[int, int, int]]],
Expand Down Expand Up @@ -421,17 +433,19 @@ def forward(

max_vid_index = max(max_vid_index, layer_num)
max_txt_seq_len_int = int(max_txt_seq_len)
# Create device-specific copy for text freqs without modifying self.pos_freqs
txt_freqs = self.pos_freqs.to(device)[max_vid_index : max_vid_index + max_txt_seq_len_int, ...]
# Use cached device-transferred freqs to avoid CPU→GPU sync every forward call
pos_freqs_device, _ = self._get_device_freqs(device)
txt_freqs = pos_freqs_device[max_vid_index : max_vid_index + max_txt_seq_len_int, ...]
vid_freqs = torch.cat(vid_freqs, dim=0)

return vid_freqs, txt_freqs

@lru_cache_unless_export(maxsize=None)
def _compute_video_freqs(self, frame, height, width, idx=0, device: torch.device = None):
seq_lens = frame * height * width
pos_freqs = self.pos_freqs.to(device) if device is not None else self.pos_freqs
neg_freqs = self.neg_freqs.to(device) if device is not None else self.neg_freqs
pos_freqs, neg_freqs = (
self._get_device_freqs(device) if device is not None else (self.pos_freqs, self.neg_freqs)
)

freqs_pos = pos_freqs.split([x // 2 for x in self.axes_dim], dim=1)
freqs_neg = neg_freqs.split([x // 2 for x in self.axes_dim], dim=1)
Expand All @@ -452,8 +466,9 @@ def _compute_video_freqs(self, frame, height, width, idx=0, device: torch.device
@lru_cache_unless_export(maxsize=None)
def _compute_condition_freqs(self, frame, height, width, device: torch.device = None):
seq_lens = frame * height * width
pos_freqs = self.pos_freqs.to(device) if device is not None else self.pos_freqs
neg_freqs = self.neg_freqs.to(device) if device is not None else self.neg_freqs
pos_freqs, neg_freqs = (
self._get_device_freqs(device) if device is not None else (self.pos_freqs, self.neg_freqs)
)

freqs_pos = pos_freqs.split([x // 2 for x in self.axes_dim], dim=1)
freqs_neg = neg_freqs.split([x // 2 for x in self.axes_dim], dim=1)
Expand Down
Loading