Thanks for opening this issue! @sayakpaul, is there a specific reason why we essentially cache the results of _compute_video_freqs again with self.rope_cache when not compiling? As far as I can tell self.rope_cache is only used for this purpose.

diffusers/src/diffusers/models/transformers/transformer_qwenimage.py

Lines 218 to 223 in 041501a

    
           if not torch.compiler.is_compiling(): 
        
               if rope_key not in self.rope_cache: 
        
                   self.rope_cache[rope_key] = self._compute_video_freqs(frame, height, width, idx) 
        
               video_freq = self.rope_cache[rope_key] 
        
           else: 
        
               video_freq = self._compute_video_freqs(frame, height, width, idx)

I believe the current implementation was introduced in #12061. See this comment from the PR, which raises a very similar issue.

	if not torch.compiler.is_compiling():
	if rope_key not in self.rope_cache:
	self.rope_cache[rope_key] = self._compute_video_freqs(frame, height, width, idx)
	video_freq = self.rope_cache[rope_key]
	else:
	video_freq = self._compute_video_freqs(frame, height, width, idx)

Redundant lru_cache in QwenEmbedRope #12401

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions