Skip to content

[Bugfix] Fix Qwen-Image SP and TeaCache incompatibility #2101

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
gcanlin merged 2 commits into from
Mar 23, 2026
+15 −3
Merged

[Bugfix] Fix Qwen-Image SP and TeaCache incompatibility #2101

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
16e724e
fix qwen-image
wtomin Mar 23, 2026
23bda66
revert
wtomin Mar 23, 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
18 changes: 15 additions & 3 deletions vllm_omni/diffusion/cache/teacache/extractors.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -188,8 +188,21 @@ def extract_qwen_context(
# ============================================================================
# PREPROCESSING (Qwen-specific)
# ============================================================================
hidden_states = module.img_in(hidden_states)
# Call image_rope_prepare instead of img_in + pos_embed directly.
# This ensures the SequenceParallelSplitHook registered on image_rope_prepare
# fires when SP is enabled, correctly sharding hidden_states and vid_freqs.
hidden_states, vid_freqs, txt_freqs = module.image_rope_prepare(hidden_states, img_shapes, txt_seq_lens)
image_rotary_emb = (vid_freqs, txt_freqs)

timestep = timestep.to(device=hidden_states.device, dtype=hidden_states.dtype)

# Call modulate_index_prepare instead of handling timestep directly.
# For zero_cond_t=False: timestep unchanged, modulate_index=None.
# For zero_cond_t=True: timestep is doubled, modulate_index is created and
# sharded by the SequenceParallelSplitHook on modulate_index_prepare so that
# its sequence dimension matches the already-sharded hidden_states.
timestep, modulate_index = module.modulate_index_prepare(timestep, img_shapes)
Copy link
Copy Markdown

@chatgpt-codex-connector chatgpt-codex-connector Bot Mar 23, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Badge Keep TeaCache extractor aligned with zero_cond_t batching

For Qwen checkpoints with zero_cond_t=True (the image-edit variants), ModulateIndexPrepare.forward() doubles timestep here (qwen_image_transformer.py:141-154), so temb becomes 2 * batch. The rest of extract_qwen_context() still consumes that embedding as if it were batch: block.img_norm1(hidden_states, img_mod1) is still called without modulate_index, and postprocess() never chunks temb back down before module.norm_out, unlike QwenImageTransformer2DModel.forward() (qwen_image_transformer.py:1062-1065). With TeaCache enabled, Qwen edit models will therefore fail on the first forward with a batch-dimension mismatch instead of running the transformer.

Useful? React with 👍 / 👎.


encoder_hidden_states = module.txt_norm(encoder_hidden_states)
encoder_hidden_states = module.txt_in(encoder_hidden_states)

Expand All @@ -202,8 +215,6 @@ def extract_qwen_context(
else module.time_text_embed(timestep, guidance, hidden_states, additional_t_cond)
)

image_rotary_emb = module.pos_embed(img_shapes, txt_seq_lens, device=hidden_states.device)

# ============================================================================
# EXTRACT MODULATED INPUT (for cache decision)
# ============================================================================
Expand Down Expand Up @@ -249,6 +260,7 @@ def run_transformer_blocks():
temb=temb,
image_rotary_emb=image_rotary_emb,
joint_attention_kwargs=attention_kwargs,
modulate_index=modulate_index,
hidden_states_mask=hidden_states_mask,
)
return (h, e)
Expand Down
Loading