Skip to content

[diffusion] Skip negative prompt encoding when guidance_scale <= 1.0 or negative_prompt is None #16919

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
mickqian merged 5 commits into from
Jan 21, 2026
Merged

[diffusion] Skip negative prompt encoding when guidance_scale <= 1.0 or negative_prompt is None #16919

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
fb7c26b
[diffusion]remove neg text_encoder for guidance_scale<=1.0
zhuyuhua-v Jan 12, 2026
8729378
Merge branch 'main' into yuhua/neg_text_encode_cond
zhuyuhua-v Jan 12, 2026
961918b
Merge branch 'main' into yuhua/neg_text_encode_cond
zhuyuhua-v Jan 21, 2026
a29aa93
update guidance condition for classifier free guidance
zhuyuhua-v Jan 21, 2026
1c4098a
update neg_image_inputs and kwargs
zhuyuhua-v Jan 21, 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
37 changes: 20 additions & 17 deletions python/sglang/multimodal_gen/runtime/pipelines_core/stages/image_encoding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -125,15 +125,16 @@ def forward(
elif self.text_encoder:
# if a text encoder is provided, e.g. Qwen-Image-Edit
# 1. neg prompt embeds
neg_image_processor_kwargs = (
server_args.pipeline_config.prepare_image_processor_kwargs(
batch, neg=True
if batch.do_classifier_free_guidance:
neg_image_processor_kwargs = (
server_args.pipeline_config.prepare_image_processor_kwargs(
batch, neg=True
)
)
)

neg_image_inputs = self.image_processor(
images=image, return_tensors="pt", **neg_image_processor_kwargs
).to(cuda_device)
neg_image_inputs = self.image_processor(
images=image, return_tensors="pt", **neg_image_processor_kwargs
).to(cuda_device)

with set_forward_context(current_timestep=0, attn_metadata=None):
outputs = self.text_encoder(
Expand All @@ -143,20 +144,22 @@ def forward(
image_grid_thw=image_inputs.image_grid_thw,
output_hidden_states=True,
)
neg_outputs = self.text_encoder(
input_ids=neg_image_inputs.input_ids,
attention_mask=neg_image_inputs.attention_mask,
pixel_values=neg_image_inputs.pixel_values,
image_grid_thw=neg_image_inputs.image_grid_thw,
output_hidden_states=True,
)
if batch.do_classifier_free_guidance:
neg_outputs = self.text_encoder(
input_ids=neg_image_inputs.input_ids,
attention_mask=neg_image_inputs.attention_mask,
pixel_values=neg_image_inputs.pixel_values,
image_grid_thw=neg_image_inputs.image_grid_thw,
output_hidden_states=True,
)
batch.prompt_embeds.append(
self.encoding_qwen_image_edit(outputs, image_inputs)
)

batch.negative_prompt_embeds.append(
self.encoding_qwen_image_edit(neg_outputs, neg_image_inputs)
)
if batch.do_classifier_free_guidance:
batch.negative_prompt_embeds.append(
self.encoding_qwen_image_edit(neg_outputs, neg_image_inputs)
)

self.offload_model()

Expand Down
Loading