Skip to content

Add cache_context wrapping to all Flux pipelines #14923

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
Reis-McMillan wants to merge 2 commits into
base: main
Choose a base branch
from
+307 −202
Open

Add cache_context wrapping to all Flux pipelines #14923

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e3c689d
added TaylorSeer cache context to Flux Kontext pipeline
Reis-McMillan Oct 1, 2026
52e4e53
extend TaylorSeer caching to all Flux pipleines
Reis-McMillan Oct 1, 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
23 changes: 12 additions & 11 deletions src/diffusers/pipelines/flux/pipeline_flux_control.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -809,17 +809,18 @@ def __call__(
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timestep = t.expand(latents.shape[0]).to(latents.dtype)

noise_pred = self.transformer(
hidden_states=latent_model_input,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]
with self.transformer.cache_context("cond"):
noise_pred = self.transformer(
hidden_states=latent_model_input,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]

# compute the previous noisy sample x_t -> x_t-1
latents_dtype = latents.dtype
Expand Down
23 changes: 12 additions & 11 deletions src/diffusers/pipelines/flux/pipeline_flux_control_img2img.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -892,17 +892,18 @@ def __call__(
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timestep = t.expand(latents.shape[0]).to(latents.dtype)

noise_pred = self.transformer(
hidden_states=latent_model_input,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]
with self.transformer.cache_context("cond"):
noise_pred = self.transformer(
hidden_states=latent_model_input,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]

# compute the previous noisy sample x_t -> x_t-1
latents_dtype = latents.dtype
Expand Down
23 changes: 12 additions & 11 deletions src/diffusers/pipelines/flux/pipeline_flux_control_inpaint.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1046,17 +1046,18 @@ def __call__(
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timestep = t.expand(latents.shape[0]).to(latents.dtype)

noise_pred = self.transformer(
hidden_states=latent_model_input,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]
with self.transformer.cache_context("cond"):
noise_pred = self.transformer(
hidden_states=latent_model_input,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]

# compute the previous noisy sample x_t -> x_t-1
latents_dtype = latents.dtype
Expand Down
44 changes: 23 additions & 21 deletions src/diffusers/pipelines/flux/pipeline_flux_controlnet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1124,30 +1124,13 @@ def __call__(
)
guidance = guidance.expand(latents.shape[0]) if guidance is not None else None

noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
controlnet_block_samples=controlnet_block_samples,
controlnet_single_block_samples=controlnet_single_block_samples,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
controlnet_blocks_repeat=controlnet_blocks_repeat,
)[0]

if do_true_cfg:
if negative_image_embeds is not None:
self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds
neg_noise_pred = self.transformer(
with self.transformer.cache_context("cond"):
noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=negative_pooled_prompt_embeds,
encoder_hidden_states=negative_prompt_embeds,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
controlnet_block_samples=controlnet_block_samples,
controlnet_single_block_samples=controlnet_single_block_samples,
txt_ids=text_ids,
Expand All @@ -1156,6 +1139,25 @@ def __call__(
return_dict=False,
controlnet_blocks_repeat=controlnet_blocks_repeat,
)[0]

if do_true_cfg:
if negative_image_embeds is not None:
self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds
with self.transformer.cache_context("uncond"):
neg_noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=negative_pooled_prompt_embeds,
encoder_hidden_states=negative_prompt_embeds,
controlnet_block_samples=controlnet_block_samples,
controlnet_single_block_samples=controlnet_single_block_samples,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
controlnet_blocks_repeat=controlnet_blocks_repeat,
)[0]
noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)

# compute the previous noisy sample x_t -> x_t-1
Expand Down
29 changes: 15 additions & 14 deletions src/diffusers/pipelines/flux/pipeline_flux_controlnet_image_to_image.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -961,20 +961,21 @@ def __call__(
)
guidance = guidance.expand(latents.shape[0]) if guidance is not None else None

noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
controlnet_block_samples=controlnet_block_samples,
controlnet_single_block_samples=controlnet_single_block_samples,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
controlnet_blocks_repeat=controlnet_blocks_repeat,
)[0]
with self.transformer.cache_context("cond"):
noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
controlnet_block_samples=controlnet_block_samples,
controlnet_single_block_samples=controlnet_single_block_samples,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
controlnet_blocks_repeat=controlnet_blocks_repeat,
)[0]

latents_dtype = latents.dtype
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
Expand Down
29 changes: 15 additions & 14 deletions src/diffusers/pipelines/flux/pipeline_flux_controlnet_inpainting.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1139,20 +1139,21 @@ def __call__(
else:
guidance = None

noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
controlnet_block_samples=controlnet_block_samples,
controlnet_single_block_samples=controlnet_single_block_samples,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
controlnet_blocks_repeat=controlnet_blocks_repeat,
)[0]
with self.transformer.cache_context("cond"):
noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
controlnet_block_samples=controlnet_block_samples,
controlnet_single_block_samples=controlnet_single_block_samples,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
controlnet_blocks_repeat=controlnet_blocks_repeat,
)[0]

# compute the previous noisy sample x_t -> x_t-1
latents_dtype = latents.dtype
Expand Down
23 changes: 12 additions & 11 deletions src/diffusers/pipelines/flux/pipeline_flux_fill.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -958,17 +958,18 @@ def __call__(
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timestep = t.expand(latents.shape[0]).to(latents.dtype)

noise_pred = self.transformer(
hidden_states=torch.cat((latents, masked_image_latents), dim=2),
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]
with self.transformer.cache_context("cond"):
noise_pred = self.transformer(
hidden_states=torch.cat((latents, masked_image_latents), dim=2),
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]

# compute the previous noisy sample x_t -> x_t-1
latents_dtype = latents.dtype
Expand Down
38 changes: 20 additions & 18 deletions src/diffusers/pipelines/flux/pipeline_flux_img2img.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -990,32 +990,34 @@ def __call__(
self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timestep = t.expand(latents.shape[0]).to(latents.dtype)
noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]

if do_true_cfg:
if negative_image_embeds is not None:
self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds
neg_noise_pred = self.transformer(
with self.transformer.cache_context("cond"):
noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=negative_pooled_prompt_embeds,
encoder_hidden_states=negative_prompt_embeds,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]

if do_true_cfg:
if negative_image_embeds is not None:
self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds
with self.transformer.cache_context("uncond"):
neg_noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=negative_pooled_prompt_embeds,
encoder_hidden_states=negative_prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]
noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)

# compute the previous noisy sample x_t -> x_t-1
Expand Down
38 changes: 20 additions & 18 deletions src/diffusers/pipelines/flux/pipeline_flux_inpaint.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1146,32 +1146,34 @@ def __call__(
self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
timestep = t.expand(latents.shape[0]).to(latents.dtype)
noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]

if do_true_cfg:
if negative_image_embeds is not None:
self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds
neg_noise_pred = self.transformer(
with self.transformer.cache_context("cond"):
noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=negative_pooled_prompt_embeds,
encoder_hidden_states=negative_prompt_embeds,
pooled_projections=pooled_prompt_embeds,
encoder_hidden_states=prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]

if do_true_cfg:
if negative_image_embeds is not None:
self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds
with self.transformer.cache_context("uncond"):
neg_noise_pred = self.transformer(
hidden_states=latents,
timestep=timestep / 1000,
guidance=guidance,
pooled_projections=negative_pooled_prompt_embeds,
encoder_hidden_states=negative_prompt_embeds,
txt_ids=text_ids,
img_ids=latent_image_ids,
joint_attention_kwargs=self.joint_attention_kwargs,
return_dict=False,
)[0]
noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)

# compute the previous noisy sample x_t -> x_t-1
Expand Down
Loading
Loading