Skip to content

fix: error on device for lpw_stable_diffusion_xl pipeline if pipe.enable_sequential_cpu_offload() enabled #5885

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
merged 2 commits into from
Nov 27, 2023
Merged

fix: error on device for lpw_stable_diffusion_xl pipeline if pipe.enable_sequential_cpu_offload() enabled #5885

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c37292a
fix: set device for pipe.enable_sequential_cpu_offload()
VicGrygorchyk Nov 21, 2023
c1ae364
Merge branch 'main' into fix_device_for_lpw_sdxl
VicGrygorchyk Nov 26, 2023
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
24 changes: 14 additions & 10 deletions examples/community/lpw_stable_diffusion_xl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -250,6 +250,7 @@ def get_weighted_text_embeddings_sdxl(
neg_prompt: str = "",
neg_prompt_2: str = None,
num_images_per_prompt: int = 1,
device: Optional[torch.device] = None,
):
"""
This function can process long prompt with weights, no length limitation
Expand All @@ -262,10 +263,13 @@ def get_weighted_text_embeddings_sdxl(
neg_prompt (str)
neg_prompt_2 (str)
num_images_per_prompt (int)
device (torch.device)
Returns:
prompt_embeds (torch.Tensor)
neg_prompt_embeds (torch.Tensor)
"""
device = device or pipe._execution_device

if prompt_2:
prompt = f"{prompt} {prompt_2}"

Expand Down Expand Up @@ -330,17 +334,17 @@ def get_weighted_text_embeddings_sdxl(
# get prompt embeddings one by one is not working.
for i in range(len(prompt_token_groups)):
# get positive prompt embeddings with weights
token_tensor = torch.tensor([prompt_token_groups[i]], dtype=torch.long, device=pipe.device)
weight_tensor = torch.tensor(prompt_weight_groups[i], dtype=torch.float16, device=pipe.device)
token_tensor = torch.tensor([prompt_token_groups[i]], dtype=torch.long, device=device)
weight_tensor = torch.tensor(prompt_weight_groups[i], dtype=torch.float16, device=device)

token_tensor_2 = torch.tensor([prompt_token_groups_2[i]], dtype=torch.long, device=pipe.device)
token_tensor_2 = torch.tensor([prompt_token_groups_2[i]], dtype=torch.long, device=device)

# use first text encoder
prompt_embeds_1 = pipe.text_encoder(token_tensor.to(pipe.device), output_hidden_states=True)
prompt_embeds_1 = pipe.text_encoder(token_tensor.to(device), output_hidden_states=True)
prompt_embeds_1_hidden_states = prompt_embeds_1.hidden_states[-2]

# use second text encoder
prompt_embeds_2 = pipe.text_encoder_2(token_tensor_2.to(pipe.device), output_hidden_states=True)
prompt_embeds_2 = pipe.text_encoder_2(token_tensor_2.to(device), output_hidden_states=True)
prompt_embeds_2_hidden_states = prompt_embeds_2.hidden_states[-2]
pooled_prompt_embeds = prompt_embeds_2[0]

Expand All @@ -357,16 +361,16 @@ def get_weighted_text_embeddings_sdxl(
embeds.append(token_embedding)

# get negative prompt embeddings with weights
neg_token_tensor = torch.tensor([neg_prompt_token_groups[i]], dtype=torch.long, device=pipe.device)
neg_token_tensor_2 = torch.tensor([neg_prompt_token_groups_2[i]], dtype=torch.long, device=pipe.device)
neg_weight_tensor = torch.tensor(neg_prompt_weight_groups[i], dtype=torch.float16, device=pipe.device)
neg_token_tensor = torch.tensor([neg_prompt_token_groups[i]], dtype=torch.long, device=device)
neg_token_tensor_2 = torch.tensor([neg_prompt_token_groups_2[i]], dtype=torch.long, device=device)
neg_weight_tensor = torch.tensor(neg_prompt_weight_groups[i], dtype=torch.float16, device=device)

# use first text encoder
neg_prompt_embeds_1 = pipe.text_encoder(neg_token_tensor.to(pipe.device), output_hidden_states=True)
neg_prompt_embeds_1 = pipe.text_encoder(neg_token_tensor.to(device), output_hidden_states=True)
neg_prompt_embeds_1_hidden_states = neg_prompt_embeds_1.hidden_states[-2]

# use second text encoder
neg_prompt_embeds_2 = pipe.text_encoder_2(neg_token_tensor_2.to(pipe.device), output_hidden_states=True)
neg_prompt_embeds_2 = pipe.text_encoder_2(neg_token_tensor_2.to(device), output_hidden_states=True)
neg_prompt_embeds_2_hidden_states = neg_prompt_embeds_2.hidden_states[-2]
negative_pooled_prompt_embeds = neg_prompt_embeds_2[0]

Expand Down