Skip to content

Fix LCM distillation bug when creating the guidance scale embeddings using multiple GPUs. #6279

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
patil-suraj merged 2 commits into from
Dec 27, 2023
Merged

Fix LCM distillation bug when creating the guidance scale embeddings using multiple GPUs. #6279

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
bb7ffde
Fix bug when creating the guidance embeddings using multiple GPUs.
dg845 Dec 21, 2023
4e34e25
Merge branch 'main' into lcm-distill-fix-unet-time-cond-multi-gpu
sayakpaul Dec 24, 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
3 changes: 2 additions & 1 deletion examples/consistency_distillation/train_lcm_distill_sd_wds.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -889,6 +889,7 @@ def main(args):
# Add `time_cond_proj_dim` to the student U-Net if `teacher_unet.config.time_cond_proj_dim` is None
if teacher_unet.config.time_cond_proj_dim is None:
teacher_unet.config["time_cond_proj_dim"] = args.unet_time_cond_proj_dim
time_cond_proj_dim = teacher_unet.config.time_cond_proj_dim
unet = UNet2DConditionModel(**teacher_unet.config)
# load teacher_unet weights into unet
unet.load_state_dict(teacher_unet.state_dict(), strict=False)
Expand Down Expand Up @@ -1175,7 +1176,7 @@ def compute_embeddings(prompt_batch, proportion_empty_prompts, text_encoder, tok

# 5. Sample a random guidance scale w from U[w_min, w_max] and embed it
w = (args.w_max - args.w_min) * torch.rand((bsz,)) + args.w_min
w_embedding = guidance_scale_embedding(w, embedding_dim=unet.config.time_cond_proj_dim)
w_embedding = guidance_scale_embedding(w, embedding_dim=time_cond_proj_dim)
w = w.reshape(bsz, 1, 1, 1)
# Move to U-Net device and dtype
w = w.to(device=latents.device, dtype=latents.dtype)
Expand Down
3 changes: 2 additions & 1 deletion examples/consistency_distillation/train_lcm_distill_sdxl_wds.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -948,6 +948,7 @@ def main(args):
# Add `time_cond_proj_dim` to the student U-Net if `teacher_unet.config.time_cond_proj_dim` is None
if teacher_unet.config.time_cond_proj_dim is None:
teacher_unet.config["time_cond_proj_dim"] = args.unet_time_cond_proj_dim
time_cond_proj_dim = teacher_unet.config.time_cond_proj_dim
unet = UNet2DConditionModel(**teacher_unet.config)
# load teacher_unet weights into unet
unet.load_state_dict(teacher_unet.state_dict(), strict=False)
Expand Down Expand Up @@ -1273,7 +1274,7 @@ def compute_embeddings(

# 5. Sample a random guidance scale w from U[w_min, w_max] and embed it
w = (args.w_max - args.w_min) * torch.rand((bsz,)) + args.w_min
w_embedding = guidance_scale_embedding(w, embedding_dim=unet.config.time_cond_proj_dim)
w_embedding = guidance_scale_embedding(w, embedding_dim=time_cond_proj_dim)
w = w.reshape(bsz, 1, 1, 1)
# Move to U-Net device and dtype
w = w.to(device=latents.device, dtype=latents.dtype)
Expand Down