Skip to content

⚡️ Speed up method LoRALayer.fuse_weights by 8% #138

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
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

⚡️ Speed up method LoRALayer.fuse_weights by 8% #138

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
30 changes: 20 additions & 10 deletions invokeai/backend/patches/layers/lora_layer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,20 +67,30 @@ def fuse_weights(self, up: torch.Tensor, down: torch.Tensor) -> torch.Tensor:
the LoRA weights have different ranks. This function handles the fusion of these differently sized
matrices.
"""
up_rows, up_cols = up.shape
down_rows, down_cols = down.shape
fused_lora = torch.zeros((up_rows, down_cols), device=down.device, dtype=down.dtype)
rank_diff = down_rows / up_cols

fused_lora = torch.zeros((up.shape[0], down.shape[1]), device=down.device, dtype=down.dtype)
rank_diff = down.shape[0] / up.shape[1]
# optimization: coalesce operation with torch.stack and .sum to avoid python-level looping

if rank_diff > 1:
rank_diff = down.shape[0] / up.shape[1]
w_down = down.chunk(int(rank_diff), dim=0)
for w_down_chunk in w_down:
fused_lora = fused_lora + (torch.mm(up, w_down_chunk))
# split down along the row dimension and compute mm for each chunk
num_chunks = int(rank_diff)
w_down = down.chunk(num_chunks, dim=0)
# if num_chunks == 1, fallback to simpler matmul
if num_chunks == 1:
fused_lora = torch.mm(up, down)
else:
fused_lora = torch.stack([torch.mm(up, w) for w in w_down], dim=0).sum(dim=0)
else:
rank_diff = up.shape[1] / down.shape[0]
w_up = up.chunk(int(rank_diff), dim=0)
for w_up_chunk in w_up:
fused_lora = fused_lora + (torch.mm(w_up_chunk, down))
rank_diff = up_cols / down_rows
num_chunks = int(rank_diff)
w_up = up.chunk(num_chunks, dim=0)
if num_chunks == 1:
fused_lora = torch.mm(up, down)
else:
fused_lora = torch.stack([torch.mm(w, down) for w in w_up], dim=0).sum(dim=0)

return fused_lora

Expand Down