DiffusionAi/stable-diffusion-webui
1
0
Fork 0
mirror of https://github.com/AUTOMATIC1111/stable-diffusion-webui synced 2025-01-09 09:57:34 +08:00
Code Issues Packages Projects Releases Wiki Activity

faster by calculating R in updown and using cached R in forward

Browse Source
This commit is contained in:
v0xie 2023-10-18 04:56:53 -07:00
parent 853e21d98e
commit eb01d7f0e0
1 changed files with 8 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
extensions-builtin/Lora/network_oft.py
View File

@ -58,17 +58,18 @@ class NetworkModuleOFT(network.NetworkModule):
def calc_updown(self, orig_weight):
# this works
R = self.R
# R = self.R
self.R = self.get_weight(self.multiplier())
# this causes major deepfrying i.e. just doesn't work
# sending R to device causes major deepfrying i.e. just doesn't work
# R = self.R.to(orig_weight.device, dtype=orig_weight.dtype)
if orig_weight.dim() == 4:
weight = torch.einsum("oihw, op -> pihw", orig_weight, R)
else:
weight = torch.einsum("oi, op -> pi", orig_weight, R)
# if orig_weight.dim() == 4:
# weight = torch.einsum("oihw, op -> pihw", orig_weight, R)
# else:
# weight = torch.einsum("oi, op -> pi", orig_weight, R)
updown = orig_weight @ R
updown = orig_weight @ self.R
output_shape = self.oft_blocks.shape
## this works