|
enable_conv = true |
|
|
|
|
|
unet_target_module = [ |
|
"Transformer2DModel", |
|
"ResnetBlock2D", |
|
"Downsample2D", |
|
"Upsample2D", |
|
] |
|
unet_target_name = [ |
|
".*time_embed\\..+", |
|
".*label_emb\\..+", |
|
".*input_blocks\\.0.+", |
|
"^out\\..+" |
|
] |
|
text_encoder_target_module = [ |
|
"CLIPAttention", |
|
"CLIPMLP", |
|
] |
|
text_encoder_target_name = [ |
|
|
|
] |
|
|
|
|
|
[module_algo_map] |
|
[module_algo_map.CrossAttention] |
|
algo = "lokr" |
|
dim = 10000 |
|
alpha = 10000 |
|
factor = 4 |
|
|
|
[module_algo_map.FeedForward] |
|
algo = "lokr" |
|
dim = 10000 |
|
alpha = 10000 |
|
factor = 2 |
|
|
|
[module_algo_map.CLIPAttention] |
|
algo = "lokr" |
|
dim = 10000 |
|
alpha = 10000 |
|
factor = 8 |
|
|
|
[module_algo_map.CLIPMLP] |
|
algo = "lokr" |
|
dim = 10000 |
|
alpha = 10000 |
|
factor = 8 |