Upload config.yml
Browse files- shortcut-09/config.yml +104 -0
shortcut-09/config.yml
ADDED
@@ -0,0 +1,104 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
model:
|
2 |
+
checkpoint_path: "./models/aura_flow_0.3.bnb_nf4.safetensors"
|
3 |
+
pretrained_model_name_or_path: fal/AuraFlow-v0.3
|
4 |
+
|
5 |
+
dtype: bfloat16
|
6 |
+
|
7 |
+
denoiser:
|
8 |
+
use_flash_attn: true
|
9 |
+
|
10 |
+
use_shortcut: true
|
11 |
+
|
12 |
+
flow_matching_ratio: 0.75
|
13 |
+
shortcut_max_steps: 128
|
14 |
+
|
15 |
+
timestep_sampling_type: "sigmoid"
|
16 |
+
|
17 |
+
peft:
|
18 |
+
type: lora
|
19 |
+
rank: 4
|
20 |
+
alpha: 4.0
|
21 |
+
dropout: 0.0
|
22 |
+
|
23 |
+
dtype: bfloat16
|
24 |
+
|
25 |
+
# include the AdaLN-Zero modulation layers
|
26 |
+
include_keys: [
|
27 |
+
".attn.",
|
28 |
+
".mlp.",
|
29 |
+
".mlpC.",
|
30 |
+
".mlpX.",
|
31 |
+
#
|
32 |
+
".modC.",
|
33 |
+
".modX.",
|
34 |
+
".modCX.",
|
35 |
+
]
|
36 |
+
exclude_keys: ["text_encoder", "vae", "t_embedder", "final_linear", ".modF."]
|
37 |
+
|
38 |
+
dataset:
|
39 |
+
folder: "data/pexels-1k-random"
|
40 |
+
num_repeats: 2
|
41 |
+
batch_size: 2
|
42 |
+
|
43 |
+
bucket_base_size: 1024
|
44 |
+
step: 128
|
45 |
+
min_size: 384
|
46 |
+
do_upscale: false
|
47 |
+
|
48 |
+
caption_processors: []
|
49 |
+
|
50 |
+
optimizer:
|
51 |
+
name: "schedulefree.RAdamScheduleFree"
|
52 |
+
args:
|
53 |
+
lr: 0.0005
|
54 |
+
|
55 |
+
tracker:
|
56 |
+
project_name: "auraflow-shortcut-1"
|
57 |
+
loggers:
|
58 |
+
- wandb
|
59 |
+
|
60 |
+
saving:
|
61 |
+
strategy:
|
62 |
+
per_epochs: 1
|
63 |
+
per_steps: null
|
64 |
+
save_last: true
|
65 |
+
|
66 |
+
callbacks:
|
67 |
+
- type: "hf_hub"
|
68 |
+
# - type: "safetensors"
|
69 |
+
name: "shortcut-09"
|
70 |
+
save_dir: "./output/shortcut-09"
|
71 |
+
|
72 |
+
hub_id: "p1atdev/afv03-lora"
|
73 |
+
dir_in_repo: "shortcut-09"
|
74 |
+
|
75 |
+
preview:
|
76 |
+
strategy:
|
77 |
+
per_epochs: 1
|
78 |
+
per_steps: 100
|
79 |
+
|
80 |
+
callbacks:
|
81 |
+
# - type: "local"
|
82 |
+
# save_dir: "./output/shortcut-08/preview"
|
83 |
+
|
84 |
+
- type: "discord"
|
85 |
+
url: "masked"
|
86 |
+
|
87 |
+
data:
|
88 |
+
path: "./projects/shortcut/preview.yml"
|
89 |
+
|
90 |
+
seed: 42
|
91 |
+
num_train_epochs: 20
|
92 |
+
|
93 |
+
trainer:
|
94 |
+
# debug_mode: "1step"
|
95 |
+
|
96 |
+
gradient_checkpointing: true
|
97 |
+
gradient_accumulation_steps: 16
|
98 |
+
|
99 |
+
torch_compile: true
|
100 |
+
torch_compile_args:
|
101 |
+
mode: max-autotune
|
102 |
+
fullgraph: true
|
103 |
+
|
104 |
+
fp32_matmul_precision: "medium"
|