serkandyck commited on
Commit
b5c0fe6
1 Parent(s): 540507b

Upload train_lora_flux_24gb.yaml with huggingface_hub

Browse files
Files changed (1) hide show
  1. train_lora_flux_24gb.yaml +98 -0
train_lora_flux_24gb.yaml ADDED
@@ -0,0 +1,98 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ job: extension
3
+ config:
4
+ # this name will be the folder and filename name
5
+ name: "my_first_flux_lora_v1"
6
+ process:
7
+ - type: 'sd_trainer'
8
+ # root folder to save training sessions/samples/weights
9
+ training_folder: "output"
10
+ # uncomment to see performance stats in the terminal every N steps
11
+
12
+ # performance_log_every: 1000
13
+ device: cuda:0
14
+ # if a trigger word is specified, it will be added to captions of training data if it does not already exist
15
+ # alternatively, in your captions you can add [trigger] and it will be replaced with the trigger word
16
+
17
+ # trigger_word: "p3r5on"
18
+ network:
19
+ type: "lora"
20
+ linear: 16
21
+ linear_alpha: 16
22
+ save:
23
+ dtype: float16 # precision to save
24
+ save_every: 250 # save every this many steps
25
+ max_step_saves_to_keep: 4 # how many intermittent saves to keep
26
+ push_to_hub: false #change this to True to push your trained model to Hugging Face.
27
+ # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in
28
+ # hf_repo_id: your-username/your-model-slug
29
+ # hf_private: true #whether the repo is private or public
30
+
31
+ datasets:
32
+ # datasets are a folder of images. captions need to be txt files with the same name as the image
33
+ # for instance image2.jpg and image2.txt. Only jpg, jpeg, and png are supported currently
34
+ # images will automatically be resized and bucketed into the resolution specified
35
+ # on windows, escape back slashes with another backslash so
36
+ # "C:\\path\\to\\images\\folder"
37
+ - folder_path: "/workspace/ai-toolkit/images"
38
+ caption_ext: "txt"
39
+ caption_dropout_rate: 0.05 # will drop out the caption 5% of time
40
+ shuffle_tokens: false # shuffle caption order, split by commas
41
+ cache_latents_to_disk: true # leave this true unless you know what you're doing
42
+ resolution: [512, 768, 1024] # flux enjoys multiple resolutions
43
+ train:
44
+ batch_size: 1
45
+ steps: 2000 # total number of steps to train 500 - 4000 is a good range
46
+ gradient_accumulation_steps: 1
47
+ train_unet: true
48
+ train_text_encoder: false # probably won't work with flux
49
+ gradient_checkpointing: true # need the on unless you have a ton of vram
50
+ noise_scheduler: "flowmatch" # for training only
51
+ optimizer: "adamw8bit"
52
+ lr: 1e-4
53
+ # skip_first_sample: true
54
+ # uncomment to completely disable sampling
55
+ # disable_sampling: true
56
+ # uncomment to use new vell curved weighting. Experimental but may produce better results
57
+ # linear_timesteps: true
58
+
59
+ # ema will smooth out learning, but could slow it down. Recommended to leave on.
60
+ ema_config:
61
+ use_ema: true
62
+ ema_decay: 0.99
63
+ # will probably need this if gpu supports it for flux, other dtypes may not work correctly
64
+ dtype: bf16
65
+ model:
66
+ # huggingface model name or path
67
+ name_or_path: "black-forest-labs/FLUX.1-dev"
68
+ is_flux: true
69
+ quantize: true # run 8bit mixed precision
70
+ # low_vram: true # uncomment this if the GPU is connected to your monitors. It will use less vram to quantize, but is slower.
71
+ sample:
72
+ sampler: "flowmatch" # must match train.noise_scheduler
73
+ sample_every: 250 # sample every this many steps
74
+ width: 1024
75
+ height: 1024
76
+ prompts:
77
+ # - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
78
+ - "woman with red hair, playing chess at the park, bomb going off in the background"
79
+ # you can add [trigger] to the prompts here and it will be replaced with the trigger word
80
+
81
+ - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
82
+ - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
83
+ - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
84
+ - "a bear building a log cabin in the snow covered mountains"
85
+ - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
86
+ - "hipster man with a beard, building a chair, in a wood shop"
87
+ - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
88
+ - "a man holding a sign that says, 'this is a sign'"
89
+ - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
90
+ neg: "" # not used on flux
91
+ seed: 42
92
+ walk_seed: true
93
+ guidance_scale: 4
94
+ sample_steps: 20
95
+ # you can add any additional meta info here. [name] is replaced with config name at top
96
+ meta:
97
+ name: "[name]"
98
+ version: '1.0'