Upload configs/train_high.toml with huggingface_hub
Browse files- configs/train_high.toml +56 -0
configs/train_high.toml
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Wan 2.2 I2V 14B - HIGH NOISE LoRA training
|
| 2 |
+
# High noise model handles motion, jumpcut timing, overall structure
|
| 3 |
+
|
| 4 |
+
output_dir = '/home/u/work/output_high'
|
| 5 |
+
dataset = '/home/u/work/diffusion-pipe/configs/dataset_blink.toml'
|
| 6 |
+
|
| 7 |
+
# training settings
|
| 8 |
+
epochs = 1000
|
| 9 |
+
max_steps = 15000
|
| 10 |
+
micro_batch_size_per_gpu = 2
|
| 11 |
+
pipeline_stages = 1
|
| 12 |
+
gradient_accumulation_steps = 1
|
| 13 |
+
gradient_clipping = 1.0
|
| 14 |
+
warmup_steps = 100
|
| 15 |
+
|
| 16 |
+
# eval settings
|
| 17 |
+
eval_every_n_epochs = 1
|
| 18 |
+
eval_before_first_step = true
|
| 19 |
+
eval_micro_batch_size_per_gpu = 1
|
| 20 |
+
eval_gradient_accumulation_steps = 1
|
| 21 |
+
|
| 22 |
+
# misc settings
|
| 23 |
+
save_every_n_steps = 500
|
| 24 |
+
checkpoint_every_n_minutes = 60
|
| 25 |
+
activation_checkpointing = true
|
| 26 |
+
compile = true
|
| 27 |
+
partition_method = 'parameters'
|
| 28 |
+
save_dtype = 'bfloat16'
|
| 29 |
+
caching_batch_size = 1
|
| 30 |
+
steps_per_print = 1
|
| 31 |
+
video_clip_mode = 'single_beginning'
|
| 32 |
+
|
| 33 |
+
[model]
|
| 34 |
+
type = 'wan'
|
| 35 |
+
ckpt_path = '/home/u/work/models/Wan2.2-I2V-A14B'
|
| 36 |
+
transformer_path = '/home/u/work/models/comfyui/split_files/diffusion_models/wan2.2_i2v_high_noise_14B_fp16.safetensors'
|
| 37 |
+
llm_path = '/home/u/work/models/comfyui/split_files/text_encoders/umt5_xxl_fp16.safetensors'
|
| 38 |
+
dtype = 'bfloat16'
|
| 39 |
+
transformer_dtype = 'float8'
|
| 40 |
+
timestep_sample_method = 'logit_normal'
|
| 41 |
+
# High noise timestep range for I2V (boundary = 0.9)
|
| 42 |
+
min_t = 0.9
|
| 43 |
+
max_t = 1
|
| 44 |
+
|
| 45 |
+
[adapter]
|
| 46 |
+
type = 'lora'
|
| 47 |
+
rank = 16
|
| 48 |
+
dtype = 'bfloat16'
|
| 49 |
+
init_from_existing = '/home/u/work/output_high/20260225_19-24-02/step2000'
|
| 50 |
+
|
| 51 |
+
[optimizer]
|
| 52 |
+
type = 'adamw_optimi'
|
| 53 |
+
lr = 2e-5
|
| 54 |
+
betas = [0.9, 0.99]
|
| 55 |
+
weight_decay = 0.01
|
| 56 |
+
eps = 1e-8
|