Upload qwen_toml.toml
Browse files- qwen_toml.toml +32 -0
qwen_toml.toml
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Output path for training runs. Each training run makes a new directory in here.
|
| 2 |
+
output_dir = '/diffusion_pipe_working_folder/output_folder/qwen_lora'
|
| 3 |
+
|
| 4 |
+
save_every_n_epochs = 20
|
| 5 |
+
epochs = 80
|
| 6 |
+
pipeline_stages = 1
|
| 7 |
+
micro_batch_size_per_gpu = 1
|
| 8 |
+
gradient_accumulation_steps = 1
|
| 9 |
+
activation_checkpointing = true
|
| 10 |
+
dataset = 'examples/dataset.toml'
|
| 11 |
+
|
| 12 |
+
[model]
|
| 13 |
+
type = 'qwen_image'
|
| 14 |
+
diffusers_path = '/diffusion_pipe_working_folder/models/Qwen-Image'
|
| 15 |
+
dtype = 'bfloat16'
|
| 16 |
+
transformer_dtype = 'float8'
|
| 17 |
+
timestep_sample_method = 'logit_normal'
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
[adapter]
|
| 21 |
+
type = "lora"
|
| 22 |
+
rank = 32
|
| 23 |
+
dtype = "bfloat16"
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
[optimizer]
|
| 28 |
+
type = 'adamw_optimi'
|
| 29 |
+
lr = 2e-4
|
| 30 |
+
betas = [0.9, 0.99]
|
| 31 |
+
weight_decay = 0.01
|
| 32 |
+
eps = 1e-8
|