[training] use the lr when using 8bit adam. (huggingface#9796)

* use the lr when using 8bit adam.

* remove lr as we pack it in params_to_optimize.

---------

Co-authored-by: Linoy Tsaban <57615435+linoytsaban@users.noreply.github.com>

Author: sayakpaul

examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py

@@ -1778,15 +1778,10 @@ def load_model_hook(models, input_dir):
         if not args.enable_t5_ti:
             # pure textual inversion - only clip
             if pure_textual_inversion:
-                params_to_optimize = [
-                    text_parameters_one_with_lr,
-                ]
+                params_to_optimize = [text_parameters_one_with_lr]
                 te_idx = 0
             else:  # regular te training or regular pivotal for clip
-                params_to_optimize = [
-                    transformer_parameters_with_lr,
-                    text_parameters_one_with_lr,
-                ]
+                params_to_optimize = [transformer_parameters_with_lr, text_parameters_one_with_lr]
                 te_idx = 1
         elif args.enable_t5_ti:
             # pivotal tuning of clip & t5
@@ -1809,9 +1804,7 @@ def load_model_hook(models, input_dir):
                 ]
                 te_idx = 1
     else:
-        params_to_optimize = [
-            transformer_parameters_with_lr,
-        ]
+        params_to_optimize = [transformer_parameters_with_lr]
 
     # Optimizer creation
     if not (args.optimizer.lower() == "prodigy" or args.optimizer.lower() == "adamw"):
@@ -1871,7 +1864,6 @@ def load_model_hook(models, input_dir):
             params_to_optimize[-1]["lr"] = args.learning_rate
         optimizer = optimizer_class(
             params_to_optimize,
-            lr=args.learning_rate,
             betas=(args.adam_beta1, args.adam_beta2),
             beta3=args.prodigy_beta3,
             weight_decay=args.adam_weight_decay,

examples/advanced_diffusion_training/train_dreambooth_lora_sd15_advanced.py

@@ -1358,10 +1358,7 @@ def load_model_hook(models, input_dir):
             else args.adam_weight_decay,
             "lr": args.text_encoder_lr if args.text_encoder_lr else args.learning_rate,
         }
-        params_to_optimize = [
-            unet_lora_parameters_with_lr,
-            text_lora_parameters_one_with_lr,
-        ]
+        params_to_optimize = [unet_lora_parameters_with_lr, text_lora_parameters_one_with_lr]
     else:
         params_to_optimize = [unet_lora_parameters_with_lr]
 
@@ -1423,7 +1420,6 @@ def load_model_hook(models, input_dir):
 
         optimizer = optimizer_class(
             params_to_optimize,
-            lr=args.learning_rate,
             betas=(args.adam_beta1, args.adam_beta2),
             beta3=args.prodigy_beta3,
             weight_decay=args.adam_weight_decay,

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -1794,7 +1794,6 @@ def load_model_hook(models, input_dir):
 
         optimizer = optimizer_class(
             params_to_optimize,
-            lr=args.learning_rate,
             betas=(args.adam_beta1, args.adam_beta2),
             beta3=args.prodigy_beta3,
             weight_decay=args.adam_weight_decay,

examples/cogvideo/train_cogvideox_image_to_video_lora.py

@@ -947,7 +947,6 @@ def get_optimizer(args, params_to_optimize, use_deepspeed: bool = False):
 
         optimizer = optimizer_class(
             params_to_optimize,
-            lr=args.learning_rate,
             betas=(args.adam_beta1, args.adam_beta2),
             beta3=args.prodigy_beta3,
             weight_decay=args.adam_weight_decay,

examples/cogvideo/train_cogvideox_lora.py

@@ -969,7 +969,6 @@ def get_optimizer(args, params_to_optimize, use_deepspeed: bool = False):
 
         optimizer = optimizer_class(
             params_to_optimize,
-            lr=args.learning_rate,
             betas=(args.adam_beta1, args.adam_beta2),
             beta3=args.prodigy_beta3,
             weight_decay=args.adam_weight_decay,

examples/dreambooth/train_dreambooth_flux.py

@@ -1226,10 +1226,7 @@ def load_model_hook(models, input_dir):
             "weight_decay": args.adam_weight_decay_text_encoder,
             "lr": args.text_encoder_lr if args.text_encoder_lr else args.learning_rate,
         }
-        params_to_optimize = [
-            transformer_parameters_with_lr,
-            text_parameters_one_with_lr,
-        ]
+        params_to_optimize = [transformer_parameters_with_lr, text_parameters_one_with_lr]
     else:
         params_to_optimize = [transformer_parameters_with_lr]
 
@@ -1291,7 +1288,6 @@ def load_model_hook(models, input_dir):
 
         optimizer = optimizer_class(
             params_to_optimize,
-            lr=args.learning_rate,
             betas=(args.adam_beta1, args.adam_beta2),
             beta3=args.prodigy_beta3,
             weight_decay=args.adam_weight_decay,

examples/dreambooth/train_dreambooth_lora_flux.py

@@ -1335,10 +1335,7 @@ def load_model_hook(models, input_dir):
             "weight_decay": args.adam_weight_decay_text_encoder,
             "lr": args.text_encoder_lr if args.text_encoder_lr else args.learning_rate,
         }
-        params_to_optimize = [
-            transformer_parameters_with_lr,
-            text_parameters_one_with_lr,
-        ]
+        params_to_optimize = [transformer_parameters_with_lr, text_parameters_one_with_lr]
     else:
         params_to_optimize = [transformer_parameters_with_lr]
 
@@ -1400,7 +1397,6 @@ def load_model_hook(models, input_dir):
 
         optimizer = optimizer_class(
             params_to_optimize,
-            lr=args.learning_rate,
             betas=(args.adam_beta1, args.adam_beta2),
             beta3=args.prodigy_beta3,
             weight_decay=args.adam_weight_decay,

examples/dreambooth/train_dreambooth_lora_sd3.py

@@ -1468,7 +1468,6 @@ def load_model_hook(models, input_dir):
 
         optimizer = optimizer_class(
             params_to_optimize,
-            lr=args.learning_rate,
             betas=(args.adam_beta1, args.adam_beta2),
             beta3=args.prodigy_beta3,
             weight_decay=args.adam_weight_decay,

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -1402,7 +1402,6 @@ def load_model_hook(models, input_dir):
 
         optimizer = optimizer_class(
             params_to_optimize,
-            lr=args.learning_rate,
             betas=(args.adam_beta1, args.adam_beta2),
             beta3=args.prodigy_beta3,
             weight_decay=args.adam_weight_decay,

examples/dreambooth/train_dreambooth_sd3.py

@@ -1328,7 +1328,6 @@ def load_model_hook(models, input_dir):
 
         optimizer = optimizer_class(
             params_to_optimize,
-            lr=args.learning_rate,
             betas=(args.adam_beta1, args.adam_beta2),
             beta3=args.prodigy_beta3,
             weight_decay=args.adam_weight_decay,