Multiply lr scheduler steps by `num_processes`. (#3983)

* Multiply lr scheduler steps by `num_processes`.

* Stop multiplying steps by gradient accumulation.

examples/controlnet/train_controlnet.py

@@ -897,8 +897,8 @@ def main(args):
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )

examples/custom_diffusion/train_custom_diffusion.py

@@ -1007,8 +1007,8 @@ def main(args):
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
     )
 
     # Prepare everything with our `accelerator`.

examples/dreambooth/train_dreambooth.py

@@ -1075,8 +1075,8 @@ def main(args):
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )

examples/dreambooth/train_dreambooth_lora.py

@@ -1039,8 +1039,8 @@ def main(args):
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )

examples/instruct_pix2pix/train_instruct_pix2pix.py

@@ -690,8 +690,8 @@ def main():
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
     )
 
     # Prepare everything with our `accelerator`.

examples/research_projects/dreambooth_inpaint/train_dreambooth_inpaint.py

@@ -600,8 +600,8 @@ def main():
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
     )
 
     if args.train_text_encoder:

examples/research_projects/dreambooth_inpaint/train_dreambooth_inpaint_lora.py

@@ -644,8 +644,8 @@ def main():
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
     )
 
     # Prepare everything with our `accelerator`.

examples/research_projects/intel_opts/textual_inversion/textual_inversion_bf16.py

@@ -481,8 +481,8 @@ def main():
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
     )
 
     text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(

examples/research_projects/intel_opts/textual_inversion_dfq/textual_inversion.py

@@ -588,8 +588,8 @@ def main():
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
     )
 
     if not train_unet:

examples/research_projects/lora/train_text_to_image_lora.py

@@ -701,8 +701,8 @@ def main():
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
     )
 
     # Prepare everything with our `accelerator`.

examples/research_projects/mulit_token_textual_inversion/textual_inversion.py

@@ -690,8 +690,8 @@ def main():
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
     )
 
     # Prepare everything with our `accelerator`.

examples/research_projects/multi_subject_dreambooth/train_multi_subject_dreambooth.py

@@ -970,8 +970,8 @@ def main(args):
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )

examples/research_projects/onnxruntime/text_to_image/train_text_to_image.py

@@ -732,8 +732,8 @@ def main():
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
     )
 
     # Prepare everything with our `accelerator`.

examples/research_projects/onnxruntime/textual_inversion/textual_inversion.py

@@ -741,8 +741,8 @@ def main():
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
     )
 
     # Prepare everything with our `accelerator`.

examples/text_to_image/train_text_to_image.py

@@ -819,8 +819,8 @@ def main():
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
     )
 
     # Prepare everything with our `accelerator`.

examples/text_to_image/train_text_to_image_lora.py

@@ -662,8 +662,8 @@ def main():
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
     )
 
     # Prepare everything with our `accelerator`.

examples/textual_inversion/textual_inversion.py

@@ -737,9 +737,9 @@ def main():
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * args.gradient_accumulation_steps,
-        num_training_steps=args.max_train_steps * args.gradient_accumulation_steps,
-        num_cycles=args.lr_num_cycles * args.gradient_accumulation_steps,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_cycles=args.lr_num_cycles,
     )
 
     # Prepare everything with our `accelerator`.