From 76d1430895316690f5fbda6f74e3b4fbb184fd22 Mon Sep 17 00:00:00 2001
From: Pepijn <pepijn@huggingface.co>
Date: Wed, 24 Sep 2025 10:19:57 +0200
Subject: [PATCH] remove lr scaling

---
 src/lerobot/configs/train.py |  1 -
 src/lerobot/scripts/train.py | 14 +-------------
 2 files changed, 1 insertion(+), 14 deletions(-)

diff --git a/src/lerobot/configs/train.py b/src/lerobot/configs/train.py
index 5b3b0fdb8..f2d07cd7f 100644
--- a/src/lerobot/configs/train.py
+++ b/src/lerobot/configs/train.py
@@ -67,7 +67,6 @@ class TrainPipelineConfig(HubMixin):
     use_accelerate: bool = False
     gradient_accumulation_steps: int = 1
     mixed_precision: str = "no"  # Options: "no", "fp16", "bf16"
-    scale_lr_with_num_gpus: bool = True  # Automatically scale learning rate with number of GPUs
 
     def __post_init__(self):
         self.checkpoint_path = None
diff --git a/src/lerobot/scripts/train.py b/src/lerobot/scripts/train.py
index 5a79e8a93..90e7807ba 100644
--- a/src/lerobot/scripts/train.py
+++ b/src/lerobot/scripts/train.py
@@ -248,17 +248,6 @@ def train(cfg: TrainPipelineConfig):
         logging.info("Creating optimizer and scheduler")
     optimizer, lr_scheduler = make_optimizer_and_scheduler(cfg, policy)
 
-    # Scale learning rate for multi-GPU training
-    if accelerator is not None and accelerator.num_processes > 1 and cfg.scale_lr_with_num_gpus:
-        # Scale learning rate linearly with number of GPUs
-        original_lr = optimizer.param_groups[0]["lr"]
-        for param_group in optimizer.param_groups:
-            param_group["lr"] *= accelerator.num_processes
-        if accelerator.is_main_process:
-            logging.info(
-                f"Scaled learning rate by {accelerator.num_processes}x for multi-GPU training: {original_lr:.2e} -> {optimizer.param_groups[0]['lr']:.2e}"
-            )
-
     grad_scaler = GradScaler(device.type, enabled=cfg.policy.use_amp)
 
     step = 0  # number of policy updates (forward + backward + optim)
@@ -292,10 +281,9 @@ def train(cfg: TrainPipelineConfig):
         if accelerator is not None:
             logging.info(f"Per-GPU batch size: {cfg.batch_size}")
             logging.info(f"Effective batch size (total): {cfg.batch_size * accelerator.num_processes}")
-            logging.info(f"Learning rate (scaled): {optimizer.param_groups[0]['lr']:.2e}")
         else:
             logging.info(f"Batch size: {cfg.batch_size}")
-            logging.info(f"Learning rate: {optimizer.param_groups[0]['lr']:.2e}")
+        logging.info(f"Learning rate: {optimizer.param_groups[0]['lr']:.2e}")
 
     # create dataloader for offline training
     if hasattr(cfg.policy, "drop_n_last_frames"):