diff --git a/lerobot/scripts/train.py b/lerobot/scripts/train.py
index 59744c7e..8648eb1b 100644
--- a/lerobot/scripts/train.py
+++ b/lerobot/scripts/train.py
@@ -57,12 +57,16 @@ from lerobot.common.datasets.video_utils import (
 # let's define a custom fn
 
 def custom_collate_fn(batch):
+    # always in the cuda, getitem is on cpu, 
+    # then implement mixed 
     """
     Custom collate function that decodes videos on GPU/CPU.
     Converts the batch to a dictionary with keys representing each field.
     Returns a tensor for video frames instead of a list.
     """
+    # know when it is called
     final_batch = {}
+    is_main_process = torch.utils.data.get_worker_info() is None
 
     # the batch is given as a list, we need to return a dict
     for item in batch:
@@ -71,14 +75,17 @@ def custom_collate_fn(batch):
             for vid_key, video_path in item["video_paths"].items():
                 # decode video frames based on timestamps
                 timestamps = item["query_timestamps"][vid_key]
-                frames = decode_video_frames_torchvision(
+                
+                # ✅ Use CUDA only in the main process
+                device = "cuda" if is_main_process else "cpu"
+                frames = decode_video_frames_torchcodec(
                     video_path=Path(video_path),
                     timestamps=timestamps,
                     tolerance_s=0.02,
-                    backend="pyav",
+                    # backend="pyav",
                     log_loaded_timestamps=False,
+                    device=device,  # ✅ Keeps CUDA safe
                 )
-
                 # stack frames for this video key and add directly to the item
                 item[vid_key] = frames