Upload run_cloud_training.py with huggingface_hub

run_cloud_training.py

@@ -603,7 +603,7 @@ def train(config_path, dataset_name, output_dir):
         # For L40S GPU, we can use a larger batch size and shard model across the single GPU
         if "L40S" in gpu_info.name or gpu_info.total_memory > 40e9:  # Check if it's L40S (>40GB VRAM)
             logger.info("Detected L40S GPU - optimizing for high-memory GPU")
-            per_device_train_batch_size = training_config.get("per_device_train_batch_size", 6)
+            per_device_train_batch_size = training_config.get("per_device_train_batch_size", 4)
             logger.info(f"Using optimized batch size for L40S: {per_device_train_batch_size}")
         else:
             # Default to a smaller batch size for other GPUs