Upload run_cloud_training.py with huggingface_hub

run_cloud_training.py

@@ -515,6 +515,15 @@ def train(config_path, dataset_name, output_dir):
         # Initialize model with our safe loading function
         logger.info("Loading pre-quantized model safely")
         dtype = torch.float16 if hardware_config.get("fp16", True) else None
+        
+        # Force eager attention implementation
+        os.environ["TRANSFORMERS_NO_FLASH_ATTENTION"] = "1"
+        logger.info("Flash attention has been DISABLED globally via environment variable")
+        
+        # Update hardware config to ensure eager attention
+        hardware_config["attn_implementation"] = "eager"
+        hardware_config["use_flash_attention"] = False
+        
         model, tokenizer = load_model_safely(model_name, max_seq_length, dtype)
         
         # Disable generation capabilities for research training