Upload run_cloud_training.py with huggingface_hub

run_cloud_training.py

@@ -259,8 +259,9 @@ def train(config_path, dataset_name, output_dir):
         logger.info("Initializing model with unsloth (preserving 4-bit quantization)")
         max_seq_length = training_config.get("max_seq_length", 2048)
         
-        # Create LoRA config
-        peft_config = LoraConfig(
+        # Create LoRA config directly
+        logger.info("Creating LoRA configuration")
+        lora_config_obj = LoraConfig(
             r=lora_config.get("r", 16),
             lora_alpha=lora_config.get("lora_alpha", 32),
             lora_dropout=lora_config.get("lora_dropout", 0.05),
@@ -273,14 +274,24 @@ def train(config_path, dataset_name, output_dir):
         dtype = torch.float16 if hardware_config.get("fp16", True) else None
         model, tokenizer = load_model_safely(model_name, max_seq_length, dtype)
         
-        # Apply LoRA
+        # Apply LoRA - correctly passing lora_config_obj directly
         logger.info("Applying LoRA to model")
-        model = FastLanguageModel.get_peft_model(
-            model,
-            peft_config=peft_config,
-            tokenizer=tokenizer,
-            use_gradient_checkpointing=hardware_config.get("gradient_checkpointing", True)
-        )
+        try:
+            logger.info("Attempting to apply LoRA with unsloth API")
+            model = FastLanguageModel.get_peft_model(
+                model,
+                lora_config=lora_config_obj,  # Pass lora_config directly instead of peft_config
+                tokenizer=tokenizer,
+                use_gradient_checkpointing=hardware_config.get("gradient_checkpointing", True)
+            )
+        except Exception as e:
+            logger.warning(f"Error applying LoRA with unsloth: {e}")
+            logger.info("Falling back to standard PEFT method")
+            
+            # Try with standard PEFT approach if unsloth fails
+            from peft import get_peft_model
+            model = get_peft_model(model, lora_config_obj)
+            logger.info("Successfully applied LoRA with standard PEFT")
         
         # No need to format the dataset - it's already pre-tokenized
         logger.info("Using pre-tokenized dataset - skipping tokenization step")