Use sequential device_map to fix key naming conflicts during LoRA merge

app.py

@@ -108,6 +108,7 @@ class ModelMerger:
             # Load base model with explicit multi-GPU configuration
             progress(0.25, desc="Loading base model (this may take several minutes)...")
             logger.info(f"Loading base model: {BASE_MODEL_NAME}")
+            logger.info(f"Note: For merging, we'll use a simpler device_map to avoid key naming issues")
             
             if use_8bit:
                 logger.info(f"Using 8-bit quantization for memory efficiency (~50% memory reduction)")
@@ -117,13 +118,13 @@ class ModelMerger:
                 precision_desc = "bfloat16"
             
             try:
+                # For merging, use sequential device map to avoid complex key nesting
+                # This ensures consistent key names between training and merging
                 load_kwargs = {
                     "trust_remote_code": True,
                     "low_cpu_mem_usage": True,
-                    "device_map": "auto",
+                    "device_map": "sequential",  # Changed from "auto" to avoid key nesting issues
                     "max_memory": max_memory,
-                    "offload_folder": "/tmp/offload",
-                    "offload_state_dict": True,
                 }
                 
                 if use_8bit: