diff --git a/training/scripts/train.py b/training/scripts/train.py
index 947c61b..423afd3 100755
--- a/training/scripts/train.py
+++ b/training/scripts/train.py
@@ -59,8 +59,8 @@ def train(config_path):
     except Exception as e:
         print(f"✗ Failed: {e}")
         
-        # Strategy 1: Load 4-bit model AS-IS (no quantization, no DeepSpeed)
-        print("\n[1/2] Trying: 4-bit model AS-IS...")
+        # Strategy 1: 4-bit AS-IS with auto device_map
+        print("\n[1/6] Trying: 4-bit AS-IS...")
         try:
             model = AutoModelForCausalLM.from_pretrained(
                 config["base_model"],
@@ -68,24 +68,52 @@ def train(config_path):
                 device_map="auto",
                 trust_remote_code=True,
             )
-            print("✓ Success: 4-bit model loaded")
+            print("✓ Success: 4-bit AS-IS")
         except Exception as e:
             print(f"✗ Failed: {e}")
             
-            # Strategy 2: bf16 to CPU
-            print("\n[2/2] Trying: bf16 model to CPU...")
+            # Strategy 2: 4-bit with bf16 compute
+            print("\n[2/6] Trying: 4-bit with bf16 compute...")
             try:
-                model = AutoModelForCausalLM.from_pretrained(
-                    config["base_model"],
-                    torch_dtype=torch.bfloat16,
-                    device_map="cpu",
-                    low_cpu_mem_usage=True,
-                    trust_remote_code=True,
-                )
-                print("✓ Success: bf16 model loaded to CPU")
+                from transformers import BitsAndBytesConfig
+                bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True)
+                model = AutoModelForCausalLM.from_pretrained(config["base_model"], quantization_config=bnb_config, device_map="auto", trust_remote_code=True)
+                print("✓ Success: 4-bit with bf16 compute")
             except Exception as e:
                 print(f"✗ Failed: {e}")
-                raise RuntimeError("All loading strategies failed!")
+                
+                # Strategy 3: 4-bit to CPU
+                print("\n[3/6] Trying: 4-bit to CPU...")
+                try:
+                    model = AutoModelForCausalLM.from_pretrained(config["base_model"], torch_dtype=torch.float16, device_map="cpu", trust_remote_code=True)
+                    print("✓ Success: 4-bit to CPU")
+                except Exception as e:
+                    print(f"✗ Failed: {e}")
+                    
+                    # Strategy 4: 4-bit fp32
+                    print("\n[4/6] Trying: 4-bit fp32...")
+                    try:
+                        model = AutoModelForCausalLM.from_pretrained(config["base_model"], torch_dtype=torch.float32, device_map="auto", trust_remote_code=True)
+                        print("✓ Success: 4-bit fp32")
+                    except Exception as e:
+                        print(f"✗ Failed: {e}")
+                        
+                        # Strategy 5: bf16 to CPU
+                        print("\n[5/6] Trying: bf16 to CPU...")
+                        try:
+                            model = AutoModelForCausalLM.from_pretrained(config["base_model"], torch_dtype=torch.bfloat16, device_map="cpu", low_cpu_mem_usage=True, trust_remote_code=True)
+                            print("✓ Success: bf16 to CPU")
+                        except Exception as e:
+                            print(f"✗ Failed: {e}")
+                            
+                            # Strategy 6: bf16 auto
+                            print("\n[6/6] Trying: bf16 auto...")
+                            try:
+                                model = AutoModelForCausalLM.from_pretrained(config["base_model"], torch_dtype=torch.bfloat16, device_map="auto", low_cpu_mem_usage=True, trust_remote_code=True)
+                                print("✓ Success: bf16 auto")
+                            except Exception as e:
+                                print(f"✗ Failed: {e}")
+                                raise RuntimeError("All loading strategies failed!")
     
     # Prepare model for k-bit training
     from peft import prepare_model_for_kbit_training