云服务器版

finetune_kopa.py

@@ -141,13 +141,13 @@ def train(
         data_path: str = "/root/shared-nvme/dataset/olive_dataset.json",
         output_dir: str = "output",
         # training hyperparams
-        batch_size: int = 16,
+        batch_size: int = 8,
-        micro_batch_size: int = 16,
+        micro_batch_size: int = 4,
         num_epochs: int = 2,
-        learning_rate: float = 1e-4,
+        learning_rate: float = 1e-5,
         cutoff_len: int = 512,
         val_set_size: int = 0,
-        num_prefix: int = 1,
+        num_prefix: int = 10,
         # llm hyperparams
         train_on_inputs: bool = True,  # if False, masks out inputs in loss
         add_eos_token: bool = False,
@@ -453,54 +453,36 @@ def train(
         final_model = untie_shared_weights(final_model)
         print(f"[INFO] Saving model to {output_dir}")
 
-        # 确保输出目录存在
+        # Save the main model components
-        os.makedirs(output_dir, exist_ok=True)
+        if hasattr(final_model, "save_model"):
-        
+            final_model.save_model(output_dir)
-        # 如果是分布式训练，只在主进程保存
+        else:
-        if int(os.environ.get("LOCAL_RANK", 0)) == 0:
+            # Fallback if save_model method doesn't exist
-            # 将模型移到CPU上保存
+            os.makedirs(output_dir, exist_ok=True)
+
+            # Save model configuration
+            if hasattr(final_model, "config"):
+                final_model.config.save_pretrained(output_dir)
+
+            # Save model state dict (avoiding shared weights)
             model_to_save = final_model.module if hasattr(final_model, "module") else final_model
-            model_to_save = model_to_save.cpu()
+            torch.save(model_to_save.state_dict(), os.path.join(output_dir, "pytorch_model.bin"))
-            
-            try:
-                # Save the main model components
-                if hasattr(final_model, "save_model"):
-                    final_model.save_model(output_dir)
-                else:
-                    # Save model configuration
-                    if hasattr(final_model, "config"):
-                        final_model.config.save_pretrained(output_dir)
 
-                    # Save model state dict
+        # Save embeddings separately if they exist
-                    torch.save(model_to_save.state_dict(), os.path.join(output_dir, "pytorch_model.bin"))
+        if hasattr(final_model, "embeddings"):
-                    print(f"[INFO] Successfully saved model state dict")
+            torch.save(final_model.embeddings, os.path.join(output_dir, "embeddings.pth"))
 
-                    # Save embeddings separately if they exist
+        # Save PEFT model components
-                    if hasattr(final_model, "embeddings"):
+        if hasattr(final_model, "model") and hasattr(final_model.model, "save_pretrained"):
-                        torch.save(final_model.embeddings, os.path.join(output_dir, "embeddings.pth"))
+            peft_save_dir = os.path.join(output_dir, "peft_model")
-                        print(f"[INFO] Successfully saved embeddings")
+            os.makedirs(peft_save_dir, exist_ok=True)
-
+            final_model.model.save_pretrained(peft_save_dir)
-                    # Save PEFT model components
+            print(f"[INFO] PEFT model saved to {peft_save_dir}")
-                    if hasattr(final_model, "model") and hasattr(final_model.model, "save_pretrained"):
-                        peft_save_dir = os.path.join(output_dir, "peft_model")
-                        os.makedirs(peft_save_dir, exist_ok=True)
-                        final_model.model.save_pretrained(peft_save_dir)
-                        print(f"[INFO] PEFT model saved to {peft_save_dir}")
-                
-                # 保存完成后将模型移回原设备
-                model_to_save = model_to_save.to(device)
-                
-            except Exception as e:
-                print(f"[ERROR] Error during model saving: {str(e)}")
-                import traceback
-                traceback.print_exc()
-                raise e
 
     except Exception as e:
-        print(f"[ERROR] Error in save process: {str(e)}")
+        print(f"[ERROR] Error saving model: {e}")
         import traceback
         traceback.print_exc()
-        raise e
 
 
 def inspect_model_structure(model):

kopa.py

@@ -16,7 +16,7 @@ class KoPA(nn.Module):
             param.requires_grad = False
 
         # Only keep gradients for the adapter parts
-        # self.num_prefix = num_prefix
+        self.num_prefix = num_prefix
         hidden_size = model.config.hidden_size
         self.embeddings = nn.Embedding(100, 4096)
         for param in model.parameters():