Fix resume adapter training (no peft_config with PeftModel)

app.py

@@ -517,10 +517,12 @@ def run_training(model_name, chunk_steps, max_total_steps, max_minutes, auto_con
         _try_download_adapter(add_log)
 
         # Resume LoRA adapter if present
+        resume_adapter = False
         if os.path.isdir(ADAPTER_DIR) and os.path.exists(os.path.join(ADAPTER_DIR, "adapter_config.json")):
             add_log("Loading existing LoRA adapter (resume)...")
             model = PeftModel.from_pretrained(base_model, ADAPTER_DIR, is_trainable=True)
             add_log("✓ Adapter loaded")
+            resume_adapter = True
         else:
             model = base_model
 
@@ -609,15 +611,18 @@ def run_training(model_name, chunk_steps, max_total_steps, max_minutes, auto_con
                 num_generations=4,
             )
 
-            trainer = GRPOTrainer(
-                model=model,
-                args=config,
-                train_dataset=dataset,
-                reward_funcs=perf_takehome_reward_fn,
-                peft_config=lora_config,
-                processing_class=tokenizer,
-                callbacks=[VLIWCallback()],
-            )
+            trainer_kwargs = {
+                "model": model,
+                "args": config,
+                "train_dataset": dataset,
+                "reward_funcs": perf_takehome_reward_fn,
+                "processing_class": tokenizer,
+                "callbacks": [VLIWCallback()],
+            }
+            if not resume_adapter:
+                trainer_kwargs["peft_config"] = lora_config
+
+            trainer = GRPOTrainer(**trainer_kwargs)
 
             train_result = trainer.train()
             metrics = train_result.metrics