修复ppo不能被整除问题

backend/app/engines/text_engine.py

@@ -328,9 +328,11 @@ class TextEngine(BaseEngine):
                 param.requires_grad = False
 
             # TRL 0.9.x PPOConfig 只接受 PPO 专用参数，不支持 HuggingFace Trainer 参数
+            # mini_batch_size 必须满足：batch_size % (mini_batch_size * gradient_accumulation_steps) == 0
             ppo_config = PPOConfig(
                 learning_rate=learning_rate,
                 batch_size=batch_size,
+                mini_batch_size=1,
                 gradient_accumulation_steps=gradient_accumulation,
                 ppo_epochs=ppo_epochs,
                 vf_coef=vf_coef,