feat : 数据流水拼接策略支持 position_ids 预计算

- OutputConfig.position_ids_mode 三种模式控制边界策略
- pipeline._flush() 按配置生成扁平 position_ids 数组
- SFTDataset 在 __getitem__ 中返回 position_ids
- SFTStrategy 将 position_ids 传入 model.forward()

astrai/trainer/strategy.py

@@ -180,14 +180,15 @@ class SFTStrategy(BaseStrategy):
 
     def compute_loss(self, batch: Dict[str, Tensor]) -> Tensor:
         batch = move_to_device(batch, self.device)
-        input_ids, target_ids, loss_mask = (
+        input_ids, target_ids, position_ids, loss_mask = (
             batch["input_ids"],
             batch["target_ids"],
+            batch["position_ids"],
             batch["loss_mask"],
         )
 
         ignore_index = -100
-        logits = self.model(input_ids=input_ids)["logits"]
+        logits = self.model(input_ids=input_ids, position_ids=position_ids)["logits"]
         target_ids = target_ids.masked_fill(loss_mask == 0, ignore_index)
 
         loss = F.cross_entropy(