瀏覽代碼

Enable rotary embeddings

Aleksandr Borzunov 3 年之前
父節點
當前提交
d0985de540
共有 1 個文件被更改,包括 1 次插入1 次删除
  1. 1 1
      task.py

+ 1 - 1
task.py

@@ -82,7 +82,7 @@ class TrainingTask:
                 attn_dropout=0,
                 shared_attn_ids=shared_layer_ids,
                 shared_ff_ids=shared_layer_ids,
-                rotary_emb=False,  # FIXME: Fix RuntimeError when True
+                rotary_emb=True,
                 reversible=True,
             )
             self.model = ModelWrapper(dalle)