[Bug Fix] fix allreduce tensor dtype (PaddlePaddle#7876)

* [Bug Fix] fix allreduce tensor dtype Reason: some CCL not support bool dtype * update int8 to int32
xysheng-baidu · Feb 22, 2024 · ed42f16 · ed42f16
1 parent 724b524
commit ed42f16
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/paddlenlp/trainer/trainer.py b/paddlenlp/trainer/trainer.py
@@ -631,7 +631,7 @@ def train(
         # The resume_from_checkpoint could be None in some machine node.
         # Here we reset None to temp directory.
         if args.world_size > 1:
-            is_resume_from_checkpoint = paddle.to_tensor([resume_from_checkpoint is not None])
+            is_resume_from_checkpoint = paddle.to_tensor([resume_from_checkpoint is not None], dtype="int32")
             paddle.distributed.all_reduce(is_resume_from_checkpoint)
             is_resume_from_checkpoint = is_resume_from_checkpoint.item()
             if is_resume_from_checkpoint > 0 and is_resume_from_checkpoint < paddle.distributed.get_world_size():