数据并行可能导致的模型训练报错 (#2139)

pointerhacker

2024-10-15 70645e48072bf193fbf069949f1d2b10fddac8a3

 funasr/bin/train_ds.py

@@ -84,9 +84,6 @@
        dist.init_process_group(
            backend=kwargs.get("backend", "nccl"),
            init_method="env://",
            find_unused_parameters=kwargs.get("train_conf", {}).get(
                "find_unused_parameters", False
            ),
        )
        torch.cuda.set_device(local_rank)