Runtime error：数据类型报错

[sqqiao]

作者好，我在复现RRHF时碰到变量类型报错：
我配置fsdp_config进行分布式训练，当我使用--bf16混合精度时，报错：
return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
RuntimeError: Expected tensor for argument #1 'indices' to have one of the following scalar types: Long, Int; but got CUDABFloat16Type instead (while checking arguments for embedding)

如果不使用bf16和tf32，报错：
return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
RuntimeError: Expected tensor for argument #1 'indices' to have one of the following scalar types: Long, Int; but got torch.cuda.FloatTensor instead (while checking arguments for embedding)

我的fsdp_config配置如图

使用的模型是llama3-8b，或者是tokenizer需要重新配置一下吗？