AttributeError: 'tuple' object has no attribute 'is_cuda'

[Tron1994]

/mpu/transformer.py", line 422, in custom_forward
x_ = layer(x_, inputs[1])
File "/home/jovyan/miniconda3/envs/cpm2/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
return forward_call(*input, **kwargs)
File "/home/jovyan/xz_nlp/icode/CPM-2-Finetune-master/mpu/transformer.py", line 304, in forward
layernorm_output = self.input_layernorm(hidden_states)
File "/home/jovyan/miniconda3/envs/cpm2/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
return forward_call(*input, **kwargs)
File "/home/jovyan/miniconda3/envs/cpm2/lib/python3.7/site-packages/apex/normalization/fused_layer_norm.py", line 290, in forward
if not input.is_cuda:
AttributeError: 'tuple' object has no attribute 'is_cuda'

报错怎么解

[t1101675]

有没有训练的脚本？

[Tron1994]

有没有训练的脚本？

为了使用past_key_values，我用https://github.com/TsinghuaAI/CPM-1-Generate中的model/gpt2_modeling.py和mpu/transformer.py做了替换后，就出现了上面的错误

[t1101675]

原来应该也是可以使用 past_key_values 的

[Tron1994]

原来应该也是可以使用 past_key_values 的
cpm1-finetune：https://github.com/TsinghuaAI/CPM-1-Finetune/blob/main/mpu/transformer.py
cpm1-generate：https://github.com/TsinghuaAI/CPM-1-Generate/blob/main/mpu/transformer.py

cpm1-generate做了更新：TsinghuaAI/CPM-1-Generate@2730b1f#diff-a7922243f0695439497bd5bdae2069ec8fa268d10d549afc26fcd2765646a36e

我仔细对比没发现有什么问题，但是在layernorm_output = self.input_layernorm(hidden_states) 会报 is_cuda的错误

[t1101675]

那用 CPM-1-Generate 也会报错？

[Tron1994]

那用 CPM-1-Generate 也会报错？

刚跑了可以通，我把cpm1和cpm2放一块了，可能是哪里起冲突了吧