Hello, thanks a lot for this repo.
I'm face with such a code error:
I print the dictionary func_inputs:
Only the values of input_ids and attention_mask are node names, which are legal. Other values of func_inputs are the same initail value, which is 0.
I think there maybe something wrong with the algo 1 "Search Node Dependency" .
The values of func_inputs should have been changed by line 7 in algo 1, but actually not.
There is also may something wrong with the torch.fx.Tracer.trace(), which wrongly inserted "placeholder" op.
I used cuda 11.7, torch 1.13.1 and four tesla v100.
I'm looking forward to your advice. Thank you.
@huils20 I am sorry to reply so late. Please try to downgrade PyTorch to 1.10 and transformers to 4.15.0 , and we will try torch 1.13 very soon and fix it in next PR.
Hello, thanks a lot for this repo. I'm face with such a code error:
I print the dictionary func_inputs:
Only the values of input_ids and attention_mask are node names, which are legal. Other values of func_inputs are the same initail value, which is
0
.I think there maybe something wrong with the algo 1 "Search Node Dependency" .![image](https://github.com/HPDL-Group/Merak/assets/89234295/e14aa521-3cb7-403c-9389-ba79f72deff5)
The values of func_inputs should have been changed by line 7 in algo 1, but actually not.
There is also may something wrong with the torch.fx.Tracer.trace(), which wrongly inserted "placeholder" op.
I used cuda 11.7, torch 1.13.1 and four tesla v100. I'm looking forward to your advice. Thank you.